Feature intersection-based service recommendation method, system, device and storage medium

Abstract

The present application relates to the technical field of data prediction recommendation in the business field, in particular to a service recommendation method, system and device based on feature cross and a storage medium; in order to solve the problem of low service recommendation accuracy in the prior art, the present application first constructs a user interaction graph structure, a service interaction graph structure and a user-service interaction graph structure, which are respectively processed by graph convolution to obtain user interaction features, service interaction features and user-service interaction features; then, the target user features, target service features and user-service interaction features are processed by feature cross to obtain user-service cross enhanced features; subsequently, the user interaction features, service interaction features and user-service cross enhanced features are aggregated to obtain comprehensive cross features; finally, based on the comprehensive cross features, a recommendation value is calculated, and when the recommendation value is greater than a recommendation value threshold, the target service is recommended to the target user, thereby enhancing the accuracy of service recommendation and improving user experience.

Description

Technical Field

[0001] The present invention relates to the field of business data prediction and recommendation technology, and specifically to a service recommendation method, system, device and storage medium based on feature intersection. Background Art

[0002] Modern people pursue experiences and relaxation, preferring to enrich their lives and relieve stress through travel. This has led to the emergence of commercial platforms that offer travel service recommendations. These platforms integrate a vast amount of travel service information and offer recommendations to customers, improving travel convenience and choice. However, with the increasing volume of travel service data, customers face a dilemma when faced with this vast amount of data. This increased difficulty in making choices can reduce the efficiency of travel service order fulfillment, diminish user experience, and reduce consumption of travel commercial content.

[0003] To ensure the quality of tourism services, existing technologies often tend to recommend highly rated services to customers. While highly rated services may be of higher quality, they may not necessarily meet the current customer preferences. This recommendation model also suffers from uneven resource allocation, which can easily lead to crowding at offline tourism destinations and diminish the user experience. Furthermore, some recommendation methods attempt to improve the targeted nature of service recommendations by learning the relationship between users and services through neural networks. However, these methods typically rely on shallow interactions between features and fail to capture deeper, implicit relationships between users and services, such as user preferences and potential service attributes. This results in low service recommendation accuracy, diminished user interest, and a negative impact on user experience. Summary of the Invention

[0004] The purpose of the present invention is to provide a service recommendation method, system, device and storage medium based on feature intersection.

[0005] The technical solutions of the present invention are as follows:

[0006] A service recommendation method based on feature intersection includes the following operations:

[0007] S1. The users and services that have interacted with the target user are regarded as similar users and similar services respectively; the graph structures of the target user and similar users, the graph structure of the target service and similar services, and the graph structure of the target user, similar users, target service and similar services are respectively constructed to obtain the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure; the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure are respectively processed by graph convolution to obtain user interaction features, service interaction features and user-service interaction features; during the graph convolution processing, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity and user location similarity between users; the adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance and content similarity between services; and the adjacency matrix of the user-service interaction graph structure is obtained based on the user's rating of the service;

[0008] S2. The target user and target service are embedded to obtain target user features and target service features. Among the target user features, target service features, and user-service interaction features, features with dimensions exceeding the dimension threshold are subjected to dimensionality reduction processing and then subjected to feature cross processing to obtain user-service cross-enhanced features.

[0009] S3. Aggregate user interaction features, service interaction features, and user-service cross-enhancement features to obtain comprehensive cross-features; obtain a recommendation value based on the comprehensive cross-features, target user features, and target service features; if the recommendation value is greater than the recommendation value threshold, recommend the target service to the target user.

[0010] In the adjacency matrix of the user interaction graph structure in S1, the value of each position is the social relationship degree between users, and the social relationship degree is the weighted sum of user social intimacy, user behavior similarity and user location similarity between users; user social intimacy is obtained based on the mutual attention behavior and social interaction behavior between users; user behavior similarity is obtained based on the historical service query behavior and historical service usage behavior between users; user location similarity is obtained based on the real-time geographic location and permanent geographic location between users.

[0011] In the adjacency matrix of the service interaction graph structure in S1, the value of each position is the social correlation between services. The social correlation is the weighted sum of the collaborative filtering relationship, functional correlation and content similarity between services. The collaborative filtering relationship is based on the frequency of simultaneous use of services by users and the correlation of user ratings. The functional correlation is based on the functional complementarity and upstream and downstream relationships between services. The content similarity is based on the topic similarity and service attribute similarity between services.

[0012] The specific operation of feature cross-processing in S2 is: cross-aggregate the target user features, target service features and user-service interaction features after dimensionality reduction to obtain cross-aggregation features; after the cross-aggregation features are processed by the ReLU activation function several times, they are spliced ​​with the cross-aggregation features to obtain user-service cross-enhanced features.

[0013] The cross-aggregation operation can be achieved through the following formula:

[0014]

[0015] H z is the cross-aggregation feature, b0 is the global bias term, α is the cross coefficient, X i is the feature of the i-th dimension in the target user feature, X j is the feature of the jth dimension in the target service feature, X k is the feature of the kth dimension in the user-service interaction feature, I, J, and K are the total number of dimensions of the target user feature, target service feature, and user-service interaction feature, respectively. ij is the latent vector of the feature of the i-th dimension in the target user feature and the feature of the j-th dimension in the target service feature, v ik is the latent vector of the feature of the i-th dimension in the target user feature and the feature of the k-th dimension in the user-service interaction feature, v jk is the latent vector of the feature of the j-th dimension in the target service feature and the feature of the k-th dimension in the user service interaction feature, 〈〉 is the dot product processing of the vector, w1, w2, w3, w4 are the first cross weight, second cross weight, third cross weight and fourth cross weight.

[0016] The aggregation operation in S3 is implemented through the following formula: f =(U·V)⊙H z , H f is a comprehensive cross feature, U, V, H z They are user interaction features, service interaction features and user-service cross-enhancement features respectively.

[0017] In S3, the recommended value is obtained through the following formula:

[0018] y=f(W1X u +W2X s +W3(X u ⊙X s )+W4Z+B),

[0019] y is the recommended value, Z, X u 、X sare comprehensive cross-features, target user features, and target service features, respectively; W1, W2, W3, and W4 are the first recommendation weight, the second recommendation weight, the third recommendation weight, and the fourth recommendation weight, respectively; B is the recommendation bias; and f() is the activation function.

[0020] A feature-crossing-based service recommendation system, used to implement the feature-crossing-based service recommendation method, includes:

[0021] The user interaction feature, service interaction feature and user-service interaction feature generation modules are used to identify users and services that have interacted with the target user as similar users and similar services, respectively; the graph structures of the target user and similar users, the graph structure of the target service and similar services, and the graph structure of the target user, similar users, target service and similar services are respectively constructed to obtain the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure; the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure are respectively processed by graph convolution to obtain user interaction features, service interaction features and user-service interaction features; during the graph convolution processing, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity and user location similarity between users; the adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance and content similarity between services; and the adjacency matrix of the user-service interaction graph structure is obtained based on the user's rating of the service;

[0022] The user-service cross-enhanced feature generation module is used to embed the target user and target service to obtain target user features and target service features; after performing dimensionality reduction processing on the features of the target user features, target service features, and user-service interaction features whose dimensions exceed the dimensionality threshold, the features are subjected to feature cross-processing to obtain user-service cross-enhanced features;

[0023] The recommendation module is used to aggregate user interaction features, service interaction features and user-service cross-enhancement features to obtain comprehensive cross-features; based on the comprehensive cross-features, target user features and target service features, a recommendation value is obtained; if the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user.

[0024] A service recommendation device based on feature intersection includes a processor and a memory, wherein the processor implements the above-mentioned service recommendation method based on feature intersection when executing a computer program stored in the memory.

[0025] A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the service recommendation method based on feature intersection is implemented.

[0026] The beneficial effects of the present invention are:

[0027] The present invention provides a service recommendation method based on feature intersection. First, similar users and similar services that can indirectly reflect the personality characteristics and preference characteristics of the target user are obtained, and the target user and target service are combined to construct a user interaction graph structure, a service interaction graph structure, and a user-service interaction graph structure that reflect the interaction relationship between users, between user services, and between services. Then, graph convolution processing is performed respectively to obtain user interaction features that are convenient for identifying the social relationship and influence between users, and convenient for improving personalized recommendation and social network analysis effects, and service interaction features that can analyze the relationship between services, reveal the strength of the relationship between the target service and similar users, and facilitate analysis of the attraction strength of the target service to the target user, as well as to capture the complex relationship between users and services, helping to understand users. Then, the features whose dimensions exceed the dimension threshold among the target user features, target service features and user service interaction features are subjected to dimensionality reduction processing, and then feature cross processing is performed to cross-combine information of different dimensions to capture the potential relationship between features, and obtain user service cross-enhanced features that can more accurately identify the target user's preferences; Subsequently, the user interaction features, service interaction features and user service cross-enhanced features are aggregated to achieve the second feature cross and obtain comprehensive cross features; Finally, based on the comprehensive cross features, the recommendation value of the target service relative to the target user is calculated. When the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user, thereby enhancing the accuracy of service recommendations and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] By reading the detailed description of the preferred embodiment below, the solutions and advantages of the present application will become clear to those skilled in the art. The accompanying drawings are only for illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0029] In the attached figure:

[0030] Figure 1 In the embodiment, a schematic flow chart of the method recommended in this embodiment;

[0031] Figure 2 2 is a comparison chart of loss rate results of different recommended methods in the embodiment. DETAILED DESCRIPTION

[0032] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0033] This embodiment provides a service recommendation method based on feature intersection. Figure 1 , including the following operations:

[0034] S1. The users and services that have interacted with the target user are regarded as similar users and similar services respectively; the graph structures of the target user and similar users, the graph structure of the target service and similar services, and the graph structure of the target user, similar users, target service and similar services are respectively constructed to obtain the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure; the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure are respectively processed by graph convolution to obtain user interaction features, service interaction features and user-service interaction features; during the graph convolution processing, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity and user location similarity between users; the adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance and content similarity between services; and the adjacency matrix of the user-service interaction graph structure is obtained based on the user's rating of the service;

[0035] S2. The target user and target service are embedded to obtain target user features and target service features. Among the target user features, target service features, and user-service interaction features, features with dimensions exceeding the dimension threshold are subjected to dimensionality reduction processing and then subjected to feature cross processing to obtain user-service cross-enhanced features.

[0036] S3. Aggregate user interaction features, service interaction features, and user-service cross-enhancement features to obtain comprehensive cross-features; obtain a recommendation value based on the comprehensive cross-features, target user features, and target service features; if the recommendation value is greater than the recommendation value threshold, recommend the target service to the target user.

[0037] S1. The users and services that have interacted with the target user are regarded as similar users and similar services respectively; the graph structures of target users and similar users, the graph structures of target services and similar services, and the graph structures of target users, similar users, target services and similar services are constructed respectively to obtain the user interaction graph structure, service interaction graph structure and user-service interaction graph structure; the user interaction graph structure, service interaction graph structure and user-service interaction graph structure are processed by graph convolution respectively to obtain user interaction features, service interaction features and user-service interaction features.

[0038] Obtain similar users and similar services that can indirectly reflect the personality traits and preference characteristics of the target user, and combine the target user and target service to construct a user interaction graph structure, a service interaction graph structure, and a user-service interaction graph structure that reflect the interaction relationship between users, between user services, and between services. Then, they are processed by graph convolution to obtain user interaction features that are convenient for identifying the social relationships and influence between users, and convenient for improving personalized recommendations and social network analysis effects; service interaction features that can analyze the relationship between services, reveal the strength of the relationship between the target service and similar users, and facilitate analysis of the attraction of the target service to the target user; and user-service interaction features that are convenient for capturing the complex relationship between users and services, helping to understand user preferences and service satisfaction, which will help to obtain more accurate recommendation results in the future.

[0039] First, similar users who may share the target user's interests are identified to extract more potential personality information. Specifically, users who have interacted with the target user are considered similar users. Similar users are those who have linked accounts with the target user and / or have reviewed the same service as the target user.

[0040] Similarly, to understand the target user's preferences, services that the target user has interacted with are considered similar services and analyzed in conjunction with the target service. This indirectly analyzes the match between the target service and the target user based on the strength of the relationship between the target service and similar users. Similar services are services that the target user has used before.

[0041] Then, in order to facilitate the analysis of the relationship between users, between services, and between users and services, we construct the graph structure of target users and similar users, the graph structure of target services and similar services, and the graph structure of target users, similar users, target services and similar services respectively, and obtain the user interaction graph structure, service interaction graph structure, and user-service interaction graph structure.

[0042] In the user interaction graph structure, user-service interaction graph structure, and service interaction graph structure, nodes are users and services, edges are interaction relationships between users, between users and services, and between services, and edge weights are the social relationship between users, user ratings of services, and social relevance between services.

[0043] Finally, in order to obtain the complex relationships and high-order connection information between users, between users and services, and between services, the user interaction graph structure, the user-service interaction graph structure, and the service interaction graph structure are respectively subjected to graph convolution processing to obtain user interaction features, user-service interaction features, and service interaction features.

[0044] The processing of graph convolution can be achieved by the following formula:

[0045] h l+1 =σ(D-1 / 2 AD -1 / 2 h l W l ),

[0046] h l+1 is the output of the l+1th layer, σ is the Sigmoid function, D is the diagonal degree matrix of A, A is the adjacency matrix, h l is the output of the lth layer, h 0 is the initial graph structure, W l is the weight of the lth layer.

[0047] During the above graph convolution processing, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity, and user location similarity between users. The adjacency matrix of the user-service interaction graph structure is obtained based on the user's rating of the service. The adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance, and content similarity between services.

[0048] In the adjacency matrix of the user interaction graph structure, the value of each position is the social relationship degree between users, which is the weighted sum of user social intimacy, user behavior similarity, and user location similarity. The adjacency matrix constructed in this way can capture richer social interaction information between users.

[0049] User social intimacy is determined based on mutual following and social interaction between users. Mutual following between users is related to whether or not users follow each other. For example, if user A and user B follow each other, the mutual following behavior is assigned a value of 1; if user A and user B follow each other unilaterally, the mutual following behavior is assigned a value of 0.5; if user A and user B do not follow each other, the mutual following behavior is assigned a value of 0. User social interaction is related to likes, comments, and shared experiences. If user A frequently likes travel guides posted by user B, then user A and user B have more frequent social interactions, and the higher the corresponding value, the higher the user social intimacy.

[0050] The user behavior similarity is obtained based on historical service query behaviors and historical service use behaviors between users. The historical service query behaviors between users are related to keywords used by users when searching for services or / and service introduction pages browsed by users. If user A and user B often browse the same type of service (tourist attraction) introduction pages and search for similar keywords (such as "family amusement park" and "cultural relic in-depth tour"), the historical service query behavior similarity between user A and user B is higher, and the user behavior similarity is higher. The historical service use behaviors between users are related to service use time, method and location of users. For example, if the historical service locations (tourist destination preferences), historical service use time (tourist time selection) and historical service use method (tourist mode, such as self-driving tour and group tour) of user A and user B are closer, the historical service use behaviors of user A and user B are more similar, and the user behavior similarity between user A and user B is higher.

[0051] The user location similarity is obtained based on real-time geographic locations and resident geographic locations between users. The closer the real-time geographic locations and resident geographic locations between users are, the higher the user location similarity is.

[0052] Meanwhile, in the adjacency matrix of the service interaction graph structure, the value of each position is the social correlation degree between services, which is the weighted sum of the collaborative filtering relationship degree, the functional correlation degree and the content similarity between services. The adjacency matrix constructed in this way can help understand the quality correlation, functional complementary characteristics and content contact density between services, so that services that are mutually matched in quality are recommended to users in the recommendation process.

[0053] The collaborative filtering relationship degree is obtained based on the user simultaneous use frequency and user rating correlation between services. The user simultaneous use frequency between services is related to the number of times that services are simultaneously used by the same user. For example, if user A uses service C (books hotel C) and service D (books tickets for tourist attraction D) at the same time, the more times the services are simultaneously used, the stronger the collaborative relationship between service C and service D is, the higher the user simultaneous use frequency is, and the higher the collaborative filtering relationship degree is. The user rating correlation between services is related to ratings given by the same user within a neighborhood time. For example, if user A gives a rating greater than a first rating threshold to service C (hotel C) and a rating greater than a second rating threshold to service D (tourist attraction D) within the same day, it indicates that the user rating correlation between service C and service D is stronger, and the collaborative filtering relationship degree is higher.

[0054] Functional relevance is derived from the functional complementarity between services and the upstream and downstream relationships between them. The functional complementarity between services is related to whether the functions of the services complement each other. For example, if Service C (tourism transportation service) and Service D (tourism accommodation service) are complementary, and User A needs to use Service C before using Service D (User A needs to take transportation to the tourism transportation service point), this indicates that the functional complementarity between Service C and Service D is strong, and the functional relevance is higher. The upstream and downstream relationship between services is related to the industrial chain in which the services are located. For example, if Service C and Service D are in the same industrial chain, Service C is the upstream product (Service C is a tourism product supplier) and Service D is the downstream product (Service D is a travel agency), then the upstream and downstream relationship between Service C and Service D is strong, and the functional relevance is higher.

[0055] Content similarity is based on thematic and service attribute similarity between services. If services C and D share a common theme, such as cultural themes, then services C and D have stronger thematic similarity and, therefore, higher content similarity. Service attribute similarity between services is related to their target audiences, price ranges, and quality levels. The closer the target audience, price range, or quality level of service C is to that of service D, the stronger the service attribute similarity and the higher the content similarity.

[0056] S2. The target user and target service are embedded to obtain target user features and target service features. Among the target user features, target service features, and user-service interaction features, features with dimensions exceeding the dimension threshold are subjected to dimensionality reduction processing, and then subjected to feature cross-processing to obtain user-service cross-enhancement features.

[0057] After dimensionality reduction processing is performed on the target user features, target service features, and user-service interaction features, the features whose dimensions exceed the dimensionality threshold and reduce the computational efficiency are then subjected to feature cross-processing. Information from different dimensions is cross-combined to capture the potential relationship between features, thereby obtaining user-service cross-enhanced features that can more accurately identify the target user's preferences.

[0058] First, the target user and target service are embedded and converted into a vector form that is easy to calculate and process, and the target user characteristics and target service characteristics are obtained.

[0059] Then, among the target user features, target service features, and user-service interaction features, those with dimensions exceeding the dimensionality threshold undergo dimensionality reduction processing and then undergo feature cross-processing to obtain user-service cross-enhanced features. For example, if the dimension of the target user feature exceeds the dimensionality threshold, the target user feature undergoes dimensionality reduction processing.

[0060] The above dimensionality reduction can be achieved by the following formula: O=max(0,Wx I +b I ),x O is the feature after dimensionality reduction, W is the dimensionality reduction weight matrix, x I is the target user feature, or target service feature, or user-service interaction feature, b I is the dimensionality reduction compensation.

[0061] The specific operation of feature cross-processing is: cross-aggregate the target user features, target service features and user-service interaction features after dimensionality reduction to obtain cross-aggregation features; after the cross-aggregation features are processed by the ReLU activation function several times, they are spliced ​​with the cross-aggregation features to obtain user-service cross-enhanced features.

[0062] The above cross-aggregation operation can be achieved through concatenation or multiplication.

[0063] The above cross-polymerization operation can also be achieved through the following formula:

[0064]

[0065] H z is the cross-aggregation feature, b0 is the global bias term, α is the cross coefficient, X i is the feature of the i-th dimension in the target user feature, X j is the feature of the jth dimension in the target service feature, X k is the feature of the kth dimension in the user-service interaction feature, I, J, and K are the total number of dimensions of the target user feature, target service feature, and user-service interaction feature, respectively. ij is the latent vector of the feature of the i-th dimension in the target user feature and the feature of the j-th dimension in the target service feature, v ik is the latent vector of the feature of the i-th dimension in the target user feature and the feature of the k-th dimension in the user-service interaction feature, v jk is the latent vector of the feature of the j-th dimension in the target service feature and the feature of the k-th dimension in the user service interaction feature, 〈〉 is the dot product processing of the vector, w1, w2, w3, w4 are the first cross weight, second cross weight, third cross weight and fourth cross weight.

[0066] S3. Aggregate user interaction features, service interaction features, and user-service cross-enhancement features to obtain comprehensive cross-features; obtain a recommendation value based on the comprehensive cross-features, target user features, and target service features; if the recommendation value is greater than the recommendation value threshold, recommend the target service to the target user.

[0067] The user interaction features, service interaction features and user-service cross-enhancement features are aggregated to achieve a second feature cross-over and obtain comprehensive cross-features, which can identify more complex interaction information between users and services. Then, based on the comprehensive cross-features, the recommendation value of the target service relative to the target user is calculated. When the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user, thereby enhancing the target user's satisfaction.

[0068] First, user interaction features, service interaction features and user-service cross-enhancement features are aggregated to obtain comprehensive cross-features.

[0069] The aggregation operation is realized by the following formula: f =(U·V)⊙H z , H f is a comprehensive cross feature, U, V, H z They are user interaction features, service interaction features and user-service cross-enhancement features respectively.

[0070] The operation of aggregation processing can also be achieved through splicing processing.

[0071] Then, based on the comprehensive cross-features, target user features, and target service features, the recommendation value is obtained.

[0072] The operation of obtaining the recommended value is achieved through the following formula: y = f (W1X u +W2X s +W3(X u ⊙X s )+W4Z+B), y is the recommended value, Z, X u 、X s are comprehensive cross-features, target user features, and target service features, respectively. W1, W2, W3, and W4 are the first recommendation weight, the second recommendation weight, the third recommendation weight, and the fourth recommendation weight, respectively. B is the recommendation bias. f() is the activation function, which can be either a Relu activation function or a Sigmoid activation function.

[0073] Finally, if the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user.

[0074] In order to verify the effectiveness of the method recommended in this embodiment, the following experiment was performed.

[0075] Experimental setup and dataset. The experimental environment is as follows: Python 3.8, deep learning framework Pytorch1.13.1, and GPU is NVIDIA GeForce RTX 3060. The evaluation indicators include accuracy, mean square error value, and mean square logarithmic error. In the experiment, the ml-25m of the MoviesLens movie public dataset is selected, which contains 25000095 ratings and 1093360 tags, about 60,000 movies, and 162541 users. Each user scores at least 20 movies. The data set address is: https: / / grouplens.org / datasets / movielens / 25m / .

[0076] Comparison of existing methods. In the experiment, four personalized recommendation models are selected as comparison methods, as follows: NGCF (Neural Graph Collaborative Filtering) method: This method is a collaborative filtering model based on graph neural network. Its core idea is to represent the interaction data of users and services as a graph structure, and capture the high-order relationship between users and services through graph convolution operation, thereby improving the recommendation performance; LightGCN (Light Graph Convolutional Networks) method: This method is a simplified and optimized version of graph convolutional network (GCN) in the recommendation system. It removes complex nonlinear transformation and feature transformation operations to improve the computational efficiency and performance of the model; WDCN method: This method combines width, depth, and cross-feature learning to provide a more accurate ROP prediction method, especially in complex drilling data environments. It has shown significant advantages in improving the stability and generalization ability of the model; Deep Crossing method: This method stacks multiple residual network modules to effectively capture the complex relationships between features. Compared with traditional feature engineering methods, Deep Crossing method can automatically learn more rich feature representations, improving the accuracy and efficiency of click-through rate prediction and recommendation tasks.

[0077] Performance comparison in experimental results. Table 1 is a summary of the performance comparison results of each method, in which the accuracy of the method of the present embodiment is the highest, the mean square error value and the mean square logarithmic error value are the lowest, and the recommendation performance is the best.

[0078] Table 1: Summary of performance comparison results of each method

[0079]

[0080] Comparison of convergence speed of each method. In the experiment, the log loss function of each method is compared, and the log loss function is as follows:

[0081]

[0082] loloss is the loss rate, N is the total number of samples, y i is the label of the i-th sample, p i is the recommended value of the i-th sample, and the experimental results are as follows Figure 2 , it can be found that the loss rate of the method in this embodiment is the smallest and the convergence speed is the fastest.

[0083] This embodiment further provides a feature-crossing-based service recommendation system, which is used to implement the feature-crossing-based service recommendation method described above, including:

[0084] The user interaction feature, service interaction feature and user-service interaction feature generation modules are used to identify users and services that have interacted with the target user as similar users and similar services, respectively; the graph structures of the target user and similar users, the graph structure of the target service and similar services, and the graph structure of the target user, similar users, target service and similar services are respectively constructed to obtain the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure; the user interaction graph structure, the service interaction graph structure and the user-service interaction graph structure are respectively processed by graph convolution to obtain user interaction features, service interaction features and user-service interaction features; during the graph convolution processing, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity and user location similarity between users; the adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance and content similarity between services; and the adjacency matrix of the user-service interaction graph structure is obtained based on the user's rating of the service;

[0085] The user-service cross-enhanced feature generation module is used to embed the target user and target service to obtain target user features and target service features; after performing dimensionality reduction processing on the features of the target user features, target service features, and user-service interaction features whose dimensions exceed the dimensionality threshold, the features are subjected to feature cross-processing to obtain user-service cross-enhanced features;

[0086] The recommendation module is used to aggregate user interaction features, service interaction features and user-service cross-enhancement features to obtain comprehensive cross-features; based on the comprehensive cross-features, target user features and target service features, a recommendation value is obtained; if the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user.

[0087] This embodiment further provides a service recommendation device based on feature intersection, including a processor and a memory, wherein the processor implements the above-mentioned service recommendation method based on feature intersection when executing a computer program stored in the memory.

[0088] This embodiment further provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned service recommendation method based on feature intersection.

[0089] This embodiment provides a service recommendation method based on feature intersection. First, similar users and similar services that can indirectly reflect the personality characteristics and preference characteristics of the target user are obtained, and combined with the target user and target service, a user interaction graph structure, a service interaction graph structure, and a user-service interaction graph structure that reflect the interaction relationship between users, between user services, and between services are constructed. Then, graph convolution processing is performed respectively to obtain user interaction features that are convenient for identifying the social relationship and influence between users, and convenient for improving the effect of personalized recommendation and social network analysis, and service interaction features that can analyze the relationship between services, reveal the strength of the relationship between the target service and similar users, and facilitate analysis of the attraction of the target service to the target user, as well as to capture the complex relationship between users and services, helping to understand the user. The method uses the user service interaction characteristics of user preferences and service satisfaction as the basis for the optimization of the target user characteristics, target service characteristics and user service interaction characteristics. Then, the characteristics whose dimensions exceed the dimension threshold in the target user characteristics, target service characteristics and user service interaction characteristics are subjected to dimensionality reduction processing, and then feature cross processing is performed to cross-combine information of different dimensions and capture the potential relationship between features to obtain user service cross-enhanced characteristics that can more accurately identify the target user preferences. Subsequently, the user interaction characteristics, service interaction characteristics and user service cross-enhanced characteristics are aggregated to achieve the second feature cross and obtain comprehensive cross-features. Finally, based on the comprehensive cross-features, the recommendation value of the target service relative to the target user is calculated. When the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user, thereby enhancing the accuracy of service recommendation and improving user experience.

Claims

1. A service recommendation method based on feature intersection, characterized in that: The following operations are included: S1. Regulate users and services that have interacted with the target user as similar users and similar services, respectively. Build graph structures of target users and similar users, target services and similar services, and target users, similar users, target services and similar services, respectively. This yields user interaction graph structures, service interaction graph structures, and user-service interaction graph structures. The user interaction graph structure, service interaction graph structure and user-service interaction graph structure are processed by graph convolution respectively to obtain user interaction features, service interaction features and user-service interaction features; During the graph convolution process, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity, and user location similarity between users; The adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance and content similarity between services; Collaborative filtering relationship is based on the frequency of simultaneous use of services and the correlation of user ratings; functional relevance is based on the functional complementarity between services and the upstream and downstream relationships between services; Content similarity is obtained based on the subject similarity and service attribute similarity between services; The adjacency matrix of the user-service interaction graph structure is obtained based on users’ ratings of services; S2. The target user and target service are embedded to obtain target user features and target service features. Among the target user features, target service features, and user-service interaction features, features with dimensions exceeding the dimension threshold are subjected to dimensionality reduction processing and then subjected to feature cross processing to obtain user-service cross-enhanced features. S3. Aggregate the user interaction features, service interaction features, and user-service cross-enhancement features to obtain comprehensive cross-features; obtain a recommendation value based on the comprehensive cross-features, target user features, and target service features; if the recommendation value is greater than a recommendation value threshold, recommend the target service to the target user; The recommended value is obtained through the following formula: y=f(W 1 X u +W 2 X s +W 3 (X u ⊙X s )+W 4 Z+B) , y is the recommended value, Z 、 X u 、 X s They are comprehensive cross-features, target user features, and target service features. W 1 、 W 2 、 W 3 、 W 4 They are the first recommendation weight, the second recommendation weight, the third recommendation weight and the fourth recommendation weight, B is the recommended offset, f( ) is the activation function.

2. The service recommendation method based on feature intersection according to claim 1, characterized in that: In S1, in the adjacency matrix of the user interaction graph structure, the value of each position is the social relationship degree between users, and the social relationship degree is the weighted sum of user social intimacy, user behavior similarity, and user location similarity between users; User social intimacy is based on the mutual attention and social interaction behaviors between users; User behavior similarity is based on historical service query behavior and historical service usage behavior between users; User location similarity is based on the real-time geographic location and permanent geographic location of users.

3. The service recommendation method based on feature intersection according to claim 1, characterized in that: In S1, in the adjacency matrix of the service interaction graph structure, the value of each position is the social correlation between services, and the social correlation is the weighted sum of the collaborative filtering relationship, functional correlation and content similarity between services.

4. The service recommendation method based on feature intersection according to claim 1, characterized in that: The operation of feature cross processing in S2 is specifically as follows: The target user features, target service features and user-service interaction features that have undergone dimensionality reduction are cross-aggregated to obtain cross-aggregated features. The cross-aggregated features are processed by the ReLU activation function several times and then concatenated with the cross-aggregated features to obtain user-service cross-enhanced features.

5. The service recommendation method based on feature intersection according to claim 4, characterized in that: The cross-aggregation operation can be achieved through the following formula: ， H z is the cross-aggregation feature, b 0 is the global bias term, α is the cross coefficient, X i The first of the target user characteristics i The characteristics of the dimensions, X j The target service feature j The characteristics of the dimensions, X k The first interactive feature for user services k The characteristics of the dimensions, I 、 J 、 K are the total number of dimensions of target user characteristics, target service characteristics, and user-service interaction characteristics, v ij The first of the target user characteristics i The characteristics of the dimension and the target service characteristics j The latent vector of the feature of dimension, v ik The first of the target user characteristics i The characteristics of the dimension and the user service interaction characteristics k The latent vector of the feature of dimension, v jk The target service feature j The characteristics of the dimension and the user service interaction characteristics k The latent vector of the feature of dimension, 〈 〉 is the dot product processing of vectors, w 1 、 w 2 、 w 3 、 w 4 are the first cross weight, the second cross weight, the third cross weight and the fourth cross weight.

6. The service recommendation method based on feature intersection according to claim 1, characterized in that: In S3, the aggregation operation is implemented by the following formula: H f =(U·V)⊙H z ， H f For comprehensive cross-features, U 、 V 、 H z They are user interaction features, service interaction features and user-service cross-enhancement features.

7. A service recommendation system based on feature intersection, used to implement the service recommendation method based on feature intersection according to claim 1, characterized in that: include: User interaction feature, service interaction feature and user-service interaction feature generation modules are used to identify users and services that have interacted with the target user as similar users and similar services, respectively; Build the graph structure of target users and similar users, the graph structure of target services and similar services, and the graph structure of target users, similar users, target services and similar services respectively, and obtain the user interaction graph structure, service interaction graph structure and user service interaction graph structure; The user interaction graph structure, service interaction graph structure, and user-service interaction graph structure are processed by graph convolution respectively to obtain user interaction features, service interaction features, and user-service interaction features. During the graph convolution process, the adjacency matrix of the user interaction graph structure is obtained based on the user social intimacy, user behavior similarity, and user location similarity between users. The adjacency matrix of the service interaction graph structure is obtained based on the collaborative filtering relationship, functional relevance, and content similarity between services. The adjacency matrix of the user-service interaction graph structure is obtained based on the user's ratings of the services. The user-service cross-enhanced feature generation module is used to embed the target user and target service to obtain target user features and target service features; after performing dimensionality reduction processing on the features of the target user features, target service features, and user-service interaction features whose dimensions exceed the dimensionality threshold, the features are subjected to feature cross-processing to obtain user-service cross-enhanced features; The recommendation module is used to aggregate user interaction features, service interaction features, and user-service cross-enhancement features to obtain comprehensive cross-features; Based on the comprehensive cross-features, target user features and target service features, the recommended value is obtained; If the recommendation value is greater than the recommendation value threshold, the target service is recommended to the target user.

8. A service recommendation device based on feature intersection, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the service recommendation method based on feature intersection according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that Used to store a computer program, wherein when the computer program is executed by a processor, the service recommendation method based on feature intersection according to any one of claims 1 to 6 is implemented.

Images