CN106776881A - A kind of realm information commending system and method based on microblog - Google Patents

Abstract

The invention discloses a domain information recommendation system and method based on a microblog platform, including: a data acquisition and preprocessing module, a domain keyword extraction module, a user-defined keyword expansion module, a linear merging module, similarity calculation and personality Recommendation module and topic acquisition module; the present invention designs and implements a domain information recommendation method for the characteristics of the microblog platform, seamlessly combines keyword extraction and keyword expansion, thereby ensuring both domain feature extraction and The dynamics of the recommendation results are guaranteed, and the effectiveness of this method is verified through the experiment of the corresponding system based on Sina Weibo. The invention can assist enterprise microblog marketing and effectively improve the efficiency of enterprise microblog marketing.

Description

A domain information recommendation system and method based on microblog platform

technical field

The invention relates to a field microblog recommendation system and method under a microblog platform, which supports unguided field feature extraction and user-defined keywords, and belongs to the field of computer technology.

Background technique

As the Internet enters the post-WEB2.0 era, social functions have become a model for the great changes in the Internet. Major social networking sites have sprung up and quickly occupied the dominance of the Internet. As early as March 2010, the famous American social networking site Facebook ^TM surpassed Google in terms of visits and became the largest website in the United States. In China, emerging social media such as Sina Weibo are also rising rapidly. As of May 16, 2012, the number of users of Sina Weibo has reached 300 million, which is second only to the number of users accumulated over ten years in the domestic Internet market. Tencent QQ products. On the other hand, according to the 2011 IT Industry Top Ten Strategic Technology Report released by Gartner, the world's most authoritative IT research and consulting company, technologies directly related to social networking accounted for two, namely Social Communications and Collaboration and Social Communications. Analytics. It is also the only technology category that occupies two seats among all technologies. This shows people's attention to social products and their prospects.

On the other hand, the popularization of Internet use and the enhancement of people's stickiness to Internet use have made big data a hot topic in the IT industry in recent years. However, in order to transform big data into useful values for human beings, the support of related technologies such as data mining is needed. Therefore, the popularity of data mining and analysis has also soared in recent years. Especially data mining and analysis for enterprises, because it can bring direct benefits to enterprises. Moreover, the analysis results based on a large number of real user data are more reliable and convincing than traditional analysis techniques, and this is a real-time analysis method that can better adapt to market changes and better capture the market opportunity.

Although the Weibo platform contains a variety of domain information and responds quickly to domain events, there are still many difficulties in obtaining more comprehensive domain information on it. The rise of the Weibo platform and the rapid growth of users have brought about the problem of information overload. As the number of users’ attention increases, more and more domain-independent content appears in the Weibo subscribed by users; at the same time, if users only Focusing on a small number of highly relevant users in the field will inevitably lead to failure to obtain comprehensive field information in a timely manner. That is to say, there is a problem that the accuracy rate and the recall rate of the user's data acquisition on the Weibo platform cannot be achieved at the same time. Moreover, the microblog platform itself has the characteristics of topic dispersion and information fragmentation, which requires a method that can identify the domain interests of enterprise microblog users, and extract and recommend microblog information according to the domain relevance. However, most of the existing social media management and analysis software only use user-defined keywords for simple text matching in domain information extraction, and this method has great defects. First of all, individual keywords cannot fully describe the information needs of the domain; second, the richness of language limits the effect of simple text matching.

Both the traditional method of modeling user interest through keyword extraction and the method of expressing user interest through user-defined keywords have their own defects.

The keyword extraction method is mainly manifested in: this is an unguided extraction algorithm, and the user cannot dynamically adjust the algorithm results, so it cannot meet the user's temporary dynamic needs for a topic in a certain field. Moreover, the algorithm cannot perform domain interest modeling when the user's history of Weibo is small, which is the so-called cold start phenomenon.

The defects of user-defined keywords are mainly manifested in: first, it is difficult to ensure that user-defined keywords completely cover all the information in this field; Sex will make it impossible to extract a lot of information that is synonymous but uses different words.

On the other hand, the application of related research work also has its own problems in this application scenario. First of all, the rise of social networks provides new application scenarios with great value for data mining and data analysis related technologies. Social networks provide new perspectives for data mining and analysis, and also make traditional data mining technologies face new challenges.

To sum up, when the above method is applied to the real system of Weibo platform, there are several major problems as follows:

(1) Both keyword extraction and keyword expansion have their own problems, and neither single strategy can well meet the needs of relevant information recommendation in the microblog platform field.

(2) In related research work, due to the characteristics of information fragmentation on social platforms, a class of algorithms need to rely on external corpus for training corpus. This greatly reduces the practicality and portability of the algorithm.

(3) Another type of algorithm builds a model based on the global corpus, which is computationally expensive and not universally applicable.

Therefore, there is a need for a recommendation method that can provide personalized recommendations for domain information based on keywords, so as to help users obtain domain-related information quickly and accurately. This method needs to solve the contradiction that the accuracy rate and the recall rate can not have both in the process of obtaining domain-related information by enterprise users on the Weibo platform. At the same time, based on the requirements of practicability, the algorithm cannot rely on external corpus, and needs to be calculated based on the user's local data. Therefore, proposing a recommendation method with the above properties is the focus of the present invention.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art and provide a domain information recommendation system and method based on the microblog platform. The method of modeling user interests not only ensures the comprehensive identification of domain information but also enables users to dynamically adjust their domain interests according to their needs; the graph-based keyword extraction algorithm TextRank is adopted, which does not depend on other corpus and avoids the extraction results being affected by language. Influenced by the phenomenon of Zif's law in the model, an optimized P-IOW algorithm is proposed to achieve better expansion of keywords. This method ensures that the user's dynamic interest needs can be met in real time and greatly enhances the expression ability of user-defined keywords. By linearly merging the results of keyword extraction and expansion according to user-defined weights, it can personalize microblog information and topics in related fields for users, and help users obtain field-related information quickly and accurately.

Technical solution of the present invention: a domain information recommendation system based on a microblog platform, including: data acquisition and preprocessing module, domain keyword extraction module, user-defined keyword expansion module, linear merge module, similarity calculation and personality Optimized recommendation module and topic acquisition module; among them:

Data acquisition and preprocessing module: Obtain user-related microblog information data and perform preprocessing; preprocessing includes data stop word filtering, word segmentation and part-of-speech tagging; the preprocessing result is the user's historical microblog data, which is passed to the domain Keyword extraction module; if the user defines domain interest keywords, the preprocessing result is simultaneously passed to the user-defined keyword expansion module;

Field keyword extraction module: Based on the preprocessing results, the keyword extraction is carried out unguided by the TextRank for Weibo algorithm modified based on the TextRank algorithm. first stage; in the construction stage of the undirected graph based on the co-occurrence relationship, firstly, the word segmentation appearing in the user’s historical Weibo is converted into the corresponding node; when constructing the connecting edge between the nodes, whether there is an edge between the nodes and the edge The weight determines the co-occurrence composition by the number of co-occurrences of two words in the same Weibo. The weight of the edge is the number of co-occurrences of words in the same Weibo. If two words co-occur in a user’s Weibo , then the weight of the edge between the nodes corresponding to the two words is increased by 1, and the final weight of the edge is the number of co-occurrences of the corresponding two words in Weibo; then based on the node weight calculation stage of the graph, the iterative calculation The weight of each stage, until the variation of the node weight converges to a certain threshold; after the iteration, the weight of each node is the importance of the participle it represents, and all the participle of the user are sorted according to the importance That is, the result of keyword extraction is obtained, so as to automatically identify the characteristics of the field where the user is located;

User-defined keyword expansion module: Calculate the similarity between keywords based on the co-occurrence and distribution of keywords and the attribute information of their users, and use highly relevant words as the expansion results of target keywords; this module supports The user enters multiple custom keywords, and for each custom keyword, the extended word vectors obtained from the keywords will be linearly summed to obtain the final extended vector; the user-defined keyword expansion function ensures the user's Dynamic interest needs can be met in real time, and at the same time, the expression ability of user-defined keywords is greatly enhanced;

Linear merge module: After the automatic extraction of domain keywords and the expansion based on user-defined keywords are completed, the maximum value normalization method is used to normalize the two result vectors, so that the results of keyword extraction and keyword expansion The vector is mapped to a unified value range; after normalization, the two normalized vectors are linearly merged, and the merge process supports user-defined keyword extraction and keyword expansion weights; the module outputs a A word vector representing the user's final domain interests;

Correlation calculation and personalized recommendation module: After the linear merging module describes the keyword vectors of user interests in the field, word segmentation and word frequency statistics are performed on each to-be-filtered microblog to generate word frequency vectors, and then the user interest keyword vectors, to-be-recommended The word frequency vector generated by the microblog and the IDF information vector are dot-multiplied to obtain the correlation between the microblog and the user's interest, which is the domain correlation of the microblog. By calculating the domain relevance of each user's microblog, sorting the domain correlation from high to low, presenting the microblog information to the user, and realizing the personalized domain microblog recommendation for the user;

Topic Acquisition Module: Train the LDA model with the domain microblog text recommended to the user as input, and cluster the terms into topics according to the distribution of terms in the topic; combine the subject term set with the user domain interest keywords obtained in the linear combination module Items are calculated to obtain the importance of topics, and presented to users in order of importance, thereby completing topic discovery and recommendation.

The implementation process of the data acquisition and preprocessing module is as follows:

(1) After the user logs in to the Weibo system, the user verification is first performed. After the verification is passed, the Weibo platform credentials associated with the user are automatically used to interact with the Weibo platform to verify the legitimacy of the user's identity on the Weibo platform;

(2) Obtain user-related microblog information data, and use the local database to persist the acquired data in a structured manner so that it can be read at any time;

(3) Preprocessing the persistent microblog text, including stop word filtering, word segmentation and part-of-speech tagging; according to the characteristics of microblog text, the stop words are first filtered by pattern matching method, and then Optimized Chinese word segmentation and part-of-speech tagging for the Weibo scene, use the word segmentation product ICTCLAS5.0 for word segmentation and part-of-speech tagging, and filter the part-of-speech results of user Weibo word segmentation before keyword extraction and keyword expansion. Only keep nouns. The preprocessing result is the user's historical microblog data.

In the domain keyword extraction module, the weight of each stage is calculated iteratively according to the algorithm idea of PageRank, and the formula is as follows:

Among them: V _i is the i-th node, TR(V _i ) is the weight of node V _i , w _ij is the weight of the edge between nodes V _i and V _j ; E(V _i ) is the weight of the edge connected by V _i Set; d is the damping coefficient of the iteration, which is set to 0.85. You can start the iteration with any initial value and iterate until it converges. The convergence condition is that the absolute difference between the sum of the weights of each node between this iteration and the last iteration is less than the specified value.

In the user-defined keyword expansion module, the improved P-IOW algorithm is used to calculate the similarity between keywords, and the implementation process is as follows:

For a given user-defined keyword s, the calculation method of the domain correlation of word t with respect to s is as follows:

in:

Among them: s is a user-defined keyword; wf(t) is the number of microblogs containing the word t; t _M is the word contained in the most microblogs in the domain-related corpus, is the number of microblogs where t _M is located; is the number of microblogs that do not contain word s; wf(t∧s) is the number of microblogs that contain both word t and word s; N is the total number of user microblogs; s _p is a smoothing coefficient with a range of (0, 1) make Divisor by zero does not occur when it is zero. From the formula, it can be concluded that the value range of P-IOLogW is between log(s _p ) and log(1/s _p ) before formula (1) is multiplied by the down-weighting factor. The larger the value of the calculation result P-IOW, the higher the domain similarity between the word w and the user-defined keyword s. For the user-defined keyword itself, the upper limit of the value range of P-IOLogW will be assigned, that is, log(1/s _p ).

In the user-defined keyword expansion module, the extended word vectors expanded by the keywords are linearly summed, so as to obtain the final extended vector, the specific algorithm of the expanded process is as follows:

(1) for each user-defined keyword collection keyword, at first, calculate all expansion word weights of keyword based on P-IOW;

(2) Map the extended word weight vector related to the keyword to the space of domain-related corpus word segmentation to form an extended word vector;

(3) Linearly accumulate all the expanded word vectors to obtain the final expanded vector, which is the output result of the keyword expanded module.

The implementation process of the linear merging module is as follows:

(1) Use the maximum value normalization method to normalize the vectors generated by keyword extraction and keyword expansion;

(2) For each component in the vector, the specific normalization method is as follows:

v _normal = v/v _max

Among them: v is the initial value of a certain component of the vector; v _max is the maximum value of all components of the vector. After normalization by the maximum value, all components of the vector are between (0, 1], and all components are non- Zero, followed by vector linear weighted merging:

V _combine ＝r×V _kw-extract +(1-r)×V _kw-expand

Among them: r is the user-defined combined weight ratio, V _kw-extract is the result vector of keyword extraction, V _kw-expand is the result vector generated by keyword expansion, and the combined result is the keyword vector V _combine that describes the interest of the user field .

In the similarity calculation and personalized recommendation module, the specific formula for domain correlation calculation is as follows:

Wherein: the keyword vector of interest in the user field is V _combine ; the word frequency vector corresponding to microblog T to be recommended is W; the IDF vector of word segmentation is V _IDF ; L is the total number of vector space dimensions after word segmentation of user history microblogs; with are the components corresponding to t _i above V _combine and W respectively; IDF(t _i ) is the IDF value of t _i .

A domain information recommendation method based on the Weibo platform, which is divided into six steps: data acquisition and preprocessing, domain keyword extraction, user-defined keyword expansion, linear merger, similarity calculation and personalized recommendation, and topic acquisition. as follows:

(1) Obtain user-related microblog information for data preprocessing; preprocessing includes filtering stop words using pattern matching; using word segmentation system ICTCLAS5.0 for word segmentation and part-of-speech tagging; the preprocessing result is the user's history Weibo data.

(2) The preprocessing result carries out the field keyword extraction, and the keyword extraction will use the TextRank for Weibo algorithm modified based on the TextRank algorithm of the present invention to carry out without guidance; There are two stages of node weight calculation; the construction stage of the undirected graph based on the co-occurrence relationship first converts the word segmentations that appear in the user's historical microblogs into corresponding nodes; The meta-segment pair is constructed as an edge, and the weight of the edge is the number of co-occurrences of the corresponding word pair in historical microblogs;

The graph-based node weight calculation stage iteratively calculates the weight of each stage according to the algorithm idea of PageRank until the change of node weight converges to a certain threshold;

(3) If the user defines the domain interest keywords, the preprocessing results will be passed to the keyword expansion module; the keyword expansion will be performed using the improved P-IOW algorithm based on the P-IOLog algorithm, using co-occurrence based on keywords , distribution, and the attributes of the users they belong to to calculate the similarity between keywords, and use highly relevant words as the extended results of target keywords; at the same time, users are supported to input multiple custom keywords; for each custom Keywords, the keyword expansion module will linearly add the expanded word vectors to obtain the final expanded vector;

(4) The results of the two are normalized according to the user-defined weights, and the two result vectors are normalized by the maximum value normalization method, so that they are mapped to a unified value range; after normalization, the The two normalized vectors are linearly merged. The merge process supports user-defined keyword extraction and keyword expansion weights. The merged results are used for correlation comparison by related recommendation modules to generate correlation scores for recommended microblogs;

(5) Segment the microblogs to be recommended subscribed by the user and vectorize them according to the word frequency, and then according to the identified user interest, point the user interest keyword vector, the word frequency vector generated by the microblog to be recommended, and the IDF information vector Multiplication operation, using the method of vector space point multiplication to calculate the domain correlation;

(6) Take the field microblog text recommended to the user as the input set, realize term clustering based on the LDA strategy, complete the topic discovery, and then combine the subject term set with the user domain interest keywords obtained in the linear merge module Correlation calculation, establishing the importance of topics, and recommending users based on the importance of topics.

The advantage of the present invention compared with prior art is:

(1) In terms of the source of domain corpus, there are respective problems in keyword extraction and keyword expansion methods, and a single strategy cannot well satisfy the problem of domain-related information recommendation on the microblog platform. Using graph-based keyword extraction and user-defined keyword expansion technology based on co-occurrence information to model users' domain interests ensures that users' dynamic interest needs can be met in real time and greatly enhances user-defined The expressive ability of keywords takes into account the comprehensive description of domain information and the dynamic changes of user interests, and solves the contradiction between accuracy and recall when domain users obtain domain-related information on the Weibo platform.

(2) Based on the requirements of practicality, the algorithm in the present invention does not depend on external corpus, and does not need to use large-scale corpus for analysis, but calculates based on the user's local data, that is, the user's historical microblog data, so the algorithm has Faster response times for greater utility and portability.

(3) Topic modeling is performed on the domain microblog text recommended to users to filter out the interference of domain-independent topics, realize more accurate topic recommendation, and further improve the accuracy of domain information recommendation.

Description of drawings

Fig. 1 is an architecture diagram of the present invention;

Fig. 2 is the overall flow chart of the system of the present invention;

Fig. 3 is a system sequence diagram of the present invention;

Fig. 4 is a system frame diagram of the present invention;

Fig. 5 is an excerpt of the main part of the UML diagram of the keyword extraction submodule based on diagrams of the present invention;

Fig. 6 is an excerpt of the main part of the UML diagram of the keyword expansion sub-module based on co-occurrence information in the present invention.

detailed description

The present invention will be described in detail below in conjunction with specific embodiments and accompanying drawings.

As shown in Figure 1, a field information recommendation system based on the microblog platform of the present invention includes: data acquisition and preprocessing module, field keyword extraction module, user-defined keyword expansion module, linear merging module, similarity Calculation and personalized recommendation module and topic acquisition module.

Data acquisition and preprocessing module: for enterprise microblog users on the Sina Weibo platform, by analyzing the historical microblog texts of relevant microblog accounts in several designated fields, use the Sina Weibo open platform to obtain user-related historical microblog information At the same time, the data preprocessing work is completed, that is, the method of pattern matching is used to filter the stop words; the word segmentation system ICTCLAS5.0 is used for word segmentation and part-of-speech tagging. The preprocessing result is the user's historical microblog data, which is passed to the field keyword extraction module. If the user defines the field interest keywords, the preprocessing result is also passed to the user-defined keyword expansion module.

Domain keyword extraction module: According to the preprocessing results, first construct an undirected graph based on the co-occurrence relationship, convert the word segmentations that appear in the user's historical microblogs into corresponding nodes, and analyze all the occurrences of the word segmentation results in each microblog Edges are constructed for binary word pairs, and the weight of edges is the number of co-occurrences of corresponding word pairs in historical microblogs. Then calculate the node weight based on the graph, iteratively calculate the weight of each stage according to the PageRank algorithm idea, until the change of the node weight converges to a certain threshold. After the iteration, the weight of each node is the importance of the word segment it represents. The results of keyword extraction can be obtained by sorting all the user's word segmentation according to the importance, so as to automatically identify the characteristics of the user's field.

User-defined keyword expansion module: use the improved P-IOW (Probabilistic Inside-OutsideLog for Weibo) method to calculate the similarity between keywords based on information such as the co-occurrence and distribution of keywords and the attributes of users to which they belong. Use highly relevant words as the expanded results of the target keywords. This module supports users to input multiple custom keywords. For each custom keyword, the keyword expansion module will linearly add the expanded word vectors to obtain the final expanded vector. The user-defined keyword expansion function ensures that the dynamic interest needs of users can be met in real time, and at the same time greatly enhances the expression ability of user-defined keywords.

Linear merge module: After the automatic extraction of domain keywords and the expansion based on user-defined keywords are completed, the maximum value normalization method is used to normalize the two result vectors so that they are mapped to a unified value range among. After normalization, linearly merge the two normalized vectors. The merging process supports user-defined weights for keyword extraction and keyword expansion. The module outputs a word vector representing the user's final domain interests.

Relevance Calculation and Personalized Recommendation Module: After generating keyword vectors that characterize the user’s interest in the field, the Relevance Calculation and Personalized Recommendation module will perform word segmentation and word frequency statistics on each microblog to be filtered to generate word frequency vectors, and then user The keyword vector of interest, the word frequency vector generated by the microblog to be recommended, and the IDF information vector are subjected to dot multiplication to obtain the correlation between the microblog and the user's interest. By calculating the domain relevance of each user's microblogs and sorting the domain correlations from high to low, personalized domain microblog recommendations for users are realized.

Topic Acquisition Module: Train the LDA model with the domain microblog text recommended to the user as input, and cluster the terms into topics according to the distribution of terms in the topic; combine the subject term set with the user domain interest keywords obtained in the linear combination module Items are calculated to obtain the importance of topics, and presented to users in order of importance, thereby completing topic discovery and recommendation.

As shown in Figure 2, the implementation process of the data acquisition and preprocessing module is as follows:

(1) After the user logs into the Weibo system, user verification is first performed. After the verification is passed, the system will automatically use the Sina Weibo open platform OAuth2.0 credentials associated with the user to interact with the open platform to verify the legitimacy of the user's identity on the Sina Weibo platform. If the credentials have not expired, user authentication is complete. If the certificate does not exist or has expired, the system will automatically go to the OAuth2.0 verification page of the open platform, which requires the user to enter their username and password on Sina Weibo. After the user enters the correct information, the open platform will send the updated credential back to the system, and the system will persist the credential to ensure that within the validity period of the credential, the user can log in to the open platform and obtain Weibo with only the user name and password of the system. data and related operations.

(2) The obtained microblog text data is persisted in a structured way by using the local database, so that the upper analysis module can read it at any time. In terms of data update, this module supports incremental update method. That is, each update only transmits the newly added relevant microblog information of the user, thereby improving the response speed of the system and saving network bandwidth to the greatest extent.

(3) Next, preprocess the persistent microblog text, including stop word filtering, word segmentation and part-of-speech tagging. Weibo stop words mainly include the following formats: hashtags in the format of "#话话词#", directional notifications to a user in the format of "@username", emoticons in the format of "[emoji text]" and microblogs containing URL links, etc. According to the characteristics of microblog text, the method of pattern matching is used to filter these stop words first. Then optimize the Chinese word segmentation and part-of-speech tagging for the Weibo scene. Word segmentation is a text processing procedure specific to a few languages like Chinese, which does not have distinct delimiters like most other languages. In natural language processing, word segmentation is an inevitable process of converting text into a format that computers can understand. In particular, the present invention adopts a space vector model, and word segmentation is an indispensable step. Moreover, the quality of word segmentation results will directly affect the quality of algorithm results. Part-of-speech tagging refers to assigning correct lexical tags to each word in a given sentence, which is a very useful preprocessing process for subsequent natural language processing. This module uses the word segmentation product ICTCLAS5.0 for word segmentation and part-of-speech tagging, because the imported new word data also includes part-of-speech tagging information, so it will not affect the accuracy of part-of-speech tagging. A survey of domain-related microblogs on the Sina Weibo platform found that nouns can more accurately describe users’ domain interests, and words with other parts of speech as keywords often introduce ambiguity, resulting in a decline in recommendation accuracy. Therefore, before keyword extraction and keyword expansion, part-of-speech filtering is performed on the results of user microblog word segmentation, and only nouns are retained. The preprocessing result is the user's historical microblog data.

As shown in Figure 2, the implementation process of the domain keyword extraction module is as follows:

(1) The present invention compares and selects the comparatively advanced TextRank algorithm at present through experiments in terms of keyword extraction, and the reasons are as follows: the TextRank algorithm overcomes the defect of the TFIDF method, and it does not need to calculate TF information, so it is not necessary to merge microblogs, and its Does not rely on external IDF information. In the calculation of TextRank, the assumption that words co-occurring with keywords with high domain relevance have high domain relevance is considered, so that the calculation of keyword weight is no longer linear calculation, thus overcoming the limitation of TFMF algorithm to a certain extent. Power law distribution problem. Therefore, it is more suitable for this application scenario than other algorithms.

(2) This module absorbs the idea of TextRank keyword extraction algorithm, and puts forward the TextRank for Weibo algorithm in combination with the characteristics of Weibo application scenarios. The TextRank for Weibo algorithm is a graph-based keyword extraction algorithm. Its inspiration comes from the idea of PageRank algorithm. In terms of graph construction, the traditional TextRank algorithm defines the weight of the connection edges between words based on the number of co-occurrences of words in a fixed-length sliding window in the document. Considering the characteristics of short microblog texts, the present invention uses the co-occurrence times of words in a microblog as the weight of the edges between words. Then use the PageRank algorithm on this undirected graph to calculate the weight of each word as a keyword. Each node in the definition diagram of the present invention represents a participle. If two words co-occur in a certain microblog of the user, the weight of the edge between the corresponding nodes is added by 1, and the final weight between the nodes is its Corresponding to the co-occurrence times of two words in Weibo.

After the undirected graph is determined, an algorithm similar to PageRank is used to iteratively generate the weight of each node. The weight of node V _i is updated according to the following formula:

Among them: V _i is the i-th node, TR(V _i ) is the weight of node V _i , w _ij is the weight of the edge between nodes V _i and V _j ; E(V _i ) is the weight of the edge connected by V _i Set; d is the damping coefficient of the iteration, which is set to 0.85. You can start the iteration with any initial value and iterate until it converges. The convergence condition is that the absolute difference between the sum of the weights of each node between this iteration and the last iteration is less than the specified value.

The main part of the UML class diagram of this module is shown in Figure 5. It is mainly divided into two stages: the construction of the undirected graph based on the co-occurrence relationship and the node weight calculation based on the graph.

The construction stage of the undirected graph based on the co-occurrence relationship first converts the word segmentations that appear in the user's historical microblogs into corresponding nodes. Then, in the word segmentation results of each microblog, all the binary word pairs that appear are constructed as edges, and the weight of the edges is the number of co-occurrences of the corresponding word pairs in the historical microblogs.

As shown in Figure 2, the implementation process of the user-defined keyword expansion module is as follows:

(1) Based on the defect of insufficient dynamicity of keyword extraction, the present invention introduces user-defined keywords to enhance the dynamicity of interest modeling. At the same time, in order to solve the problem of insufficient expression ability of self-defined keywords, the present invention proposes a method combining keyword extraction and user-defined keyword expansion to model user interests.

(2) Considering that the application scenario requirement of the present invention is: under the background of the domain-related corpus, several domain-related keywords defined by the user are expanded based on the domain correlation degree. Therefore, the present invention uses the co-occurrence information of words in domain-related corpus to calculate the similarity between words. This method is based on the assumption that words co-occurring in the same Weibo with user-defined keywords have a strong domain similarity with the keyword.

(3) The present invention absorbs the extended method P-IOLog for topic tags, but considering that the confidence degree of the extended word weight generated by this method is directly proportional to the size of the sample space, that is, the frequency of self-defined keywords in the corpus to be analyzed , that is, when the number of microblogs containing user-defined keywords is small, the error of the extended result generated by the P-IOLog algorithm is relatively large. Therefore, the present invention improves P-IOLog for this application scenario, and introduces a weight reduction factor, which increases as the frequency of user-defined keywords increases; at the same time, the consideration of the topic layer in the original method is removed , proposed the Probabilistic Inside-Outside Log for Weibo (referred to as P-IOW) method.

(4) The specific method is as follows. For a given user-defined keyword s, the calculation method of the domain correlation of word t with respect to s is as follows:

in:

Among them: s is a user-defined keyword; wf(t) is the number of microblogs containing the word t; t _M is the word contained in the most microblogs in the domain-related corpus, is the number of microblogs where t _M is located; is the number of microblogs that do not contain word s; wf(t∧s) is the number of microblogs that contain both word t and word s; N is the total number of user microblogs; s _p is a smoothing coefficient with a range of (0, 1) make Divisor by zero does not occur when it is zero. From the formula, it can be concluded that the value range of P-IOLogW is between log(s _p ) and log(1/s _p ) before formula (1) is multiplied by the down-weighting factor. The larger the value of the calculation result P-IOW, the higher the domain similarity between the word w and the user-defined keyword s. For the user-defined keyword itself, the upper limit of the value range of P-IOLogW will be assigned, that is, log(1/s _p ).

(5) At the same time, the present invention supports the user to input multiple self-defined keywords. For each custom keyword, the keyword expansion module will linearly add the expanded word vectors to obtain the final expanded vector. The specific algorithm of the expansion process is as follows:

1) For each user-defined keyword collection keyword, at first, calculate all expansion word weights of keyword based on P-IOW;

2) Map the extended word weight vector related to the keyword to the space of domain-related corpus word segmentation to form the extended word vector;

3) Linearly accumulate all the expanded word vectors to obtain the final expanded vector, which is the output result of the keyword expanded module.

The main part of the UML class diagram of this module is shown in Figure 6. Based on the co-occurrence information of other word segmentation and user-defined keywords, the user-defined keywords are extended, and the extended results are expressed in the form of vectors, so that they can be seamlessly integrated with key words. Word extraction results are combined.

As shown in Figure 2, the implementation process of the linear merge module is as follows:

(1) Linear merging according to user-defined weights refers to the process of merging two result vectors after automatic keyword extraction and expansion based on user-defined keywords are completed.

Both the traditional method of modeling user interest through keyword extraction and the method of expressing user interest through user-defined keywords have their own defects. Keyword extraction is mainly manifested in that users cannot manually adjust the algorithm results, and cannot meet users' temporary dynamic needs for topics in a certain field. Moreover, the algorithm cannot perform domain interest modeling when the user's history of Weibo is small, which is the so-called cold start phenomenon. The defects of user-defined keywords are mainly manifested in: first, it is difficult to ensure that user-defined keywords completely cover all the information in this field; Sex will make it impossible to extract a lot of information that is synonymous but uses different words.

Considering the defects of the above two methods, the present invention proposes a method of combining automatic field keyword extraction and user-defined keywords, and expands the word-based co-occurrence information of user-defined keywords. This method not only takes into account the comprehensive identification of domain information, but also dynamically adjusts keywords at any time according to changes in user interests; at the same time, user-defined keywords can solve the cold start phenomenon to a certain extent.

(2) Because the result vectors of keyword extraction and keyword expansion are not in the same value range, the two vectors need to be mapped to the same value range. This module uses the maximum value normalization method to normalize the two result vectors so that they are mapped to a unified value range. After normalization, linearly merge the two normalized vectors. The merging process supports user-defined weights for keyword extraction and keyword expansion. The module outputs a word vector representing the user's final domain interests.

For each component in the vector, the specific normalization method is as follows:

v _normal = v/v _max

Among them: v is the initial value of a certain component of the vector; v _max is the maximum value of all components of the vector. After normalization by the maximum value, all components of the vector are between (0, 1], and all components are non-zero. Then the vectors are linearly weighted and merged:

V _combine ＝r×V _kw-extract +(1-r)×V _kw-expand

Among them: r is the user-defined combination weight ratio, V _kw-exract is the result vector of keyword extraction, V _kw-expand is the result vector generated by keyword expansion, and the combined result is the keyword vector V _combine that describes the interest of the user field .

As shown in Figure 2, the implementation process of the similarity calculation and personalized recommendation module is as follows:

(1) After completing the generation of the keyword vector V _combine that characterizes the user's interest in the field, the correlation comparison module will perform word segmentation and word frequency statistics on each microblog to be filtered to generate a word frequency vector, and then combine the user interest keyword vector, The term frequency vector generated by the microblog to be recommended and the IDF information vector are subjected to dot multiplication to obtain the correlation between the microblog and the user's interest. The specific formula for calculating the correlation is as follows:

Wherein: the keyword vector of interest in the user field is V _combine ; the word frequency vector corresponding to microblog T to be recommended is W; the IDF vector of word segmentation is V _IDF ; L is the total number of vector space dimensions after word segmentation of user history microblogs; with are the components corresponding to t _i above V _combine and W respectively; IDF(t _i ) is the IDF value of t _i .

(2) Finally, the model calculates the domain relevance of each user's microblog. The system can recommend microblogs according to the obtained domain correlation. The form of recommendation is not limited, and it can be sorted according to the domain correlation degree from high to low, or filter out the microblog information whose domain correlation degree is greater than the specified threshold, etc. The present invention adopts the first presentation method.

As shown in Figure 2, the implementation process of the topic acquisition module is as follows:

(1) Perform text preprocessing on the microblogs in the recommended domain. The preprocessing work is the same as module 1, including stop word filtering, word segmentation and part-of-speech tagging.

(2) Use the processed data as input to train the LDA model, and cluster the terms into topics according to the term distribution of the topics;

(3) Calculate the correlation between the set of subject terms and the interest keywords in the user domain obtained in the linear combination module. The calculation basis is to use the co-occurrence information between the terms to measure the correlation. Finally, all subject term sets containing this keyword are considered relevant, and topics that do not contain any field interest keywords are regarded as having little relevance or irrelevance. The specific algorithm uses the correlation calculation method of module 5 .

(4) The degree of relevance represents the importance of the topic, and the topics are sorted according to the importance and presented to the user.

The present invention has designed a prototype system based on the Sina Weibo open platform to verify the results. The system sequence diagram of the present invention is shown in Figure 3, which shows the overall use process of the present invention:

1) First, after the user logs in to the system to verify the username and password, the system will call the OAuth verification module of Sina Weibo to verify the authorization of the Weibo account; if the user has not bound the Weibo account, it will use OAuth to bind the Weibo account Certainly.

2) After the login and authorization verification are completed, the system will update the user data, and locally persist the user-related Weibo information to ensure the timeliness of subsequent analysis work.

3) After the data is persisted locally, word segmentation and part-of-speech filters will be used to perform word segmentation and part-of-speech filtering on user-related microblogs in the database, so as to facilitate subsequent keyword extraction and keyword expansion. The user-related field microblog here mainly refers to the microblog information sent by the bound enterprise microblog, or it can be the microblog information sent by several field-related users specified by the user, or even external field-related corpus. The present invention does not limit the word breaker and part-of-speech tagger, and theoretically Chinese word breakers and part-of-speech taggers can be used. The ICTCLAS5.0 word segmentation and part-of-speech tagging system developed by the Computing Institute of the Chinese Academy of Sciences is used in the system corresponding to the present invention, and the external thesaurus data of the stuttering word segmenter is introduced to solve the problem that the thesaurus of the word segmenter is relatively old. The preprocessing result is the user's historical microblog data.

4) The field-related Weibo after word segmentation and part-of-speech filtering will be handed over to the keyword extraction module in the form of word segmentation and word frequency for keyword extraction. The system uses the method of TextRank for Weibo to extract keywords. The specific extraction results represent the interests of the field feature.

5) If the user defines the keyword of interest, in order to further enhance the recall ability of the user-defined keyword, the system will use the P-IOW keyword expansion method to expand the keyword based on domain information according to the user-defined keyword. The expansion results are expressed in the form of word segmentation vectors. If the user does not define interest keywords, this step will output an empty result.

6) After the keyword extraction based on TextRank for Weibo and the keyword expansion based on P-IOW are completed, the results will be linearly weighted and merged based on user-defined weights. Users can customize the proportion of keyword extraction and keyword expansion results. The output form is the merged word segmentation vector, which is used to represent the user's final domain interest.

7) In this step, the similarity calculation between the user interest and the microblog to be recommended is performed. First, the microblog to be recommended needs to be subjected to word segmentation and part-of-speech filtering (same as step 3), so as to convert it into the form of a word-segmented word frequency vector. Then calculate the dot product of the user domain interest vector, the converted vector of the microblog to be recommended, and the IDF value of the word segmentation, see the above discussion for details. The product is the similarity between the Weibo to be recommended and the user's domain interests.

8) Use the microblog text in the recommended domain as input, use the LDA model to implement clustering of terms, form a collection of subject terms, and finally calculate the importance of subject terms, and present them to users according to their importance.

The architecture diagram of the implementation system corresponding to the present invention is shown in Figure 4. The system is based on the technology stack structure of MySQL+JSP+Servlet+Twitter Bootstrap, and is divided into a data acquisition and preprocessing module, a field keyword extraction module, and a user-defined key Word expansion module, linear merge module, similarity calculation and personalized recommendation module, and topic acquisition module are composed of six components. The target users of the present invention are enterprise microblog users on the Sina microblog platform. By analyzing the historical microblog texts of their designated microblog accounts in several fields, the system can automatically identify the characteristics of the field where the user is located. At the same time, the system also supports users to dynamically define keywords to express their own dynamic needs in the field. The system will expand keywords according to the co-occurrence information of keywords in historical microblogs to enhance the semantic expression ability of keywords. Then, users can model fully personalized domain interests according to their own defined keyword weights. Finally, according to the interests of the field, the filter recommendation function and topic recommendation function for users to subscribe to Weibo are realized. It prevents users from viewing a large number of irrelevant microblogs one by one, which improves the work efficiency of enterprise marketing personnel.

Claims (8)

1. A domain information recommendation system based on a microblog platform, characterized in that it includes: data acquisition and preprocessing module, domain keyword extraction module, user-defined keyword expansion module, linear merging module, similarity calculation and personalization Recommendation module and topic acquisition module; among them: Data acquisition and preprocessing module: Obtain user-related microblog information data and perform preprocessing; preprocessing includes data stop word filtering, word segmentation and part-of-speech tagging; the preprocessing result is the user's historical microblog data, which is passed to the domain Keyword extraction module; if the user defines domain interest keywords, the preprocessing result is simultaneously passed to the user-defined keyword expansion module; Field keyword extraction module: Based on the preprocessing results, the keyword extraction is carried out unguided by the TextRank for Weibo algorithm modified based on the TextRank algorithm. first stage; in the construction stage of the undirected graph based on the co-occurrence relationship, firstly, the word segmentation appearing in the user’s historical Weibo is converted into the corresponding node; when constructing the connecting edge between the nodes, whether there is an edge between the nodes and the edge The weight determines the co-occurrence composition by the number of co-occurrences of two words in the same Weibo. The weight of the edge is the number of co-occurrences of words in the same Weibo. If two words co-occur in a user’s Weibo , then the weight of the edge between the nodes corresponding to the two words is increased by 1, and the final weight of the edge is the number of co-occurrences of the corresponding two words in Weibo; then based on the node weight calculation stage of the graph, the iterative calculation The weight of each stage, until the variation of the node weight converges to a certain threshold; after the iteration, the weight of each node is the importance of the participle it represents, and all the participle of the user are sorted according to the importance That is, the result of keyword extraction is obtained, so as to automatically identify the characteristics of the field where the user is located; User-defined keyword expansion module: Calculate the similarity between keywords based on the co-occurrence and distribution of keywords and the attribute information of their users, and use highly relevant words as the expansion results of target keywords; this module supports The user enters multiple custom keywords, and for each custom keyword, the extended word vectors obtained from the keywords will be linearly summed to obtain the final extended vector; the user-defined keyword expansion function ensures the user's Dynamic interest needs can be met in real time, and at the same time, the expression ability of user-defined keywords is greatly enhanced; Linear merge module: After the automatic extraction of domain keywords and the expansion based on user-defined keywords are completed, the maximum value normalization method is used to normalize the two result vectors, so that the results of keyword extraction and keyword expansion The vector is mapped to a unified value range; after normalization, the two normalized vectors are linearly merged, and the merge process supports user-defined keyword extraction and keyword expansion weights; the module outputs a A word vector representing the user's final domain interests; Correlation calculation and personalized recommendation module: After the linear merging module describes the keyword vectors of user interests in the field, word segmentation and word frequency statistics are performed on each to-be-filtered microblog to generate word frequency vectors, and then the user interest keyword vectors, to-be-recommended The word frequency vector generated by the microblog and the IDF information vector are dot-multiplied to obtain the correlation between the microblog and the user's interest. The correlation is the domain correlation of the microblog. By calculating the domain of each user's microblog Relevance, sorted from high to low according to the domain relevance, presents Weibo information to users, and realizes personalized domain Weibo recommendations for users; Topic Acquisition Module: Train the LDA model with the domain microblog text recommended to the user as input, and cluster the terms into topics according to the distribution of terms in the topic; combine the subject term set with the user domain interest keywords obtained in the linear combination module Items are calculated to obtain the importance of topics, and presented to users in order of importance, thereby completing topic discovery and recommendation. 2. The field information recommendation system based on microblog platform according to claim 1, characterized in that: the implementation process of the data acquisition and preprocessing module is as follows: (1) After the user logs in to the Weibo system, the user verification is first performed. After the verification is passed, the Weibo platform credentials associated with the user are automatically used to interact with the Weibo platform to verify the legitimacy of the user's identity on the Weibo platform; (2) Obtain relevant microblog texts that users follow and subscribe, and use the local database to persist the acquired data in a structured manner so that they can be read at any time; (3) Preprocessing the persistent microblog text, including stop word filtering, word segmentation and part-of-speech tagging; according to the characteristics of microblog text, the stop words are first filtered by pattern matching method, and then Optimized Chinese word segmentation and part-of-speech tagging for the Weibo scene, use the word segmentation product ICTCLAS5.0 for word segmentation and part-of-speech tagging, and filter the part-of-speech results of user Weibo word segmentation before keyword extraction and keyword expansion. Only keep nouns. The preprocessing result data is the user's historical microblog data. 3. The domain information recommendation system based on microblog platform according to claim 1, characterized in that: in the domain keyword extraction module, based on historical microblog data, the weight of each stage is iterated according to the algorithm idea of PageRank Calculated as follows: $\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; \&times;</span>\underset{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&Sigma;</span>}\frac{{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; k</span>}}}\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>$ Among them: V _i is the i-th node, TR(V _i ) is the weight of node V _i , w _ij is the weight of the edge between nodes V _i and V _j ; E(V _i ) is the weight of the edge connected by V _i Set; d is the damping coefficient of the iteration, which is set to 0.85. You can start the iteration with any initial value and iterate until it converges. The convergence condition is that the absolute difference between the sum of the weights of each node between this iteration and the last iteration is less than the specified value. 4. the domain information recommender system based on microblog platform according to claim 1, is characterized in that: in described user-defined keyword extension module, adopts improved P-IOW algorithm to calculate the similarity between keywords, The implementation process is as follows: For a given user-defined keyword s, the calculation method of the domain correlation of word t with respect to s is as follows: $\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\frac{\mathrm{<span\; class="notranslate">\; l</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; w</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; l</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; WF</span>}}_{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; g</span>}\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\\ \mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; g</span>}\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$ in: Among them: s is a user-defined keyword; wf(t) is the number of microblogs containing the word t; t _M is the word contained in the most microblogs in the domain-related corpus, is the number of microblogs where t _M is located; is the number of microblogs that do not contain word s; wf(t∧s) is the number of microblogs that contain both word t and word s; N is the total number of user microblogs; s _p is a smoothing coefficient with a range of (0, 1) , making When it is zero, the divisor will not be zero; it can be concluded from the formula that before formula (1) is multiplied by the weight reduction factor, the value range of P-IOLogW is between log(s _p ) and log(1/s _p ) , the larger the value of the calculation result P-IOW, the higher the field similarity between the word w and the user-defined keyword s. For the user-defined keyword itself, the upper limit of the value range of P-IOLogW will be given, that is, log (1/s _p ). 5. the field information recommendation system based on microblog platform according to claim 1, is characterized in that: in the user-defined keyword expansion module, the extended word vector that keyword is expanded is carried out linear addition, thereby obtains When the final expansion vector is obtained, the specific algorithm of the expansion process is as follows: (1) for each user-defined keyword collection keyword, at first, calculate all expansion word weights of keyword based on P-IOW; (2) Map the extended word weight vector related to the keyword to the space of domain-related corpus word segmentation to form an extended word vector; (3) Linearly accumulate all the expanded word vectors to obtain the final expanded vector, which is the output result of the keyword expanded module. 6. The field information recommendation system based on microblog platform according to claim 1, characterized in that: the implementation process of the linear merging module is as follows: (1) Use the maximum value normalization method to normalize the vectors generated by keyword extraction and keyword expansion; (2) For each component in the vector, the specific normalization method is as follows: v _normal = v/v _max Among them: v is the initial value of a certain component of the vector; v _max is the maximum value of all components of the vector. After normalization by the maximum value, all components of the vector are between (0, 1], and all components are non- Zero, followed by vector linear weighted merging: V _combine ＝r×V _kw-extract +(1-r)×V _kw-expand Among them: r is the user-defined combined weight ratio, V _kw-extract is the result vector of keyword extraction, V _kw-expand is the result vector generated by keyword expansion, and the combined result is the keyword vector V _combine that describes the interest of the user field . 7. The field information recommendation system based on microblog platform according to claim 1, characterized in that: in the described similarity calculation and personalized recommendation module, the specific formula for field correlation calculation is as follows: ${\mathrm{<span\; class="notranslate">\; Relevance</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; I</span>}\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span>\underset{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}_{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}\mathrm{<span\; class="notranslate">\; I</span>}\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$ Wherein: the keyword vector of interest in the user field is V _combine ; the word frequency vector corresponding to microblog T to be recommended is W; the IDF vector of word segmentation is V _IDF ; L is the total number of vector space dimensions after word segmentation of user history microblogs; with are the components corresponding to t _i above V _combine and W respectively; IDF(t _i ) is the IDF value of t _i . 8. A method for recommending domain information based on a microblog platform, characterized in that it is divided into data acquisition and preprocessing, domain keyword extraction, user-defined keyword expansion, linear merging, similarity calculation and personalized recommendation, and theme Get six steps, implemented as follows: (1) Obtain user-related microblog information for data preprocessing; preprocessing includes filtering stop words using pattern matching; using word segmentation system ICTCLAS5.0 for word segmentation and part-of-speech tagging; the preprocessing result is the user's history Weibo data; (2) Preprocessing results carry out field keyword extraction, keyword extraction will use the TextRank for Weibo algorithm modified based on TextRank algorithm of the present invention to carry out without guidance, and TextRank for Weibo algorithm is divided into the structure of the undirected graph based on co-occurrence relationship and The graph-based node weight calculation has two stages; the construction stage of the undirected graph based on the co-occurrence relationship first converts the word segmentations that appear in the user's historical microblogs into corresponding nodes; then for each microblog word segmentation result, all occurrences The edge is constructed for the binary word pair, and the weight of the edge is the number of co-occurrences of the corresponding word pair in historical microblogs; The graph-based node weight calculation stage iteratively calculates the weight of each stage according to the algorithm idea of PageRank until the change of node weight converges to a certain threshold; (3) If the user defines the domain interest keywords, the preprocessing results will be passed to the keyword expansion module; the keyword expansion will be performed using the improved P-IOW algorithm based on the P-IOLog algorithm, using co-occurrence based on keywords , distribution, and the attributes of the users they belong to to calculate the similarity between keywords, and use highly relevant words as the extended results of target keywords; at the same time, users are supported to input multiple custom keywords; for each custom Keywords, the keyword expansion module will linearly add the expanded word vectors to obtain the final expanded vector; (4) The results of the two are normalized according to the user-defined weights, and the two result vectors are normalized by the maximum value normalization method, so that they are mapped to a unified value range; after normalization, the The two normalized vectors are linearly merged. The merge process supports user-defined keyword extraction and keyword expansion weights. The merged results are used for correlation comparison by related recommendation modules to generate correlation scores for recommended microblogs; (5) Segment the microblogs to be recommended subscribed by the user and vectorize them according to the word frequency, and then according to the identified user interest, point the user interest keyword vector, the word frequency vector generated by the microblog to be recommended, and the IDF information vector Multiplication operation, using the method of vector space point multiplication to calculate the domain correlation; (6) Take the field microblog text recommended to the user as the input set, realize term clustering based on the LDA strategy, complete the topic discovery, and then combine the subject term set with the user domain interest keywords obtained in the linear merge module Correlation calculation, establishing the importance of topics, and recommending users based on the importance of topics.