Multi-behavior commonality fusion recommendation method and system

Abstract

The present application relates to a kind of multi-behavior commonality fusion recommendation method, comprising the following steps: step A: obtaining the behavior data of user, contain user ID, target item ID, interaction type and interaction time, and the data obtained is preprocessed, and then training set is constructed;Step B: construct deep learning network model based on multi-behavior commonality fusion, the commonality information of different types of behavior is obtained and fused into final preference representation, while learning is assisted by multi-angle contrast learning Multi representation learning;Training set is used to train deep learning network model, and the required deep learning network model is trained by minimizing loss function on training set;Step C: the preprocessed user to be recommended and candidate item data are input into trained deep learning network model, and the probability that current user generates target behavior to target item is output.The method and system are beneficial to improve the accuracy of recommending interested items to users.

Description

Technical Field

[0001] This invention belongs to the field of recommendation systems, specifically relating to a multi-behavior commonality fusion recommendation method and system. Background Technology

[0002] In an era of information overload, extracting information of interest to users from massive amounts of data is a core principle. Recommendation systems, as an application of data mining, aim to achieve this by accurately mining each user's interests and preferences to provide them with information tailored to their individual needs. Personalized recommendation systems have been widely adopted in video platforms, e-commerce, and social media to alleviate the chain reaction of information overload. Most existing recommendation models typically only utilize behavioral data related to their target task. However, other types of behavior often contain underlying representations of user interests. For example, on online retail platforms, users may engage in various types of behavior with items, such as browsing, adding to cart, and purchasing. Different types of behavior correspond to different dimensions of user preferences. Using only target behavior to represent overall user preferences presents a suboptimal problem; effectively capturing and integrating user preferences from different perspectives corresponding to various types of behavior is challenging. Therefore, recent models have begun to incorporate multi-type behavioral data, modeling the correlations within different types of data and the dependencies between different types of data, and constructing multi-behavioral recommendation models using a two-paradigm approach. Furthermore, recent work has begun to introduce contrastive learning to assist model learning.

[0003] However, despite the performance improvements achieved by existing models to some extent, they still face several major weaknesses. First, the models suffer from data sparsity and semantic realism. Although various auxiliary behavioral data are used to assist in modeling user preferences, their impact on the target task is relatively small compared to the target behavior, and the sparsity of the target data persists. Furthermore, the fusion of semantics between multiple types of behaviors may compromise the true semantic meaning of the user's actions; therefore, properly guiding the fusion of behavioral semantics and addressing sparsity is crucial. Second, existing multi-behavior recommendation models typically model the contextual relationships within each behavior and the dependencies between behaviors, but lack the extraction of commonalities in overall user behavior. Behavioral commonalities can reflect the user's level of attention to each item and indirectly indicate whether the user will perform the target behavior for that item. Therefore, effectively capturing potential behavioral commonalities and integrating them into the overall user preferences is an important issue. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-behavioral commonality fusion recommendation method and system, which helps to improve the accuracy of recommending items of interest to users.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a multi-behavior commonality fusion recommendation method, comprising the following steps:

[0006] Step A: Obtain user behavior data, including user ID, target item ID, interaction type, and interaction time, and preprocess the obtained data before constructing a training set;

[0007] Step B: Construct a deep learning network model G based on the fusion of common behaviors. The deep learning network model G acquires common information of different types of behaviors and fuses it into the final preference representation. At the same time, it assists in the learning of multiple representations through multi-angle comparative learning. Train the deep learning network model G using the training set. The desired deep learning network model G is obtained by training on the training set by minimizing the loss function.

[0008] Step C: Input the preprocessed user and candidate item data into the trained deep learning network model G, and output the probability that the current user will perform the target behavior for the target item.

[0009] Furthermore, step A specifically includes the following steps:

[0010] Step A1: Obtain an existing general multi-behavior recommendation dataset; each sample in the multi-behavior recommendation dataset contains a user ID, a target item ID, the type of interaction behavior, whether an interaction occurred, and the interaction time;

[0011] Step A2: Preprocess each multi-behavior recommendation dataset, including: removing outlier data; removing users with fewer than 5 interactions and items with fewer than 10 interactions to reduce the number of users and items with fewer interactions; and forming a training set from the preprocessed multi-behavior recommendation dataset.

[0012] Furthermore, step B specifically includes the following steps:

[0013] Step B1: Divide the training set according to different behavior types and construct user-item interaction graphs G of different types. k ;

[0014] Step B2: Create user-item interaction graphs G for different types of users. k The input is fed into the corresponding initialization embedding layer to obtain the user initialization embedding for the k-th behavior type. Embedded with item initialization Based on multiple types of initialization embeddings, the overall initialization embedding e corresponding to the user is obtained. u The overall initialization embedding e corresponding to the item i ;

[0015] Step B3: Construct a behavior-aware relationship network, embedding the overall user initialization obtained in Step B2 into e. u Embedded with the overall initialization of the item i As input to the first layer of the behavior-aware relationship network, and respectively fed into the behavior-internal relationship encoder to capture specific contextual information within the same type, the behavior-internal relationship representation of users and items is obtained. and Then represent the relationships within all types of user behavior. Relationship between behavior and object The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining the overall behavior relationship representation for users and items. and The relationships between behaviors are represented and aggregated using multiple types of behavior relationships. The output of each layer is used as the input of the next layer for iterative aggregation. A fully connected layer is then used to fuse the behavior relationship representations of users and items across multiple layers, resulting in... and

[0016] Step B4: Based on the user-item interaction graph for each type, construct the overall user-item interaction graph G with heterogeneous relationships. all Construct a multi-view meta-network based on the overall user-item interaction graph G. all Using the behavioral relationship representations obtained from each layer of the network in step B3, construct a meta-feature view specific to each user layer. Meta-feature view of each item's unique type This captures common information across different types; the different types of meta-feature views obtained from each layer are input into personalized metanetworks F1 and F2 respectively, thereby capturing adaptive transformation matrices; and the adaptive transformation matrices are used to aggregate the common information of different types in each layer to obtain personalized common representations of users and items at each layer. and A feature debiasing network is constructed to enhance the realism and robustness of the extracted common representations. For the common representations obtained from each layer, a proportional fusion method is used to fuse them to obtain a multi-layer fused common representation. and Represent the user behavior relationships obtained in step B3 Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation

[0017] Step B5: Construct a multi-angle contrastive learning paradigm, using uniform noise augmentation to generate corresponding augmented views for the user-item behavior relationship representation at each layer in Step B3; and perform contrastive learning with the original views to obtain the contrastive loss based on uniform noise augmentation for both users and items. and Using user final preference representation that incorporates common information Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and

[0018] Step B6: Randomly negative sample users in a batch to generate corresponding positive sample interaction sets and negative sample interaction sets; calculate the prediction loss using the pairwise loss function; then fuse it with multiple contrastive losses to obtain the final optimization target loss; then calculate the gradient of each parameter in the deep learning network model G using the backpropagation method, and update each parameter using the Adam optimization algorithm to constrain the loss magnitude.

[0019] Step B7: Train the model using the final preference representation of users and items. When the loss value generated by the deep learning network model is less than the set threshold or the maximum number of iterations is reached, terminate the training of the deep learning network model G.

[0020] Furthermore, step B2 specifically includes the following steps:

[0021] Step B21: Input the user-item interaction graph into the corresponding initialization embedding layer to obtain the user's specific type of initialization embedding. Initialization embedding of specific item types

[0022] Step B22: Aggregate the initialization embeddings of each type obtained in step B21 into a list to obtain the overall initialization embedding for the user. and the overall initialization embedding of items

[0023] Furthermore, step B3 specifically includes the following steps:

[0024] Step B31: Construct a behavior-aware relationship network, which includes an intra-behavior relationship encoder and an inter-behavior relationship encoder. These are constructed using a memory-weighted network and a multi-head self-attention network with residual connections, respectively. The network layers are iterated to learn multi-hop information about users and items. The overall user initialization obtained in step B22 is then embedded into e. u and the overall initialization of the item embedding e i As the first layer input to the network, and using the output of each layer as the input to the next layer, the first layer first uses a behavior-internal relation encoder to capture contextual information for each type, in order to obtain the behavior-internal relation representation between the user and the item. and

[0025]

[0026]

[0027] Among them, e u With e i The IntraEnc() function, which consists of a memory-weighted network, performs overall initialization embeddings for users and items, and is mainly used for transformation learning between different dimensions.

[0028] Step B32: Represent the user-item behavior relationship obtained in Step B31. and The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining different types of behavior relationship representations. and And sum to obtain the overall relationship between user and item behaviors. and

[0029]

[0030]

[0031] IntraEnc() is composed of a multi-head self-attention mechanism, which preserves the original behavior-internal relation representation information;

[0032] Step B33: Repeat steps B31 and B32, the number of times corresponding to the number of layers in the network; thus obtaining a list of relationships between multi-layered overall behaviors corresponding to users and items. and

[0033] Step B34: Based on the list of multi-level behavioral relationship representations of users and items obtained in Step B33, aggregate the multi-level representations using a fully connected layer:

[0034]

[0035]

[0036] Here, concat() represents a concatenation function that concatenates representations from multiple different layers, while Mlp1() and Mlp2() represent fully connected layers with activation functions.

[0037] Furthermore, step B4 specifically includes the following steps:

[0038] Step B41: Based on the user-item interaction graph for each type [G] 1 G 2 ,…,G K Construct a holistic user-item interaction graph G with heterogeneous relationships. all It includes all types of interaction edges that exist between users and items;

[0039] Step B42: Based on the overall user-item interaction graph G obtained in step B41 all The list of relationships between user and item behaviors obtained from the l-th layer network in step B32 and Construct user- and item-specific meta-feature views separately, and use these meta-feature views to capture common behavioral information across different types:

[0040]

[0041]

[0042]

[0043]

[0044] in, The target representation of common behaviors aggregates multiple different types of behavioral relationship representations in the same layer through residual connections. Mlp3() represents a fully connected layer with an activation function, and concat() represents a concatenation function that concatenates multiple representations to form a specific type of meta-information view, which provides input for subsequent networks that capture common behaviors.

[0045] Step B43: Use two adaptive meta-transformation matrix learners to process the user's different types of meta-feature view lists obtained in step B42 for each layer. List of meta-feature views for different types of items Traverse and capture the adaptive transformation matrices of users and items under different types:

[0046]

[0047]

[0048] Here, F1 and F2 represent meta-learners that learn adaptive transformation matrices, which are constructed from two fully connected layers with activation functions;

[0049] Step B44: Using the adaptive transformation matrix obtained in step B34, aggregate the common information of different types in each layer to obtain personalized common representations of users and items at each layer:

[0050]

[0051]

[0052] Where δ() represents the activation function, which learns the potential behavioral common transformation representation under a specific type of representation through two adaptive transformation matrices;

[0053] Step B45: Enhance the realism of the extracted common representations and improve their robustness by utilizing the constructed feature-debiased network to capture enhanced common representations of users and items:

[0054]

[0055]

[0056]

[0057]

[0058] Among them, Feature_Decorr() represents the feature debiasing network, which learns unique dimension debiasing weights by removing noise between dimensions and enhances the original common representation in the form of residual connections.

[0059] Step B46: For the enhanced common representations of users and items obtained in step B45, fuse them using a proportional fusion method to obtain a multi-layered enhanced common representation, and then pass it through a fully connected layer to obtain an aggregated behavioral common representation:

[0060]

[0061]

[0062]

[0063]

[0064] Where Mlp4 represents a fully connected layer with an activation function, α m This represents the hyperparameter controlling the fusion ratio of each layer;

[0065] Step B47: Represent the relationships between user behaviors obtained in step B3. Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation

[0066]

[0067]

[0068] Where, α f The control commonality represents the proportion of the fusion hyperparameter.

[0069] Furthermore, step B5 specifically includes the following steps:

[0070] Step B51: Generate corresponding enhanced views for the user-item behavior relationship representation at each layer in Step B3 using a uniform noise enhancement method. and

[0071]

[0072]

[0073]

[0074]

[0075] in, and The enhancement representations are obtained by averaging L representations based on controllable uniform noise enhancement, thus avoiding the randomness of a single enhancement representation; where and Let α represent the feature representations initialized with a uniform distribution from 0 to 1 to achieve uniform enhancement. eps ∈R represents a hyperparameter that controls the size of the augmented representation;

[0076] Step B52: Perform a contrastive learning operation using the original view and the enhanced view. The contrastive learning loss paradigm based on uniform noise enhancement for both the user end and the item end is as follows:

[0077]

[0078]

[0079] in, A similarity function is adopted, which uses the inner product as the similarity function; τ represents the temperature coefficient, and u' and i' represent negative samples with different indices, respectively; based on the contrast loss defined above, the consistency between the original view and the enhanced view in each layer can be maximized, and the differences between different users and items can be maximized, thereby enhancing the robustness of the view after the fusion of multiple behavioral relationships;

[0080] Step B53: Utilize the fusion of common information to represent the user's final preferences Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and

[0081]

[0082]

[0083] in, A similarity function is adopted, which uses the inner product as the similarity function; τ represents the temperature coefficient, and u' and i' represent negative samples with different indices, respectively. Based on the contrast loss defined above, the similarity between the multi-relation fusion representation view and the behavior relationship representation view can be maximized, and the differences between different users and items can be maximized, thereby enhancing the robustness of the multi-relation fusion representation view.

[0084] Furthermore, step B6 specifically includes the following steps:

[0085] Step B61: Sample the corresponding negative samples from a small batch of users to form the corresponding positive sample interaction set. interaction set with negative samples

[0086] Step B62: Use the pairwise loss function Calculate the prediction loss, given u i Given a set of positive samples and a set of negative samples, the prediction loss is defined as follows:

[0087]

[0088] in, Indicates user u i The corresponding s-th interactive item in the positive sample item sequence, and Indicates user u i The corresponding negative sample item sequence contains the s-th non-interactive item, where θ represents the learnable parameters in the network, and parameters λ1 and L2 regularization are used to alleviate the overfitting problem.

[0089] Step B63: Update the learning rate using the Adam gradient optimization algorithm, iteratively update the model parameters using backpropagation, and train model G to minimize the loss function; fuse the contrastive loss obtained in steps B52 and B53 with the predictive loss obtained in step B61 to define the final recommendation loss:

[0090]

[0091] Where λ2 and λ3 represent the hyperparameters that control the intensity of the contrastive learning task.

[0092] The present invention also provides a multi-behavior commonality fusion recommendation system, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the above-described method steps.

[0093] Compared with existing technologies, this invention has the following advantages: This invention provides a multi-behavioral commonality fusion recommendation method and system. Unlike previous methods, this invention fully considers the common behavioral information present in multi-behavioral data, explicitly captures its features, and fuses them into the final preference representation; furthermore, it constructs a multi-angle comparative learning paradigm to promote effective learning of the representation from two perspectives. Experimental results show that the performance of this invention is superior to existing commonly used models, and it can effectively improve the accuracy of recommending items of interest to users. Attached Figure Description

[0094] Figure 1 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0095] Figure 2 This is an architecture diagram of a deep learning network model G based on the fusion of multiple behaviors in an embodiment of the present invention. Detailed Implementation

[0096] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0097] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0098] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0099] like Figure 1 As shown, this embodiment provides a multi-behavior commonality fusion recommendation method, including the following steps:

[0100] Step A: Obtain user behavior data, including user ID, target item ID, interaction type, and interaction time, and preprocess the obtained data before constructing a training set;

[0101] Step B: Construct a deep learning network model G based on the fusion of common behaviors. The deep learning network model G acquires common information of different types of behaviors and fuses it into the final preference representation. At the same time, it assists in the learning of multiple representations through multi-angle comparative learning. Train the deep learning network model G using the training set. The desired deep learning network model G is obtained by training on the training set by minimizing the loss function.

[0102] Step C: Input the preprocessed user and candidate item data into the trained deep learning network model G, and output the probability that the current user will perform the target behavior for the target item.

[0103] In this embodiment, step A specifically includes the following steps:

[0104] Step A1: Obtain an existing general multi-behavior recommendation dataset; each sample in the multi-behavior recommendation dataset contains a user ID, a target item ID, the type of interaction behavior, whether an interaction occurred, and the interaction time;

[0105] Step A2: Preprocess each multi-behavior recommendation dataset, including: removing outlier data; removing users with fewer than 5 interactions and items with fewer than 10 interactions to reduce the number of users and items with fewer interactions; and forming a training set from the preprocessed multi-behavior recommendation dataset.

[0106] Figure 2 This is an architecture diagram of the deep learning network model G based on multi-behavior commonality fusion in this embodiment. The specific implementation steps of step B in this embodiment are as follows.

[0107] Step B1: Divide the training set according to different behavior types and construct user-item interaction graphs G of different types. k .

[0108] Step B2: Create user-item interaction graphs G for different types of users. k The input is fed into the corresponding initialization embedding layer to obtain the user initialization embedding for the k-th behavior type. Embedded with item initialization Based on multiple types of initialization embeddings, the overall initialization embedding e corresponding to the user is obtained. u The overall initialization embedding e corresponding to the item i .

[0109] Step B2 specifically includes the following steps:

[0110] Step B21: Input the user-item interaction graph into the corresponding initialization embedding layer to obtain the user's specific type of initialization embedding. Initialization embedding of specific item types

[0111] Step B22: Aggregate the initialization embeddings of each type obtained in step B21 into a list to obtain the overall initialization embedding for the user. and the overall initialization embedding of items

[0112] Step B3: Construct a behavior-aware relationship network, embedding the overall user initialization obtained in Step B2 into e. u Embedded with the overall initialization of the item i As input to the first layer of the behavior-aware relationship network, and respectively fed into the behavior-internal relationship encoder to capture specific contextual information within the same type, the behavior-internal relationship representation of users and items is obtained. and Then represent the relationships within all types of user behavior. Relationship between behavior and object The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining the overall behavior relationship representation for users and items. and The relationships between behaviors are represented and aggregated using multiple types of behavior relationships. The output of each layer is used as the input of the next layer for iterative aggregation. A simple fully connected layer is then used to fuse the behavior relationship representations of users and items across multiple layers, resulting in... and

[0113] Step B3 specifically includes the following steps:

[0114] Step B31: Construct a behavior-aware relationship network, which includes an intra-behavior relationship encoder and an inter-behavior relationship encoder. These are constructed using a memory-weighted network and a multi-head self-attention network with residual connections, respectively. The network layers are iterated to learn multi-hop information about users and items. The overall user initialization obtained in step B22 is then embedded into e. u and the overall initialization of the item embedding e i As the first layer input to the network, and using the output of each layer as the input to the next layer, the first layer first uses a behavior-internal relation encoder to capture contextual information for each type, in order to obtain the behavior-internal relation representation between the user and the item. and

[0115]

[0116]

[0117] Among them, e u With e i InterEnc() consists of a memory-weighted network and is used for overall initialization embedding of users and items, respectively. It mainly performs transformation learning between different dimensions.

[0118] Step B32: Represent the user-item behavior relationship obtained in Step B31. and The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining different types of behavior relationship representations. and And sum to obtain the overall relationship between user and item behaviors. and

[0119]

[0120]

[0121] InterEnc() is composed of a multi-head self-attention mechanism, which preserves the original behavior-internal relation representation information.

[0122] Step B33: Repeat steps B31 and B32, the number of times corresponding to the number of layers in the network; thus obtaining a list of relationships between multi-layered overall behaviors corresponding to users and items. and

[0123] Step B34: Based on the list of multi-level behavioral relationship representations of users and items obtained in Step B33, aggregate the multi-level representations using a fully connected layer:

[0124]

[0125]

[0126] Here, concat() represents a concatenation function that concatenates representations from multiple different layers, while Mlp1() and Mlp2() represent fully connected layers with activation functions.

[0127] Step B4: Based on the user-item interaction graph for each type, construct the overall user-item interaction graph G with heterogeneous relationships. all Construct a multi-view meta-network based on the overall user-item interaction graph G. all Using the behavioral relationship representations obtained from each layer of the network in step B3, construct a meta-feature view specific to each user layer. Meta-feature view of each item's unique type This captures common information across different types; the different types of meta-feature views obtained from each layer are input into personalized metanetworks F1 and F2 respectively, thereby capturing adaptive transformation matrices; and the adaptive transformation matrices are used to aggregate the common information of different types in each layer to obtain personalized common representations of users and items at each layer. and A feature debiasing network is constructed to enhance the realism and robustness of the extracted common representations. For the common representations obtained from each layer, a proportional fusion method is used to fuse them to obtain a multi-layer fused common representation. and Represent the user behavior relationships obtained in step B3 Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation

[0128] Step B4 specifically includes the following steps:

[0129] Step B41: Based on the user-item interaction graph for each type [G] 1 G 2 ,…,G K Construct a holistic user-item interaction graph G with heterogeneous relationships. all It contains all types of interaction edges that exist between users and items.

[0130] Step B42: Based on the overall user-item interaction graph G obtained in step B41all The list of relationships between user and item behaviors obtained from the l-th layer network in step B32 and Construct user- and item-specific meta-feature views separately, and use these meta-feature views to capture common behavioral information across different types:

[0131]

[0132]

[0133]

[0134]

[0135] in, The target representation of behavioral commonality aggregates multiple different types of behavioral relationship representations in the same layer through residual connections. Mlp3() represents a fully connected layer with an activation function, and concat() represents a concatenation function that concatenates multiple representations to form a specific type of meta-information view, which provides input for subsequent networks that capture behavioral commonality.

[0136] Step B43: Use two adaptive meta-transformation matrix learners to process the user's different types of meta-feature view lists obtained in step B42 for each layer. List of meta-feature views for different types of items Traverse and capture the adaptive transformation matrices of users and items under different types:

[0137]

[0138]

[0139] Here, F1 and F2 represent meta-learners that learn the adaptive transformation matrix, which are constructed from two fully connected layers with activation functions.

[0140] Step B44: Using the adaptive transformation matrix obtained in step B34, aggregate the common information of different types in each layer to obtain personalized common representations of users and items at each layer:

[0141]

[0142]

[0143] Here, δ() represents the activation function, which learns the potential common behavioral transformation representation under a specific type of representation through two adaptive transformation matrices.

[0144] Step B45: Enhance the realism of the extracted common representations and improve their robustness by utilizing the constructed feature-debiased network to capture enhanced common representations of users and items:

[0145]

[0146]

[0147]

[0148]

[0149] Feature_Decorr() represents the feature debiasing network, which learns unique dimension debiasing weights by removing noise between dimensions and enhances the original common representation in the form of residual connections.

[0150] Step B46: For the enhanced common representations of users and items obtained in step B45, fuse them using a proportional fusion method to obtain a multi-layered enhanced common representation, and then pass it through a fully connected layer to obtain an aggregated behavioral common representation:

[0151]

[0152]

[0153]

[0154]

[0155] Where Mlp4 represents a fully connected layer with an activation function, α m This represents the hyperparameter controlling the fusion ratio of each layer. Step B47: Represent the user behavior relationships obtained in step B3. Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation

[0156]

[0157]

[0158] Where, α f The control commonality represents the proportion of the fusion hyperparameter.

[0159] Step B5: Construct a multi-angle contrastive learning paradigm, using uniform noise augmentation to generate corresponding augmented views for the user-item behavior relationship representation at each layer in Step B3; and perform contrastive learning with the original views to obtain the contrastive loss based on uniform noise augmentation for both users and items. and Using user final preference representation that incorporates common information Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and

[0160] Step B5 specifically includes the following steps:

[0161] Step B51: Generate corresponding enhanced views for the user-item behavior relationship representation at each layer in Step B3 using a uniform noise enhancement method. and

[0162]

[0163]

[0164]

[0165]

[0166] in, and The enhancement representations are obtained by averaging L representations based on controllable uniform noise enhancement, thus avoiding the randomness of a single enhancement representation; where and Let α represent the feature representations initialized with a uniform distribution from 0 to 1 to achieve uniform enhancement. eps ∈R represents the hyperparameter that controls the size of the augmented representation, which is adjusted by α. eps We can avoid the problem of large deviations between the augmented view and the original view, which would lead to the loss of true semantics. The augmented view can retain most of the information of the original representation and also retain some variance through this data augmentation method. Based on this augmentation method, we can construct an effective and simple comparison view.

[0167] Step B52: Perform a contrastive learning operation using the original view and the enhanced view. The contrastive learning loss paradigm based on uniform noise enhancement for both the user end and the item end is as follows:

[0168]

[0169]

[0170] in, A similarity function is adopted, which uses the inner product as the similarity function; τ represents the temperature coefficient, and u' and i' represent negative samples with different indices, respectively; based on the contrast loss defined above, the consistency between the original view and the enhanced view in each layer can be maximized, and the differences between different users and items can be maximized, thereby enhancing the robustness of the view after the fusion of multiple behavioral relationships.

[0171] Step B53: Utilize the fusion of common information to represent the user's final preferences Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and

[0172]

[0173]

[0174] in, A similarity function is adopted, which uses the inner product as the similarity function; τ represents the temperature coefficient, and u' and i' represent negative samples with different indices, respectively. Based on the contrast loss defined above, the similarity between the multi-relation fusion representation view and the behavior relationship representation view can be maximized, and the differences between different users and items can be maximized, thereby enhancing the robustness of the multi-relation fusion representation view.

[0175] Step B6: Randomly negative sample users in a batch to generate corresponding positive sample interaction sets and negative sample interaction sets; calculate the prediction loss using the pairwise loss function; then fuse it with multiple contrastive losses to obtain the final optimization target loss; then calculate the gradient of each parameter in the deep learning network model G using the backpropagation method, and use the Adam optimization algorithm to update each parameter to constrain the loss magnitude.

[0176] Step B6 specifically includes the following steps:

[0177] Step B61: Sample the corresponding negative samples from a small batch of users to form the corresponding positive sample interaction set. interaction set with negative samples

[0178] Step B62: Use the pairwise loss function Calculate the prediction loss, given u i Given a set of positive samples and a set of negative samples, the prediction loss is defined as follows:

[0179]

[0180] in, Indicates user u i The corresponding s-th interactive item in the positive sample item sequence, and Indicates user u i The corresponding negative sample item sequence contains the s-th non-interactive item, where θ represents the learnable parameters in the network, and parameters λ1 and L2 regularization are used to alleviate the overfitting problem.

[0181] Step B63: Update the learning rate using the Adam gradient optimization algorithm, iteratively update the model parameters using backpropagation, and train model G to minimize the loss function; fuse the contrastive loss obtained in steps B52 and B53 with the predictive loss obtained in step B61 to define the final recommendation loss:

[0182]

[0183] Where λ2 and λ3 represent the hyperparameters that control the intensity of the contrastive learning task.

[0184] Step B7: Train the model using the final preference representation of users and items. When the loss value generated by the deep learning network model is less than the set threshold or the maximum number of iterations is reached, terminate the training of the deep learning network model G.

[0185] In this embodiment, the proposed method (MMNCL) was tested on the same platform as commonly used existing methods, and the results are shown in Table 1. The test results show that the performance of the proposed method is superior to that of existing methods.

[0186] Table 1. Comparison of test results between this method and existing methods.

[0187]

[0188] This embodiment also provides a multi-behavior commonality fusion recommendation system, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the above-described method steps.

[0189] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multi-behavior commonality fusion recommendation method, characterized in that, Includes the following steps: Step A: Obtain user behavior data, including user ID, target item ID, interaction type, and interaction time, and preprocess the obtained data before constructing a training set; Step B: Construct a deep learning network model based on the fusion of common behaviors. The deep learning network model The common information of different types of behavior is obtained and fused into the final preference representation. At the same time, multi-angle comparative learning is used to assist the learning of multiple representations. A deep learning network model is trained using a training set. The desired deep learning network model is obtained by training on the training set by minimizing the loss function. ; Step C: Input the preprocessed user and candidate item data to be recommended into the trained deep learning network model. The output shows the probability that the current user will perform the target action on the target item. Step B specifically includes the following steps: Step B1: Divide the training set according to different behavior types and construct user-item interaction graphs for different types. ; Step B2: Create user-item interaction graphs for different types of users. The input is fed into the corresponding initialization embedding layer to obtain the corresponding first... User initialization embedding for each behavior type Embedded with item initialization Based on multiple types of initialization embeddings, the overall initialization embedding corresponding to the user is obtained. Overall initialization embedding corresponding to the item ; Step B3: Construct a behavior-aware relationship network, embedding the overall user initialization obtained in Step B2. Embedded with the overall initialization of the item As input to the first layer of the behavior-aware relationship network, and respectively fed into the behavior-internal relationship encoder to capture specific contextual information within the same type, the behavior-internal relationship representation of users and items is obtained. and Then represent the relationships within all types of user behavior. Relationship between behavior and object The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining the overall behavior relationship representation for users and items. and The relationships between behaviors are represented and aggregated using multiple types of behavior relationships. The output of each layer is used as the input of the next layer for iterative aggregation. A fully connected layer is then used to fuse the behavior relationship representations of users and items across multiple layers, resulting in... and ; Step B4: Based on the user-item interaction graph for each type, construct a heterogeneous overall user-item interaction graph. Construct a multi-view meta-network based on the overall user-item interaction graph. Using the behavioral relationship representations obtained from each layer of the network in step B3, construct a meta-feature view specific to each user layer. Meta-feature view of each item's unique type This allows for the capture of common information across different types; the resulting meta-feature views of different types in each layer are then input into a personalized meta-network. and This allows for the capture of an adaptive transformation matrix; and the adaptive transformation matrix is ​​used to aggregate common information of different types in each layer to obtain personalized common representations of users and items at each layer. and A feature debiasing network is constructed to enhance the realism of the extracted common representations and improve their robustness. For the common representations obtained from each layer, a proportional fusion method is used to fuse them to obtain a multi-layer fused common representation. and Represent the relationships between user behaviors obtained in step B3. Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation ; Step B5: Construct a multi-angle contrastive learning paradigm, using uniform noise augmentation to generate corresponding augmented views for the user-item behavior relationship representation at each layer in Step B3; and perform contrastive learning with the original views to obtain the contrastive loss based on uniform noise augmentation for both users and items. and ; Utilizing user final preference representations that incorporate common information Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and ; Step B6: Randomly negative sample users from a batch to generate corresponding positive and negative sample interaction sets; calculate the prediction loss using a pairwise loss function; then fuse it with multiple contrastive losses to obtain the final optimized target loss, and finally calculate the deep learning network model using the backpropagation method. The gradients of each parameter are calculated, and the Adam optimization algorithm is used to update each parameter to constrain the magnitude of the loss. Step B7: Train the model using the final user and item preference representations. Terminate the deep learning network model when the loss value generated is less than a set threshold or the maximum number of iterations is reached. Training.

2. The multi-behavior commonality fusion recommendation method according to claim 1, characterized in that, Step A specifically includes the following steps: Step A1: Obtain an existing general multi-behavior recommendation dataset; each sample in the multi-behavior recommendation dataset contains a user ID, a target item ID, the type of interaction behavior, whether an interaction occurred, and the interaction time; Step A2: Preprocess each multi-behavior recommendation dataset, including: removing outlier data; removing users with fewer than 5 interactions and items with fewer than 10 interactions to reduce the number of users and items with fewer interactions; and forming a training set from the preprocessed multi-behavior recommendation dataset.

3. The multi-behavior commonality fusion recommendation method according to claim 2, characterized in that, Step B2 specifically includes the following steps: Step B21: Input the user-item interaction graph into the corresponding initialization embedding layer to obtain the user's specific type of initialization embedding. Initialization embedding of specific item types ; Step B22: Aggregate the initialization embeddings of each type obtained in step B21 into a list to obtain the user's overall initialization embedding. = and the overall initialization embedding of items = .

4. The multi-behavior commonality fusion recommendation method according to claim 3, characterized in that, Step B3 specifically includes the following steps: Step B31: Construct a behavior-aware relationship network, which includes an intra-behavior relationship encoder and an inter-behavior relationship encoder. These are constructed using a memory-weighted network and a multi-head self-attention network with residual connections, respectively. The network layers are iterated to learn multi-hop information about users and items. The overall user initialization obtained in step B22 is then embedded into this network. and the overall initialization embedding of items As the first layer input to the network, and using the output of each layer as the input to the next layer, the first layer first uses a behavior-internal relation encoder to capture contextual information for each type, in order to obtain the behavior-internal relation representation between the user and the item. and : in, and The overall initialization embeddings for users and items are respectively performed. It consists of a memory-weighted network and mainly performs conversion learning between different dimensions; Step B32: Represent the user-item behavior relationship obtained in Step B31. and The inputs are fed into the behavior relationship encoder to capture the relationship dependencies between different types of users and items, thus obtaining different types of behavior relationship representations. and And sum to obtain the overall relationship between user and item behaviors. and : in, It consists of a multi-head self-attention mechanism, which retains the original behavioral relationship representation information; Step B33: Repeat steps B31 and B32, the number of times corresponding to the number of layers in the network; thus obtaining a list of relationships between multi-layered overall behaviors corresponding to users and items. and ; Step B34: Based on the list of multi-level behavioral relationship representations of users and items obtained in Step B33, aggregate the multi-level representations using a fully connected layer: in, This indicates a concatenation function that joins together representations from multiple different layers. and This represents a fully connected layer with an activation function.

5. The multi-behavior commonality fusion recommendation method according to claim 4, characterized in that, Step B4 specifically includes the following steps: Step B41: Based on the user-item interaction graph for each type Construct a holistic user-item interaction graph with heterogeneous relationships. It includes all types of interaction edges that exist between users and items; Step B42: Based on the overall user-item interaction graph obtained in step B41 With step B32 The list of relationships between user and item behaviors obtained from the layer network and Construct user- and item-specific meta-feature views respectively, and use these meta-feature views to capture common behavioral information existing in different types: in, The target representation, which represents the commonalities of behavior, aggregates the relationship representations between multiple different types of behavior in the same layer through residual connections. This represents a fully connected layer with an activation function. The concatenation function concatenates multiple representations to form a specific type of meta-information view, which provides input for subsequent networks that capture behavioral commonalities. Step B43: Use two adaptive meta-transformation matrix learners to process each of the matrices obtained in step B42. List of different types of meta-feature views for users in the layer List of meta-feature views for different types of items Traverse and capture the adaptive transformation matrices of users and items under different types: in, and This represents a meta-learner that learns an adaptive transformation matrix, which consists of two fully connected layers with activation functions; Step B44: Using the adaptive transformation matrix obtained in step B34, aggregate the common information of different types in each layer to obtain personalized common representations of users and items at each layer: in, The activation function learns the underlying common behavioral transformation representation under a specific type of representation through two adaptive transformation matrices. Step B45: Enhance the realism of the extracted common representations and improve their robustness by utilizing the constructed feature-debiased network to capture enhanced common representations of users and items: in, The feature debiasing network learns unique dimension debiasing weights by removing noise between dimensions and enhances the original common representation in the form of residual connections. Step B46: For the enhanced common representations of users and items obtained in step B45, fuse them using a proportional fusion method to obtain a multi-layered enhanced common representation, and then pass it through a fully connected layer to obtain an aggregated behavioral common representation: in, This represents a fully connected layer with an activation function. This represents the hyperparameter controlling the fusion ratio of each layer; Step B47: Represent the relationships between user behaviors obtained in step B3. Relationship between the behavior of objects Commonality representation of users obtained in step B4 Common characteristics of objects The user preference representation is obtained by aggregating the data using a proportional fusion method. and item preference representation : in, The control commonality represents the proportion of the fusion hyperparameter.

6. The multi-behavior commonality fusion recommendation method according to claim 5, characterized in that, Step B5 specifically includes the following steps: Step B51: Generate corresponding enhanced views for the user-item behavior relationship representation at each layer in Step B3 using a uniform noise enhancement method. and : in, and By averaging An enhancement representation based on controllable uniform noise enhancement is obtained, avoiding the randomness of a single enhancement representation; where and These represent feature representations initialized with a uniform distribution from 0 to 1 to achieve uniform enhancement. Hyperparameters that control the size of the augmented representation; Step B52: Perform a contrastive learning operation using the original view and the enhanced view. The contrastive learning loss paradigm based on uniform noise enhancement for both the user end and the item end is as follows: in, A similarity function is used, which employs the inner product as the similarity function; Indicates the temperature coefficient. and These represent negative samples with different indices. Based on the contrastive loss defined above, the consistency between the original view and the enhanced view in each layer can be maximized, as well as the differences between different users and items, thereby enhancing the robustness of the view after the fusion of multiple behavioral relationships. Step B53: Utilize the fusion of common information to represent the user's final preferences Relationship between user behavior and item final preference representation Relationship between behavior and object Contrastive learning based on feature fusion is performed separately to obtain contrastive losses for users and items. and : in, A similarity function is used, which employs the inner product as the similarity function; Indicates the temperature coefficient. and Let represent negative samples with different indices. Based on the contrastive loss defined above, we can maximize the similarity between the multi-relation fusion representation view and the behavior-relation representation view, and maximize the differences between different users and items, thereby enhancing the robustness of the multi-relation fusion representation view.

7. The multi-behavior commonality fusion recommendation method according to claim 6, characterized in that, Step B6 specifically includes the following steps: Step B61: Sample the corresponding negative samples from a small batch of users to form the corresponding positive sample interaction set. Interaction set with negative samples ; Step B62: Use the pairwise loss function Calculate the prediction loss, given Given a set of positive samples and a set of negative samples, the prediction loss is defined as follows: in, Indicates user The corresponding positive sample item sequence in the first An interactive item, and Indicates user The corresponding negative sample item sequence in the first One uninteracted item, among which This represents the learnable parameters in the network, and uses the parameters... as well as Regularization is used to alleviate the overfitting problem; Step B63: Update the learning rate using the Adam gradient optimization algorithm, and iteratively update the model parameters using backpropagation to train the model by minimizing the loss function. The contrast loss obtained in steps B52 and B53 is fused with the prediction loss obtained in step B61 to define the final recommendation loss: in, , ...

8. A multi-behavior commonality fusion recommendation system, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, which, when executed by the processor, enable the implementation of the steps of the method as described in any one of claims 1-7.

Images