A service recommendation system based on personality perception and complex relationship interaction

Abstract

The application discloses a service recommendation system based on character perception and complex relationship interaction, comprising an information acquisition module, a user character subgraph generation module, an information propagation module, an information fusion module and a personalized recommendation module; the application considers the influence of user character on user preference, constructs a user character graph, and improves the precision of user representation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of service recommendation, and in particular to a service recommendation system based on personality perception and complex relationship interaction. BACKGROUND

[0002] Personalized recommendation, as a hot issue in the field of recommendation system, has attracted attention and research in many fields, whether in the industry or in academia, personalized recommendation has shown its great research value. In recent years, with the rapid development of information technology, social media represented by QQ, WeChat and microblog have been welcomed by the public, which has promoted the explosive growth of Internet information, but at the same time, it has also led to the problem that target information is easily submerged in massive data. In order to better utilize the social relationship between users to filter out appropriate information or products for users, and deeply mine user preferences, social recommendation has emerged as the times require.

[0003] Traditional recommendation systems have data sparsity and cold start problems because they only obtain data from the interaction information between users and items. According to the social influence theory, people who are connected to each other will influence each other, and the closer the relationship between them, the more frequent the communication, and the greater the influence. Therefore, social recommendation, as a recommendation system that uses social domain information to improve the prediction accuracy of user behavior domain, can to some extent alleviate the impact of inherent problems such as data sparsity in the behavior domain. According to the development of social recommendation algorithm, it has experienced from traditional collaborative filtering-based methods to graph embedding-based social recommendation algorithms. In recent years, scholars in the field of social recommendation have applied existing machine learning models to social recommendation, such as using graph convolutional network (GNN) model as the basis for social recommendation system, which has achieved good results.

[0004] The prior art discloses a social recommendation method and system for enhancing neighbor relationship, as shown in Figure 1 The method proposes a social recommendation model for enhancing neighbor relationship. The model calculates the intimacy of the neighbors of users and items by establishing a mapping between users and items, samples the neighbors of users and items based on the intimacy, and calculates the weight of the sampled neighbors according to the intimacy through a self-attention mechanism, then calculates the social space representation of the items and the social space representation of the users, and combines the feature representation of the items and the feature representation of the users to obtain the feature representation of the items and the feature representation of the users. This method weakens the influence of invalid neighbors and improves the accuracy of social recommendation.

[0005] Some scholars provide a social recommendation method of a heterogeneous graph convolutional network combining social and interest information, as shown in Figure 2The method effectively combines the social information and interest information of the user, integrates the two into a model, captures high-order correlation information between users and between users and items, and maximizes the use of user data by using a hierarchical attention mechanism training, so as to realize more accurate commodity recommendation for the user.

[0006] The above methods respectively enhance the neighbor relationship, combine the user social and interest information to represent the user more accurately, and although the user data sparsity problem is alleviated to a certain extent, the influence of personality on user preference is not considered. SUMMARY

[0007] The purpose of the present application is to provide a service recommendation system based on personality perception and complex relationship interaction, which comprises an information acquisition module, a user personality subgraph generation module, an information propagation module, an information fusion module and a personalized recommendation module.

[0008] The information acquisition module randomly generates user vector information and service vector information and transmits them to the user personality subgraph generation module and the information propagation module.

[0009] The information acquisition module acquires a service representation set and transmits it to the personalized recommendation module.

[0010] The user personality subgraph generation module processes the user vector information and the service vector information to generate a user personality subgraph and transmits it to the information propagation module.

[0011] The user personality subgraph includes a user-service interaction personality subgraph and a user social-interest personality subgraph N represents the number of personality types. respectively represent the RNth user-service interaction personality subgraph and the SNth user social-interest personality subgraph.

[0012] According to the user personality subgraph, the information propagation module processes the user vector information to obtain a user social representation set and a user interest representation set at different levels, and transmits them to the information fusion module.

[0013] The information fusion module fuses the user social representation set and the user interest representation set to obtain a user representation set and transmits it to the personalized recommendation module.

[0014] The personalized recommendation module matches the user representation set and the service representation set with each other, and when the matching degree between a user and a service is greater than or equal to a threshold value ε, the corresponding service is recommended to the user.

[0015] Further, the information acquisition module generates user vector information and service vector information using edge structure information or standard normal distribution random generation n and m respectively represent the number of users and the number of services, and d represents the dimension of the vector.

[0016] Further, the user personality subgraph generation module processes the user vector information and the service vector information to generate a user-service interaction personality subgraph, and the steps include:

[0017] a1) The user personality subgraph generation module fuses the user initial vector information and the first-order user-service interaction personality vector to obtain the fusion embedding vector F of the user u u1 , that is:

[0018]

[0019]

[0020] In the formula, N u , N u represent the user node set and the service node set;

[0021] a2) The fusion embedding vector F u1 is input into a two-layer fully connected neural network for personality classification to obtain the user-service interaction personality representation parameter U o1 , that is:

[0022] U h1 = σ (W 11 F u1 + b 11 ) (3)

[0023] U o1 = (W 12 U h1 + b 12 ) (4)

[0024] In the formula, W 11 , W 12 respectively represent weight matrices; b 11 , b12 respectively represent weight biases; σ(·) is an activation function; U h1 is the output of the first layer fully connected neural network;

[0025] a3) according to the user-service interaction personality characteristic parameter U o1 , all users are divided into different user-service interaction personality subgraphs.

[0026] Further, the activation function is a LeakyReLU activation function.

[0027] Further, the user personality subgraph generation module processes user vector information and service vector information , and the step of generating a user social-interest personality subgraph includes:

[0028] b1) obtaining the first layer embedding of user u in the user social graph

[0029] fuse the user initial vector information and the first layer embedding to obtain the fusion embedding vector F of user u u2 , that is:

[0030]

[0031] b2) input the fusion embedding vector F u2 into the two-layer fully connected neural network for personality classification to obtain the user social-interest personality characteristic parameter U o2 , that is:

[0032] U h2 = σ (W 21 F u2 + b 21 ) (6)

[0033] U o2 = (W 22 U h2 + b 22 ) (7)

[0034] In the formula, U h2 is the first layer output of the two-layer fully connected neural network for personality classification; W 21 , W 22 respectively represent weight matrices; b 21 , b 22 respectively represent weight biases; σ(·) is an activation function;

[0035] b3) according to the user-service interaction personality characterization parameter U o2 all users are divided into different user social-interest personality subgraphs.

[0036] Further, the information propagation module processes the user vector information to obtain a set of user personality representations at different levels The steps include:

[0037] c1) input the user vector information into the graph convolution network of the information propagation module;

[0038] c2) the graph convolution network iterates K times on the user social graph on the user vector information to obtain a set of user social representations at different levels

[0039] wherein the k+1th level user social representation is as follows:

[0040]

[0041] In the formula, denotes the set of adjacent users of user u on the social graph; denotes the set of adjacent users of user v on the social graph;

[0042] The graph convolution network iterates K times on the user-service interaction personality subgraph on the user vector information to obtain a set of user interest representations at different levels

[0043] wherein the k+1th level user interest representation is as follows:

[0044]

[0045] In the formula, denotes the set of adjacent nodes of user u on the user service graph; denotes the set of adjacent nodes of service i on the user service graph.

[0046] Further, the information fusion module fuses the set of user social representations and the set of user interest representations The steps include:

[0047] d1) multiply the elements in the set of user social representations and the set of user interest representations to obtain:

[0048]

[0049] wherein, ⊙ represents element-level multiplication between two vectors; is a user personality representation, a user interest representation; X ij is a product result;

[0050] d2) inputting the product result into a three-layer neural network to generate a user representation set i.e.

[0051] X1 = σ(ξ1X ij +a1)

[0052] X2 = σ(ξ2X1+a2)

[0053] X3 = ξ3X2+a3

[0054]

[0055] wherein, X1, X2, X3 represent user representations in the first, second and third layers of the neural network respectively; X o is a final user representation after fusing the outputs of the three-layer neural network. ξ 1、 ξ2, ξ3 are weights; a1, a2, a3 are biases.

[0056] Further, the personalized recommendation module performs the steps of matching the user representation set and the service representation set two by two, which includes:

[0057] e1) calculating a user vector representation e u and a service vector representation e v i.e.

[0058]

[0059]

[0060] wherein, α k is a parameter;

[0061] e2) calculating a matching degree between a target user u and a service i i.e.

[0062]

[0063] wherein, e u , e v represent row vectors of E u , E v respectively;

[0064] e3) judging whether the matching degree is greater than or equal to a threshold value, and if yes, recommending the service i to the user u.

[0065] The technical effectiveness of this invention is undeniable. By capturing user personality types, it matches users with similar personality traits to their preferences. For example, Type A users lack patience and don't pay attention to service details; compared to other users, they are less likely to spend time comparing similar services. Type B users, on the other hand, are patient and detail-oriented; these users have higher expectations for the service itself and are more likely to accept recommendations for similar services. Therefore, this invention takes into account the influence of user personality on user preferences, constructs a user personality map, and improves the accuracy of user representation.

[0066] This invention incorporates user personality into the factors affecting recommendation accuracy. In addition to the user-service interaction graph, it also constructs a user-service personality graph and a user social-interest personality graph. User representations are propagated on these two graphs respectively. User personality data is used to enhance user and service representations, and element-level multiplication is used to fully integrate user and service representations, ensuring that the information learned by users in the two graphs does not conflict, thus improving recommendation performance. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the prior art (I).

[0068] Figure 2 Schematic diagram II of the prior art;

[0069] Figure 3 Workflow diagram for the service recommendation system;

[0070] Figure 4 A model diagram for a service recommendation system;

[0071] Figure 5 For comparison of experimental values ​​in lastfm and ciao. Detailed Implementation

[0072] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0073] Example 1:

[0074] See Figures 3 to 5 A service recommendation system based on personality perception and complex relationship interaction includes an information acquisition module, a user personality subgraph generation module, an information dissemination module, an information fusion module, and a personalized recommendation module.

[0075] The information acquisition module randomly generates user vector information. and service vector information and transmit to the user personality subgraph generation module and the information propagation module;

[0076] The information acquisition module acquires service representation set and transmit to the personalized recommendation module;

[0077] The user personality subgraph generation module processes the user vector information and service vector information to generate a user personality subgraph and transmit to the information propagation module;

[0078] The user personality subgraph includes a user-service interaction personality subgraph and a user social-interest personality subgraph N represents the number of personality types; respectively represent the RNth user-service interaction personality subgraph and the SNth user social-interest personality subgraph;

[0079] According to the user personality subgraph, the information propagation module processes the user vector information to obtain different levels of user social representation set and user interest representation set and transmit to the information fusion module;

[0080] The information fusion module fuses the user social representation set and the user interest representation set to obtain a user representation set and transmit to the personalized recommendation module;

[0081] The personalized recommendation module matches the user representation set and the service representation set two by two, and when the matching degree between the user and the service is greater than or equal to the threshold value ε, the corresponding service is recommended to the user.

[0082] The information acquisition module uses edge structure information or standard normal distribution to randomly generate user vector information and service vector information n and m respectively represent the number of users and the number of services, and d represents the dimension of the vector.

[0083] The user personality subgraph generation module processes the user vector information and service vector information The steps of generating a user-service interaction personality subgraph include:

[0084] a1) The user personality subgraph generation module processes the user initial vector information and a first-order user-service interaction personality vector fusion, obtaining a fusion embedding vector F of the user u u1 , that is:

[0085]

[0086]

[0087] In the formula, N u , N i represent the user node set and the service node set;

[0088] a2) input the fusion embedding vector F u1 into a two-layer fully connected neural network for personality classification, to obtain a user-service interaction personality representation parameter U o1 , that is:

[0089] U h1 = σ (W 11 F u1 + b 11 ) (3)

[0090] U o1 = (W 12 U h1 + b 12 ) (4)

[0091] In the formula, W 11 , W 12 respectively represent the weight matrix; b 11 , b 12 respectively represent the weight bias; σ(·) is an activation function; U h1 is the output of the first layer fully connected neural network;

[0092] a3) according to the user-service interaction personality representation parameter U o1 , all users are divided into different user-service interaction personality subgraphs.

[0093] The activation function is a LeakyReLU activation function.

[0094] The user personality subgraph generation module processes user vector information and service vector information , and the steps of generating a user social-interest personality subgraph include:

[0095] b1) obtaining a first layer embedding of the user u in the user social graph

[0096] The user initial vector information and the first layer embedding are fused to obtain the fusion embedding vector F of the user u u2 , that is:

[0097]

[0098] b2) input the fusion embedding vector F into a two-layer fully connected neural network for personality classification to obtain the user social-interest personality representation parameter U u2 o2 , that is:

[0099] U h2 = σ (W 21 F u2 + b 21 ) (6)

[0100] U o2 = (W 22 U h2 + b 22 ) (7)

[0101] In the formula, U h2 is the output of the first layer of the two-layer fully connected neural network for personality classification; W 21 , W 22 respectively represent weight matrices; b 21 , b 22 respectively represent weight biases; σ(·) is an activation function;

[0102] b3) according to the user-service interaction personality representation parameter U o2 , all users are divided into different user social-interest personality subgraphs.

[0103] The information propagation module processes the user vector information to obtain a user personality representation set of different levels , and the steps include:

[0104] c1) input the user vector information into the graph convolution network of the information propagation module;

[0105] c2) the graph convolution network iterates K times on the user social graph to obtain a user social representation set of different levels

[0106] Wherein, the k+1 layer user social representation is as follows:

[0107] ​

[0108] wherein, denotes the set of adjacent users of user u on the social graph; denotes the set of adjacent users of user v on the social graph;

[0109] The graph convolution network iterates the user vector information on the user-service interaction graph K times to obtain a set of user interest representations of different levels

[0110] wherein, the k+1th level user interest representation is as follows:

[0111]

[0112] wherein, denotes the set of adjacent nodes of user u on the user-service graph; denotes the set of adjacent nodes of service i on the user-service graph.

[0113] The information fusion module fuses the set of user social representations and the set of user interest representations The steps of fusion include:

[0114] d1) multiplying the elements in the set of user social representations and the set of user interest representations to obtain:

[0115]

[0116] wherein, denotes the element-level multiplication between two vectors; is the user personality representation, the user interest representation; X ij is the multiplication result;

[0117] d2) inputting the multiplication result into a three-layer neural network to generate a set of user representations i.e.:

[0118] X1=σ(ξ1X ij +a1)

[0119] X2=σ(ξ2X1+a2)

[0120] X3=ξ3X2+a3

[0121]

[0122] wherein, X1, X2, X3 respectively denote the user representations in the first, second and third layers of neural network; X ois the final user representation after the fusion of the outputs of the three-layer neural network. ξ1, ξ2, ξ3 are weights; a1, a2, a3 are biases.

[0123] The personalized recommendation module matches the set of user representations and the set of service representations pairwise, including:

[0124] e1) calculating the user vector representation e u and the service vector representation e v , i.e.:

[0125]

[0126]

[0127] wherein α k is an impact factor;

[0128] e2) calculating the matching degree between the target user u and the service i , i.e.:

[0129]

[0130] wherein e u , e v respectively represent the row vectors of E u , E v ;

[0131] e3) judging whether the matching degree is greater than or equal to a threshold value, and if so, recommending the service i to the user u.

[0132] Embodiment 2:

[0133] A service recommendation system based on personality perception and complex relationship interaction, comprising an information acquisition module, a user personality subgraph generation module, an information propagation module, an information fusion module and a personalized recommendation module;

[0134] The information acquisition module randomly generates user vector information and service vector information and transmits them to the user personality subgraph generation module and the information propagation module;

[0135] The information acquisition module acquires a set of service representations and transmits them to the personalized recommendation module;

[0136] The user personality subgraph generation module processes the user vector information and the service vector information , generates a user personality subgraph, and transmits it to the information propagation module;

[0137] The user personality subgraph includes a user-service interaction personality subgraph. Social interaction with users - interest and personality subgraph N represents the number of personality types; Let represent the user-service interaction personality subgraph of the RN-th user and the social-interest personality subgraph of the SN-th user, respectively.

[0138] Based on the user personality subgraph, the information dissemination module processes user vector information. The data is processed to obtain sets of user social representations at different levels. and user interest representation set And transmit it to the information fusion module;

[0139] The information fusion module analyzes the user's social representation set. and user interest representation set By merging, a set of user representations is obtained. And transmit it to the personalized recommendation module;

[0140] The personalized recommendation module represents the user representation set. Service representation set Perform pairwise matching. If the matching degree between a user and a service is greater than or equal to the threshold ε, then recommend the corresponding service to that user.

[0141] Example 3:

[0142] A service recommendation system based on personality perception and complex relationship interaction is described in Embodiment 2. The information acquisition module uses side-structure information or a standard normal distribution to randomly generate user vector information. service vector information n and m represent the number of users and the number of services, respectively, and d represents the dimension of the vector.

[0143] Example 4:

[0144] A service recommendation system based on personality perception and complex relationship interaction, the main contents of which are described in Embodiment 2, wherein the user personality subgraph generation module processes user vector information service vector information The steps for processing and generating a user-service interaction personality subgraph include:

[0145] a1) The user personality subgraph generation module will generate user initial vector information. First-order user-service interaction personality vector The fusion is performed to obtain the fusion embedding vector F of user u. u1 ,Right now:

[0146]

[0147]

[0148] wherein N u , N i represents the user node set and the service node set;

[0149] a2) inputting the fusion embedding vector F u1 into a two-layer fully connected neural network for personality classification to obtain a user-service interaction personality characterization parameter U o1 , that is:

[0150] U h1 = σ(W 11 F u1 +b 11 ) (3)

[0151] U o1 = (W 12 U h1 +b 12 ) (4)

[0152] wherein W 11 , W 12 respectively represent weight matrices; b 11 , b 12 respectively represent weight biases; σ(·) is an activation function; U h1 is the output of the first layer fully connected neural network;

[0153] a3) according to the user-service interaction personality characterization parameter U o1 , all users are divided into different user-service interaction personality subgraphs.

[0154] Embodiment 5:

[0155] A service recommendation system based on personality perception and complex relationship interaction, the main content of which is seen in Embodiment 2, wherein the activation function is a LeakyReLU activation function.

[0156] Embodiment 6:

[0157] A service recommendation system based on personality perception and complex relationship interaction, the main content of which is seen in Embodiment 2, wherein the user personality subgraph generation module processes user vector information and service vector information to generate a user social-interest personality subgraph, and the steps include:

[0158] b1) obtaining a first layer embedding of a user u in a user social graph

[0159] fusing the user initial vector information and the first layer embedding to obtain a fusion embedding vector F of the user u u2 , that is,

[0160]

[0161] b2) inputting the fusion embedding vector F u2 into a two-layer meta fully connected neural network for personality classification to obtain a user social-interest personality representation parameter U o2 , that is,

[0162] U h2 = σ (W 21 F u2 + b 21 ) (6)

[0163] U o2 = (W 22 U h2 + b 22 ) (7)

[0164] In the formula, U h2 is the output of the first layer of the two-layer fully connected neural network for personality classification; W 21 , W 22 respectively represent weight matrices; b 21 , b 22 respectively represent weight biases; σ(·) is an activation function;

[0165] b3) according to the user-service interaction personality representation parameter U o2 , all users are divided into different user social-interest personality subgraphs.

[0166] Embodiment 7:

[0167] A service recommendation system based on personality perception and complex relationship interaction, the main content of which is shown in Embodiment 2, wherein the information propagation module processes user vector information to obtain a user personality representation set of different levels.

[0168] c1) inputting user vector information into a graph convolution network of the information propagation module;

[0169] c2) the graph convolution network iterates the user vector information on the user social graph K times to obtain a user social representation set of different levels

[0170] wherein the k+1th level user social representation as follows:

[0171]

[0172] wherein, denotes the adjacent user set of user u on the social graph; denotes the adjacent user set of user v on the social graph;

[0173] the graph convolution network iterates the user vector information on the user-service interaction graph K times to obtain a user interest representation set of different levels

[0174] wherein the k+1th level user interest representation as follows:

[0175]

[0176] wherein, denotes the adjacent node set of user u on the user service graph; denotes the adjacent node set of service i on the user service graph.

[0177] Embodiment 8:

[0178] A service recommendation system based on personality perception and complex relationship interaction, the main content is seen in embodiment 2, wherein the information fusion module fuses the user social representation set and the user interest representation set The step includes:

[0179] d1) the elements in the user social representation set and the user interest representation set are multiplied to obtain:

[0180]

[0181] wherein, ⊙ denotes the element-level multiplication between two vectors; is the user personality representation, the user interest representation; X ij is the product result;

[0182] d2) the product result is input into a three-layer neural network to generate a user representation set that is:

[0183] X1=σ(ξ1Xij +a1)

[0184] X2= σ(ξ2X1+a2)

[0185] X3= ξ3X2+a3

[0186]

[0187] wherein X1, X2, X3 represent user representations in the first, second, and third neural networks, respectively; X o is the final user representation after the outputs of the three-layer neural network are fused. ξ 1、 ξ2, ξ3 are weights; a1, a2, a3 are biases.

[0188] Embodiment 9:

[0189] A service recommendation system based on personality perception and complex relationship interaction, the main content of which is shown in Embodiment 2, wherein the personalized recommendation module performs a step of matching the user representation set and the service representation set twice as follows:

[0190] e1) calculating user vector representation e u and service vector representation e v , i.e.

[0191]

[0192]

[0193] wherein α k is an influence factor;

[0194] e2) calculating the matching degree between target user u and service i , i.e.

[0195]

[0196] wherein e u , e v represent row vectors of E u , E v , respectively;

[0197] e3) judging whether the matching degree is greater than or equal to a threshold value, and if so, recommending service i to user u.

[0198] Embodiment 10:

[0199] A service recommendation system based on personality perception and complex relationship interaction, comprising an information acquisition module, a user personality division and capture module, an information propagation module, an information fusion module, and a personalized recommendation module.

[0200] Information acquisition.

[0201] The information acquisition module randomly generates user vector information and service vector information Wherein, the information acquisition module uses edge structure information or standard normal distribution to randomly generate user vector information and service vector information n and m represent the number of users and the number of services respectively, and d represents the dimension of the vector.

[0202] User personality subgraph generation.

[0203] The user personality subgraph generation module aims to divide users with similar personalities into the same group and generate user personality subgraphs based on this.

[0204] Specifically, the user-service interaction graph is represented by graph Users with similar personalities are divided into a group, and a user-service interaction personality subgraph is generated The user social graph is represented by graph The personality subgraph module groups users with similar interests into the same social network, generating a user social-interest personality subgraph Where N represents the number of personality types.

[0205] The specific steps for obtaining the user-service interaction personality subgraph and the user social-interest personality subgraph are as follows:

[0206] User-service personality subgraph generation.

[0207] The user initial vector information and the first-order user-service interaction personality vector are input into the fusion module to generate the fusion embedding vector of user u:

[0208]

[0209] Where F u represents the fusion embedding of u, F u (·) represents a new fusion module. The new fusion vector is input into a two-layer fully connected neural network to classify the user's personality:

[0210] U h = σ(W1F u +b1)

[0211] Uo = (W2U h +b2) (2)

[0212] where W1, W2 and b1, b2 represent weight matrix and weight bias respectively, and the weight matrix weight bias is LeakyReLU activation function. The user group is segmented into different personality subgraphs by unsupervised classification. The F u similar, and U o also similar, and U o The index of the maximum value represents the number of the personality subgraph to which the user u belongs.

[0213] User social-interest personality subgraph generation.

[0214] represents the first layer embedding of user u in the user social graph. Like the user-service personality subgraph, the user initial vector information and is input into the fusion module to generate the fusion embedding vector of user u:

[0215]

[0216] The new fusion vector is input into a two-layer fully connected neural network to classify the personality of the user:

[0217] U h = σ(W1F u +b1)

[0218] U o = (W2U h +b2) (4)

[0219] Where the target user u will only appear in one user-service interaction personality subgraph and one user social-interest personality subgraph. At the same time, the one-hop neighbors of the target user u, i.e. the projects or friends directly connected to the user u, will be added to the same subgraph as the user u.

[0220] Information propagation

[0221] The information propagation module has a user service interaction subgraph, a user social relationship subgraph and a graph convolutional neural network. The user service interaction subgraph is a personality subgraph composed of users with the same personality and their interactive services; the user social relationship subgraph is a personality subgraph of the social relationship of users with the same personality.

[0222] The steps of obtaining a set of user personality representations at different levels comprise:

[0223] 1) input user vector information to a graph convolution network, which iterates K times on the user social graph to obtain a set of user social representations at different levels

[0224] 2) input user vector information to a graph convolution network, which iterates K times on the user-service interaction graph to obtain a set of user interest representations at different levels

[0225] where the k+1th level user social representation is as follows:

[0226]

[0227] where, denotes the set of adjacent users of user u on the social graph. denotes the set of adjacent users of user v on the social graph.

[0228] 3) the k+1th level user interest representation is as follows:

[0229]

[0230] Information fusion

[0231] The information fusion module, the present application adopts a new feature fusion algorithm to fuse the user representation obtained from the user-service social subgraph and the user social-interest subgraph. Specifically, there are user representations h i and h i from different subgraphs or different connection layers, and element-level multiplication is performed on them:

[0232] X ij = [h i , h j , h i ⊙ h j ] (7)

[0233] where ⊙ denotes element-level multiplication between two vectors. Input X ij to a three-layer neural network to generate the final representation:

[0234] X1= σ(ξ1X ij +a1)

[0235] X2= σ(ξ2X1+a2)

[0236] X3= ξ3X2+a3

[0237]

[0238] The information fusion module, wherein X1, X2, X3 respectively represent user representations in a three-layer neural network, X o is the final representation after the three-layer neural network output is fused.

[0239] Personalized recommendation

[0240] The personalized recommendation module determines the matching degree between the user and the item according to the item representation set and the user representation set When the matching degree reaches the threshold value ε, the target user is recommended. The specific steps are as follows:

[0241] Calculate the user vector representation e u and the service vector representation e v .

[0242]

[0243]

[0244] Calculate the matching degree between the target user u and the item i

[0245]

[0246] In the formula, e u , e v respectively represent the row vectors of E u , E v .

[0247] Example 11:

[0248] The verification test of the service recommendation system based on personality perception and complex relationship interaction is as follows:

[0249] 1) Obtain the data set as shown in Table 1.

[0250] Table 1: Data set used in the experiment

[0251]

[0252] 2) Use BPR, LightGCN, and IMP-GCN three models to process the data in Table 1 respectively, and obtain the experimental comparison values in Table 2, which respectively compare the recall rate, normalized discounted gain (NDCG), and accuracy (precision).

[0253] Table 2: Comparison of experimental values of each model in the original data set

[0254]

[0255] The application discloses a service recommendation method based on personality perception and complex relationship interaction, comprising an information acquisition module, a user personality division and capture module, an information propagation module, an information fusion module and a personalized recommendation module. For the first time, the user personality is considered in the factors affecting the recommendation accuracy. In addition to the user-service interaction graph, the application further constructs a user-service personality graph and a user social-interest personality graph. The user representation is propagated on the two graphs respectively, the user and service representations are enhanced by using the user personality data, the user and service representations are fully fused by using the element-level product, the information learned by the user in the two graphs is ensured not to conflict with each other, and the recommendation effect is ensured to be improved.

Claims

1.A service recommendation system based on personality perception and complex relationship interaction, characterized in that, The information acquisition module, the user personality subgraph generation module, the information propagation module, the information fusion module and the personalized recommendation module are included. The information acquisition module randomly generates user vector information and service vector information and transmits to the user personality subgraph generation module and the information dissemination module The information acquisition module acquires a service representation set and transmits to the personalized recommendation module; The user character subgraph generation module processes the user vector information and service vector information to generate a user character subgraph and transmits the user character subgraph to the information dissemination module. The user personality sub-graphs include a user-service interaction personality sub-graph and a user social-interest personality sub-graph N represents the number of personality types; respectively represent the RNth user-service interaction personality sub-graph, the SNth user social-interest personality sub-graph; According to the user personality subgraph, the information propagation module processes the user vector information to obtain a user social representation set of different levels and a user interest representation set and transmits them to the information fusion module; The information fusion module analyzes the user's social representation set. and user interest representation set By merging, a set of user representations is obtained. And transmit it to the personalized recommendation module; The personalized recommendation module represents the user representation set. Service representation set Perform pairwise matching. If the matching degree between a user and a service is greater than or equal to the threshold ε, then recommend the corresponding service to that user. The user character subgraph generation module processes the user vector information and the service vector information e i (0) to generate a user-service interaction character subgraph. 1) The user personality subgraph generation module fuses the user vector information and the first-order user-service interaction personality vector to obtain the fused embedding vector F of the user u u1 , that is: In the formula, N u , N i represents the user node set and the service node set; 2) the fusion embedding vector F u1 is input into a two-layer fully connected neural network for personality classification, to obtain a user-service interaction personality characterization parameter U o1 That is: U h1 = σ(W 11 F u1 +b 11 ) (3) U o1 = (W 12 U h1 + b 12 ) (4) wherein W 11 , W 12 represent weight matrices, respectively; b 11 , b 12 represent weight biases, respectively; σ(·) is an activation function; U h1 is the output of the first fully connected neural network. 3) according to the user-service interaction personality characterization parameter U o1 all users are divided into different user-service interaction personality subgraphs; The user character subgraph generation module processes the user vector information and service vector information to generate a user social-interest character subgraph, and the steps include: s1 ) obtaining a first layer embedding of a user u in a user social graph The user vector information and the first layer embedding are fused to obtain a fused embedding vector F of the user u u2 That is: s2) embedding the fusion vector F u2 into a two-layer fully connected neural network for personality classification, to obtain a user social-interest personality characterization parameter U o2 i.e.: U h2 = σ(W 21 F u2 +b 21 ) (6) U o2 = (W 22 U h2 + b 22 ) (7) In the formula, U h2 is the output of the first layer of the two-layer fully connected neural network for personality classification; W 21 , W 22 respectively represent weight matrices; b 21 , b 22 respectively represent weight biases; σ(·) is an activation function; s3) according to the user-service interaction personality characterization parameter U o2 all users are divided into different user social-interest personality subgraphs. 2.The service recommendation system based on personality perception and complex relationship interaction according to claim 1, wherein, The information acquisition module adopts side structure information or standard normal distribution to randomly generate user vector information and service vector information n and m respectively represent the number of users and the number of services, and d represents the dimension of the vector. 3.The service recommendation system based on personality perception and complex relationship interaction according to claim 1, characterized in that, The activation function is a LeakyReLU activation function. 4.The service recommendation system based on personality perception and complex relationship interaction according to claim 1, wherein, The information propagation module processes the user vector information to obtain a user representation set of different levels The steps include: 1) inputting the user vector information to a graph convolution network of the information propagation module; 2) the graph convolution network iterates the user vector information on the user social graph K times to obtain a set of user social representations at different levels wherein the (k+1)th layer user social representation As shown below: wherein denotes the set of adjacent users of user u on the social graph; denotes the set of adjacent users of user v on the social graph; The graph convolution network iterates K times on the user-service interaction lattice graph to obtain a set of user interest representations at different levels wherein the (k+1)th layer user interest representation as follows: wherein denotes the set of adjacent nodes of a user u on the user-service graph; denotes the set of adjacent nodes of a service i on the user-service graph. 5.The service recommendation system based on personality perception and complex relationship interaction according to claim 1, wherein, The information fusion module fuses a set of user social representations and a set of user interest representations The step of fusing includes: 1) the product of the elements in the set of user social representations and the set of user interest representations results in: In the formula, ⊙ represents an element-level multiplication between two vectors; is a user social representation, a user interest representation; X ij is a product result; 2) input the product result into a three-layer neural network to generate a set of user representations That is: X1 = σ(ξ1X ij + a1) X2 = σ(ξ2X1 + a2) X3 = ξ3X2 + a3 In the formula, X1, X2, X3 respectively represent user representations in the first, second and third layer neural networks; X o is the final user representation after the output of the three-layer neural network is fused; ξ1, ξ2, ξ3 are weights; a1, a2, a3 are biases; and σ(·) is an activation function. 6.The service recommendation system based on personality perception and complex relationship interaction according to claim 1, wherein, The personalized recommendation module pairs the set of user representations with the set of service representations and the set of service representations The step of pairing each other includes: 1) Compute user vector representation e u and service vector representation e i i.e.: In the formula, α k is a parameter; 2) calculating the matching degree between the target user u and the service i That is: wherein e u , e v respectively represent row vectors; 3) judging the matching degree whether it is greater than or equal to a threshold value, and if so, recommending the service i to the user u.

Images