A service recommendation method, system, device and storage medium based on graph convolution

Abstract

The present application relates to the technical field of user profile-based and personalized custom search, and specifically to a service recommendation method, system, device and storage medium based on graph convolution. In order to solve the problems of narrow service recommendation range, low service accuracy and reduced user experience in the prior art, the present application first divides multiple negative sample services that have not interacted with the target user into several service groups to be recommended, and performs multi-layer graph convolution processing on each service group to be recommended, the positive sample service and the neighboring users to obtain several service group features to be recommended that can facilitate the analysis of the characteristics of the negative sample services. Finally, the recommendation value is calculated based on the service group features to be recommended and the target user features, and the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold is recommended to the target user as a high-quality service group, thereby improving the accuracy of service recommendation and the user experience on the basis of enhancing the richness of service recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of user profile-based and personalized custom search, and in particular to a service recommendation method, system, device and storage medium based on graph convolution. Background Art

[0002] In the era of big data, the explosive growth of information has caused rapid changes in users' needs and preferences, making it difficult for service recommendation software in existing business fields to quickly predict services that can meet customer needs and make recommendations. It is also difficult to implement customized searches based on user profiles and personalization, which can easily reduce user experience and commercial content consumption.

[0003] In the existing technology, in order to improve the efficiency of service recommendation, services that users have interacted with are often recommended to users repeatedly. Repeatedly recommending the same service may cause users to feel tired and bored, and users may have new needs that have not been expressed. Repeatedly recommending old services may miss the opportunity to identify and meet these potential needs, thereby reducing user interest and participation. The user-item recommendation method based on ratings can expand the user's service selection range. This method often recommends items with higher ratings to users based on previous user ratings of items. Items with higher ratings may be of better quality, but they may not necessarily meet the user's preference needs, resulting in lower accuracy of service recommendations and lower 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 graph convolution.

[0005] The technical solution of the present invention is as follows:

[0006] A service recommendation method based on graph convolution includes the following operations:

[0007] S1. Randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; obtain several other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample service and the negative sample service are the services that have interacted with the target user and the services that have not interacted with the target user, respectively;

[0008] S2. Several service groups to be recommended and corresponding target samples, as well as neighbor user sets, are processed by multi-layer graph convolution to obtain several service group features to be recommended; in the process of multi-layer graph convolution: in the current service group to be recommended, each negative sample service feature and its corresponding target sample feature are processed by several layers of graph convolution. After the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer and different corresponding negative sample services; from each convolution layer data set, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; the multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group features to be recommended;

[0009] S3. Based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service scores, obtain the recommendation value of each service group to be recommended; and recommend the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold as a high-quality service group to the target user.

[0010] In S2, during the current layer of graph convolution processing, the target sample convolution features of the previous layer and the negative sample convolution injection features of the previous layer are respectively processed by graph convolution and weighted processing, and then summed to obtain the current layer of negative sample convolution injection features, which are used to perform the next layer of graph convolution operations; the current negative sample service and the corresponding target sample are respectively processed by the first layer of graph convolution to obtain the first layer of negative sample convolution features and the first layer of target sample convolution features; the first layer of negative sample convolution features and the first layer of target sample convolution features are respectively weighted and summed to obtain the first layer of negative sample convolution injection features, which are used to perform the second layer of graph convolution operations.

[0011] In S2, the key negative sample convolution injection feature is the negative sample convolution injection feature corresponding to the maximum value of the pooled aggregation value of the target user feature in the corresponding convolution layer dataset.

[0012] In S2, the specific operations of the aggregation processing are as follows: the target user features and each neighbor user feature are processed by multi-layer graph convolution respectively to obtain several layers of target user convolution features and several layers of neighbor user convolution features, forming a multi-layer target user convolution feature set and several multi-layer neighbor user convolution feature sets respectively; the multi-layer negative sample convolution feature set, the multi-layer target user convolution feature set, and all multi-layer neighbor user convolution feature sets are sequentially processed by performing feature inner product processing according to the corresponding layer and summing up the features of all layers, and then divided by the total number of users to obtain the current service group features to be recommended.

[0013] In S3, the recommended value is obtained by the following formula:

[0014]

[0015] Q is the recommended value, E is the feature vector of the current service group to be recommended, and e u is the target user feature vector, e i is the i-th negative sample service vector in the current service group to be recommended, I is the total number of negative sample services, λ is the adjustment coefficient, b is the feature deviation compensation value, and p i is the score of the i-th negative sample service, q1, q2, q3 are the first coefficient, the second coefficient and the third coefficient, and σ is the sigmoid function.

[0016] Furthermore, all negative sample services are sorted in descending order of the number of repetitions in several high-quality service groups; and according to the number of recommended requirements of the target user, the negative sample services with the number of recommended requirements before sorting are recommended to the target user.

[0017] User features are composed of user address information, and service sample features are composed of service address information, service interaction information, and service interaction score information.

[0018] A graph convolution-based service recommendation system, used to implement the above-mentioned graph convolution-based service recommendation method, comprising:

[0019] The service group and user set generation module is used to randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; obtain several other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample service and the negative sample service are the services that have interacted with the target user and the services that have not interacted with the target user, respectively;

[0020] The feature generation module of the service group to be recommended is used for several service groups to be recommended and corresponding target samples, as well as neighbor user sets, and obtains several features of the service group to be recommended through multi-layer graph convolution processing; in the process of multi-layer graph convolution processing: in the current service group to be recommended, each negative sample service feature and the corresponding target sample feature are processed through several layers of graph convolution, and after the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer and different corresponding negative sample services; from each convolution layer data set, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; the multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group feature to be recommended;

[0021] The service recommendation module is used to obtain the recommendation value of each service group to be recommended based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service score; the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold is recommended to the target user as a high-quality service group.

[0022] A graph convolution-based service recommendation device comprises a processor and a memory, wherein the processor implements the above-mentioned graph convolution-based service recommendation method when executing a computer program stored in the memory.

[0023] A computer-readable storage medium is used to store a computer program, wherein the computer program implements the above-mentioned graph convolution-based service recommendation method when executed by a processor.

[0024] The beneficial effects of the present invention are:

[0025] The present invention provides a service recommendation method based on graph convolution. First, multiple negative sample services that have not interacted with a target user are divided into several service groups to be recommended. Each service group to be recommended, positive sample services and neighbor users are subjected to multi-layer graph convolution processing, and the positive sample service information is injected into the negative sample service to enhance the negative sample service feature expression capability; then, the service group to be recommended is combined with the neighbor users to obtain several service group features that can facilitate the analysis of the characteristics of the negative sample service; finally, based on the service group features to be recommended, the target user features, the negative sample service features in the service group to be recommended and the negative sample service score, the recommendation value of each service group to be recommended is calculated, and the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold, that is, the service group to be recommended with a high correlation with the target user, is recommended to the target user as a high-quality service group, thereby improving the accuracy of service recommendation and the user experience on the basis of enhancing the richness of service recommendation. DETAILED DESCRIPTION

[0026] This embodiment provides a service recommendation method based on graph convolution, including the following operations:

[0027] S1. Randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; obtain several other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample service and the negative sample service are the services that have interacted with the target user and the services that have not interacted with the target user, respectively;

[0028] S2. Several service groups to be recommended and corresponding target samples, as well as neighbor user sets, are processed by multi-layer graph convolution to obtain several service group features to be recommended; in the process of multi-layer graph convolution: in the current service group to be recommended, each negative sample service feature and its corresponding target sample feature are processed by several layers of graph convolution. After the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer and different corresponding negative sample services; from each convolution layer data set, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; the multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group features to be recommended;

[0029] S3. Based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service scores, obtain the recommendation value of each service group to be recommended; and recommend the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold as a high-quality service group to the target user.

[0030] S1. Randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; obtain several other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample service and the negative sample service are the services that have interacted with the target user and the services that have not interacted with the target user, respectively.

[0031] The positive sample service information and neighbor user information that have interacted with the target user are combined with the negative sample services to be recommended that have not interacted with the target user to enhance the feature expression ability of the negative sample services, thereby facilitating the selection of appropriate negative sample services to recommend to users, enhancing the richness of service recommendations, and enhancing user experience.

[0032] First, several negative sample services to be recommended to the target user are randomly combined to form several groups of services to be recommended. Negative sample services are services that the target user has not interacted with.

[0033] Then, a positive sample service is assigned to each negative sample service in the group of services to be recommended as the corresponding target sample, which is used to inject positive sample information into the negative sample service to improve the expression ability of the negative sample service. The allocation method includes but is not limited to random allocation, and the positive sample service is the service that has interacted with the target user.

[0034] At the same time, several other users who have interactive relationships with the target user are obtained to form a neighbor user set, which is used for subsequent combination with negative sample services to enhance the feature expression ability of negative sample services, which is conducive to indirectly analyzing the relationship between negative sample services and target users and enhancing the accuracy of service recommendations.

[0035] S2, several service groups to be recommended and corresponding target samples, as well as neighbor user sets, are processed by multi-layer graph convolution to obtain several features of the service groups to be recommended.

[0036] Each service group to be recommended and the corresponding target sample, as well as all neighboring users, form a graph structure with services and users as nodes and interaction relationships as edges; the corresponding graph structures of several service groups to be recommended, corresponding target samples and neighboring user sets are subjected to multi-layer graph convolution processing respectively, and the target sample information is injected into the negative sample service to enhance the feature expression ability of the negative sample service, and then combined with the neighboring users to obtain several features of the service group to be recommended that can facilitate the analysis of the characteristics of the negative sample service.

[0037] Taking the current service group to be recommended as an example, the process of multi-layer graph convolution processing of the current service group to be recommended is as follows.

[0038] First, in the current service group to be recommended, each negative sample service feature and its corresponding target sample feature are processed by several layers of graph convolution. After the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, several negative sample convolution injection features are obtained; the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets.

[0039] In each convolutional layer dataset, all negative sample convolution injection features have the same corresponding layer, but the corresponding negative sample services are different.

[0040] Among them, in the current layer of graph convolution processing, when the target sample features are injected into the corresponding negative sample service features, the target sample convolution features of the previous layer and the negative sample convolution injection features of the previous layer are respectively subjected to graph convolution processing and weighted processing, and then summed to obtain the current layer of negative sample convolution injection features, which are used to perform the next layer of graph convolution operations.

[0041] In the first layer of graph convolution processing, when the target sample features are injected into the corresponding negative sample service features: the current negative sample service and the corresponding target sample are respectively processed by the first layer of graph convolution to obtain the first layer of negative sample convolution features and the first layer of target sample convolution features; the first layer of negative sample convolution features and the first layer of target sample convolution features are respectively weighted and summed to obtain the first layer of negative sample convolution injection features, which are used to perform the second layer of graph convolution operations.

[0042] The operation of injecting the target sample features into the corresponding negative sample service features can be achieved through the following formula:

[0043]

[0044] Inject features into the i-th negative sample convolution of the l-th layer, is the service feature of the jth positive sample corresponding to the i-th negative sample in the l-th layer, Inject features into the i-th negative sample convolution in the l-1th layer, α l It is the mixing coefficient of the lth layer of graph convolution. In the graph convolution process, the mixing coefficient value of each layer is randomly generated in (0, 1) by uniform sampling. GCN() is the graph convolution process.

[0045] Then, a key negative sample convolution injection feature is selected from each convolution layer data set to form a multi-layer negative sample convolution feature set. The key negative sample convolution injection feature is the negative sample convolution injection feature corresponding to the maximum pooling aggregation value of the target user feature in the corresponding convolution layer data set. The pooling aggregation value can be obtained by pooling the target user feature and the negative sample convolution injection feature and then calculating the inner product.

[0046] Finally, the multi-layer negative sample convolution feature set, the target user features, and the neighbor user features corresponding to the neighbor user set are aggregated to obtain the features of the current service group to be recommended.

[0047] Among them, the specific operation of the aggregation processing is: the target user features and each neighbor user feature are processed by multi-layer graph convolution respectively to obtain several layers of target user convolution features and several layers of neighbor user convolution features, forming a multi-layer target user convolution feature set and several multi-layer neighbor user convolution feature sets respectively; the multi-layer negative sample convolution feature set, the multi-layer target user convolution feature set, and all multi-layer neighbor user convolution feature sets are sequentially processed according to the corresponding layers for feature inner product processing and all layer feature summation processing, and then divided by the total number of users to obtain the current service group features to be recommended.

[0048] The characteristics of the current service group to be recommended can be realized by the following formula:

[0049]

[0050] E is the characteristic of the current service group to be recommended, Inject features into the lth layer of key negative sample convolution, is the convolutional feature of the target user at the lth layer, is the convolutional feature of the vth neighbor user in the lth layer, and N is the total number of users.

[0051] S3. Based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service scores, obtain the recommendation value of each service group to be recommended; and recommend the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold as a high-quality service group to the target user.

[0052] Based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended and the negative sample service scores, the recommendation value of each service group to be recommended is calculated; the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold, that is, the service group to be recommended with a high correlation with the target user, is recommended to the target user as a high-quality service group, thereby improving the accuracy of service recommendations and the user experience on the basis of enhancing the richness of service recommendations.

[0053] The recommended value is obtained by the following formula:

[0054]

[0055] Q is the recommended value, E is the feature vector of the current service group to be recommended, and e u is the target user feature vector, e i is the i-th negative sample service vector in the current service group to be recommended, I is the total number of negative sample services, λ is the adjustment coefficient, b is the feature deviation compensation value, and p i is the score of the i-th negative sample service, q1, q2, q3 are the first coefficient, the second coefficient and the third coefficient, and σ is the sigmoid function.

[0056] In order to further improve the accuracy of service recommendations and enhance user experience, all negative sample services are sorted in descending order according to the number of repetitions in several high-quality service groups; according to the number of recommended requirements of target users, negative sample services with the same number of recommended requirements before sorting are recommended to target users.

[0057] The above user features are composed of user address information, and the service sample features are composed of service address information, service interaction information and service interaction score information.

[0058] To verify the effect of the service recommendation method of this embodiment, the following experiment was conducted.

[0059] Datasets and evaluation indicators. Three public datasets, Yelp2018 dataset, Alibaba dataset, and Gowalla dataset, are used in the experiment. The datasets are divided into training set, validation set, and test set in a ratio of 7:1:2 for comparative experiments. The detailed information of the datasets is shown in Table 1:

[0060] Table 1 Summary of information of three datasets

[0061]

[0062] Experimental evaluation indicators. In the experiment, recall rate Recall@20 and normalized discounted cumulative gain NDCG@20 are used as evaluation indicators.

[0063] The calculation formula of recall rate Recall@20 is as follows:

[0064]

[0065] R(u) is the set of services recommended by the recommendation system to user u, T(u) is the set of services that user u has interacted with, and U is the total number of users. After obtaining the Recall@20 value of each user, the Recall@20 value of the entire data set can be obtained by summing and averaging them.

[0066] Normalized discounted cumulative gain (NDCG@20) is used to evaluate the accuracy of recommendation results. Services that are more closely related to users are placed higher in the list. The larger the NDCG value, the better the recommendation performance. The calculation formula is as follows:

[0067]

[0068] rel i Is whether the user likes the service at the ith position, 1 if the user likes it, and 0 if the user does not like it.

[0069] Existing methods and their characteristics for comparison. PinSage method: It is a recall algorithm implemented by Pinterest based on GraphSAGE; NGCF method: It captures the high-dimensional interaction relationship between users and services by constructing a bipartite graph of user and service interactions; LightGCN method: Based on the NGCF method, a simplified strategy is adopted, which abandons the traditional complete graph convolution process and removes feature conversion and nonlinear activation steps; RNS method: The negative sampling module selects services that have not been interacted with as negative examples to expand the training data; SGL method: On the basis of the classic supervised recommendation task, an auxiliary self-supervision task is added to strengthen node representation learning through self-identification; GTN method: A graph trend collaborative filtering technology is introduced, and a new graph trend filtering network framework is proposed to capture the adaptive reliability between users and items.

[0070] Experimental environment and parameter settings. The PyTorch machine learning framework was used in the experiment. The experiment was conducted in the environment of PyCharm 2020, NVIDIA GeForce GTX 3090, and Python 3.6. In the experiment, the embedding size M was set to 64, the learning rate was set to 0.001, the batch size was 2048, the graph convolution layer L was set to 3, the hyperparameters q = 0.5, λ = 0.025, the Xavier initialization parameters were used, and the optimizer was set to Adam.

[0071] Experimental results. As can be seen from Table 2, the method of this embodiment achieved the best recommendation effect in the three public data sets. Taking the Recall@20 index results as the benchmark, the method of this embodiment achieved an improvement rate of 8.43%, 16.27%, and 7.0% on the three data sets, respectively, verifying the superiority of the method of this embodiment.

[0072] Table 2 Summary of experimental results of the method of this embodiment and the existing method

[0073]

[0074] This embodiment further provides a graph convolution-based service recommendation system, which is used to implement the above-mentioned graph convolution-based service recommendation method, including:

[0075] The service group and user set generation module is used to randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; obtain several other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample service and the negative sample service are the services that have interacted with the target user and the services that have not interacted with the target user, respectively;

[0076] The feature generation module of the service group to be recommended is used for several service groups to be recommended and corresponding target samples, as well as neighbor user sets, and obtains several features of the service group to be recommended through multi-layer graph convolution processing; in the process of multi-layer graph convolution processing: in the current service group to be recommended, each negative sample service feature and the corresponding target sample feature are processed through several layers of graph convolution, and after the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer and different corresponding negative sample services; from each convolution layer data set, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; the multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group feature to be recommended;

[0077] The service recommendation module is used to obtain the recommendation value of each service group to be recommended based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service score; the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold is recommended to the target user as a high-quality service group.

[0078] This embodiment also provides a graph convolution-based service recommendation device, including a processor and a memory, wherein the processor implements the above-mentioned graph convolution-based service recommendation method when executing a computer program stored in the memory.

[0079] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program implements the above-mentioned graph convolution-based service recommendation method when executed by a processor.

[0080] The present embodiment provides a service recommendation method based on graph convolution. First, multiple negative sample services that have not interacted with the target user are divided into several service groups to be recommended. Each service group to be recommended, positive sample services and neighbor users are subjected to multi-layer graph convolution processing, and the positive sample service information is injected into the negative sample service to enhance the negative sample service feature expression capability; then, the service group to be recommended is combined with the neighbor users to obtain several service group features that can facilitate the analysis of the characteristics of the negative sample service; finally, based on the service group features to be recommended, the target user features, the negative sample service features in the service group to be recommended and the negative sample service score, the recommendation value of each service group to be recommended is calculated, and the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold, that is, the service group to be recommended with a high correlation with the target user, is recommended to the target user as a high-quality service group, thereby improving the accuracy of service recommendation and the user experience on the basis of enhancing the richness of service recommendation.

Claims

1. A service recommendation method based on graph convolution, characterized in that: The following operations are included: S1. Randomly combine several negative sample services to form several service groups to be recommended; assign a positive sample service to each negative sample service in the several service groups to be recommended as the corresponding target sample; Obtain a number of other users who have interactive relationships with the target user to form a neighbor user set; The positive sample services and negative sample services are services that have interacted with the target user and services that have not interacted with the target user, respectively; S2, several service groups to be recommended and corresponding target samples, as well as neighbor user sets, are processed by multi-layer graph convolution to obtain several features of the service groups to be recommended; In the process of multi-layer graph convolution processing: In the current service group to be recommended, each negative sample service feature and its corresponding target sample feature are processed by several layers of graph convolution. After the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer but different corresponding negative sample services; in the process of the current layer of graph convolution processing, the previous layer of target sample convolution features and the previous layer of negative sample convolution injection features are respectively processed by graph convolution and weighted processing, and then summed to obtain the current layer of negative sample convolution injection features, which are used to perform the next layer of graph convolution operations; the current negative sample service and the corresponding target sample are respectively processed by the first layer of graph convolution to obtain the first layer of negative sample convolution features and the first layer of target sample convolution features; the first layer of negative sample convolution features and the first layer of target sample convolution features are respectively weighted and summed to obtain the first layer of negative sample convolution injection features, which are used to perform the second layer of graph convolution operations; From each convolutional layer dataset, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; The multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group feature to be recommended; S3. Obtaining a recommendation value for each service group to be recommended based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service scores; The recommended value is obtained by the following formula: ， Q is the recommended value, E is the feature vector of the current service group to be recommended, e u is the target user feature vector, e i The first service in the current recommended service group i negative sample serving vectors, I is the total number of negative sample services, λ The adjustment factor ，b is the characteristic deviation compensation value, p i For the i The ratings of negative sample services, q 1 , q 2 , q 3 are the first coefficient, the second coefficient and the third coefficient, σ is the sigmoid function, r For the recommended power; The service group to be recommended corresponding to the recommendation value greater than the recommendation threshold is recommended to the target user as a high-quality service group.

2. The service recommendation method based on graph convolution according to claim 1, characterized in that: In S2, the key negative sample convolution injection feature is the negative sample convolution injection feature corresponding to the maximum value of the pooled aggregation value of the target user feature in the corresponding convolution layer data set.

3. The service recommendation method based on graph convolution according to claim 1, characterized in that: In S2, the specific operations of the aggregation process are: The target user features and each neighbor user features are processed by multi-layer graph convolution respectively to obtain several layers of target user convolution features and several layers of neighbor user convolution features, forming a multi-layer target user convolution feature set and several multi-layer neighbor user convolution feature sets respectively; The multi-layer negative sample convolution feature set, the multi-layer target user convolution feature set, and all multi-layer neighbor user convolution feature sets are sequentially processed according to the corresponding layers and the feature summation of all layers, and then divided by the total number of users to obtain the current service group feature to be recommended.

4. The service recommendation method based on graph convolution according to claim 1, characterized in that: Among several high-quality service groups, all negative sample services are sorted in descending order of the number of repetitions; according to the number of recommended requirements of the target user, the negative sample services with the number of recommended requirements before sorting are recommended to the target user.

5. The service recommendation method based on graph convolution according to claim 1, characterized in that: User features are composed of user address information, and service sample features are composed of service address information, service interaction information, and service interaction score information.

6. A graph convolution-based service recommendation system, used to implement the graph convolution-based service recommendation method according to claim 1, characterized in that: include: A service group and user set generation module, used to randomly combine a number of negative sample services to form a number of service groups to be recommended; assign a positive sample service to each negative sample service in the number of service groups to be recommended as a corresponding target sample; Obtain a number of other users who have an interactive relationship with the target user to form a neighbor user set; the positive sample services and negative sample services are services that have interacted with the target user and services that have not interacted with the target user, respectively; The feature generation module of the service group to be recommended is used for several service groups to be recommended and corresponding target samples, as well as neighbor user sets, and obtains several features of the service group to be recommended through multi-layer graph convolution processing; in the process of multi-layer graph convolution processing: in the current service group to be recommended, each negative sample service feature and the corresponding target sample feature are processed through several layers of graph convolution, and after the target sample feature is injected into the corresponding negative sample service feature in each layer of graph convolution, the negative sample convolution injection features of the same layer are classified into the same data set to obtain several convolution layer data sets; in each convolution layer data set, all negative sample convolution injection features have the same corresponding layer and different corresponding negative sample services; from each convolution layer data set, a key negative sample convolution injection feature is selected to form a multi-layer negative sample convolution feature set; the multi-layer negative sample convolution feature set, the target user feature, and the neighbor user feature corresponding to the neighbor user set are aggregated to obtain the current service group feature to be recommended; The service recommendation module is used to obtain the recommendation value of each service group to be recommended based on the characteristics of the service group to be recommended, the characteristics of the target user, the characteristics of the negative sample services in the service group to be recommended, and the negative sample service score; the service group to be recommended corresponding to the recommendation value greater than the recommendation threshold is recommended to the target user as a high-quality service group.

7. A service recommendation device based on graph convolution, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the graph convolution-based service recommendation method as described in any one of claims 1 to 5 is implemented.

8. 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 graph convolution-based service recommendation method according to any one of claims 1 to 5 is implemented.

Images