Bundle recommendation method and system based on cross-view enhancement

Abstract

The application provides a bundle recommendation method and system based on cross-view enhancement, which comprises the following steps: constructing a two-part graph through the interaction between users and bundles, the interaction between users and items, and the affiliation between items and bundles; learning the node representation under each view through a view representation learning module after the construction of the view; using a prediction enhancement module to predict the possible interaction between users and bundles through the bundle view and the item view, and adding the enhanced information to the user view to obtain the enhanced user view; narrowing the representation of each node in the three views through a contrast learning module; training the model, aggregating the user representation and the bundle representation of each view to obtain the final user representation and bundle representation, updating the model parameters by using a loss function; and testing the model to obtain the final user score for the bundle by using inner product and obtaining the evaluation standard. The application alleviates the problem of sparse connection between users and bundles.

Description

Technical Field

[0001] This application relates to the field of image enhancement technology, and in particular to a bundled recommendation method and system based on cross-view enhancement. Background Technology

[0002] With the development of information technology and the explosive growth of information, people often need to sift through massive amounts of data to find information of interest, leading to the emergence of recommendation systems. The background technologies of recommendation systems mainly include the following categories:

[0003] Content-based recommendation: This type of recommendation system infers a user's interests based on their past behavior or explicit preferences and recommends items that match those interests. For example, a system based on a user's movie-watching habits might recommend more science fiction movies of the same genre. Collaborative filtering recommendation: Collaborative filtering methods primarily focus on the similarity of behavior between users. If two users have very similar ratings or behaviors regarding the same set of items in the past, it is assumed that their interests are similar, and items liked by one user can be recommended to the other. Knowledge-based recommendation: This type of recommendation system uses specific domain knowledge to recommend items. For example, in book recommendations, knowledge-based recommendations might use the opinions or standards of domain experts. Hybrid recommendation: This type of recommendation system combines all the above types of recommendation methods to optimize the recommendation results and address the problems that may exist with a single recommendation method. Application of deep learning in recommendation systems: In recent years, deep learning has also been widely applied to recommendation systems. By training complex neural network models, deeper features can be extracted from user behavior and item attributes, and these features can be used for more accurate recommendations.

[0004] Bundled recommendations are a subtask of recommendation systems, aiming to recommend a series of items to users, such as playlists, game packs, and book lists. Compared to general recommendation systems, bundled recommendations face the problem of more sparsity data, and user interactions with individual items provide additional information for bundled recommendations.

[0005] Previous methods treated bundle recommendation as a general recommendation task, but due to the high sparsity of user interactions with bundles, they often failed to achieve satisfactory results. While current methods incorporate additional user interaction information about individual items into the main task to improve bundle recommendation performance, they do not effectively utilize information about the affiliation of items and bundles, thus limiting the performance of the bundle recommendation system. Summary of the Invention

[0006] This application provides a bundle recommendation method based on cross-view enhancement, which can be used to solve the technical problem of high sparsity in user interactions with bundles.

[0007] This application provides a bundled recommendation method based on cross-view enhancement, the method including:

[0008] Step 1: Construct a bipartite diagram based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships. The bipartite diagram includes a user view (UB), a project view (UI), and a bundle view (BI). Figure 3 One view;

[0009] Step 2: After constructing the view, the node representation of each view is learned through the view representation learning module;

[0010] Step 3: Using the prediction enhancement module, predict possible user and bundle interactions through the bundle view and item view, and add them to the user view as enhancement information to obtain the enhanced user view;

[0011] Step 4: Use the comparative learning module to bring the representations of each node in the three views closer together;

[0012] Step 5: Train the model, and aggregate the user representations and bundled representations of each view obtained in Steps 2 and 3 to obtain the final user representation e. u And binding means e b By using the common loss function BPRLoss, user-bundle pairs with interaction are used as positive samples, and those without interaction are used as negative samples to update the model parameters.

[0013] Step 6, Model Testing: Obtain the final user rating y for the bundle using the inner product. ub The highest score among the top k is taken as the prediction result; the Recall value and NDCG value of the test set and the prediction result are used as the evaluation criteria.

[0014] Optionally, in step 1, construct a bipartite graph based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships; including:

[0015] Represents a set of users. Represents a collection of items. This represents a set of bundled packages; where the matrix of user and project interactions is X∈{0,1}. M×N The user-bound interaction matrix is ​​Y∈{0,1} M×O The membership matrix of bundled projects is Z∈{0,1} O×N ;

[0016] The nodes in the graph are user nodes, bundle nodes, and project nodes. Edges are constructed based on the interaction relationships between user bundles, user projects, and the membership relationships between bundles and projects.

[0017] Optionally, step 2: After constructing the view, the node representations under each view are learned through the view representation learning module, including:

[0018] Using a graph convolutional neural network model, the information of adjacent nodes is aggregated for three views: user view, item view, and bundle view, to obtain the vector representations of user and bundle nodes in the user view, the representations of user and item nodes in the item view, and the representations of bundle and item nodes in the bundle view.

[0019] Optionally, a graph convolutional neural network model is used to aggregate the information of adjacent nodes for the three views: user view, item view, and bundle view, to obtain the vector representations of user and bundle nodes in the user view, the representations of user and item nodes in the item view, and the representations of bundle and item nodes in the bundle view, including:

[0020] For bipartite graphs, some edges are randomly masked, and then LightGCN is used to aggregate neighbor node information in the bipartite graph, removing nonlinear transformations and biases from the graph convolutional neural network; for UB graphs, the process is as follows:

[0021]

[0022]

[0023] in, It is the information that is propagated to the user and the bundled k-th layer. and In the UB graph, the users and bundles are neighbors; connecting the embeddings of the k-layers yields representations of the users and bundles, ultimately... and Defined as:

[0024]

[0025] The UI design is as follows:

[0026]

[0027]

[0028] in It is the information propagated to the k-th layer of users and projects. and It represents the neighbors of users and items in the UI diagram; connecting the embeddings of k layers yields a representation of the user and the bundled elements, ultimately... and Defined as:

[0029]

[0030] The BI chart is as follows:

[0031]

[0032]

[0033] in It is the information propagated to the k-th layer of the bundle and project. and In the BI diagram, bundles and items are neighbors; connecting the embeddings of the k-layers yields representations of users and bundles, ultimately... and Defined as:

[0034]

[0035] Optionally, in step 3, the bundled representation is obtained from the bundled view using the prediction enhancement module. Project view gets user representation include:

[0036] By combining the user node representations under the project view and the bundled node representations under the bundled view obtained in step 2, and using the view's relationship network as the training set for the prediction enhancement model, new prediction results are obtained; the newly obtained user bundled connection interactions are added to the user view to obtain the final view.

[0037] By taking the inner product of the two, and selecting the top k highest-scoring bundles for each user, we can obtain the set of bundles that the user might be interested in. Add the newly obtained user bundle to the user and bundle interaction in the user view to obtain the final UB interaction. The specific method for updating the node representation in the user view is as follows:

[0038]

[0039] Optionally, step 4 involves using a contrastive learning module to refine the representation of each node in the three views, including:

[0040] The user representation in the user view and the user representation in the project view are compared and learned separately. The bundled representation in the user view and the bundled representation in the bundled view are compared and learned separately. The project representation in the project view and the project representation in the bundled view are compared and learned separately. The loss of the three is summed to obtain the final comparison and learning module loss.

[0041] Optionally, the final contrastive learning module loss is determined using the following method:

[0042] The feature alignment method for the comparative learning between the user representation in the user view and the user representation in the project view is as follows:

[0043]

[0044] The feature alignment method for comparing and learning the item representations in the project view and the bundled view is as follows:

[0045]

[0046] The feature alignment method for comparative learning between the bundled representation in the user view and the bundled representation in the bundled view is as follows:

[0047]

[0048] Where s(·,·) represents the cosine similarity function, and τ represents the hyperparameter called temperature;

[0049] Final contrastive learning module loss: for:

[0050]

[0051] Optionally, step 5: Perform model training, aggregating the user representations and bundled representations of each view obtained in steps 2 and 3 to obtain the final user representation e. u And binding means e b The model parameters are updated using the BPRLoss loss function, which is commonly used in recommender systems, including:

[0052] Model training includes training the prediction augmentation module and training the overall model. The prediction augmentation module training involves obtaining user node representations for the project view and bundled node representations for the bundled view through representation learning in step 2, feeding them into the prediction model to obtain new predictions of user and bundled interactions, and adding these to the user view to obtain new view and node representations. The overall model training involves initializing the representations of each node, obtaining new node representations through representation learning in step 2 and the prediction augmentation module, comparing the loss from the comparison learning module, merging the user and bundled representations of the three views, obtaining the final prediction score through the inner product, and updating the parameters using BPRLoss.

[0053] The method for concatenating user representations and bundled representations from different views is as follows:

[0054]

[0055] Where + indicates vector concatenation; the predicted score for the user's bundle is obtained through the inner product:

[0056]

[0057] The model's loss is divided into two parts: one part is the loss of the contrastive learning module, which is the loss from step four. The other part uses BPRLoss, the method is as follows:

[0058]

[0059] in σ(·) represents the sigmoid function, and the final loss is:

[0060]

[0061] Where λ1 and λ2 are hyperparameters, and Θ is the parameter of the model.

[0062] This application also provides a bundled recommendation system based on cross-view enhancement, the system including:

[0063] Building units are used to construct bipartite diagrams based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships; the bipartite diagram includes a user view (UB diagram), a project view (UI diagram), and a bundle view (BI diagram). Figure 3 One view;

[0064] The learning unit is used to learn the node representations of each view after the view has been built, through the view representation learning module.

[0065] The enhancement unit is used to use the predictive enhancement module to predict possible user and bundle interactions through the bundled view and the item view, and add them as enhancement information to the user view to obtain the enhanced user view.

[0066] The zoom-in unit is used to zoom in on the representations of each node in the three views through the contrastive learning module;

[0067] The training unit is used for model training, aggregating the user representations and bundled representations of each view obtained in steps 2 and 3 to obtain the final user representation e. u And binding means e b By using the common loss function BPRLoss, user-bundle pairs with interaction are used as positive samples, and those without interaction are used as negative samples to update the model parameters.

[0068] The test unit, used for model testing, obtains the final user rating y for the bundle through the inner product. ub The highest score among the top k is taken as the prediction result; the Recall value and NDCG value of the test set and the prediction result are used as evaluation criteria.

[0069] This application has the following advantages: (1) It obtains node representations of three elements—user, project, and bundle—through graph learning of three views, and performs pairwise comparative learning on the three views (user view, project view, and bundle view), transmitting information from the project view and bundle view across views to the user view. The feature alignment of the comparative learning makes the information of the three elements no longer independent, but mutually reinforcing. (2) Through the additional information from the project view and bundle view, the connection of the user bundle is expanded by the prediction enhancement module, which alleviates the problem of sparse connection of the user bundle to a certain extent. At the same time, due to the convergence of features of the three views by the comparative learning, the enhancement of the user view will also be fed back to the other two views. Attached Figure Description

[0070] Figure 1 The view construction module provided in the embodiments of this application;

[0071] Figure 2 The view representation learning module and prediction enhancement module provided in the embodiments of this application;

[0072] Figure 3 The comparison learning module provided for the embodiments of this application.

[0073] Figure 4 A flowchart provided for an embodiment of this application. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0075] This method constructs user views, project views, and bundle views by considering user interactions with bundles, user interactions with projects, and the affiliation relationships between projects and bundles. For the bundle view, graph learning fully explores the affiliation relationships between projects and bundles. The latter two views are used to enhance the main method's predictive performance on bundles across views. Graph learning and comparative learning across the three views enhance the model's predictive ability.

[0076] This application provides a bundled recommendation method based on cross-view enhancement, the specific steps of which are as follows:

[0077] Step 1: Construct a bipartite diagram based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships; the bipartite diagram includes a user view (UB diagram), a project view (UI diagram), and a bundle view (BI diagram). Figure 3 One view;

[0078] in, Represents a set of users. Represents a collection of items. This represents a set of bundled packages; where the matrix of user and project interactions is X∈{0,1}.M×N The user-bound interaction matrix is ​​Y∈{0,1} M×O The membership matrix of bundled projects is Z∈{0,1} O×N ;

[0079] The nodes in the graph are user nodes, bundle nodes, and project nodes. Edges are constructed based on the interaction relationships between user bundles, user projects, and the membership relationships between bundles and projects.

[0080] Step 2: After constructing the view, learn the node representation of each view through the view representation learning module.

[0081] Graph learning is used to learn the node representations in each of the three views in step (1). Specifically, a graph convolutional neural network (GCN) model is used to aggregate the information of adjacent nodes in the three views: user view, item view, and bundle view, to obtain the user and bundle node vector representations in the user view, the user and item node representations in the item view, and the bundle and item node representations in the bundle view.

[0082] Specifically, some edges of the bipartite graph are randomly masked, and then LightGCN is used to aggregate the neighbor node information in the bipartite graph, removing nonlinear transformations and biases from the graph convolutional neural network; for the UB graph (user view), the specific steps are as follows:

[0083]

[0084]

[0085] in, It is the information that is propagated to the user and the bundled k-th layer. and In the UB graph, the users and bundles are neighbors; connecting the embeddings of the k-layers yields representations of the users and bundles, ultimately... and Defined as:

[0086]

[0087] For the UI diagram (project view), the details are as follows:

[0088]

[0089]

[0090] in It is the information propagated to the k-th layer of users and projects. and It represents the neighbors of users and items in the UI diagram; connecting the embeddings of the k-layers yields a representation of the user and the bundled elements, ultimately... and Defined as:

[0091]

[0092] For BI (Bundled View) graphs, previous methods simply used the mean of the item representations to obtain the bundle representation. This simplistic approach is insufficient to fully extract information about BI membership relationships. This application utilizes graph convolutional neural networks to learn richer representations of the nodes in the BI graph, as detailed below:

[0093]

[0094]

[0095] in It is the information propagated to the k-th layer of the bundle and project. and In the BI diagram, bundles and items are neighbors; connecting the embeddings of the k-layers yields representations of users and bundles, ultimately... and Defined as:

[0096]

[0097] Step 3: Predictive enhancement module, through the bundle representation in the bundle view. User representation in project view Predict potential user and bundled interactions, add them as enhancement information to the user view, and obtain an enhanced user view.

[0098] By combining the user node representation under the project view and the bundled node representation under the bundled view obtained in step (2), the relational network of the view is used as the training set of the prediction enhancement model to obtain new prediction results; the newly obtained user bundled connection interactions are added to the user view to obtain the final view.

[0099] By taking the inner product of the two, we can obtain the top k highest-scoring bundles for each user to determine the bundles that the user might be interested in. Add the newly obtained user bundle to the user and bundle interaction in the user view to obtain the final UB interaction. The specific method for updating the node representation in the user view is as follows:

[0100]

[0101] Step 4: To facilitate information transfer among the three views, a comparative learning approach is used to narrow down the representations of nodes within each view. This aims to improve the flow of information from one view to another, resulting in better node representations. Specifically, the user representation in the user view is compared with the user representation in the project view; the bundle representation in the user view is compared with the bundle representation in the bundle view; and the project representation in the project view is compared with the project representation in the bundle view. The losses from these three approaches are then summed to obtain the final comparative learning module loss.

[0102] The feature alignment method for the comparative learning between the user representation in the user view and the user representation in the project view is as follows:

[0103]

[0104] The feature alignment method for comparing and learning the item representations in the project view and the bundled view is as follows:

[0105]

[0106] The feature alignment method learned by comparing the bundled representations in the user view and the bundled representations in the bundled view is as follows:

[0107]

[0108] Where s(·,·) represents the cosine similarity function, and τ represents the hyperparameter called temperature;

[0109] The final contrastive learning module loss is:

[0110]

[0111] Step 5: Train the model, and aggregate the user representations and bundled representations of each view obtained in Steps 2 and 3 to obtain the final user representation e. u And binding means e b By using the BPRLoss loss function commonly used in recommender systems, user-bundle pairs with interaction are used as positive samples, and non-interactive pairs are used as negative samples to update the model parameters.

[0112] Model training includes training the prediction enhancement module and training the overall model; wherein, the prediction enhancement module training includes learning the user node representation of the project view and the bundled node representation under the bundled view through the representation learning in step (2), feeding them into the prediction model to obtain the new predicted user and bundled interaction, adding the user view to obtain the new view and node representation; the overall model training includes initializing the representation of each node, learning the new node representation through the representation learning in step (2) and the prediction enhancement module, comparing the learning module loss, merging the user and bundled representations of the three views, obtaining the final prediction score through the inner product, and updating the model parameters by combining BPRLoss;

[0113] The method for concatenating user representations and bundled representations from different views is as follows:

[0114]

[0115] Where + indicates vector concatenation; the predicted score for the user's bundle is obtained through the inner product:

[0116]

[0117] The model's loss consists of two parts: one part is the contrastive learning loss, which is the loss from step four. The other part uses BPRLoss, specifically as follows:

[0118]

[0119] in σ(·) represents the sigmoid function, and the final loss is:

[0120]

[0121] Where λ1 and λ2 are hyperparameters, and Θ is the parameter of the model.

[0122] Step 6: Model testing, obtain the final user rating y for the bundle through the inner product. ub The highest score among the top k is taken as the prediction result. The Recall and NDCG values ​​of the test set and the prediction results are used as evaluation criteria.

[0123] Specifically, the final prediction score is obtained by combining the inner product of the user and bundled representations obtained during training. The top K bundles are selected as prediction results based on the final prediction score of each user, and the Recall and NDCG values ​​are calculated by comparing them with the results in the test set.

[0124] This application has the following advantages: (1) It obtains node representations of three elements—user, project, and bundle—through graph learning of three views, and performs pairwise comparative learning on the three views (user view, project view, and bundle view), transmitting information from the project view and bundle view across views to the user view. The feature alignment of the comparative learning makes the information of the three elements no longer independent, but mutually reinforcing. (2) Through the additional information from the project view and bundle view, the connection of the user bundle is expanded by the prediction enhancement module, which alleviates the problem of sparse connection of the user bundle to a certain extent. At the same time, due to the convergence of features of the three views by the comparative learning, the enhancement of the user view will also be fed back to the other two views.

[0125] This application also provides a bundled recommendation system based on cross-view enhancement, the system including:

[0126] Building units are used to construct bipartite diagrams based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships; the bipartite diagram includes a user view (UB diagram), a project view (UI diagram), and a bundle view (BI diagram). Figure 3 One view;

[0127] The learning unit is used to learn the node representations of each view after the view has been built, through the view representation learning module.

[0128] The enhancement unit is used to use the predictive enhancement module to predict possible user and bundle interactions through the bundled view and the item view, and add them as enhancement information to the user view to obtain the enhanced user view.

[0129] The zoom-in unit is used to zoom in on the representations of each node in the three views through the contrastive learning module;

[0130] The training unit is used for model training, aggregating the user representations and bundled representations of each view obtained in steps 2 and 3 to obtain the final user representation e. u And binding means e b By using the common loss function BPRLoss, user-bundle pairs with interaction are used as positive samples, and those without interaction are used as negative samples to update the model parameters.

[0131] The test unit, used for model testing, obtains the final user rating y for the bundle through the inner product. ub The highest score among the top k is taken as the prediction result; the Recall value and NDCG value of the test set and the prediction result are used as evaluation criteria.

[0132] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the service building apparatus and service loading apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0133] The embodiments described above do not constitute a limitation on the scope of protection of this application.

Claims

1. A bundled recommendation method based on cross-view enhancement, characterized in that, The method includes: Step 1: Construct a bipartite diagram based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships. The bipartite diagram includes three views: User View (UB), Project View (UI), and Bundle View (BI). Step 2: After constructing the view, the node representation of each view is learned through the view representation learning module; Step 3: Using the prediction enhancement module, predict possible user and bundle interactions through the bundle view and item view, and add them to the user view as enhancement information to obtain the enhanced user view; Step 4: Use the comparative learning module to bring the representations of each node in the three views closer together; Step 5: Train the model and aggregate the user representations and bundled representations of each view obtained in Steps 2 and 3 to obtain the final user representation. and bundling By using the common loss function BPRLoss, user-bundle pairs with interaction are used as positive samples, and those without interaction are used as negative samples to update the model parameters. Step 6, Model Testing: Obtain the final user rating for the bundle using the inner product. The highest score among the top k is taken as the prediction result; the Recall value and NDCG value of the test set and the prediction result are used as evaluation criteria. Step 1: Construct a bipartite diagram based on user-bundle interactions, user-project interactions, and project-bundle hierarchical relationships; including: Represents a set of users. Represents a collection of items. This represents a bundle set; the matrix of user and project interactions is as follows: The user-linked interaction matrix is ​​as follows: The membership matrix of bundled projects is as follows: ; The nodes of the graph are user nodes, bundle nodes, and project nodes. Edges are constructed based on the interaction relationships between user bundles, the interaction relationships between user projects, and the membership relationships between bundles and projects. Step 2: After constructing the views, the representation learning module learns the node representations for each view, including: Using a graph convolutional neural network model, the information of adjacent nodes is aggregated for three views: user view, item view and bundle view, to obtain the vector representation of user and bundle nodes in the user view, the representation of user and item nodes in the item view, and the representation of bundle and item nodes in the bundle view. Using a graph convolutional neural network model, the information of adjacent nodes is aggregated for three views: user view, item view, and bundle view, to obtain the vector representations of user and bundle nodes in the user view, the user and item node representations in the item view, and the bundle and item node representations in the bundle view, including: For bipartite graphs, some edges are randomly masked, and then LightGCN is used to aggregate neighbor node information in the bipartite graph, removing nonlinear transformations and biases from the graph convolutional neural network; for UB graphs, the process is as follows: ； ； in, , It is the information that is propagated to the user and the bundled k-th layer. and In the UB graph, the users and bundles are neighbors; connecting the embeddings of the k-layers yields representations of the users and bundles, ultimately... and Defined as: ； The UI design is as follows: ； ； in , It is the information propagated to the k-th layer of users and projects. and In the UI diagram, users and items are neighbors; connecting the embeddings of k layers yields a representation of the user and the bundled elements, ultimately... and Defined as: ； The BI chart is as follows: ； ； in , It is the information propagated to the k-th layer of the bundle and project. and It is the neighbor of bundles and items in the BI diagram; connecting the embeddings of k layers to obtain the representation of users and bundles, ultimately and Defined as: 。 2. The method according to claim 1, characterized in that, Step 3: Using the prediction enhancement module, obtain the bundled representation through the bundled view. The project view is obtained from the user's representation. ,include: By combining the user node representations under the project view and the bundled node representations under the bundled view obtained in step 2, and using the view's relationship network as the training set for the prediction enhancement model, new prediction results are obtained; the newly obtained user bundled connection interactions are added to the user view to obtain the final view. By taking the inner product of the two, and selecting the top k highest-scoring bundles for each user, we can obtain the set of bundles that the user might be interested in. The newly obtained user bundle is added to the user and bundle interaction in the user view to obtain the final UB interaction. To update the node representation in the user view, the specific method is as follows: 。 3. The method according to claim 2, characterized in that, Step 4: Use the comparative learning module to refine the representation of each node in the three views, including: The user representation in the user view and the user representation in the project view are compared and learned separately. The bundled representation in the user view and the bundled representation in the bundled view are compared and learned separately. The project representation in the project view and the project representation in the bundled view are compared and learned separately. The loss of the three is summed to obtain the final comparison and learning module loss.

4. The method according to claim 3, characterized in that, The final contrastive learning module loss was determined using the following method: The feature alignment method for the comparative learning between the user representation in the user view and the user representation in the project view is as follows: ； The feature alignment method for comparing and learning the item representations in the project view and the bundled view is as follows: ； The feature alignment method learned by comparing the bundled representations in the user view and the bundled representations in the bundled view is as follows: ； in Represents the cosine similarity function. This refers to the hyperparameter called temperature; Final contrastive learning module loss: for: 。 5. The method according to claim 4, characterized in that, Step 5: Train the model and aggregate the user representations and bundled representations of each view obtained in Steps 2 and 3 to obtain the final user representation. and bundling The model parameters are updated using the BPRLoss loss function, which is commonly used in recommender systems, including: Model training includes training the prediction augmentation module and training the overall model. The prediction augmentation module training involves obtaining user node representations for the project view and bundled node representations for the bundled view through representation learning in step 2, feeding them into the prediction model to obtain new predictions of user and bundled interactions, and adding these to the user view to obtain new view and node representations. The overall model training involves initializing the representations of each node, obtaining new node representations through representation learning in step 2 and the prediction augmentation module, comparing the loss from the comparison learning module, merging the user and bundled representations of the three views, obtaining the final prediction score through the inner product, and updating the parameters using BPRLoss. The method for concatenating user representations and bundled representations from different views is as follows: ； Where + indicates vector concatenation; the predicted score for the user's bundle is obtained through the inner product: ； The model's loss is divided into two parts: one part is the loss of the contrastive learning module, which is the loss from step four. The other part uses BPRLoss, the method is as follows: ； in ; Let represent the sigmoid function, and the final loss is: ； in and For hyperparameters These are the parameters of the model.

6. A bundled recommendation system based on cross-view enhancement, the system being used to implement any one of the methods of claims 1 to 5, characterized in that, The system includes: Building units are used to construct bipartite diagrams based on user and bundle interactions, user and project interactions, and project and bundle affiliation relationships. The bipartite diagram includes three views: User View (UB), Project View (UI), and Bundle View (BI). The learning unit is used to learn the node representations of each view after the view has been built, through the view representation learning module. The enhancement unit is used to use the predictive enhancement module to predict possible user and bundle interactions through the bundled view and the item view, and add them as enhancement information to the user view to obtain the enhanced user view. The zoom-in unit is used to zoom in on the representations of each node in the three views through the contrastive learning module; The training unit is used for model training, aggregating the user representations and bundled representations of each view obtained in steps 2 and 3 to obtain the final user representation. and bundling By using the common loss function BPRLoss, user-bundle pairs with interaction are used as positive samples, and those without interaction are used as negative samples to update the model parameters. The test unit, used for model testing, obtains the final user rating of the bundle through the inner product. The highest score among the top k is taken as the prediction result; the Recall value and NDCG value of the test set and the prediction result are used as the evaluation criteria.

Images