CN116662564A - A service recommendation method based on deep matrix factorization and knowledge graph - Google Patents

Abstract

The invention relates to a service recommendation method based on deep matrix decomposition and a knowledge graph, which utilizes government service item data to construct a government service knowledge graph, models entity context and entity description text information of the knowledge graph through a knowledge representation method, and combines knowledge representation learning and personalized recommendation through a combined learning mode to obtain an optimal GKGR model. And finally predicting the score of each pair of users and service items, and recommending the service items with higher scores to the users in a recommendation list mode. According to the method, the neural network is utilized to conduct feature extraction on users and service matters, user behavior data is fully utilized, and the problem of data sparseness is effectively relieved; modeling is carried out on the entity context and the entity description text information through a knowledge representation method, knowledge representation tasks and personalized recommendation tasks are jointly learned, accuracy and interpretability of recommendation results are improved, and the problem of cold start of government service recommendation is effectively relieved.

Description

A service recommendation method based on deep matrix factorization and knowledge graph

technical field

The invention relates to the field of government service recommendation, in particular to a service recommendation method based on deep matrix decomposition and knowledge graph.

Background technique

"Internet + government service" combines traditional government service methods with modern Internet technology, realizes convenient interaction and information exchange between the government and citizens, improves the efficiency, transparency and fairness of government services, and also promotes digitalization Transformation and smart city building. However, with the construction of city-level one-stop service platforms, government service resources are huge and scattered, with many types and complex levels. Government services are oriented to citizen users and often require personalized information services. How to filter out the required service items for users from a large number of urban government services and recommend them to users is a pain point and difficulty faced by the one-stop city service platform. In the personalized recommendation technology, the traditional collaborative filtering recommendation algorithm is widely used and the technology is mature, but it is difficult to deal with the data sparse problem faced in the personalized recommendation scenario of government services.

Contents of the invention

The purpose of the present invention is to provide a service recommendation method based on deep matrix decomposition and knowledge graph, which aims to solve the problem of sparse user data, a large number of heterogeneous, multi-source, and loosely organized data that cannot be fully utilized when recommending government services.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solution: a service recommendation method based on depth matrix decomposition and knowledge graph, including the following steps:

S1: Obtain the initial user vector according to the behavior data generated during the interaction between the user and the service item, and output the initial user to the fully connected layer as the input user vector u _i :

According to the behavior data generated during the interaction between users and service items, the user behavior data is quantified according to the measurement rules, and the user-service item behavior matrix is constructed Each row of the matrix represents an initial user vector, and the value R _ij in the matrix represents the number of times user i clicks on service item j.

Construct the government service knowledge map G, and use the graph structure to represent the service item entities and relationships. Each service item entity is regarded as a node in the graph, and the relationship is regarded as an edge;

S2: build and train GKGR model, described GKGR model comprises:

S2-1: Obtain the service item entity vector e _s according to the entity context information defined in the government service knowledge graph;

S2-2: Obtain the second service item entity vector _ed according to the definition entity description text of the government service knowledge graph;

S2-3: Obtain the final service item vector e according to the service item entity vector e _s and the service item entity vector e _d ;

S2-4: Given user i, service item entity j, and user-service item behavior matrix R _ij , construct a user-service item preference pair <i,j,j′>, indicating that user i interacts with service item entity j , there is no interaction with service item j′, that is, user i has demand for service item j. Find triples and entity description text related to j, j′ from G. Learn e _s through the knowledge representation method of entity context, and use Bi-LSTM to learn service item entity vector _ed . The two entity vectors are fused through the gating mechanism, and the user vector u _i and the service item vector are input into the personalized ranking model.

When the objective function is the largest and no longer changes, the training ends, and the optimal GKGR model is obtained at this time;

S3: For a user, use S1 to obtain the user vector input optimal GKGR model, the optimal GKGR model calculates the degree of association between the user and all service items, and arranges them in descending order according to the value of the degree of association, and outputs the service corresponding to the value of the degree of association sequence of events.

As a preference, the process of S2-1 obtaining the service item entity vector e _s is as follows:

Entity context information C(h,r,t) includes neighbor context C _n (h) and path context C _p (h,t).

Neighborhood context C _n (h) refers to the set of other nodes directly connected to a given node.

Path context C _p (h, t) refers to the context information composed of all paths connected to a given node, that is, C (h, r, t) = C _n (h)∪C _p (h, t).

The neighbor context definition of the service item entity h is shown in the following formula, and G represents the government service knowledge map.

Among them, h, t represent different service item entities, and r represents a relationship;

The path context definitions of service item entities h and t are as follows:

Among them, p _i is the sequence of relations from h to entity t, L is the maximum length among all relation paths, r ₁ , Indicates the other relations of h reaching entity t, e ₁ , /> Indicates that h reaches other entities on the way to entity t, and l _i indicates the i-th relationship.

The probability of the triplet (h, r, t) being established is shown in the following formula.

f(h,r,t)=P((h,r,t)|C(h,r,t);θ) (3)

Among them, θ represents the parameters of the model, and the higher the score of the scoring function f( ), the greater the probability of the triplet being established.

The pre-training model transE, the triple (h, r, t) is input into transE, when the value of f(h, r, t) is the largest, the output of transE is the service item entity vector e _s .

As preferably, the scoring function f ( ) is optimized using the method:

By decomposing f(h,r,t) through conditional probability, we can get:

f(h,r,t)=P(h|C(h,r,t);θ)·P(t|C(h,r,t),h;θ)·P(r|C(h , r, t), h, t; θ) (4)

where P(h|C(h,r,t); θ) represents the conditional probability of occurrence at h. Since the entity h is contextually related to its neighbors, P(h|C(h,r,t); θ) can be directly approximated as P(h|C _n (h); θ), defined as shown in the following formula.

in Indicates the degree of association of any entity with the context of h-entity neighbors.

h' represents the head entity in the wrong triplet;

P(t|C(h,r,t),h;θ) represents the probability of entity t, and uses the path context to measure the degree of association between the head entity and the tail entity,

P(t|C(h,r,t),h;θ) is approximately expressed as P(t|C _p (h,t),h;θ), defined as shown in the following formula.

Among them, ε represents the tail entity set, and t′ represents the tail entity in the error triplet (negative example);

P(r|C(h, r, t), h, t; θ) represents the conditional probability of occurrence of relation r. The entities h and t have been determined, and the entity context has been introduced. Therefore, the entity context C(h,r,t) in P(r|C(h,r,t),h,t; θ) is omitted as shown in the following formula Show.

P(h|C(h,r,t);θ), P(t|C(h,r,t),h;θ) and P(r|C(h,r,t) in the scoring function ,h,t;θ) are approximately expressed as P(h|C _n (h);θ), P(t|C _p (g,t),g;θ) and P(r|h,t;θ) , as shown in the following formula.

f(g,r,t)≈P(h|C _n (h);θ)·P(t|C _p (h,t),h;θ)·P(r|h,t;θ) ( 8) Optimizing the vector of entity context information by maximizing the scoring function f(h,r,t)=P((h,r,t)|C(h,r,t);θ). As a preference, the process of S2-2 obtaining the service item entity vector _ed according to the description text of the government knowledge map entity is as follows:

Based on the constructed government service knowledge graph G, the entity description text is defined for the service entity. The entity description text includes the entity name, the relationship name associated with the entity, and the tail entity name;

The weight of the i-th position of the entity description text to a given relation r is defined as α _i (r), as shown in the following formula.

where, where, is the relationship vector obtained through representation learning, /> is the output of the i-th position, W _a and U _a are parameter matrices, /> is a parameter vector. e _i (r) is the correlation between z _i and relation r, and n represents the length of entity description text.

Service item entity vector _ed is defined as shown in the following formula.

x ₁ , x _n represent the position of the entity description text whose length is 1 and n respectively.

As a preference, the process of S2-3 obtaining the final service item vector e according to the service item entity vector e _s and the service item entity vector _ed is as follows:

Through the gating mechanism, e _s and ed _d are fused to obtain e, which is defined as shown in the following formula.

e=β⊙e _s +(1-β)⊙e _d (12)

where β ∈ [0,1] represents a gate that balances two representation weights.

Preferably, the objective function L is:

v _j ＝β⊙e _sj +(1-β)⊙e _dj (15)

Among them, u _i is the vector representation of user i, v _j and v _j′ are the vector representations of government service item entities j and j′ respectively, and z represents the regular term.

Among them, f(h,r,t; d _h ,d _t ) represents the scoring function of knowledge representation learning part, g _h , g _t are the gate sizes of head entity and tail entity respectively, h _s , t _s are h and t's entity context information vectors, h _d , t _d are the vectors of h and t's entity description text knowledge.

Compared with the prior art, the present invention has at least the following advantages:

The S1 constructs a user-service item behavior matrix by analyzing the characteristics of user behavior, and uses a neural network to extract features of users and service items. Make full use of user behavior data to effectively alleviate the problem of data sparseness. It solves the lack of personalization of government service recommendation methods, and the traditional collaborative filtering technology is difficult to solve the problem of sparse user-service item matrix.

The S2 constructs a government service knowledge map using government resource data, models entity context and entity description text information through knowledge representation methods, jointly learns knowledge representation tasks and personalized recommendation tasks, and improves the accuracy and interpretability of recommendation results , which effectively alleviates the cold start problem of government service recommendation. It solves the problem of heterogeneity and multiple sources of government information resources, loose organization and underutilization.

Description of drawings

Figure 1 An example of knowledge-based government service recommendation.

Figure 2 Construction steps of government ontology.

Figure 3 Government ontology structure.

Figure 4 protégé constructs the government ontology.

Figure 5 Partial RDF triples.

Fig. 6 is a structural diagram of entity description text information.

Fig. 7 is a structure diagram of the method of the present invention.

Figure 8 is an example diagram of a triplet.

Detailed ways

The present invention will be described in further detail below.

The present invention proposes a service recommendation method (BDMF) based on deep matrix decomposition and knowledge graph by modeling user behavior. The features of users and service items are extracted through the neural network, and the depth of collaborative filtering is used to predict the degree of user demand for non-interactive service items. The BDMF does not consider the impact of information on service items on user needs. In the government service scenario, each service item contains various information such as acceptance conditions, processing subject, exercise level, and service object. User needs are closely related to the above-mentioned item information, and rational use of these service item information can effectively improve the recommendation accuracy of BDMF. As shown in Figure 1, when a user clicks on the service item of "corporate social insurance registration", the acceptance condition of this item is "need to be registered with the market supervision department", and the recommendation algorithm should consider this information to recommend the implementation subject for the user to be the market supervision administration And service matters related to enterprises, such as: "Registration of Establishment of Domestic-funded Enterprises and Branches (Company Establishment Registration)", "Registration of Establishment of Foreign-funded Enterprises and Branches (Registration of Establishment of Foreign-Invested Enterprises)" and other matters. In government service recommendation, the more similar the service item information, the higher the reference value for user needs. Reasonable use of the similarity between service item information can improve the accuracy of the algorithm and effectively alleviate the cold start problem of government service recommendation.

In the government service recommendation scenario, the potential needs of users are related to service item information. However, government service resources are huge and scattered, with many types and complex levels. Government service recommendation faces the problems of heterogeneity and multiple sources of government information resources, loose organization, and insufficient utilization. As a heterogeneous network containing rich semantic information, the knowledge graph can effectively represent the entities and relationships in the knowledge graph in a low-dimensional continuous vector space by using the knowledge representation method, and can make reasonable use of government resource data while endowing the knowledge graph with The ability to integrate, reason, and apply improves the accuracy of government service recommendations.

Government Service Knowledge Graph Construction

The knowledge map can be logically divided into a schema layer and a data layer. The schema layer is the core of the knowledge graph, storing refined knowledge, and defining and standardizing data levels and categories in the domain. The ontology library is usually used to manage the schema layer of the graph, and the rules, axioms, and constraints in the ontology library are used to standardize the associations between entities, relationships, and entity types and attributes in the graph.

The data layer is responsible for the specific storage of specific triples in the knowledge graph. It is structurally under the schema layer and is the actual form of expression of the entire knowledge graph. In the data layer, triples are stored in the graph database through two expressions of <entity, relationship, entity> and <entity, attribute, value>, as shown in Figure 8.

The construction process of the knowledge map starts with the acquisition of original knowledge data, and uses knowledge processing technology (including automatic or semi-automatic) to extract the required knowledge elements from the original data. Store according to the definition of schema layer and data layer. Massive heterogeneous knowledge forms a huge entity relationship network through the structural definition of the schema layer and the processing of the data layer, thereby constructing a knowledge graph. The knowledge graph is constructed based on the prior knowledge of domain experts.

A service recommendation method based on deep matrix decomposition and knowledge graph, comprising the following steps:

S1: Obtain the initial user vector according to the behavior data generated during the interaction between the user and the service item, and output the initial user to the fully connected layer as the input user vector u _i :

User vectors are obtained based on user behavior and deep matrix factorization. According to the behavior data generated during the interaction between users and service items, the user behavior data is quantified according to the measurement rules, and the user-service item behavior matrix is constructed Each row of the matrix represents an initial user vector, and the value R _ij in the matrix represents the number of times user i clicks on service item j.

Construct the government service knowledge map G, and use the graph structure to represent the service item entities and relationships. Each service item entity is regarded as a node in the graph, and the relationship is regarded as an edge;

S2: build and train GKGR model, described GKGR model comprises:

S2-1: Obtain the service item entity vector e _s according to the entity context information defined in the government service knowledge graph;

S2-2: Obtain the second service item entity vector _ed according to the definition entity description text of the government service knowledge graph;

S2-3: Obtain the final service item vector e according to the service item entity vector e _s and the service item entity vector e _d ;

S2-4: Using the method of joint training, learn the vector representation of users, service item entities and relationships by combining user-service item behavior data and government affairs knowledge map data. Given user i, service item entity j, and user-service item behavior matrix R _ij , a user-service item preference pair <i,j,j′> is constructed, which means that user i interacts with service item entity j, and interacts with service item j' has no interaction, that is, user i has demand for service item j. Find triples and entity description text related to j, j′ from G. Learn e _s through the knowledge representation method of entity context, and use Bi-LSTM to learn service item entity vector _ed . The two entity vectors are fused through a gating mechanism, and the user vector and service item vector are input into the personalized ranking model. The personalized sorting model uses Bayesian personalized sorting to convert the recommendation problem into a sorting problem. For each user, sort all the recommended items in the system, and rank the items that the user likes as much as possible. Finally, the sequence The Top-K items are recommended to users.

When the objective function is the largest and no longer changes, the training ends, and the optimal GKGR model is obtained at this time;

S3: For a user, use S1 to obtain the user vector input optimal GKGR model, the optimal GKGR model calculates the degree of association between the user and all service items, and arranges them in descending order according to the value of the degree of association, and outputs the service corresponding to the value of the degree of association sequence of events.

Specifically, the process of S2-1 obtaining the service item entity vector e _s is as follows:

Entity context information includes neighbor context and path context. In the government service knowledge map, given a service item entity, its neighbor context nodes include the type of service item, implementation subject, exercise level, etc. The knowledge representation method that introduces the entity context encodes the semantic information associated with the entity into a high-dimensional vector representation, which improves the representation ability of the model. Secondly, the introduction of knowledge representation of entity context can better understand the semantic relationship of entities, such as the similarity and hierarchical structure between entities, and improve the interpretability of the recommendation system.

Entity context information C(h,r,t) includes neighbor context C _n (h) and path context C _p (h,t).

Neighborhood context C _n (h) refers to the set of other nodes directly connected to a given node.

Path context C _p (h, t) refers to the context information composed of all paths connected to a given node, that is, C (h, r, t) = C _n (h)∪C _p (h, t).

The neighbor context definition of the service item entity h is shown in the following formula, and G represents the government service knowledge graph.

Among them, h, t represent different service item entities, and r represents a relationship;

The path context definitions of service item entities h and t are as follows:

Among them, p _i is the sequence of relations from h to entity t, L is the maximum length among all relation paths, r ₁ , Indicates the other relations of h reaching entity t, e ₁ , /> Indicates that g reaches other entities on the way to entity t, and l _i indicates the i-th relationship. The probability of the triplet (h, r, t) being established is shown in the following formula.

f(h,r,t)=P((h,r,t)|C(h,r,t);θ) (3)

Among them, θ represents the parameters of the model, and the higher the score of the scoring function f( ), the greater the probability of the triplet being established.

The pre-training model transE, the triple (h, r, t) is input into transE, when the value of f(h, r, t) is the largest, the output of transE is the service item entity vector e _s .

Specifically, the scoring function f( ) is optimized using the method:

By decomposing f(h,r,t) through conditional probability, we can get:

f(h,r,t)=P(h|C(h,r,t);θ)·P(t|C(h,r,t),h;θ)·P(r|C(h , r, t), h, t; θ) (4)

where P(h|C(h,r,t); θ) represents the conditional probability of occurrence at h. Since the entity h is contextually related to its neighbors, P(h|C(h,r,t); θ) can be directly approximated as P(h|C _n (h); θ), defined as shown in the following formula.

in Indicates the degree of association of any entity with the context of h-entity neighbors.

h' represents the head entity in the wrong triplet (i.e. negative example);

P(t|C(h,r,t),h;θ) represents the probability of entity t, and uses the path context to measure the degree of association between the head entity and the tail entity,

P(t|C(h,r,t),h;θ) is approximately expressed as P(t|C _p (h,t),h;θ), defined as shown in the following formula.

Among them, ε represents the tail entity set, and t′ represents the tail entity in the error triplet (negative example).

P(r|C(h, r, t), h, t; θ) represents the conditional probability of occurrence of relation r. The entities h and t have been determined, and the entity context has been introduced. Therefore, the entity context C(h,r,t) in P(r|C(h,r,t),h,t; θ) is omitted as shown in the following formula Show.

P(h|C(h,r,t);θ), P(t|C(h,r,t),h;θ) and P(r|C(h,r,t) in the scoring function , h, t; θ) are approximately expressed as P(h|C _n (h); θ), P(t|C _p (h, t), h; θ) and P(r|h, t; θ) , as shown in the following formula.

f(h,r,t)≈P(h|C _n (h);θ)·P(t|C _p (h,t),h;θ)·P(r|h,t;θ) ( 8)

The vector of entity context information is optimized by maximizing the scoring function f(h,r,t)=P((h,r,t)|C(h,r,t);θ).

Specifically, the process of S2-2 obtaining the service item entity vector _ed according to the description text of the government knowledge map entity is as follows:

Based on the constructed government service knowledge graph G, the entity description text is defined for the service entity. The entity description text includes the entity name, the relationship name associated with the entity, and the tail entity name;

The weight of the i-th position of the entity description text to a given relation r is defined as α _i (r), as shown in the following formula.

where, where, is the relationship vector obtained through representation learning, /> is the output of the i-th position, W _a and U _a are parameter matrices, /> is a parameter vector. The correlation between e _i (r) z _i and relation r, n represents the length of entity description text.

Service item entity vector _ed is defined as shown in the following formula.

x ₁ , x _n represent the position of the entity description text whose length is 1 and n respectively.

Specifically, the process of S2-3 obtaining the final service item vector e according to the service item entity vector e _s and the service item entity vector e _d is as follows:

Through the gating mechanism, e _s and ed _d are fused to obtain e, which is defined as shown in the following formula.

e=β⊙e _s +(1-β)⊙e _d (12)

where β ∈ [0,1] represents a gate that balances two representation weights.

Specifically, the objective function L is:

v _j ＝β⊙e _sj +(1-β)⊙e _dj (15)

Among them, u _i is the vector representation of user i, v _j and v _j′ are the vector representations of government service item entities j and j′ respectively, and z represents the regular term.

Among them, f(h,r,t; d _h ,d _t ) represents the scoring function of knowledge representation learning part, g _h , g _t are the gate sizes of head entity and tail entity respectively, h _s , t _s are h and t's entity context information vectors, h _d , t _d are the vectors of h and t's entity description text knowledge. h _s , r, and t _s are obtained through pre-training based on the TransE representation learning method, and h _d , t _d are obtained through representation learning on entity description texts.

Table 1 GKGR training process

Experimental Design and Analysis

The present invention takes the service recommendation method based on user behavior and deep matrix decomposition as the experimental baseline, analyzes the impact of adding the knowledge representation module on the service recommendation results, and verifies the recommendation method that integrates the government service knowledge map.

1. Dataset

The construction method of the negative example triplet: randomly select a service item that the user has never interacted with as a negative example in the user-service item behavior data, find the triplet containing the entity of the service item in the government service knowledge map, and Negative sample it. The way of negative sampling is to replace the head entity or tail entity in the triplet, and obtain the triplet and entity description text that do not exist in the knowledge graph. By increasing the diversity of negative samples, the algorithm can better learn the relationship between positive triples and improve the accuracy of recommendation. Finally, the sampled negative examples are used together with the positive examples to train and test the algorithm, with a ratio of 1:3.

2. Evaluation indicators

In order to ensure the consistency of the experiment, the evaluation indicators are the same as those used in the service recommendation method experiment based on user behavior and deep matrix decomposition, namely HR, Precision, Recall and F1 value.

In the experiment, the dimension of each model vector is set to 64 dimensions. The network parameters are initialized using a Gaussian distribution with mean 0 and variance 0.001. The batch size is set to 64, the ratio of positive and negative samples is 1:3, the learning rate is set to 0.01, the regularization coefficient is 0.001, and the Adam optimizer is used for parameter optimization.

BDMF: The method proposed by the present invention.

BDMF+TransE: Use the TransE model to represent entities.

CKE: An Approach to Recommender Systems Based on Collaborative Knowledge Base Embedding.

① Compare the difference in recommendation effect between the method of the present invention and the comparison method

The experimental results of each model under the indicators of HR, Precision, Recall and F1 are shown in Table 2.

Table 2 Comparative Experimental Results

According to the analysis of the experimental results, the GKGR, CKE, and BDMF+TransE models are higher than the BDMF method in four indicators after integrating the government service knowledge map, and the GKGR model proposed by the present invention improves Precision, Recall, F1, and HR respectively. 5.08%, 17.53%, 10.58%, 22.87%. In the model that integrates the knowledge map of government services, the BDMF+TransE method is lower than the GKGR model and the CKE model in terms of related indicators because it does not model the entity description text. The CKE model introduces entity description text on the basis of TransE to represent the knowledge map of government services, and the recommendation effect has been improved to a certain extent. The GKGR model proposed by the present invention realizes the representation of entity path context and neighbor context through the knowledge representation of entity context on the basis of CKE. Compared with CKE, Precision, Recall, F1 and HR are respectively increased by 5.76%, 3.77%, and 5.6%. , 9.95%, improving the effect of government service recommendation, indicating that the fusion of government service knowledge graph entity context can further improve the performance.

In the government service recommendation scenario, if the path context and neighbor context of two government service item entities are similar in the knowledge graph, the entity vector representations will be correspondingly closer. When handling business on the government service platform, users usually find the pre-service of the service item based on the acceptance conditions of the service item, exercise level and other information. By describing the text representation of the service item entity, the semantic representation ability of the knowledge graph is enhanced. The recommendation model that integrates the government service knowledge map proposed by the present invention makes full use of the relationship between government service item entities, introduces the entity context and entity description text of the government service knowledge map, and balances the entity description text and entity context information through the gating mechanism. The weight accounted for can achieve a better recommendation effect.

②Comparing the influence of different knowledge representation methods on the recommendation effect

In order to verify the impact of knowledge representation methods on recommendation results, four common TransX series models are selected, namely TransE, TransH, TransR and TransD. And use Precision, Recall, F1, HR evaluation indicators to measure. Table 3 shows the recommendation effect after using different knowledge representation methods. The experimental results show that in government service recommendation tasks, TransH, TransR, and TransD perform better than TransE, among which the TransD knowledge representation method performs best in the two indicators of recommendation accuracy and recall rate, with a F1 value of 0.2897 and HR The value is 0.7661. However, the effect of TransE is less different from that of TransD, and the model structure is simple and easy to train, so this experiment is mainly based on the TransE model for training.

Table 3 The impact of different representation methods on each indicator

③Comparing the difference in the recommendation effect of different models for entity description text

In order to verify the impact of entity description text on recommendation results, four common models were selected for the representation of entity description text, namely Word2Vec, CNN, RNN and Bi-LSTM, and used Precision, Recall, F1, HR evaluation indicators to measure . As shown in Table 4, the recommendation effects on different characterization methods are used.

Table 4 The influence of different characterization methods on each indicator

As shown in Table 4, in the modeling of entity description text, RNN and Bi-LSTM, which are good at processing text sequences, have better results, and have a stronger ability to extract features of entity description text. Among them, Bi-LSTM introduces a gating mechanism on the basis of RNN, and has a stronger modeling ability for entity description text. In terms of F1 and HR indicators, it is 6.5% and 9.25% higher than RNN, respectively, which improves the accuracy of government service recommendations. .

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of the present invention shall be covered by the claims of the present invention.

Claims (6)

1. A service recommendation method based on depth matrix decomposition and knowledge graph is characterized by comprising the following steps:

s1: obtaining an initial user vector according to behavior data generated in the interaction process of the user and the service items, and outputting the initial user vector as an input full-connection layer to obtain a user vector u _i ：

Quantifying the user behavior data according to the measurement rules according to the behavior data generated in the interaction process of the user and the service items, and constructing a user-service item behavior matrixWherein each row of the matrix represents an initial user vector, the values R in the matrix _ij The number of clicks of the user i on the service item j is represented;

constructing a government service knowledge graph G, wherein the service event entities and the relations are represented by graph structures, each service event entity is regarded as a node in the graph, and the relations are regarded as edges;

s2: constructing and training a GKGR model, said GKGR model comprising:

s2-1: obtaining service item entity vector e according to government service knowledge graph definition entity context information _s ；

S2-2: obtaining a second service item entity vector e according to the government service knowledge graph definition entity description text _d ；

S2-3: entity according to service mattersVector e _s And service item entity vector e _d Obtaining a final service item vector e;

s2-4: given user i, service item entity j and user-service item behavior matrix R _ij User-service item preference pairs are constructed<i，j，j′>Indicating that the user i has interaction with the service item entity j and does not have interaction with the service item j ', namely that the user i has a requirement on the service item j, finding out triples and entity description text related to j and j' from G, and learning e by a knowledge representation method of entity context _s Learning service item entity vector e using Bi-LSTM _d The two entity vectors are fused through a gating mechanism, and the user vector u is obtained _i And the service item vector is input into the personalized sequencing model;

when the objective function is maximum and is not changed any more, training is finished, and an optimal GKGR model is obtained at the moment;

s3: for a user, S1 is adopted to obtain a user vector and input an optimal GKGR model, the optimal GKGR model calculates the association degree of the user and all service matters, and the service matters are arranged in descending order according to the association degree value, and the service matter sequence corresponding to the association degree value is output.

2. The service recommendation method based on depth matrix decomposition and knowledge graph as claimed in claim 1, wherein: the S2-1 obtains a service item entity vector e _s The process of (2) is as follows:

the entity context information C (h, r, t) includes neighbor context C _n (h) And path context C _p (h，t)；

Neighbor context C _n (h) Refers to a collection of other nodes directly connected to a given node;

path context C _p (h, t) refers to the context information composed of all paths connected to a given node, i.e., C (h, r, t) =c _n (h)∪C _p (h，t)；

The neighbor context definition of the service item entity h is shown in the following formula, and G represents a government service knowledge graph;

wherein h, t represent different service event entities, and r represents a relationship;

the path context definition of service instance entities h and t is as follows:

wherein p is _i Is the relation sequence of h reaching entity t, L is the maximum length in all relation paths, r ₁ ，Representing other relationships of h to the entity t pathway, e ₁ ，/>Representing h reaching other entities of the entity t pathway, l _i Representing an ith relationship;

the probability that triplet (h, r, t) is established is shown in the following formula;

f(h，r，t)＝P((h，r，t)|C(h，r，t)；θ) (3)

wherein θ represents a parameter of the model, and the higher the score of the scoring function f (·) is, the greater the probability of the triplet being established;

the pre-training model is transmitted, the triplet (h, r, t) is input into the transmitted, when the f (h, r, t) value is maximum, the output of the transmitted is the service item entity vector e _s 。

3. The service recommendation method based on depth matrix decomposition and knowledge graph as claimed in claim 2, wherein: optimizing the scoring function f (·) by adopting a method:

by decomposing f (h, r, t) with conditional probability, we can get:

f(h，r，t)＝P(h|C(h，r，t)；θ)·P(t|C(h，r，t)，h；θ)·P(r|C(h，r，t)，h，t；θ) (4)

wherein P (h|C (h, r, t); θ) represents the conditional probability of occurrence at h, defined as shown in the following formula;

wherein the method comprises the steps ofRepresenting the association degree of any entity and a h entity neighbor context;

h' represents the head entity in the wrong triplet;

p (t|C (h, r, t), h; θ) represents entity t probability, the degree of association between head and tail entities is measured by path context, and P (t|C (h, r, t), h; θ) is approximately represented as P (t|C) _p (h, t), h; θ), defined as the following formula;

where ε represents the set of tail entities and t' represents the tail entities in the error triplet (negative example);

p (r|C (h, r, t), h, t; θ) represents the conditional probability of occurrence of the relationship r; entities h and t have determined that the entity context has been introduced, and therefore omitting the entity context C (h, r, t) in P (r|C (h, r, t), h, t; θ) is shown in the following formula;

p (h|C (h, r, t); θ), P (t|C (h, r, t), h; θ) and P (r|C (h, r, t), h, t; θ) in the scoring function are approximately expressed as P (h|C) _n (h)；θ)、P(t|C _p (h, t), h; θ) and P (r|h, t; θ) as shown in the following formula;

f(h，r，t)≈P(h|C _n (h)；θ)·P(t|C _p (h，t)，h；θ)·P(r|h，t；θ) (8)

the vector of entity context information is optimized by maximizing the scoring function f (h, r, t) =p ((h, r, t) |c (h, r, t); θ).

4. The service recommendation method based on depth matrix decomposition and knowledge graph as claimed in claim 3, wherein: s2-2 obtains a service item entity vector e according to the government knowledge map entity description text _d The process of (2) is as follows:

based on the constructed government service knowledge graph G, defining an entity description text aiming at the service event entity, wherein the entity description text comprises an entity name, an entity association relationship name and a tail entity name;

the weight of the ith position of the entity description text to a given relationship r is defined as α _i (r) as shown in the following formula;

wherein, among them,is a relation vector obtained by expression learning, +.>Is the output of the i-th position, W _a And U _a Is a parameter matrix,/->Is a parameter vector; e, e _i (r) is z _i Closing deviceThe correlation of r, n represents the length of the entity description text;

service item entity vector e _d The definition is shown in the following formula;

x ₁ ，x _n representing the positions of the entity description text of length 1 and n, respectively.

5. The service recommendation method based on depth matrix decomposition and knowledge graph as claimed in claim 4, wherein: the S2-3 is based on the service item entity vector e _s And service item entity vector e _d The process of obtaining the final service event vector e is as follows:

e is controlled by a gating mechanism _s And e _d Fusing to obtain e, wherein the definition is shown in the following formula;

e＝β⊙e _s +(1-β)⊙e _d (12)

wherein β ε [0,1] represents the gate that balances two types of representation weights.

6. The service recommendation method based on depth matrix decomposition and knowledge graph as claimed in claim 5, wherein: the objective function L is:

v _j ＝β⊙e _sj +(1-β)⊙e _dj (15)

wherein u is _i Is a vector representation of user i, v _j 、v _j′ Vector table of government service event entities j and j' respectivelyShown, z represents a regularization term;

wherein f (h, r, t, d) _h ，d _t ) Scoring function, g, representing knowledge representation learning portion _h 、g _t The gating sizes of the head entity and the tail entity, h respectively _s 、t _s Vector of entity context information h and t, respectively, h _d 、t _d Vectors of textual knowledge are described for the entities of h and t.