A project information service recommendation method based on graph neural network

Abstract

The present invention discloses a project information service recommendation method based on a graph neural network, belongs to the technical field of project information service and recommendation, and comprises the following steps: step 1, obtaining a data set; step 2, constructing a bipartite graph; step 3, constructing an M‑LGACL model, comprising a feature vector initialization module, a Mixup data enhancement module, a lightweight graph convolution layer, a cross attention mechanism, and a contrast loss learning module; step 4, generating a trained M‑LGACL model by continuously iterating the M‑LGACL model constructed in step 3; step 5, evaluating the model performance by calculating the recall rate and the normalized discounted cumulative gain; the project information service recommendation method based on a graph neural network provided by the present invention not only improves the recommendation accuracy of the model, but also effectively solves the problems of data sparsity and insufficient model generalization ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of project information services and recommendations, and in particular, relates to a project information service recommendation method based on graph neural network. Background Art

[0002] Item information service is an important tool in modern retail, e-commerce and marketing. It helps merchants and consumers better understand and meet market needs by collecting, analyzing and displaying item data. With the rapid development of the Internet and mobile computing technology, big data and artificial intelligence technologies are widely used in personalized recommendation systems, which greatly improves users' online shopping experience. By analyzing users' historical behaviors and preferences, recommendation systems are able to recommend relevant items and services to users. However, although item information services and recommendation systems have achieved remarkable achievements in their respective fields, the combination of the two is still in its early stages. With the popularity of smartphones and mobile payment technologies, people are beginning to seek more personalized and context-related shopping experiences, which requires recommendation systems to consider not only users' preferences, but also their shopping habits and consumption environment. In addition, the diversification of consumer demand and the fierce market competition have further promoted the development of data-based item recommendation services and promoted the research and application of the combination of item information services and recommendation systems. Regarding system recommendation models, in the early days, they mainly relied on collaborative filtering methods, using users' historical interaction information to model, thereby achieving personalized recommendations. There are three relatively classic and representative implementations of collaborative filtering. The first is the matrix decomposition method, which is the basic paradigm of the recommendation model. The second is the neural collaborative filtering (NCF) model, which introduces a multi-layer perceptron (MLP) based on matrix decomposition to capture the nonlinear relationship between users and items. The third is the ConVNCF model, which further improves the NCF model by integrating the convolutional neural network (CNN). It uses the outer product to simulate the user-item interaction pattern, and uses CNN to capture the high-order correlation between the embedding dimensions, thereby enhancing the model's ability to capture the interaction pattern. However, collaborative filtering methods are mainly used to deal with supervised or semi-supervised learning scenarios, which rely on a large amount of accurate labeled data. However, in real applications, obtaining a large amount of accurate labeled data is time-consuming and expensive, and the reliance on labeled data poses a challenge. In addition, the sparsity of labeled data may lead to overfitting of model training, limiting the performance of the recommendation system in different scenarios; in addition, the noise present in the interaction process may distort the signal learned by the recommendation system, thereby affecting the overall performance of the system. Summary of the invention

[0003] The purpose of the present invention is to provide a project information service recommendation method based on graph neural network, so as to solve the problems existing in the prior art that the sparsity of label data leads to overfitting of model training, and the interactive noise distorts the signal learned by the recommendation system, thereby affecting the overall performance of the system.

[0004] To achieve the above object, the present invention provides a project information service recommendation method based on graph neural network, comprising the following steps:

[0005] Step 1: Get the data set;

[0006] Step 2: Construct a bipartite graph;

[0007] Step 3: Construct the M-LGACL model, including the feature vector initialization module, the Mixup data enhancement module, the lightweight graph convolution layer, the cross attention mechanism, and the contrast loss learning module;

[0008] Step 4: Generate a trained M-LGACL model by continuously iterating the M-LGACL model constructed in step 3;

[0009] Step 5: Evaluate the model performance by calculating the recall rate and normalized discounted cumulative gain.

[0010] Preferably, the specific process of constructing the bipartite graph in step 2 is as follows:

[0011] S21. Named Entity Extraction: Use natural language processing technology to identify and extract named entities from the Yelp dataset; named entities include but are not limited to project names, users, merchants, and location information;

[0012] S22, Relationship extraction: For the identified named entities, the relationships between entities are extracted by analyzing their associations in the data; for example, the relationship between a user and an item can be a comment relationship or a purchase relationship, and the relationship between a user and a merchant can be a consumption relationship, etc.;

[0013] S23, attribute extraction: After extracting the named entities and relations, further extract the attribute information of each named entity; the attribute information includes the category of the item, the price, the user's preference and the merchant's rating; these attribute information help to more comprehensively describe the named entity;

[0014] S24. Combine the named entities, relationships and attribute information to complete the construction of the bipartite graph of the user project.

[0015] Preferably, the feature vector initialization module constructs the embedding matrices of user u and item i into an overall embedding matrix, and the specific expression is as follows:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, the embedding vector Represents the features of user u, embedded vector represents the characteristics of item i, represents a set of real numbers, d represents the dimension of the embedding vector, represents the overall embedding matrix of user u, represents the overall embedding matrix of item i.

[0021] Preferably, the Mixup data enhancement module is specifically as follows: select two similar user training samples and Or similar project training samples and , generate a new training sample or ,here and are the embedding vectors of the two users, and The specific expressions of the embedding vectors of the items are as follows:

[0022] ;

[0023] ;

[0024] In the formula, is a distribution from Beta ( , ), , this parameter controls the proportion of the new sample coming from the two original samples.

[0025] Preferably, the operation process of the lightweight graph convolution layer is specifically as follows:

[0026] First, embed the user u and item i at the initial stage and To perform aggregation, the specific expression is as follows:

[0027] ;

[0028] ;

[0029] Then, by stacking multiple layers of graph convolution, neighbor information of different distances is aggregated to form the final user and item embeddings. , , the specific expression is as follows:

[0030] ;

[0031] ;

[0032] Finally, the embedded propagation process is represented in matrix form as follows:

[0033] Define the user-item interaction matrix as , where M and N represent the number of users and items respectively. If u interacts with i, then R is 1, otherwise it is 0; then the adjacency matrix of user-item interaction is as follows:

[0034] ;

[0035] Where T is the embedding size, given the embedding matrix at layer 0, the embedding propagation process It is described by the following matrix formula, as follows:

[0036] ;

[0037] In the formula, and They represent the kth layer as the embedding of user u and item i, respectively. and They represent the neighbor sets of user u and item i respectively, K is the number of graph convolution layers, are the weights in the final embedding of the kth layer, which can usually be set to equal weight or adjusted according to the number of layers, represents the node embedding matrix of the kth layer. Each column represents the embedding vector of a node (user or item) in the graph. A is the adjacency matrix of the graph, which includes self-connection, that is, the original adjacency matrix A plus the unit matrix I, the addition of this self-connection helps to retain the characteristic information of the node itself; D is a The diagonal matrix of It is a normalization operation to ensure that the scaling of information during propagation remains consistent, which helps the stability and convergence of the model.

[0038] Preferably, the cross attention mechanism is specifically as follows: for each user u and item i, define the user's query, key, and value as , , , the query, key, and value of the project are , , ; Then, the cross attention is calculated as follows, including the user query , the key of the project , the value of the item , the user's attention weight on the project, the specific calculation expression is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] In the formula, , , represents the weight matrix, Represents the characteristics of the user, Indicates the characteristics of the project, express is the key vector The dimension of is used to scale the dot product to control the scale of the gradient.

[0045] Preferably, the contrast loss learning module includes the design and optimization of Triplet Loss, specifically:

[0046] The expression of Triplet Loss is as follows:

[0047] ;

[0048] Triplet Loss is optimized through InfoNCE. The specific expression is as follows:

[0049] ;

[0050] In the formula, is a user vector output, is a positive sample item vector, indicating the items that the user likes, is a negative sample item vector, indicating items that the user dislikes, is a predefined interval parameter used to control the distance difference between positive sample pairs and negative sample pairs. d( , ) Is a distance metric function used to measure vector and 、 The distance or similarity between is the temperature parameter, controlling the scale of similarity.

[0051] Preferably, the calculation expression of the recall rate in step 5 is as follows:

[0052] ;

[0053] In the formula, represents the number of samples that are actually positive examples and predicted to be positive examples, Indicates the number of samples that are actually positive but predicted to be negative.

[0054] Preferably, the calculation expression of normalized loss cumulative gain is as follows:

[0055] ;

[0056] ;

[0057] In the formula, represents the relevance score of the recommendation result at position i, Indicates the order of the recommended results in the ranking list. is the loss factor at position i, represents the cumulative gain of loss, Represents the ideal loss cumulative gain.

[0058] Therefore, the present invention adopts the above-mentioned project information service recommendation method based on graph neural network, which has the following beneficial effects:

[0059] (1) The Mixup data enhancement, cross-attention mechanism, and contrastive loss learning are successfully integrated into the service recommendation model based on graph neural network, which not only improves the recommendation accuracy of the model, but also effectively solves the problems of data sparsity and insufficient model generalization ability;

[0060] (2) Effectiveness of data enhancement: By introducing Mixup data enhancement technology, the model can generate and utilize virtual user-service interaction data, effectively alleviating the data sparsity problem, enhancing the generalization ability of the model, and reducing the risk of overfitting;

[0061] (3) The application of the cross-attention mechanism enables the model to more accurately understand and capture the complex interactions between user preferences and service characteristics, significantly improving the quality of personalized recommendations in the recommendation system;

[0062] (4) By introducing contrastive loss learning, the distance between similar samples is shortened and the distance between dissimilar samples is increased, which significantly optimizes the model at the feature representation level, enhances the ability to capture subtle differences between users and services, and improves the accuracy and robustness of the recommendation system.

[0063] (5) Extensive experimental evaluations show that the recommendation method of this application demonstrates significant performance improvements over existing recommendation system methods and graph neural network models on multiple standard datasets, especially in alleviating the cold start problem and processing sparse data.

[0064] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is an overall flow chart of a project information service recommendation method based on graph neural network of the present invention;

[0066] Figure 2 This is a framework diagram of the M-LGACL model according to an embodiment of the present invention;

[0067] Figure 3 This is a diagram showing the Mixup enhancement technology of an embodiment of the present invention;

[0068] Figure 4 This is a model diagram of the cross-attention mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] See also Figure 1-Figure 4 , a project information service recommendation method based on graph neural network, comprising the following steps:

[0071] Step 1: Get the data set;

[0072] Step 2: Construct a bipartite graph; the specific process is as follows:

[0073] S21. Named Entity Extraction: Use natural language processing technology to identify and extract named entities from the Yelp dataset; named entities include but are not limited to project names, users, merchants, and location information;

[0074] S22, Relationship extraction: For the identified named entities, the relationships between entities are extracted by analyzing their associations in the data; for example, the relationship between a user and an item can be a comment relationship or a purchase relationship, and the relationship between a user and a merchant can be a consumption relationship, etc.;

[0075] S23, attribute extraction: After extracting the named entities and relations, further extract the attribute information of each named entity; the attribute information includes the category of the item, the price, the user's preference and the merchant's rating; these attribute information help to more comprehensively describe the named entity;

[0076] S24. Combine the named entities, relationships and attribute information to complete the construction of the bipartite graph of the user project.

[0077] Step 3: Construct the M-LGACL model, including the feature vector initialization module, the Mixup data enhancement module, the lightweight graph convolution layer, the cross attention mechanism, and the contrast loss learning module;

[0078] The feature vector initialization module builds the embedding matrices of user u and item i into an overall embedding matrix. Initializing the embedding allows learning the user's preferences and the characteristics of the product without explicit features (such as the user's personal information or the specific description of the item). As the model training progresses, these embeddings will be adjusted through the interaction between users and items (such as ratings and comments) to better reflect the user's preferences and the characteristics of the product. The specific expression is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula, the embedding vector Represents the features of user u, embedded vector represents the characteristics of item i, represents a set of real numbers, d represents the dimension of the embedding vector, represents the overall embedding matrix of user u, represents the overall embedding matrix of item i.

[0084] Mixup data enhancement module: In the recommendation system, Mixup is a data enhancement method that was originally used in the field of image processing to improve the generalization ability and robustness of the model. This method generates new training samples by linear interpolation between input data. For the recommendation system, the Mixup data enhancement method can help the model better understand and learn the complex interactive relationship between users and products, thereby improving the accuracy and robustness of recommendations; the Mixup data enhancement module is specifically: select two similar user training samples and Or similar project training samples and , generate a new training sample or ,here and are the embedding vectors of the two users, and The specific expressions of the embedding vectors of the items are as follows:

[0085] ;

[0086] ;

[0087] In the formula, is a distribution from Beta ( , ), , this parameter controls the proportion of the new sample coming from the two original samples.

[0088] like Figure 3 The figure shows the data performance after using Mixup data enhancement. The original bipartite graph is user A and item 1, user B and item 2. After using Mixup data enhancement technology, a new bipartite graph of virtual users AB and item 12 is generated. The new embedding vector generated by Mixup can increase the diversity of training data and improve the problem of data sparsity. It also helps the recommendation model learn to generalize more smoothly between user preferences. In this experiment, the Mixup data enhancement method was used to enhance 15% of the data set.

[0089] Lightweight Graph Convolution Layer: The main goal of the Lightweight Graph Convolution Layer (LightGCN) is to leverage the graph structure to update the feature representation of nodes (users and items). In the LightGCN layer, this process is simplified to only include feature aggregation of neighboring nodes without involving weight matrices or nonlinear activation functions. This simplification helps the model capture the topological structure of the graph while reducing the complexity of the model and the risk of overfitting. The purpose of the LGC layer is to create new node representations by aggregating user nodes and item nodes on the graph structure. To achieve this goal, the model combines the features of a node's neighborhood (i.e., adjacent nodes) through iterative graph convolution to update the vector representation of the node. The specific operation process of the lightweight graph convolution layer is as follows:

[0090] First, embed the user u and item i at the initial stage and To perform aggregation, the specific expression is as follows:

[0091] ;

[0092] ;

[0093] Then, by stacking multiple layers of graph convolution, neighbor information of different distances is aggregated to form the final user and item embeddings. , , the specific expression is as follows:

[0094] ;

[0095] ;

[0096] Finally, the embedding propagation process is represented in matrix form; in order to clearly show the overall picture of embedding propagation and facilitate batch processing, the propagation rules are given in matrix form. This expression can simplify the calculation process and make the application of the embedding update mechanism in the entire network more intuitive and efficient. The matrix form of propagation rules allows the model to operate more efficiently on the entire graph instead of processing each node individually, which is particularly important for large-scale data sets. The details are as follows:

[0097] Define the user-item interaction matrix as , where M and N represent the number of users and items respectively. If u interacts with i, then R is 1, otherwise it is 0; then the adjacency matrix of user-item interaction is as follows:

[0098] ;

[0099] Where T is the embedding size, given the embedding matrix at layer 0, the embedding propagation process It is described by the following matrix formula, as follows:

[0100] ;

[0101] In the formula, and They represent the kth layer as the embedding of user u and item i, respectively. and They represent the neighbor sets of user u and item i respectively, K is the number of graph convolution layers, are the weights in the final embedding of the kth layer, which can usually be set to equal weight or adjusted according to the number of layers, represents the node embedding matrix of the kth layer. Each column represents the embedding vector of a node (user or item) in the graph. A is the adjacency matrix of the graph, which includes self-connection, that is, the original adjacency matrix A plus the unit matrix I, the addition of this self-connection helps to retain the characteristic information of the node itself; D is a The diagonal matrix of It is a normalization operation to ensure that the scaling of information during propagation remains consistent, which helps the stability and convergence of the model.

[0102] Cross-attention mechanism: The cross-attention mechanism achieves effective fusion and complementation of information by evaluating the mutual influence between multiple inputs, or the influence of a single input on multiple inputs. In the field of image research, this mechanism greatly improves the model's ability to handle image tasks by extracting the contextual features of a single pixel in its row and column. This ability enables the model to understand the environmental information around the pixel, thereby significantly improving performance in a variety of tasks such as image classification and object detection. Specifically: For each user u and item i, define the user's query, key, and value as , , , the query, key, and value of the project are , , ; Then, the cross attention is calculated as follows, including the user query , project key , the value of the item , the user's attention weight on the project, the specific calculation expression is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, , , represents the weight matrix, Represents the characteristics of the user, Indicates the characteristics of the project, express is the key vector The dimension of is used to scale the dot product to control the scale of the gradient.

[0109] The contrast loss learning module includes the design and optimization of Triplet Loss, specifically:

[0110] The expression of Triplet Loss is as follows:

[0111] ;

[0112] Triplet Loss is optimized through InfoNCE. The specific expression is as follows:

[0113] ;

[0114] In the formula, is a user vector output, is a positive sample item vector, indicating the items that the user likes, is a negative sample item vector, indicating items that the user dislikes, is a predefined interval parameter used to control the distance difference between positive sample pairs and negative sample pairs. d( , ) Is a distance metric function used to measure vector and 、 The distance or similarity between is the temperature parameter, controlling the scale of similarity.

[0115] Step 4: Generate a trained M-LGACL model by continuously iterating the M-LGACL model constructed in step 3;

[0116] Step 5: Evaluate the model performance by calculating the recall rate and normalized discounted cumulative gain. The calculation expression of the recall rate is as follows:

[0117] ;

[0118] In the formula, represents the number of samples that are actually positive examples and predicted to be positive examples, Indicates the number of samples that are actually positive but predicted to be negative.

[0119] The calculation expression of normalized discounted cumulative gain is as follows:

[0120] ;

[0121] ;

[0122] In the formula, represents the relevance score of the recommendation result at position i, Indicates the order of the recommended results in the ranking list. is the loss factor at position i, represents the cumulative gain of loss, Represents the ideal loss cumulative gain.

[0123] Example

[0124] The present invention uses the Yelp 2018 and douban-book datasets, and the relevant information is shown in Table 1:

[0125] Table 1 Dataset information

[0126] ;

[0127] Among them, datasets are the names of the two datasets used; users are the people who rate items and accept items recommended by the recommendation system; items are the products we want to recommend; interactions are the number of times users enter information, use functions, and other interactive behaviors on the recommendation system.

[0128] Based on the LightGCN recommendation algorithm framework, the experimental parameters are set as follows: the TOP@N recommendation performance is evaluated by the full ranking protocol, and the average value is taken after 20 experiments; the model is optimized by the Adam optimizer, the learning rate is set to 0.001, the batch size is set to 1024, the embedding size is set to 64, the number of convolutional layers is set to 3, and the parameter in the Mixup data enhancement method is set to 0.4. After the above settings, the M-LGACL model in the present invention is tested on the Yelp 2018 and douban-book datasets with NGCF, LightGCN, and SGL, respectively. The specific results are shown in Tables 2 and 3:

[0129] Table 2 Comparison of experimental results on the Yelp dataset

[0130] ;

[0131] Table 3 Comparison of experimental results on the douban-book dataset

[0132] ;

[0133] The following conclusions were drawn:

[0134] ① By comparing the NGCF method that directly applies graph neural networks to collaborative filtering with the simplified graph neural network LightGCN, which removes nonlinear activation functions and feature conversion, the results in the table show that the performance of the latter has been improved. This suggests that the simplification and optimization of graph neural networks have a positive effect on improving system performance; by reducing model complexity, LightGCN not only maintains the ability of graph neural networks to process complex user-item interaction data, but also improves computational efficiency and model generalization capabilities;

[0135] ②Through comparative analysis with other technologies, it can be observed that on both sets of data, the self-supervised graph neural network model SGL, combined with the contrastive learning method, shows better performance than other GCN technologies. This verifies the effectiveness of self-supervised learning, such as contrastive loss learning methods, in improving collaborative filtering effects;

[0136] ③From the evaluation results shown in the above two datasets, it can be seen that the M-LGACL method has improved compared with other methods, which further confirms that the method of combining Mixup data enhancement, cross-attention mechanism and contrast loss learning is feasible.

[0137] Therefore, the present invention adopts the above-mentioned project information service recommendation method based on graph neural network, and successfully integrates Mixup data enhancement, cross-attention mechanism and contrast loss learning into the service recommendation model based on graph neural network, which not only improves the recommendation accuracy of the model, but also effectively solves the problems of data sparsity and insufficient model generalization ability.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A project information service recommendation method based on graph neural network, characterized in that: The following steps are involved: Step 1: Get the data set; Step 2: Construct a bipartite graph; Step 3: Construct the M-LGACL model, including the feature vector initialization module, the Mixup data enhancement module, the lightweight graph convolution layer, the cross attention mechanism, and the contrast loss learning module; Step 4: Generate a trained M-LGACL model by continuously iterating the M-LGACL model constructed in step 3; Step 5: Evaluate the model performance by calculating the recall rate and normalized discounted cumulative gain; The specific process of constructing the bipartite graph in step 2 is as follows: S21. Named Entity Extraction: Use natural language processing technology to identify and extract named entities from the data set; named entities include project names, users, merchants, and location information; S22, Relationship extraction: For the identified named entities, extract the relationships between them by analyzing their associations in the data; S23, attribute extraction: After extracting the named entities and relations, further extract the attribute information of each named entity; the attribute information includes the category, price, user preference and merchant rating of the item; S24, combining the named entities, relations and attribute information to complete the construction of the bipartite graph of the user project; The feature vector initialization module is to build the embedding matrix of user u and item i into an overall embedding matrix. The specific expression is as follows: In the formula, the embedding vector Represents the features of user u, embedded vector represents the characteristics of item i, represents a set of real numbers, d represents the dimension of the embedding vector, represents the overall embedding matrix of user u, represents the overall embedding matrix of item i; The Mixup data enhancement module is as follows: select two similar user training samples and Or similar project training samples and , generate a new training sample or ,here and are the embedding vectors of the two users, and The specific expressions of the embedding vectors of the items are as follows: In the formula, is a distribution from Beta ( , ), ; The operation process of the lightweight graph convolution layer is as follows: First, embed the user u and item i at the initial stage and To perform aggregation, the specific expression is as follows: Then, by stacking multiple layers of graph convolution, neighbor information of different distances is aggregated to form the final user and item embeddings. , , the specific expression is as follows: Finally, the embedded propagation process is represented in matrix form as follows: Define the user-item interaction matrix as , where M and N represent the number of users and items respectively. If u interacts with i, then R is 1, otherwise it is 0; then the adjacency matrix of user-item interaction is as follows: Where T is the embedding size, given the embedding matrix at layer 0, the embedding propagation process It is described by the following matrix formula, as follows: In the formula, and They represent the kth layer as the embedding of user u and item i, respectively. and They represent the neighbor sets of user u and item i respectively, K is the number of graph convolution layers, are the weights in the final embedding of layer k, represents the node embedding matrix of the kth layer, each column represents the embedding vector of a node in the graph, A is the adjacency matrix of the graph; D is a The diagonal matrix of It is an operation of normalization; The cross attention mechanism is as follows: for each user u and item i, define the user's query, key, and value as , , , the query, key, and value of the project are , , ; Then, the cross attention is calculated as follows, including the user query , project key , the value of the item , the user's attention weight on the project, the specific calculation expression is as follows: In the formula, , , represents the weight matrix, Represents the characteristics of the user, Indicates the characteristics of the project, express is the key vector Dimensions; The contrast loss learning module includes the design and optimization of Triplet Loss, specifically: The expression of Triplet Loss is as follows: Triplet Loss is optimized through InfoNCE. The specific expression is as follows: In the formula, is a user vector output, is a positive sample item vector, indicating the items that the user likes, is a negative sample item vector, indicating items that the user dislikes, is a predefined interval parameter used to control the distance difference between positive sample pairs and negative sample pairs. d( , ) Is a distance metric function used to measure vector and 、 The distance or similarity between is the temperature parameter, controlling the scale of similarity; The calculation expression of recall rate in step 5 is as follows: In the formula, represents the number of samples that are actually positive examples and predicted to be positive examples, Indicates the number of samples that are actually positive but predicted to be negative; The calculation expression of normalized discounted cumulative gain is as follows: In the formula, represents the relevance score of the recommendation result at position i, Indicates the position order of the recommended results in the ranking list, is the loss factor at position i, represents the cumulative gain of loss, Represents the ideal loss cumulative gain.

Images