CN110188208A - A method and system for querying and recommending information resources based on knowledge graphs - Google Patents

Abstract

The present invention proposes a method and system for querying and recommending information resources based on knowledge graphs. The method first preprocesses the knowledge graphs, uses representation learning methods to map the knowledge graphs into low-dimensional dense vector spaces, and obtains vector representations of entities ; Then calculate the user's interest in information resources according to the user's historical behavior, and construct a user interest model by combining the vectorized representation of information resources and the user's interest in information resources; by calculating the similarity between resources and resources, users and resources To achieve accurate recommendation of information resources. The present invention combines knowledge map representation learning with user interest models to provide users with personalized services, taking into account the internal relationship of knowledge and user interests, and recommending resources that are related to the query content and in line with the user's interests according to the name of the resource input by the user. Information resources make personalized query and recommendation more professional and targeted.

Description

A method and system for querying and recommending information resources based on knowledge graphs

technical field

The present invention relates to the technical field of knowledge graph and recommendation, in particular to a method and system for querying and recommending information resources based on knowledge graph.

Background technique

In recent years, the vigorous development of information technology has driven the pace of informatization in all walks of life. The Internet, Internet of Things, cloud computing, etc. have gradually integrated into people's daily life, which has brought about explosive growth of data. The huge information resource library provides users with rich information, but it also brings the problem of resource overload, which makes users spend a lot of time in selecting the information resources they are interested in. The personalized query recommendation based on the user's historical behavior data can effectively alleviate the problem of resource overload.

The recommendation system is one of the effective means to deal with information overload at present. It analyzes the user's preferences based on the user's historical behavior, and actively chooses what he likes, such as which item the user buys, which news to read, and which to listen to in the various decision-making processes. music.

The collaborative filtering algorithm was first proposed, and it is also the most researched and applied recommendation technology. It depends on the behavior of users and pays attention to the association between users and items. It is mainly divided into two different algorithms, namely user-based algorithm and user-based The algorithm of the item. The basic principle of user-based collaborative filtering is to find users with similar behaviors, and recommend resources that users like with similar interests; item-based collaborative filtering recommendation aims to recommend to users similar items that are similar to the items he was interested in. For projects, similarity does not refer to the similarity of project content, but to use users' evaluation or behavior of projects to mine the similarity between projects. However, the collaborative filtering algorithm is too dependent on user behavior, resulting in no basis for recommendation when there are new users or new items in the system. In addition, there are tens of millions of items in real life, and the items that interact with users are often a small number. Mining similar items only through user behavior on items will lead to poor collaborative filtering algorithms. In response to this problem, most of the current research methods are to introduce auxiliary information as the input of the recommendation algorithm.

The knowledge map contains rich semantic information, which aims to represent entities or concepts in the real world and the relationship between them in a structured form. Its essence is a huge semantic network map, which integrates massive knowledge into a more It is displayed in front of the user in an intuitive way, consisting of nodes and edges, where nodes represent entities or concepts, and edges represent the relationship between entities or the attributes of entities. The knowledge graph introduces more semantic relationships and provides different types of relationship connections. The introduction of the knowledge graph into the recommendation system can make full use of the rich semantic information in the knowledge graph, so that users' interests can be discovered in depth, and the recommendation results can be avoided. Based on a single type, it improves the accuracy, diversity and interpretability of the recommendation system, thereby improving user satisfaction with the recommendation results.

At present, there have been some researches on recommendation methods based on knowledge graphs, such as path-based recommendation methods, which need to construct a specific path connecting two entities, but the method of manually constructing paths is difficult to reach the optimum in practice; graph-based algorithms The recommendation method intuitively utilizes the knowledge graph as a feature of the semantic network graph, and uses algorithms such as random walk to sample nodes in the graph, but the graph algorithm has poor portability and high computational complexity. When faced with a large knowledge graph, it is difficult to achieve real-time calculate.

Contents of the invention

Purpose of the invention: Aiming at the defects and deficiencies of the existing technology, the present invention provides a method for querying and recommending information resources based on knowledge graphs, which takes into account the internal relationship of knowledge and user interests, and quickly and efficiently recommends to users according to the name of the resource entered by the user. Information resources that are relevant to the query and in line with the user's interests.

Technical solution: According to the first aspect of the present invention, a method for querying and recommending information resources based on knowledge graphs is provided, the method comprising the following steps:

(1) Use the knowledge graph representation learning method to map the knowledge graph into a low-dimensional dense vector space, and realize the vectorized semantic representation of the information resources in the knowledge graph;

(2) Calculate the user's interest in information resources according to the user's historical behavior;

(3) Combining the user's interest in information resources and the vectorized semantic representation of information resources, construct a user interest model;

(4) Calculate the similarity between the information resource and other information resources according to the information resource inquired by the user, and take the information resources with the similarity TOP-M to form a candidate resource set;

(5) Calculate the similarity between the information resources in the candidate resource set and the user, and select the information resources with similarity TOP-N from the candidate resource set to form a recommendation list.

Further, said step 1 includes:

(11) Select a specified number of triples (h, r, t) from the knowledge graph, which are called positive triples, where h and t represent the head entity and tail entity respectively, and r represents the relationship between two entities relation;

(12) Utilize the negative sampling algorithm to replace the head entity or the tail entity of the positive example triple to obtain the negative example triple;

(13) Use the representation learning model to iteratively train positive triples and negative triples until convergence, and obtain the vector representation of the entity V _i ={v ₁ ,v ₂ ...,v _m }, where m represents the dimension.

Further, said step 12 includes:

(121) In all triples of relation r, count the average number of tail entities corresponding to each head entity, which is recorded as tph; count the average number of head entities corresponding to each tail entity, which is recorded as hpt;

(122) For a positive triplet (h, r, t), extract entities to replace the head entity h and tail entity t, replace the head entity with the probability of p, replace the tail entity with the probability of 1-p, and generate a negative Example triplet, where the formula for calculating the replacement probability p is:

Further, said step 2 includes:

(21) Collect logs containing user behavior, including the name of resources browsed by users, the length of resource content, and the duration of browsing;

(22) Establish a multiple linear equation based on whether to click to browse, browsing time, and browsing speed to calculate the user's interest in the resource.

Further, the step 22 includes:

(221) The user clicks to browse a certain information resource i, and records the click interest degree as C _i ;

(222) Calculate the browsing interest degree R _i according to the user's browsing time t _i of the resource i and the user's average browsing speed:

Among them, _t1 represents the user’s minimum browsing time for resource i, _t2 represents the user’s maximum browsing time for resource i, S is the user’s average browsing speed, L is the total length of the user's browsing resources, and T is the total time of the user's browsing resources;

(223) Combining click interest and browsing interest to obtain the user's interest in resource i I _i = ω ₁ C _i + ω ₂ R _i , where ω ₁ and ω ₂ represent click interest and browsing interest in the calculation total The weight occupied by the degree of interest, and ω ₁ +ω ₂ =1.

Further, the user interest model in the step 3 is: in represents the weight of the user's past interest vector, Represents the weight of the current interest vector, U _present represents the user's current updated interest vector representation, U _previous represents the user's past interest vector representation, I _i represents the user's interest in the i-th resource, V _i represents the i-th resource A vector representation of a resource.

Further, after step 1, the method further includes: calculating the distance between resources according to the vector of information resources, judging their similarity according to the distance, gathering similar information resource entities to form a cluster, and dividing different information resource entities into in different clusters.

Further, in the step 4, the similarity between two resources is calculated by the cosine distance, and in the step 5, the similarity between the information resource and the user interest is calculated by the cosine distance.

According to the second aspect of the present invention, a knowledge graph-based information resource query recommendation system is provided, the system comprising:

The data preprocessing module is used to use the knowledge graph representation learning model to embed the knowledge graph into a low-dimensional vector space, and obtain the vectorized representation of entities, relationships, and attributes through learning;

A user interest model building block, which is used to analyze user behavior, understand user interest, and construct a user interest model; and

The query recommendation module is used to obtain a candidate resource set according to the resource input by the user for query, and select resources close to the user's interest from the candidate resource set for recommendation.

Beneficial effects: the recommendation method based on knowledge map representation learning in the present invention integrates the knowledge map as a data set with rich language and strong logical reasoning ability into the traditional recommendation algorithm, and uses representation learning to express each entity and relationship of the knowledge map as Dense low-dimensional real-valued vectors reduce the high-dimensionality of knowledge graphs, so that in low-dimensional vector spaces, the semantic connections between entities can be efficiently calculated, reducing the additional computational burden caused by the introduction of knowledge graphs, thereby enhancing the application of knowledge graphs. flexibility. Specifically reflected in:

1. Make full use of the rich semantic information in the knowledge graph, and make up for the defect that the traditional collaborative filtering algorithm does not consider the semantic information of the recommended items. Use the knowledge map representation learning to map the entities and relationships in the knowledge base into a low-dimensional dense vector space, complete the semantic representation of entities and relationships, and significantly improve the computational efficiency. The semantic similarity between entities can be measured by cosine distance , and an entity has a dense vector corresponding to it, which also alleviates the problem of data sparsity.

2. In the process of model training, the Bernoulli negative sampling algorithm is used, which effectively avoids the introduction of wrong negative triplets by setting different probabilities of replacing the head entity or tail entity.

3. Construct a user interest model based on the analysis of user historical behavior, map user interest to a low-dimensional vector space, make user interest and information resources become points in the vector space, and calculate the relationship between resources and resources, users and resources The similarity can be known by the distance, and the calculation process is concise.

4. Combining knowledge graph representation learning with user interest models to provide users with personalized services, taking into account the internal relationship of knowledge and user interests, and recommending resources that are related to the query content and in line with user interests according to the resource name entered by the user. Information resources make personalized query and recommendation more professional and targeted.

Description of drawings

FIG. 1 is an overall flowchart of a recommendation method according to an embodiment of the present invention;

Fig. 2 is a log preprocessing result diagram according to an embodiment of the present invention;

Fig. 3 is a flow chart of building a user interest model according to an embodiment of the present invention;

Fig. 4 is a block diagram of a recommendation system according to an embodiment of the present invention.

Detailed ways

The technical solution of the present invention will be further described below in conjunction with the accompanying drawings. It should be understood that the embodiments provided below are only intended to disclose the present invention in detail and completely, and fully convey the technical concept of the present invention to those skilled in the art. The present invention can also be implemented in many different forms, and does not Limited to the embodiments described herein. The terms used in the exemplary embodiments shown in the drawings do not limit the present invention.

In one embodiment, taking the query recommendation of water conservancy information resources as an example, the knowledge map of the water conservancy field is introduced as auxiliary information in the query recommendation process, and the knowledge map representation learning method is used to map the entities and relationships in the knowledge map to low-dimensional dense vectors In the space, the semantic representation of entities and relationships is realized, which makes up for the defect that traditional recommendation algorithms do not consider semantic information. Then, based on the analysis of user browsing behavior and browsing content, a low-dimensional user interest model is constructed. Finally, the knowledge graph and the user interest model are combined to build an information resource query recommendation system based on the knowledge graph in the water conservancy field, so as to accurately recommend water conservancy information resources that meet the user's interests according to the user's query.

Figure 1 is a flowchart of an information resource query and recommendation method based on knowledge graphs in the water conservancy field. As shown in Figure 1, the implementation process of this method includes the following steps:

Step 1. Use the knowledge graph representation learning model to map the knowledge graph into a low-dimensional dense vector space, and realize the vectorized semantic representation of the water conservancy information resources in the knowledge graph.

The knowledge graph uses information resources as its conceptual entity nodes, and uses the relevant attribute information of information resources as its feature label nodes. The edges of the two nodes represent the relationship between entities or the attributes of entities. Using triplet representation is (h, r, t) or (e, a, v), where h and t in (h, r, t) represent the head entity and tail entity respectively, and r represents the distance between two entities relationship, e in (e,a,v) represents the entity, and a and v represent the attribute and attribute value of the entity.

Specifically, Step 1 includes:

Step 1.1, select a certain number of triples (h, r, t) from the knowledge graph, which are called positive triples, such as the triples in the knowledge graph in the field of water conservancy (Yantan Reservoir, project scale, large Type 1);

Step 1.2, use the negative sampling algorithm to replace the head entity or tail entity of the positive triplet, and generate a negative triplet, that is, a wrong positive triplet. The specific steps are:

Step 1.2.1, in all triples of the relation r, count the average number of tail entities corresponding to each head entity, denoted as tph; count the average number of head entities corresponding to each tail entity, denote as hpt;

For example, for a knowledge graph with 3 triples (h ₁ ,r,t ₁ ), (h ₁ ,r,t ₂ ), (h ₂ ,r,t ₃ ), the number of tail entities corresponding to h ₁ is 2, the number of tail entities corresponding to h ₂ is 1, then In the same way,

Step 1.2.2, define the replacement probability formula From the above tph and hpt can be obtained

Step 1.2.3, for a positive triplet (h, r, t), extract entities in the knowledge graph to replace the head entity h or tail entity t in the positive triplet, thereby generating a new triplet group, this triplet is considered as a negative triplet. Replace the head entity with a probability of p, and replace the tail entity with a probability of 1-p, thus breaking positive triples and generating negative triples.

Step 1.3, use the representation learning model to iteratively train positive triplets and negative triplets until convergence.

Indicates that the learning model updates the parameters in the model through continuous iteration, that is, the loss function. After multiple iterations, the loss converges, that is, the minimum value of the loss function is obtained, and the entity vector, relationship vector and attribute vector converge to the optimum. The loss function It is defined as follows:

Where (h, r, t) represents a positive triplet, (h′, r, t′) represents a negative triplet, γ is the set marginal value, and the symbol [] ₊ is the hinge loss function. ‖h+rt‖ means the distance between the sum of the vectors of the head entity h and the relationship r and the difference between the vector of the tail entity t. In the model, for a positive triplet, it is expected that the value of ‖h+rt‖ should be as small as possible. , for negative triplets, it is expected that the larger the value of ‖h′+rt′‖, the better, so that in the experiment, the training model can be used to distinguish positive and negative samples.

In order to speed up the convergence, the data is initialized and normalized. First, the vectors of entities and relationships are initialized with a uniform distribution:

k represents the specified vector dimension, after initialization, it is normalized:

In each iteration process, the entity vector needs to be normalized first. In the process of model training, if all triples are iteratively trained, the cost will be very high. Therefore, in order to speed up the convergence, the small batch gradient descent algorithm is used as the training algorithm of the model, that is, at each iteration, from the water conservancy domain knowledge graph Select the small batch of training triples as a reference for training to determine the update direction of the model, and then update the parameters through a gradient step with a constant learning rate, that is, the loss function.

When the model is trained to convergence, the entities in the knowledge map can be mapped to the corresponding positions in the low-dimensional space, so that entities with the same attributes or the same relationship are close in distance, and the model training ends. Obtain the vector representation of the entity V _i ={v ₁ , v ₂ . . . , v _k }.

Step 2, clustering information resources.

According to the vector of information resources, calculate the cosine distance between resources, judge their similarity, gather similar information resource entities to form a cluster, and divide different information resource entities into different clusters, similar information resources are Information resources with the same attributes or the same relationship, while different information resources are information resources that do not exist or have less of the same attributes or the same relationship. After clustering, similar information resource entities are gathered into a cluster, which reduces the amount of calculation for the screening of candidate resource sets in subsequent steps and effectively improves efficiency.

Step 3, according to the user's historical behavior, calculate the user's interest in information resources, the specific steps are as follows:

Step 3.1, collect logs containing user behavior, including information such as the name of the resource browsed by the user, the length of the content of the resource, and the duration of browsing;

In the embodiment, the log component is used for log recording, and the js buried point technology is used to collect behavior data of users browsing water conservancy information resources. The log contains the operation of the system, such as the connection status of the database, the error information of the system, such as the error inside the server or the program, and the user-defined log output content, such as the debugging information of the program, the user's behavior data, etc. According to requirements, the system log needs to be filtered to obtain a log containing only user behavior data.

Figure 2 is the preprocessed log, where _ip is the user's ip address, which is the unique identifier of the user; _url is the current URL address; _refer is the URL address that the user visited last; _millisecond is the URL that the user entered the page The number of milliseconds, which is the number of milliseconds since 1970.1.1 in long type, which is convenient for calculating the time interval; _id is the unique identifier of the water conservancy information resource browsed by the user; _name is the name of the water conservancy information resource; _length is the content length of the information resource , which represents the length of the summary information of the information resource.

Step 3.2, starting from the three aspects of click browsing, browsing time, and browsing speed, based on multiple linear equations, abstracting user interest into numbers to calculate user interest in resources. The specific steps are:

Step 3.2.1, when the user clicks to browse a piece of information resource i, record the click interest as C _i ;

Users only have two operations on resources, clicking and not clicking, and the user's click interest is defined as:

From the log in Figure 2, it can be seen that the user clicks to browse the water conservancy information resource of "Danjiangkou Hydropower Station", and record C _i =1.

Step 3.2.2, calculate the user's browsing interest degree R _i according to the user's browsing time and browsing speed, wherein the browsing speed can be calculated according to the browsing time and resource content length;

The longer the user browses each resource, the higher the user's interest in the resource; on the contrary, the shorter the time, the lower the user's interest in the resource. In the case that the average reading speed of users is basically stable, the slower the user browses each resource, it means that they spend more time reading, and it can be judged that the higher their interest in the resource is; Low. The user's browsing interest is defined as:

Among them, t _i represents the duration of user browsing resource i, and t ₁ represents the minimum browsing time. When t _i <t ₁ , it means that the user has no interest in the resource or clicks on the resource details page by mistake; t ₂ represents the maximum browsing time, when t i When _i >t ₂ , it is considered that the user may stay on the page during the browsing process and deal with other things. In order to prevent these situations from affecting the calculation of the user interest degree, the browsing interest degree of these situations is counted as 0. S is the user's average browsing speed. Since different users have different reading abilities, their reading speed is stable. According to the user's historical behavior, the user's average browsing speed is calculated according to the following formula:

Wherein L is the total length of the user's browsing resources, and T is the total time of the user's browsing resources.

From the log in Figure 2, it can be seen that the content length "_length" of the resource is 119 characters, and the browsing time can be obtained by subtracting the time difference between the two actions before and after. The browsing time t of the resource is about 35 seconds. For example, assuming that the user's average browsing speed S is 7.5 words/second, the calculated browsing interest degree R _i is 2.21.

Step 3.2.3, combining click interest and browsing interest, the user's interest in resource i is defined as I _i =ω ₁ C _i +ω ₂ R _i , where ω ₁ and ω ₂ represent click interest and browsing interest The weight it occupies when calculating the total interest degree, and ω ₁ +ω ₂ =1.

For example, ω ₁ and ω ₂ take values of 0.2 and 0.8 respectively, and finally the user's comprehensive interest degree I _i for the resource is 1.97.

Step 4: Combining the user's degree of interest in information resources with the vectorized semantic representation of information resources, a user interest model is constructed.

The user interest model expresses user interest as a dense low-dimensional real-valued vector, whose dimension is the same as that of the information resource entity vector. The purpose is to map user interest to the low-dimensional space where the entity is located. The definition formula is U={u ₁ ,u ₂ ...,u _m }.

Fig. 3 is the construction process of the user interest model, combining the user's interest degree I _i to the information resource i and the vectorized semantic representation of the information resource i V _i ={v ₁ ,v ₂ ...,v _m } to construct the user interest model, Its formula is, in represents the weight of the user's past interest vector, Represents the weight of the current interest vector, U _present represents the user's current updated interest vector representation, U _previous represents the user's past interest vector representation, I _i represents the user's interest in the i-th resource, V _i represents the i-th resource A vector representation of a resource.

In the end, both the knowledge graph and user interests are mapped to the low-dimensional space, so that user interests and water conservancy information resource entities become points in the low-dimensional space, and the similarity can be judged by calculating the distance between resources and resources, and between resources and users.

Step 5: According to the information resource queried by the user, the similarity between the information resource and other information resources is calculated, and information resources with similarity TOP-M are selected to form a candidate resource set.

The candidate resource set refers to a collection of similar information resources, and is the source of information resources in the final recommendation list, which ensures that the final recommended information resources are related to the user's query content. In a low-dimensional vector space, the similarity between two resources can be calculated by the cosine distance, and the calculation formula is:

Where e _t is an entity other than entity e _i in the knowledge graph, and V _i and V _t represent the vectors of entity e _i and entity e _t . After calculation, M water information resources with high similarity with the water information resource entity e _i are obtained to form a candidate resource set D={d ₁ ,d ₂ ,...,d _M }.

Step 6: Calculate the similarity between the information resources in the candidate resource set and the user, and filter out information resources with a similarity of TOP-N from the candidate resource set to form a recommendation list.

The candidate resource set is the prerequisite for generating the final recommendation list, which ensures that the recommended content is related to the query content. To provide personalized recommendation services, it is also necessary to combine the user interest model to select resources close to the user's interest from the candidate resource set to generate the final recommendation. list.

In the same low-dimensional space, if the distance between the user and the resource is close, it means that the user has a high degree of liking for the resource; on the contrary, if the distance between the user and the resource is far, it means that the user does not pay attention to the resource. The cosine similarity formula is used to calculate the similarity between the user and the water conservancy information resources in the candidate resource set, and N entities with high similarity to the user's interests are screened out from the candidate resource set to generate TOP-N recommendations.

Fig. 4 is a module diagram of an information resource query recommendation system based on a knowledge graph, including a data preprocessing module, a user interest model building module, and a query recommendation module. The data preprocessing module includes a knowledge graph representation unit, and uses a knowledge graph to represent a learning model Embed the knowledge graph into a low-dimensional vector space, and obtain vectorized representations of entities, relationships, and attributes through learning; the user interest model building module analyzes user behavior, understands user interests, and constructs a user interest model; the query recommendation module It mainly obtains the candidate resource set according to the resource input by the user, and selects resources close to the user's interest from the candidate resource set.

Specifically, the knowledge graph representation unit selects a certain number of positive triples (h, r, t) from the knowledge graph, uses the negative sampling algorithm to replace the head entity or tail entity of the positive triples, and generates negative triples Tuples, use the model to iteratively train positive triples and negative triples to convergence, realize the mapping of the entities in the knowledge map to the corresponding positions in the low-dimensional space, and obtain the vector representation of the entity V _i ={v ₁ , v ₂ ..., v _k }.

As a preferred embodiment, the data preprocessing module further includes a clustering unit, which uses a clustering method to cluster the information resource entities. The clustering unit calculates the cosine distance between resources according to the vector of information resources, judges their similarity, gathers similar information resource entities to form a cluster, and divides different information resource entities into different clusters, similar Information resources are information resources with the same attributes or the same relationship, and different information resources are information resources that do not exist or have less of the same attributes or the same relationship.

The user interest model building module includes a log collection unit, a log processing unit, a log analysis unit, and a user interest model building unit. The log collection unit collects logs containing user behavior, including information such as the name of the resource browsed by the user, the length of the resource content, and the browsing time ; The log processing unit filters the system log to obtain a log containing only user behavior data; the log analysis unit starts from the three aspects of click browsing, browsing time, and browsing speed, and based on multiple linear equations, abstracts user interest into The number calculates the user's interest in resources; the user interest model construction unit combines the user's interest in information resources with the vectorized semantic representation of information resources to construct a user interest model.

The query recommendation module includes a candidate resource set unit and a recommendation list unit. The candidate resource set unit calculates the similarity between two resources through the cosine distance, and takes information resources with a similarity TOP-M to form a candidate resource set; the recommendation list unit uses the cosine distance The similarity formula calculates the similarity between the user and the information resources in the candidate resource set, and selects N entities with high similarity with the user's interests from the candidate resource set to generate TOP-N recommendations.

For the specific calculation formulas involved in the above modules, reference may be made to the corresponding formulas in the method embodiments, which will not be repeated here.

The present invention provides a knowledge graph-based information resource query recommendation method and recommendation system, which combines knowledge graph representation learning with a user interest model to achieve accurate recommendation of information resources that meet the user's interest according to the user's query. The knowledge map is a data set with rich semantics and strong logical capabilities, including structured and unstructured data, and in-depth knowledge of the internal relationship. The invention integrates connotative knowledge and user preference, making the recommendation algorithm more authoritative, professional and pertinent. Moreover, the recommendation system that incorporates knowledge graphs can also provide explanations, allowing users or system designers to know why these items are recommended, which helps to improve efficiency, persuasiveness, and user satisfaction of the recommendation system. Representation learning refers to the use of numbers, such as matrices and vectors, to express something in the real world. This expression is conducive to subsequent classification or decision-making problems, making subsequent tasks more effective with less effort. In the same way, knowledge graph representation learning aims to transform entities and relations into vectors in low-dimensional space without changing the internal structure of knowledge graphs. Entity and relationship vectorization significantly improves computational efficiency. The semantic similarity between entities can be measured by Euclidean distance or cosine distance. At the same time, an entity has a dense vector corresponding to it, which also alleviates the problem of data sparsity.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (10)

1. a kind of information resources of knowledge based map inquire recommended method, which is characterized in that the described method comprises the following steps:

(1) it indicates that learning method maps to knowledge mapping in the dense vector space of low-dimensional using knowledge mapping, realizes to knowing Know the vectorization semantic expressiveness of the information resources in map；

(2) according to user's history behavior, user is calculated to the interest-degree of information resources；

(3) it combines user to the interest-degree of information resources and the vectorization semantic expressiveness of information resources, constructs user interest model；

(4) according to the information resources of user query, the similarity of the information resources Yu other information resource is calculated, similarity is taken The information resources of TOP-M form candidate resource collection；

(5) similarity for calculating information resources and user that candidate resource is concentrated filters out similarity from candidate resource concentration The information resources of TOP-N form recommendation list.

2. the information resources of knowledge based map according to claim 1 inquire recommended method, which is characterized in that the step Rapid 1 includes:

(11) triple (h, r, t) of specified quantity, referred to as positive example triple are chosen from knowledge mapping, wherein h, t distinguish Head entity, tail entity are represented, r indicates the relationship between two entities；

(12) using the head entity or tail entity of negative sampling algorithm replacement positive example triple, negative example triple is obtained；

(13) using the extremely convergence of learning model repetitive exercise positive example triple and negative example triple is indicated, the vector table of entity is obtained Show V_i={ v₁, v₂..., v_m, wherein m indicates dimension.

3. the information resources of knowledge based map according to claim 2 inquire recommended method, which is characterized in that the step Rapid 12 include:

(121) in all triples of relationship r, the mean number of the corresponding tail entity of every head entity is counted, tph is denoted as； The mean number for counting the corresponding head entity of each tail entity, is denoted as hpt；

(122) for a positive example triple (h, r, t), entity is extracted to replace an entity h and tail entity t, is replaced with the probability of p An entity is changed, tail entity is replaced with the probability of 1-p, generates negative example triple, wherein the calculation formula of replacement Probability p are as follows:

4. the information resources of knowledge based map according to claim 2 inquire recommended method, which is characterized in that the step Rapid 2 include:

(21) log comprising user behavior is collected, resource name, resource content length, browsing duration including user's browsing；

(22) according to whether click browsing, browsing time, surfing establish multiple linear equation, user is to the emerging of resource for calculating Interesting degree.

5. the information resources of knowledge based map according to claim 4 inquire recommended method, which is characterized in that the step Rapid 22 include:

(221) user clicks certain information resources i of browsing, and remembering that it clicks interest-degree is C_i；

(222) according to user to the browsing duration t of resource i_iIts navigation interest degree R is calculated with the average surfing of user_i:

Wherein t₁Indicate minimum browsing time of the user to resource i, t₂User is indicated to the maximum browsing time of resource i, S is to use The average surfing at family,L is the total length that user browses resource, and T is the total time that user browses resource；

(223) comprehensive to click interest-degree and navigation interest degree, user is obtained to the interest-degree I of resource i_i=ω₁C_i+ω₂R_i, wherein ω₁、ω₂It represents and clicks interest-degree and the navigation interest degree weight shared when calculating total interest-degree, and ω₁+ω₂=1.

6. the information resources of knowledge based map according to claim 5 inquire recommended method, which is characterized in that the step User interest model in rapid 3 are as follows:WhereinRepresent the past interest vector of user Shared weight,Represent weight shared by current interest vector, U_presentIndicate user's currently updated interest vector It indicates, U_previousIndicate that user goes over the vector expression of interest, I_iIndicate interest-degree of the user to i-th resource, V_iIndicate i-th The vector of resource indicates.

7. the information resources of knowledge based map according to claim 1 inquire recommended method, which is characterized in that the side Method is after step 1 further include: will according to its similarity of Distance Judgment according to the distance between the vector computing resource of information resources Similar information resources entity aggregates form a cluster, and different information resources entity division is into different clusters.

8. the information resources of knowledge based map according to claim 1 inquire recommended method, which is characterized in that the step The similarity between two resources is calculated by COS distance in rapid 4, information resources are calculated by COS distance in the step 5 Similarity between user interest.

9. a kind of information resources of knowledge based map inquire recommender system, which is characterized in that the system comprises:

Data preprocessing module is led to for indicating that knowledge mapping is embedded in low-dimensional vector space by learning model using knowledge mapping The vectorization that overfitting obtains entity, relationship and attribute indicates；

User interest model constructs module, for analyzing user behavior, understands the interest of user, constructs user interest mould Type；And

Recommending module is inquired, for the resource acquisition candidate resource collection according to user input query, concentrates and screens in candidate resource The resource for interest of being close to the users out is recommended.

10. the information resources of knowledge based map according to claim 9 inquire recommender system, which is characterized in that described Data preprocessing module is also used to the distance between the vector computing resource according to information resources, according to its similarity of Distance Judgment, Similar information resources entity aggregates are formed into a cluster, different information resources entity division is into different clusters.