Self-adaptive unwrapping comparison recommendation method

Abstract

The invention provides a self-adaptive unwrapping comparison recommendation method. The method comprises the following steps: S1, performing neighborhood aggregation on a user-project interaction graph; s2, performing intention unwrapping on the potential vector space of the user and the project: enabling the user and the project to have K1 and K2 different intentions respectively, and performing intention unwrapping on the potential intentions of the user and the project respectively; s3, generating a mask subgraph: performing cosine similarity generation on the mask subgraph # imgabs0 # by using the user-item interaction matrix and the intention vector to obtain a relation matrix # imgabs1 #; S4, generating a local collaboration view and a global collaboration view with adaptive enhancement; s5, the attention mechanism is used for distributing weights to the enhanced representation after aggregation is completed, final user and item representations E (u) and E (v) are obtained, and then dot product is carried out on the final user and item representations E (u) and E (v) to obtain a final prediction score Y; in order to enable user-project information propagation not to be limited by local levels, a learnable adaptive contrast enhancer is designed to effectively relieve data sparsity while relieving the noise problem.

Description

Technical Field

[0001] The present invention relates to the field of recommendation, and in particular to an adaptive disentanglement comparison recommendation method. Background Art

[0002] Recommendation systems can effectively alleviate data overload and play an indispensable role in people's daily lives. They mainly achieve personalized and accurate recommendations for users by learning the historical interactions between users and items. Collaborative filtering recommendation based on graph neural networks (GNNs) is one of the existing mainstream methods. For example: PinSage, NGCF, LightGCN, and LR-GCCF. It mainly uses the features of iteratively aggregated neighborhood nodes and learns the propagation mechanism of target node features, providing advanced recommendation performance for graph neural collaborative filtering methods.

[0003] However, the interaction data of users and items, as the key labels of supervised recommendation models, are often highly sparse in reality, making it difficult for the model to fully utilize the limited label information for effective training. Contrastive learning, as a self-supervised learning method, learns effective representations by comparing the similarities between positive samples and the differences between negative samples during training. To this end, recent studies have begun to combine the power of graph neural networks and contrastive learning to explore unlabeled information to provide self-supervisory signals for the model. Following the principle of maximizing mutual information to keep the positive pairs consistent and the negative pairs different, graph contrastive learning strengthens the user and item representations from different perspectives by constructing interaction graph subgraphs of users and items, and has shown strong performance in solving the problem of shortage of supervised labels. For example: XSimGCL, LightGCL, AdaMCL, AutoCF.

[0004] Although these methods have shown excellent performance in the recommendation field, there are still two key problems to be solved:

[0005] (1) As the number of stacking layers of traditional collaborative filtering models based on graph neural networks increases, the rapid expansion of the neighborhood receptive field will cause each node to exchange information with more and more neighbors. The node information of users and projects will be propagated to a large number of distant nodes, making the node embedding too similar and difficult to distinguish the characteristics of different users and projects. In this process, it is difficult to avoid introducing irrelevant long-distance node information into the node representation, which affects the quality of node representation and leads to performance degradation. Most current recommendation systems based on graph neural networks have this problem. Therefore, alleviating over-smoothing and improving the modeling method of user-item interaction graphs to obtain node representations with high discrimination and quality is one of the key issues that need to be overcome in this direction.

[0006] (2) There are highly complex latent factors behind the interaction between users and items. These factors are entangled with each other due to the diversity of user preferences. Modeling user preferences with a unified intent is difficult to capture a more fine-grained representation of the interaction between users and items, resulting in suboptimal user representation. In real life, the interaction between users and items is often driven by many intentional factors, such as buying a bag or work needs or being attracted by its style; watching a movie is to kill time or like a certain actor. Obviously, different intentions have different contributions to stimulating user behavior. However, the existing recommendation method based on graph contrastive learning uses a fixed encoding representation for the learned user preferences, which fails to capture more detailed interaction patterns between users and items, which hinders the ability to capture users' real interests and provide accurate intent-aware self-supervision. To this end, a new recommendation method needs to be designed to achieve information enhancement through de-entangled contrast signals to further improve recommendation performance. Summary of the invention

[0007] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes an adaptive disentanglement comparison recommendation method.

[0008] In order to achieve the above-mentioned object of the present invention, the present invention provides an adaptive disentanglement comparison recommendation method, characterized in that it includes the following steps:

[0009] S1, perform neighborhood aggregation on the user-item interaction graph: aggregate the neighbor node information of the user and item center nodes respectively to obtain the user's embedded representation and the embedded representation of the items

[0010] S2, perform intention disentanglement on the latent vector space of users and items: let users and items have K1 and K2 different intentions respectively, and use latent intentions c u and c v Then, the potential intentions of users and items are disentangled, and the disentangled intentions of users are The intention of the project after untangling The multi-intention encoder is used to encode the and In the example, the encoder is LightGCN; at this time, the prediction target is:

[0011]

[0012] in, The representation of the user's intention after passing through the encoder;

[0013] Indicates the intended encoding representation of the project after passing through the encoder; the encoder is LightGCN;

[0014] c u 、c v Represents the intention of the user and the project respectively;

[0015] ⊙ is the dot product operation;

[0016] f(·) represents the prediction function on the encoding intention;

[0017] S3, mask subgraph generation: mask subgraph generated by cosine similarity between user-item interaction matrix and intent vector t Get the relationship matrix This generates a parameterized edge mask generator for data augmentation;

[0018] S4, using edge mask generator to generate local collaborative view and global collaborative view with adaptive enhancement, and decoupled global collaborative view T (u) Complete the graph with global information I (u) ;

[0019] S5, use the attention mechanism to assign weights to the enhanced representation after aggregation to obtain the final user and item representation E (u) and E (v) , then for the E (u) and E (v) Perform a dot product to get the final prediction score Y, and recommend the item with the highest score to the user.

[0020]

[0021] In the formula, Represents the embedded representation of the user and the embedded representation of the project, that is, the user embedded representation obtained by S1 and item embedding representation The convolutional layer expression of

[0022] Respectively represent the local level and global level enhanced representation of the user, that is, two enhanced representation views; It is calculated by formula (14). At this time, formula (14) It is calculated by formula (6); It is calculated by formula (14). At this time, formula (14) It is calculated by formula (9);

[0023] Respectively representing the enhanced representation of the local level and the global level of the project, that is, two enhanced representation views; It is calculated by formula (14). At this time, formula (14) It is calculated by formula (6); It is calculated by formula (14). At this time, formula (14) It is calculated by formula (9);

[0024] The intention representation embeddings of the user and item are represented respectively, i.e., the disentangled intentions of the user and item;

[0025]

[0026] Among them, E (u) 、E (v) Represent the final user and item embedding vectors respectively;

[0027] Attweight is the weight distribution function learned by the attention mechanism;

[0028] δ (l) Represents the attention weight assigned to the lth layer;

[0029] Represent the embedding representation vectors of users and items in the l-th layer of graph neural network respectively;

[0030] The preference scores of users and items are calculated as follows:

[0031]

[0032] in, Represents user u i and project v j Embedded representation of ;

[0033] T is the transposition symbol;

[0034] Y i,j represents the interaction score between user i and item j;

[0035] Y represents the final interaction score between user i and item j;

[0036] E (u) represents the final embedding representation of user i and item j.

[0037] Preferably, performing neighborhood aggregation on the user-item interaction graph comprises:

[0038]

[0039] Where σ(·) represents the LeakyReLU activation function;

[0040] Represents the central node u i and v jThe embedded representation obtained by aggregating the neighbor nodes of

[0041] Represent the user-item interaction matrix The i-th row and interaction matrix The jth column in ;

[0042] E (u) 、E (v) Represent the initialization embedding representation of users and items respectively; the initialization results obtained by using the xavier function on users and items.

[0043] Preferably, is the normalized adjacency matrix derived from the user-item interaction matrix:

[0044]

[0045] In the formula and Represents the angle matrices of users and items respectively;

[0046] represents the user-item interaction matrix;

[0047] and Represents user u i and project v j The number of neighbor nodes set;

[0048] || indicates normalization operation;

[0049] Represents the element in the i-th row and j-th column of the adjacency matrix of the user-item interaction graph.

[0050] Preferably, the mask subgraph generated by using the user-item interaction matrix and the cosine similarity using the intent vector t Get the relationship matrix include:

[0051]

[0052] in, represents the user-item interaction matrix;

[0053] Represents a relationship matrix;

[0054] Represents the mask sub-image, and the solution formula is as follows:

[0055]

[0056] in, Represents the mask sub-graph For each value in ;

[0057]

[0058] ||||2 represents the two-norm;

[0059] for The transpose of

[0060] denote the disentangled representations of users and items, respectively.

[0061] Formulas (11) and (12) represent the edge mask generator.

[0062] Preferably, the relationship matrix Use learnable data enhancement to improve computing efficiency:

[0063] Use the node degree pair relationship matrix Perform normalization operation, expressed as:

[0064]

[0065] in, Represents the relationship matrix after normalization operation The element value of the i-th row and j-th column;

[0066] Representing the relationship matrix The element at row i and column j;

[0067] U represents the size of the user;

[0068] I represents the size of the project;

[0069] T is the transpose symbol.

[0070] Preferably, S4, generating a local collaborative view and a global collaborative view with adaptive enhancement, includes:

[0071] First, the normalized learning relationship matrix Applied to node messages to achieve learnable message propagation, the graph structure is perturbed to generate contrastive learning views with adaptive enhancement. This enhancement with adaptive mask is expressed as:

[0072]

[0073] in, They respectively represent the graph perturbation operation on the representation embedding vector of layer l;

[0074] Represent the representation embedding vectors of users and items obtained by traversing the graph neural network at layer l;

[0075] represents the relationship matrix obtained after normalization operation,

[0076] In formula (14) The following formula (6) is used to obtain the enhanced representation of the local level of users and items:

[0077]

[0078] in, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0079] The representation embedding vector obtained by passing the aggregated information of the user and its adjacent target nodes through the l-1 layer graph neural network;

[0080] The representation embedding vector obtained by passing the aggregated information of the project and its adjacent target nodes through the l-1 layer graph neural network;

[0081] In formula (14) The following formula (9) is used to obtain the global level enhanced representation of users and items:

[0082]

[0083] In the formula, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0084] They represent the first l-1 layers from the central node u i and v j The embedded representation obtained by aggregating the neighbor nodes of

[0085] and They represent the users after the first l-1 layers of the graph neural network are stacked. and Projects Intent-aware representation, They respectively represent the disentangled representations of users and items obtained by the l-1 layer graph neural network.

[0086] Preferably, Obtained by the following formula:

[0087]

[0088] in, The disentangled representations of users and items respectively;

[0089] Indicates that in the lth layer of the graph neural network, after passing through the encoder, in the intention c u Below, the user’s representation embedding vector; It means that when the l-th layer of the graph neural network is passed through the encoder, the intention c v Below, the representation embedding vector of the item;

[0090] c u 、c v Represent the user and intention representation vector respectively;

[0091] They represent the corresponding i-th user and intention representation vector respectively;

[0092] and Represents the representation vector of users and items at level l;

[0093] Represents user u i and each intent prototype c u relevance;

[0094] Indicates that the user v j and each intent prototype c v relevance;

[0095] Preferably, the correlation S k Use intent relevance score Calculate and use formula (8) to deduce:

[0096]

[0097] in, Represents the embedding representation vector of users and items at the lth layer;

[0098] for The transpose of

[0099] represents the embedded representation vector of the user at the l-1 layer;

[0100] They all represent the representation vectors of the user's i-th and i′-th intentions, where i′ represents the index, which means it includes the traversal operation;

[0101] for The transpose of

[0102] Represents the embedding representation vector of the item at level l-1;

[0103] They represent the representation vectors of the i-th and i′-th intentions of the item respectively, where i′ represents the index, which means that the traversal operation is included.

[0104] This gives the relevance score for each intent k Performing embedding propagation on each layer allows the information that influences the user intention k to be encoded into the representation.

[0105] Preferably, the weight is obtained by the following formula:

[0106]

[0107] In the formula, δ (l) Represents the attention weight assigned to the lth layer;

[0108] L represents the number of layers of the graph convolutional network;

[0109] ⊙ is the dot product operation;

[0110] E (l) 、E (l′) Both represent the feature vectors of the l-th layer and the l′-th layer of the graph neural network;

[0111] W Att is a parameter used for attention calculation. The attention mechanism can learn different weights for the graph neural network layer to optimize the loss function and effectively alleviate the over-smoothing problem.

[0112] Preferably, the negative samples used in training are obtained through the rating function:

[0113]

[0114] Among them, r m Represents candidate negative sample i m Rating value of

[0115] | | is the absolute value symbol;

[0116] represents the transpose of the embedding representation of user u;

[0117] Represent the embedding representation of candidate negative samples;

[0118] Represents the embedding representation of the positive sample;

[0119] α, β, p are predefined hyperparameters;

[0120] After calculating the ratings of all candidate negative samples, we get a rating set exist Select r m The candidate negative sample i with the smallest value m as the final negative sample.

[0121] In summary, due to the adoption of the above-mentioned technical solutions, the present invention defines K intentions with global semantic information from a global level (users and projects have K1 and K2 different intentions respectively) in order to make the user-item information propagation not limited to the local level; in order to solve the problem that most previous disentanglement recommendation methods are carried out in a fully supervised manner, a learnable adaptive contrast enhancer is designed, which effectively alleviates the noise problem while also effectively alleviating data sparsity; in view of the common limitation of graph neural networks that they all have over-smoothing problems, a layer attention mechanism is adopted to effectively alleviate this problem; in addition, in order to achieve the adaptive selection of negative samples with corresponding hardness for the prediction scores of positive samples, a two-round sampling strategy is carried out, thereby enhancing the accuracy and interpretability of the AdaDCL method.

[0122] The specific beneficial effects are:

[0123] (1) By applying graph contrastive learning technology, the refinement of user-item interaction data is enhanced, and the intent-aware information with a global level is integrated into the collaborative filtering framework. The layer attention mechanism is used to assign attention weights to each layer, which stabilizes the deep GNN training. The double negative sampling strategy is adopted to optimize the model learning process.

[0124] (2) A new AdaDCL recommendation method is proposed. A parameterized mask generator is used to implement an adaptive graph neural network architecture with global semantic enhancement. The introduced cross-view contrastive learning task is used to enhance the representation from multiple perspectives. The designed negative sampling rating function can adaptively select the appropriate negative sample hardness for the positive samples, thereby improving the robustness and generalization of the model.

[0125] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0127] Figure 1 This is the AdaDCL system architecture diagram of the present invention.

[0128] Figure 2 This is a diagram showing the impact of the number of intents on performance. Figure 2 (a) is the recall @20 and recall @40 values ​​on the Amazon-book dataset; Figure 2 (b) is the value of normalized discounted cumulative gain @20 and normalized discounted cumulative gain @40 on the Amazon-book dataset; Figure 2 (c) are the values ​​of recall@20 and recall@40 on the Tmall dataset; Figure 2 (d) is the value of normalized discounted cumulative gain @20 and normalized discounted cumulative gain @40 on the Tmall dataset; Figure 2 (e) is the value of recall@20 and recall@40 on the Gowalla dataset; Figure 2 (f) is the value of normalized discounted cumulative gain @20 and normalized discounted cumulative gain @40 on the Gowalla dataset.

[0129] Figure 3 It is a schematic diagram of the effect of β value on model performance when α=1; Figure 3 (a) is to explore the effect of β on recall@20, recall@40, normalized cumulative loss gain@20 and normalized cumulative loss gain@40 on the Amazon-book dataset when α=1.0; Figure 3 (b) is the exploration of the effect of β on recall @20, recall @40, normalized cumulative loss gain @20, and normalized cumulative loss gain @40 on the Gowalla dataset when α = 1.0; Figure 3 (c) is the effect of β on recall@20, recall@40, normalized discounted cumulative gain@20 and normalized discounted cumulative gain@40 when α=1.0 on the Tmall dataset.

[0130] Figure 4 It is a schematic diagram of the effect of α value on model performance when β = 0.1; Figure 4 (a) is to explore the effect of α value on recall @20, recall @40, normalized cumulative loss gain @20 and normalized cumulative loss gain @40 on the Amazon-book dataset when β = 0.1; Figure 3 (b) is the exploration of the effect of α value on recall @20, recall @40, normalized discounted cumulative gain @20, and normalized discounted cumulative gain @40 on the Gowalla dataset when β = 0.1; Figure 3(c) is the effect of α value on recall@20, recall@40, normalized discounted cumulative gain@20 and normalized discounted cumulative gain@40 on the Tmall dataset when β=0.1.

[0131] Figure 5 This is a schematic diagram of the impact of different sparsity on model performance. Figure 5 (a) is the value of normalized discounted cumulative gain @20 explored on the Amazon-book dataset according to sparsity; Figure 5 (b) explores the recall @20 value on the Amazon-book dataset based on sparsity; Figure 5 (c) is the value of normalized discounted cumulative gain @20 explored on the Gowalla dataset based on sparsity; Figure 5 (d) explores the recall@20 value on the Gowalla dataset based on sparsity.

[0132] Figure 6 It is a schematic diagram of the MAD indicator of AdaDCL and its variants and the baseline; Figure 6 (a) is the average distance of user representations on the Amazon-book dataset; Figure 6 (b) is the average distance of item representations on the Amazon-book dataset; Figure 6 (c) is the average distance of user representations on the Tmall dataset; Figure 6 (d) is the average distance of item representations on the Tmall dataset. DETAILED DESCRIPTION

[0133] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0134] This paper proposes an adaptive disentangled contrast recommendation method, AdaDCL, which realizes the intention decoupling enhanced by adaptation and self-supervision. By designing an intention prototype with local information and global level to transmit and aggregate the intention-aware information, it effectively realizes the multi-intention encoding representation of users and projects from the global dependency relationship; at the same time, in order to prevent the self-supervision information with serious noise from being extracted, it is necessary to be able to identify and utilize the key information of the graph structure. For this purpose, a parameterized edge mask generator is designed to capture the implicit relationship between users and projects, and incorporate the global dependency of intention awareness. The graph structure masker designed by this method can adaptively evaluate the importance of each user-project interaction; four contrast enhancement views are generated in the contrast learning stage to enhance the user-project representation; in the enhanced representation after aggregation, we introduce the attention mechanism to learn an Attweight weight distribution function, dynamically allocate weights according to the relevance of each layer to the central node, ensure that the model focuses on the information transmission of key nodes, and maintain the diversity of node features; finally, two rounds of negative sampling sampling strategies are used in the model training stage to realize the adaptive contrast enhancement of user-project relationships.

[0135] In order to make the user-item information propagation not limited to the local level, K intentions with global semantic information are defined at the global level; in order to solve the problem that most previous disentanglement recommendation methods are performed in a fully supervised manner, a learnable adaptive contrast enhancer is designed, which effectively alleviates the noise problem while also effectively alleviating data sparsity; in view of the common limitation of graph neural networks that they all have over-smoothing problems, a layer attention mechanism is used to effectively alleviate this problem; in addition, in order to achieve the adaptive selection of negative samples with corresponding hardness for the prediction scores of positive samples, a two-round sampling strategy is implemented, thereby enhancing the accuracy and interpretability of the AdaDCL method. The system architecture of AdaDCL is shown in the figure. Figure 1 As shown in the figure, it is mainly divided into three core parts: global disentanglement representation, contrastive multi-view enhanced representation, and adaptive hard negative sampling.

[0136] 1. Disentanglement of Intention and Reconstruction of Expression

[0137] 1.1 Modeling latent intention factors

[0138] Following the common collaborative filtering paradigm, using Represents a set of users, represents a set of items, where and Represents the size of users and items. The observed user-item interaction matrix can be expressed as If user u i With Project I jIf interaction occurs, The value of is set to 1, otherwise AdaDCL aims to predict the possibility of candidate users interacting with items based on historical user-item interactions, using P(y|u i ,v j ) represents the interaction probability, and y represents the learned prediction score.

[0139] The interaction between users and items is usually driven by diverse intentions, such as a special preference for a certain color. The intention of an item can be understood as its attribute, such as chocolate being labeled as romantic or blessing. In order to capture these diverse intentions, assume that users have K1 different intentions and items have K2 different intentions, respectively. and Thus, the score prediction of users and items can be transformed into the formula (1).

[0140]

[0141] Among them, K is K1 and K2, which represent the total number of different intentions of users and items (items) respectively;

[0142] P(y|u i ,v j ) represents the interaction probability between users and items;

[0143] represents the interaction probability between the user and the item driven by the kth intention;

[0144] Respectively represent the integration operation of user intention and item intention;

[0145] K represents the total number of intents of users and items;

[0146] The user-item interaction probability is given by the latent intent c u and c v We can further derive formula (2).

[0147]

[0148] Among them, K is K1 and K2, which represent the total number of different intentions of users and items respectively;

[0149] represents the interaction probability between the user and the item driven by the kth intention. It is just a change of expression;

[0150] represents the probability of interaction between the user and the item driven by the kth intention;

[0151] represents the probability that the user will interact with this item under the k intentions of the item;

[0152] Indicates the intention encoding representation of users and items after passing through the encoder;

[0153] P(y|c u ,c v ) represents the probability of interaction between users and items driven by all intents.

[0154] Using f(·) to represent the prediction function on the coding intent, the prediction target of the AdaDCL method is formula (3).

[0155]

[0156] in, The representation of the user's intention after passing through the encoder;

[0157] It represents the intended encoding representation of the item after passing through the encoder; the encoder in this article is LightGCN;

[0158] c u 、c v Represents the intention of the user and the project respectively;

[0159] ⊙ is the dot product operation.

[0160] 1.2 Multi-dimensional Intent Representation Using Global Information

[0161] Although existing recommendation systems have explored the disentangled representation and diversity of intent, they have largely ignored the intention-aware collaborative relationship at the global level. User interests and preferences usually have multidimensional characteristics. Modeling user (item) dependencies from a global level can not only strengthen the message passing based on graph neural networks and effectively deal with data sparsity, so that the information propagation between users and items is no longer limited to direct local relationships, but also help the model better capture the potential relationships hidden behind the direct interaction between users and items. It can also provide richer contextual information from a global perspective and enhance the stability and reliability of the model. To this end, in the process of information transmission, the disentangled collaborative relationship between users and items is represented at both the local and global levels.

[0162] (1) Message passing based on graph neural network

[0163] The powerful representation learning ability of graph neural networks has made it the mainstream paradigm for capturing collaborative filtering signals, such as LightGCN, HGCF, and SVD-GCN. This inspired us to use lightweight graph convolution (LightGCN) as the core architecture of the model, which abandons feature transformation and nonlinear activation parts and directly aggregates the feature information of neighbors. Use d to represent the embedding dimension. They represent the embedding matrices of users and items respectively, so the local graph embedding propagation can be shown as formula (4).

[0164]

[0165] Where σ(·) represents the LeakyReLU activation function;

[0166] Represents the central node u i and v j The embedded representation obtained by aggregating the neighbor nodes of

[0167] Respectively The interaction matrix of the i-th row in and the interaction matrix of the j-th column in;

[0168] E (u) 、E (v) Represent the initialization embedding representation of users and items respectively; the initialization results obtained by using the xavier function on users and items.

[0169] is the normalized adjacency matrix derived from the user-item interaction matrix:

[0170]

[0171] In the formula and Represents the angle matrices of users and items respectively;

[0172] represents the user-item interaction matrix;

[0173] and Represents user u i and project v j The number of neighbor nodes set;

[0174] || indicates normalization operation;

[0175] Represents the element in the i-th row and j-th column of the adjacency matrix of the user-item interaction graph;

[0176] Therefore, the information propagation from layer l-1 to layer l can be expressed as formula (6).

[0177]

[0178] in, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0179] The representation embedding vector obtained by passing the aggregated information of the user and its adjacent target nodes through the l-1 layer graph neural network;

[0180] The representation embedding vector obtained by passing the aggregated information of the item and its adjacent target nodes through the l-1 layer graph neural network;

[0181] In addition, to emphasize the semantics of central nodes and alleviate the over-smoothing problem, residual connections are applied in the aggregation stage.

[0182] (2) Intent-aware information fusion strategy

[0183] The user / item intention perception information with global dependencies is integrated into the collaborative filtering framework based on graph neural network. The multi-intent encoder in the previous section saves the user / item preferences in and K1 and K2 global intents are defined for users and projects respectively. and Through these learnable intent embeddings, we use formula (7) to aggregate K different intent prototype information at the lth layer that contains global semantic information.

[0184]

[0185] in, The disentangled representations of users and items respectively;

[0186] Indicates that in the lth layer of the graph neural network, after passing through the encoder, in the intention c u Below, the user’s representation embedding vector; It means that when the l-th layer of the graph neural network is passed through the encoder, the intention c v Below, the representation embedding vector of the item;

[0187] c u 、c v Represent the user and intention representation vector respectively;

[0188] They represent the corresponding i-th user and intention representation vector respectively;

[0189] and Represents the representation vector of users and items at level l;

[0190] Represents user u i and each intent prototype c u relevance;

[0191] Indicates that the user v j and each intent prototype c v relevance;

[0192] Correlation k Use intent relevance score Calculated, use formula (8) to derive the calculation.

[0193]

[0194] in, Represents the embedding representation vector of users and items at the lth layer;

[0195] for The transpose of

[0196] represents the embedded representation vector of the user at the l-1 layer;

[0197] They all represent the representation vectors of the user's i-th and i′-th intentions, where i′ represents the index, which means it includes the traversal operation;

[0198] for The transpose of

[0199] Represents the embedding representation vector of the item at level l-1;

[0200] They represent the representation vectors of the i-th and i′-th intentions of the item, respectively, where i′ represents the index, which means it includes the traversal operation;

[0201] This gives the relevance score for each intent k Performing embedding propagation on each layer allows the information that influences the user intention k to be encoded into the representation.

[0202] Therefore, the local collaborative filtering signal Untangling signals with global collaboration After executing message propagation, it can be further refined into formula (9).

[0203]

[0204] In the formula, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0205] They represent the first l-1 layers from the central node u i and v j The embedding representation obtained by aggregating neighbor nodes and Represents the users after the layers are stacked and Projects Intent-aware representation.

[0206] Decomposing intent into a graph neural network architecture enables the learned representation to effectively parse the underlying factors driving complex user-item interaction behaviors.

[0207] 2 Disentanglement Contrastive Learning

[0208] Explore the combination of contrast enhancement and intent disentanglement to address the challenge of data sparsity. Existing self-supervisory signals are mainly achieved by maximizing the consistency between positive sample pairs in contrasting views, but are easily disturbed by noisy data, such as accidental clicks. Noise contrast regularization may mislead the self-supervised learning process. For example, forcing learning to make embeddings consistent at nodes with noisy interactions may cause the model to focus on irrelevant or misleading intents, resulting in bias in the final generated embeddings, which in turn affects the matching degree between users and items.

[0209] To overcome this challenge, a learnable adaptive contrast enhancer is designed, which can take into account both local collaborative relationships and global disentangled dependencies of users / items. Through such a design, the learnable contrast enhancer can adaptively learn the disentangled self-supervised learning signal. In addition, in view of the common limitation of the negative sampling strategy adopted by most existing recommendation methods that only negative samples with fixed hardness levels can be selected, which is prone to false positives and false negatives, a new strategy of adaptive hardness negative sampling is used to effectively realize the adaptive selection of negative samples with different hardness levels for different positive samples. It is worth noting that the existing negative sampling method can be regarded as a more relaxed case of the negative sampling method adopted by AdaDCL.

[0210] 2.1 Disentanglement Data Augmentation

[0211] In order to make the enhancer adaptive to each connection jump, a learnable relationship matrix is ​​introduced To encode the implicit relationship between users and items at each layer, and filter out the edge information irrelevant to the task in the graph neural network layer, it is expected to generate a mask subgraph With the help of mask subgraph, element-by-element multiplication is used to obtain the relationship matrix as shown in formula (10).

[0212]

[0213] in, represents the user-item interaction matrix;

[0214] represents the mask sub-image;

[0215] Represents a relationship matrix;

[0216] (1) Mask sub-graph: Mask sub-graph Each value in It reflects the degree to which the interaction between users and items is masked. The closer its value is to 1, the more important the interaction is, and vice versa. and Projects The untangle representation is derived from To maintain the intention-aware interaction mode. Specifically, the cosine similarity between node embedding representations is used to measure the importance of the interaction as shown in formula (11).

[0217]

[0218] || ||2 represents the two-norm;

[0219] for The transpose of

[0220] The disentangled representations of users and items respectively;

[0221] Note that formula (11) may produce negative values, so formula (12) is used to convert it into a normalized cosine distance so that the mask value similarity range is linearly transformed to [0, 1].

[0222]

[0223] in, Represents the mask sub-graph For each value in ;

[0224] (2) Learnable data enhancement: When there is no interaction between users and items, the interaction matrix Since the relationship matrix The value of is and It is obtained by element-by-element multiplication. Therefore, in order to improve the computational efficiency, only the mask value with interaction needs to be calculated. Using the learnable relationship matrix, the node degree is used for normalization operation as shown in formula (13).

[0225]

[0226] in, Represents the relationship matrix after normalization operation The element value of the i-th row and j-th column;

[0227] Representing the relationship matrix The element at row i and column j;

[0228] U represents the size of the user;

[0229] I indicates the size of the item;

[0230] T is the transpose symbol;

[0231] In order to combine the designed learnable adaptive contrast enhancer with the information transfer strategy, the normalized learning relation matrix Applied to node messages to achieve learnable message propagation, contrastive learning views with adaptive enhancement are generated by perturbing the graph structure. This enhancement with adaptive mask can be expressed as shown in Equation (14).

[0232]

[0233] in, They respectively represent the graph perturbation operation on the representation embedding vector of layer l;

[0234] Represent the representation embedding vectors of users and items obtained by traversing the graph neural network at layer l;

[0235] represents the relationship matrix obtained after normalization operation,

[0236] To generate diverse comparative views, we consider both local collaborative signals and global disentanglement relations. Specifically, in the local embedding of user-item encodings Global embedding disentangled from intent Two learnable mask matrices are used to implement the enhancement operation, specifically using Get the two mask values.

[0237] After performing these operations, the enhanced perceptual message passing paradigm can be further refined as Equation (15).

[0238]

[0239] In the formula Respectively represent the local level and global level enhanced representation of the user, that is, two enhanced representation views; It is calculated by formula (14). At this time, formula (14) It is calculated by formula (6); It is calculated by formula (14). At this time, formula (14) It is calculated by formula (9);

[0240] Respectively representing the enhanced representation of the local level and the global level of the project, that is, two enhanced representation views; It is calculated by formula (14). At this time, formula (14) It is calculated by formula (6); It is calculated by formula (14). At this time, formula (14) It is calculated by formula (9);

[0241] Intent representation embeddings representing users and items respectively;

[0242] (3) Layer Attention Mechanism

[0243] The deep collaborative filtering structure of the graph makes the node representation diverse, but it is also very easy to produce over-smoothing. This phenomenon is common in many collaborative filtering recommendation systems based on graph neural networks, which limits the improvement of recommendation effects. To effectively alleviate this problem, a layer attention mechanism is designed to learn a weight distribution function Attweight to distribute weights to each layer, so that the central node u i Graph convolution layers with higher correlations have higher attention weights, which effectively alleviate the over-smoothing problem while increasing high-order neighbor information:

[0244]

[0245] Among them, E (u) Represents the representation embedding vector obtained by traversing the graph neural network at layer l of the item;

[0246] Attweight is the weight distribution function learned by the attention mechanism;

[0247] E (L-1)Represents the embedded representation vector of the L-1th layer user;

[0248] L represents the number of layers of the graph convolutional network;

[0249] δ (l) represents the attention weight assigned to the lth layer, which can be specifically expressed as formula (17).

[0250]

[0251] Where L represents the number of layers of the graph convolutional network;

[0252] ⊙ is the dot product operation;

[0253] E (l) 、E (l′) They all represent the feature vectors of the lth layer and the l′th layer, where l′ is an index variable used to indicate the meaning of traversal.

[0254] W Att is a parameter used for attention calculation. The attention mechanism can learn different weights for the graph neural network layer to optimize the loss function and effectively alleviate the over-smoothing problem.

[0255] 2.2 Contrastive Learning

[0256] (1) Multi-view generation

[0257] By considering users / items matching the original collaborative view embedding as positive pairs and different encoding nodes as negative pairs, multiple contrastive views are generated from four different perspectives: i) Local collaborative view with adaptive enhancement (Q β,(u) ); ii) Decoupled global collaboration view (T (u) ); iii) Adaptive Enhanced View (Q γ,(u) );iv) Global information completion graph (I (u) ). Finally, the InfoNCE loss function is used to generate the contrastive self-supervisory signal, which is calculated as shown in formula (18).

[0258]

[0259] Where m represents the view z∈Z with the initial representation obtained from the graph neural network encoding (u) ;

[0260] n represents the samples h obtained from the four enhanced embedding representations β ∈H β,(u) ,T (u) ,h γ ∈H γ,(u) ,I (u) ;

[0261] I(m,n) represents the contrast loss value;

[0262] L represents the number of layers of the graph convolutional network;

[0263] denote the unperturbed encoded representation of user i and the perturbed encoded representation of user i respectively;

[0264] τ represents the temperature constant, which is used to adjust the smoothness of the Softmax function;

[0265] η(·) represents the cosine similarity function.

[0266] Therefore, the contrastive learning loss can be expressed as Equation (19).

[0267]

[0268] Among them, z represents the original view;

[0269] t represents the intention representation vector of users and items;

[0270] q β ,q γ represent the enhanced representation at the local level and the global level respectively;

[0271] After stacking L graph neural layers, the embeddings of different layers are aggregated to get the final representation:

[0272]

[0273] Among them, E (u) 、E (v) Represent the final user and item embedding vectors respectively;

[0274] Represent the embedding representation vectors of users and items at the lth layer respectively;

[0275] The preference scores of users and items are calculated as formula (21).

[0276]

[0277] in, Represents user u i and itemsv j Embedded representation of ;

[0278] T is the transpose symbol;

[0279] Y i,j represents the interaction score between user i and item j;

[0280] Y represents the final interaction score between user i and item j;

[0281] E (u) Represents the final embedding representation of user i and item j;

[0282] (2) Adaptive hardness negative sampling

[0283] Although the existing negative sampling method has achieved good results, it can only select negative samples with a fixed hardness level during sampling, which is very likely to cause false positive and false negative problems. In order to alleviate such restrictions and achieve adaptive selection of negative samples with appropriate hardness to obtain better recommendation results, we adopted a more rigorous and scientific sampling method. Specifically, for each positive sample pair (u, i + ), following the traditional sampling method but with a two-round sampling strategy. First, a fixed-size set of items is randomly selected from the items that the user has not interacted with to form a negative sample candidate set C. The size of the candidate set S is usually smaller than the total number of items N, ensuring the effectiveness and feasibility of sampling. In the second round of sampling, negative samples are carefully selected from the candidate set according to the predefined rating function and sampling rules. The rating function and sampling rules of this step play a decisive role in the final determination of negative samples and are the research focus of all hard negative sampling methods. To this end, for each candidate negative sample i m ∈C, we carefully designed the negative sampling rating function as shown in formula (22):

[0284]

[0285] Among them, r m Represents candidate negative sample i m Rating value of

[0286] || is the absolute value symbol;

[0287] represents the transpose of the embedding representation of user u;

[0288] Represent the embedding representation of candidate negative samples;

[0289] Represents the embedding representation of the positive sample;

[0290] α, β, and p are predefined hyperparameters, which are mainly used to adjust the hardness of the selected negative samples. Through this negative sampling method, the implicit collaborative filtering model can obtain a larger normalized discounted cumulative gain (NDCG) than fixed hardness negative sampling. Other negative sampling is a broader expression of our use of the negative sampling model. After calculating the ratings of all candidate negative samples, a rating set is obtained. exist Select r m The i with the smallest value mAs the final negative sampling, as shown in formula (23).

[0291]

[0292] (2) Model training

[0293] During the model training phase, the classic ranking model Bayesian Personalized Ranking (BPR) recommendation loss is followed to optimize the model so that items that users prefer are ranked higher than those that they dislike:

[0294]

[0295] Where σ(·) represents the sigmoid activation function;

[0296] represents the training data;

[0297] Represent the interaction between the user and the positive sample and the negative sample respectively;

[0298] j represents the sample that has never interacted with user u.

[0299] Finally, the self-supervision loss and BPR loss are integrated into multi-task learning to obtain the final loss function expressed as formula (25).

[0300]

[0301] Where λ1, λ2 and λ3 are adjustable hyperparameters;

[0302] represents the BPR loss function;

[0303] Represents the contrastive loss between users and items;

[0304] and

[0305] Θ1 and Θ2 represent the initial embedding representation of users and items and the initial embedding representation of the intentions of users and items;

[0306] |||| F represents the F-norm;

[0307] To demonstrate the effectiveness of this method, experiments were conducted on three public datasets: Gowalla, Tmall, and Amazon-book to answer the following questions and verify the effectiveness of the model:

[0308] RQ1: How does the performance of the proposed AdaGCL recommendation method compare to some current advanced recommendation models?

[0309] RQ2: Are the key components of model design conducive to the representation learning of AdaGCL and thus improve the performance of the model?

[0310] RQ3: How does the variation of hyperparameters in experiments affect the performance of the model?

[0311] RQ4: Is the AdaGCL recommendation method helpful in alleviating the sparsity problem of data in disentangled self-supervisory signals?

[0312] RQ5: Is the proposed negative sampling strategy for adaptive hardness universal and adaptable to the experimental settings of other recommendation models?

[0313] (1) Experimental setup

[0314] (1.1) Dataset

[0315] Three representative real-world datasets were selected for experiments to test the performance of the model. Gowalla: A real-world dataset widely used to study social networks and recommendation systems. The dataset contains user check-in records, location information, timestamps, etc. Tmall: It comes from Tmall, a large Chinese e-commerce platform. It includes user purchase records, browsing behaviors, item evaluations, item attributes, timestamps, and other related user behavior data. It is an important data source for recommendations. Amazon-book: It is built based on the Amazon e-commerce platform. The dataset covers a large number of interactions between users and books, and is suitable for the study of user behavior modeling and book recommendation systems. Table 1 shows the detailed statistical information of the dataset.

[0316] Table 1 Detailed information of the experimental dataset

[0317] Dataset #Users #Items #Interactions Density Gowalla 50，821 57，440 1，172，425 <![CDATA[4.0e -4 ]]> Amazon-book 78，578 77，801 2，240，156 <![CDATA[3.7e -4 ]]> Tmall 47，939 41，390 2，357，450 <![CDATA[1.2e -4 ]]>

[0318] (1.2) Evaluation indicators

[0319] In order to ensure the fairness of the evaluation, we follow the conventional settings and use the two most commonly used evaluation indicators in the recommendation field: Recall @K and Normalized Discounted Cumulative Gain (NDCG) @K to evaluate the recommendation performance of the Top-K model. Unless otherwise specified, the default K value is 20. The calculation methods of Recall and Normalized Discounted Cumulative Gain (NDCG) are as follows:

[0320] 1) Recall: The recall rate measures the proportion of user-item pairs that the recommendation system can successfully predict among all the user-item pairs that actually interact, reflecting the coverage of the recommendation system for the items that the user is actually interested in. R(u) represents the Top-K recommendation list provided to the user based on the learned user behavior; T(u) represents the set of items that the user actually selects after the recommendation system recommends to the user.

[0321]

[0322] 2) CG (Cumulative Gain): Accumulate the relevance scores of each recommended result in the recommendation list. It simply accumulates the relevance without considering the position information of the recommended result in the list. In some cases, high-scoring results may be ranked lower, while low-scoring results may be ranked higher.

[0323]

[0324] 3) DCG (Discounted Cumulative Gain): DCG introduces the position factor based on CG. By considering the sorting order, the top-ranked items gain more and the bottom-ranked items lose more.

[0325]

[0326] 4) NDCG (Normalized Discounted Cumulative Gain): When evaluating a recommendation system, we do not evaluate only the recommendation list of a single user, but the recommendation lists of the entire test set of users as a whole. Therefore, the evaluation scores of the recommendation lists of different users need to be normalized. NDCG is the DCG of the current result normalized with IDCG, which is used to indicate the degree to which the current result is close to the most ideal result. Among them, IDCG is the best recommendation result list returned by the recommendation system for a specific user.

[0327]

[0328] (1.3) Hyperparameter setting

[0329] The number of potential intent K values ​​is set in the range of {32, 64, 128, 256}, and the default K is 128. In addition, the hyperparameters α and β affect the hardness of the selected negative samples. For this reason, the values ​​of α and β are set in the range of {0.1, 0.2, ..., 0.9, 1.0} and {0.1, 0.2, 0.3, 0.4, 0.5} respectively to explore their impact on model performance. In order to ensure a baseline performance with fair settings, for all compared methods, all models are optimized using the Adam optimizer, and the potential embedding dimension d and batch size are set to 32 and 10240; for graph-based models, the number of propagation layers L is selected from {1, 2, 3, 4}.

[0330] (1.4) Baseline approach

[0331] A total of 12 methods were comprehensively compared in Recall@K and NDCG@K based on the following four groups of baselines, as shown below.

[0332] (I) GCN-based models

[0333] NGCF: This approach proposes a neural graph collaborative filtering framework, which significantly improves the learning of user and item representations by propagating embeddings on the user-item graph structure, thereby improving the performance of recommendation systems.

[0334] LightGCN: This approach simplifies the design of GCN, retaining only the core component of neighborhood aggregation, and implements a lightweight and high-performance GCN model.

[0335] SGCF: This is a linear model that calculates the network embedding on the compressed graph based on the removal of nonlinearity to obtain the embedding initialization of each node in the graph, alleviating the over-smoothing problem when aggregating sparse interaction data.

[0336] (II) Collaborative filtering model based on graph filtering

[0337] Turbo-CF: This method constructs an item-item similarity graph and a polynomial low-pass filter. This design effectively avoids the expensive matrix decomposition problem, fully utilizes computer components, and achieves extremely fast and accurate recommendations.

[0338] (III) Based on contrastive learning model

[0339] HCCF: This method proposes a new self-supervised recommendation framework that jointly captures the local and global collaborative relationships of users and enhances the discriminability and quality of node representation through a hypergraph-enhanced contrastive learning structure.

[0340] LightGCL: This is a lightweight graph contrastive learning norm that uses singular value decomposition and reconstruction to achieve graph data enhancement, improving the problems of existing methods in semantic structure preservation and noise resistance.

[0341] KPA-GCL: This method uses a relational graph convolutional neural network to optimize the item-item interaction representation and introduces a variety of adaptive contrast loss optimization functions to alleviate the cold start and noise problems in knowledge graph recommendation.

[0342] (IV) Model based on multi-intention disentanglement

[0343] DGCF: This method generates an intent-aware graph by separating user intent from user-item relations and modeling them, thereby generating a more informative representation.

[0344] DGCL: This method proposes a factorized discriminative objective to learn disentangled representations, uses it to learn nodes of disentangled representations, and uses inner products for user-item interaction prediction.

[0345] DisenHAN: This method can effectively exploit high-order connections between meta-relation decomposition nodes by learning disentangled user / item representations in heterogeneous information networks.

[0346] DisenGCN: This method proposes a novel neighborhood routing mechanism to dynamically identify latent factors and learn disentangled node representations.

[0347] DCCF: This self-supervised model framework proposes a disentangled contrastive collaborative filtering framework, which achieves intent disentanglement and alleviates noise interference through self-supervised enhancement in an adaptive manner.

[0348] Table 2 Experimental results of AdaDCL and 12 baselines on 3 datasets

[0349]

[0350]

[0351] (2) Performance comparison (RQ1)

[0352] The proposed AdaDCL method is validated and logically compared with twelve different baselines on three datasets. To ensure that the experimental results are statistically significant, AdaDCL and the best performing baselines are retrained five times. Table 2 shows the results in detail and summarizes the following observations:

[0353] ·AdaDCL consistently outperforms all baselines on all datasets. AdaDCL uses adaptive disentangled contrastive learning to fine-tune interactive motivations, achieves a more sophisticated embedding representation of users and items, and improves the accuracy and robustness of recommendations. This significant performance gain is attributed to the following key points of the AdaDCL method: (1) By extracting disentangled self-supervisory signals as supplementary training tasks, four subgraphs are constructed based on global and local learnable enhancements, which effectively alleviates the problem of data sparsity. (2) We propose the use of a parameterized graph mask generator that can adaptively implement self-supervision of data noise redundancy, thereby further improving the robustness of the representation. (3) The two-layer negative sampling strategy used achieves the adaptive selection of negative samples with appropriate hardness for positive samples, effectively alleviating the current problems of false positives and false negatives, and further improving the performance of the implicit collaborative filtering model.

[0354] Although some current methods based on self-supervised learning signals (e.g., CL4SRec, AutoCF, HCCF) have proposed data augmentation techniques, they ignore the fact that user and item interactions are driven by many potential intentions. It is difficult to finely model users through such coarse-grained learning enhancement representations, which will lead to suboptimal recommendations. For this reason, the AdaDCL recommendation method is far superior to them. In addition, a large number of studies have found that most models based on self-supervised learning signals perform better than pure graph neural network-based models (e.g., LightGCN, HMLET). AdaDCL still promotes this progress through adaptive disentanglement.

[0355] Compared with other disentangled recommender systems (e.g., DGCF, DisenGCN, FDGNN, SDHID), AdaDCL verifies that the AdaDCL method is not limited by the label shortage problem. The combination of intention disentanglement and contrastive learning leads to a significant performance improvement. Most existing disentanglement learning methods often have difficulty in generating information embeddings due to overfitting effects when training labels are insufficient. For example, although DGCL attempts to use contrastive learning to encode latent factors into enhanced representations, its non-adaptive contrast view generation makes it susceptible to noise perturbations.

[0356] Table 3 Ablation analysis experiment (evaluation indicators are Recall@20 and NDCG@20)

[0357]

[0358]

[0359] (3) Ablation experiment (RQ2)

[0360] Through ablation comparative analysis experiments, we explore the role of different components in improving performance on various data sets and verify their effectiveness. Specifically, we examine the advantages of specific components of the AdaGCL method from three aspects: (i) disentangled multi-intention encoding (DME), (ii) self-supervised learning (SSL), and (iii) adaptive hardness sampling (AHNS). Table 3 shows the experimental results of the performance, and the detailed experimental steps are as follows:

[0361] Disentangled Multi-Intent Encoding (DME): In order to verify that the introduced multi-intent encoding module can improve the user-item information representation, the disentangled multi-intent encoding module (-Disen) is removed, and the importance of disentangled multi-intent encoding is verified by comparing the performance of the models before and after removal.

[0362] Self-supervised learning (SSL): We explore the impact of disentangled contrastive learning on the performance of the proposed disentangled model by introducing a self-supervised optimization objective. We create two variants by removing the original graph with auxiliary augmented views: (i) disentangled global collaborative subgraph (-GloG) and (ii) all views with adaptive augmentation (-AllG). Experiments show that AdaGCL has the best performance, which emphasizes the contribution of adaptive data augmentation of the global view of intent-aware collaborative relations by integrating auxiliary self-supervised learning signals.

[0363] Adaptive hard negative sampling: By removing the adaptive hard negative sampling (-AHNS), it is replaced by the traditional hard negative sampling method. Experimental results further show that these variants will lead to a decrease in model performance, and AdaDCL achieves the best performance, which fully proves the importance of understanding the modules of entangled multi-intention coding, self-supervised learning, and adaptive hard negative sampling.

[0364] (4) Further analysis of the model (RQ 3 and RQ 4)

[0365] (4.1) Impact of hyperparameters

[0366] In order to explore the impact of the number of intents K, α, and β on experimental performance, experiments were conducted under different experimental configurations. The larger the K value, the finer the granularity of the user intent, so its value range is set to {32, 64, 128, 256}. α and β are used to adjust the hardness of negative samples, so the values ​​of α and β are set to {0.2, 0.4, 0.6, 0.8, 1.0} and {0.1, 0.2, 0.3, 0.4, 0.5} respectively, and the model is retrained under different experimental configurations. The experimental results of the impact of hyperparameters on model performance are shown in Figure 2. Figure 2 to Figure 4 As shown. Figure 2 It can be found that as the number of intents gradually increases, the expressiveness of the model also increases. However, when the number of intents increases from 128 to 256, the performance increases slowly or even decreases. This is because the granularity of the latent factors is too fine, resulting in intent redundancy and introducing noise in the learned representation. Figure 3 By setting the value of α to 1.0 and exploring the effect of parameter β on model performance, we found that although a β value of 0.2 showed good performance on the Gowalla dataset, overall, a β value of 0.3 achieved good results on all three datasets. After that, as the β value increased, the performance of the model remained stable or showed a downward trend, indicating that the optimal β value was 0.3. Figure 4By setting the value of β to 0.1 and exploring the impact of parameter α on model performance, it can be found that when the value of α increases to 0.6, the model can achieve the best performance. As the value increases, the model performance always remains in a relatively stable state. Therefore, comprehensive analysis shows that when the hyperparameters K = 128, α = 0.6, and β = 0.3, ideal experimental performance can be obtained.

[0367] (4.2) Sparse expressiveness

[0368] Data sparsity is a common problem in the real world, so alleviating sparsity has become an important indicator for evaluating recommendation systems. To further verify that the proposed method can effectively combat data sparsity, AdaGCL is compared with two representative baselines, LightGCN and SGL, on the Amazon-book and Gowalla datasets. The users are divided into five groups according to the interaction frequency between the users and the items, namely [0, 15), [15, 20), [20, 25), [25, 30), [30, 35). The recommendation accuracy is measured based on Recall@20 and NDCG@20. The sparsity experimental analysis results of AdaGCL, LightGCN and SGL are shown in the figure. Figure 5 As shown in the figure, it can be found that AdaGCL is always better than LightGCN and SGL, which shows that the proposed method has certain advantages over self-supervised learning with stronger performance, which fully verifies the effectiveness of adaptive data enhancement methods for learning user preferences with fewer interactive labels. In addition, the experimental results also fully show that SGL is prone to losing important information of inactive users by randomly discarding central nodes or edges, which aggravates the sparseness of data.

[0369] (4.3) Oversmoothing robustness

[0370] In order to verify the effectiveness of the designed layer attention module in dealing with the over-smoothing problem and stabilize the training process of deep graph neural networks (GNNs), a series of experiments were conducted. Specifically, the average distance (MAD) of user / item embedding encodings of AdaDCL and its variants DdaDCL-AT (with the attention mechanism removed) and DdaDCL-CL (with the contrast view module removed)

[58] was calculated to verify the model's ability to combat over-smoothing; at the same time, the average distance of AdaDCL was compared with some baseline methods, including deep graph contrastive learning (DGCL), disentangled graph convolutional network (DisenGCN) and lightweight graph convolutional network (LightGCN). To ensure the fairness of the experiment, all embedding representations have been normalized before calculating MAD. The experimental results of the MAD values ​​of AdaDCL and its variants and the three baselines are shown in Figure 2. Figure 6As shown in the figure. The size of the MAD value reflects the severity of the over-smoothing phenomenon: the smaller the value, the more obvious the over-smoothing phenomenon. From the experimental results, it can be seen that among all the comparison methods, AdaDCL shows the highest MAD value, and after removing the contrastive learning and attention modules, the MAD value drops significantly. This finding strongly proves the effectiveness of the AdaDCL recommendation method in alleviating the over-smoothing problem when stacking deep graph neural networks.

[0371] (5) Negative Sampling Extension

[0372] The negative sampling method used can simultaneously meet the following three key criteria, which is different from most adopted hard negative sampling: first, negative samples of appropriate hardness are identified according to the corresponding positive samples, which is different from the traditional ones; second, the hardness of the selected negative samples should be negatively correlated with the prediction score of their corresponding positive samples: for positive samples with higher prediction scores, items with lower hardness should be selected as negative samples, which can effectively alleviate the false negative problem. Conversely, for positive samples with lower prediction scores, items with higher hardness should be selected as negative samples, which can accelerate the optimization of positive samples and make negative samples with higher hardness fully updated, thereby alleviating the false positive problem. Finally, in order to cover various practical recommendation scenarios, such as different datasets or evaluation metrics, the hardness of the selected negative samples should be adjustable. For this reason, the existing negative sampling method can be regarded as a more relaxed case of the proposed negative sampling.

[0373] In summary, this application proposes an adaptive disentangled contrastive recommendation method for user recommendation, which explores the potential factors in the user-item interaction graph. The graph structure learning layer introduced by AdaDCL can adaptively enhance the interaction between nodes according to the learned intent-aware dependencies; based on this, a contrastive learning method is further proposed, which uses an extended intent-aware graph structure to effectively remove the entangled information in the self-supervisory signal, and generates four contrast-enhanced views from multiple different angles based on the original interaction view. In the model training stage, different from the traditional hard negative sampling method, a two-time sampling strategy is used to achieve adaptive selection of negative sample hardness, which alleviates the false negative and false positive problems commonly found in existing sampling.

[0374] We conducted multiple comparative analyses and ablation experiments on three public datasets, Gowalla, Amazon-book, and Tmall, and 12 current mainstream recommendation methods, and used the commonly used evaluation indicators Recall@K and NDCG@K for recommendation systems to verify the effectiveness and accuracy of the model. Compared with the latest DCCF recommendation method on the Amazon-book dataset, the proposed AdaDCL improved the Recall@20 indicator by 1.5% and the NDCG@20 indicator by 1.9%, which fully demonstrated the superiority of AdaDCL.

[0375] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. An adaptive disentanglement comparison recommendation method, characterized in that: The following steps are involved: S1, perform neighborhood aggregation on the user-item interaction graph: aggregate the neighbor node information of the user and item center nodes respectively to obtain the embedded representation of the user and the embedded representation of the item; S2, perform intention disentanglement on the latent vector space of users and items: let users and items have K1 and K2 different intentions respectively, and use latent intentions c u and c v Representation; then the potential intentions of users and items are disentangled respectively, and the disentangled intentions of users are The intention of the project after untangling The multi-intention encoder is used to encode the and middle; S3, masked subgraph generation: masked subgraph generated by cosine similarity between user-item interaction matrix and intent vector Get the relationship matrix S4, generating local collaborative views and global collaborative views with adaptive enhancement; S5, use the attention mechanism to assign weights to the enhanced representation after aggregation to obtain the final user and item representation E (u) and E (v) , then for the E (u) and E (v) Perform dot product to get the final prediction score Y; In the formula, Represent the embedded representation of users and the embedded representation of items; Respectively represent the local-level and global-level enhanced representations of users; Respectively, the enhanced representations of the local and global levels of the project; The intention representation embeddings of the user and item are represented respectively, i.e., the disentangled intentions of the user and item; Among them, E (u) 、E (v) Represent the final user and item embedding vectors respectively; Attweight is the weight distribution function learned by the attention mechanism; δ (l) Represents the attention weight assigned to the lth layer; Represent the embedding representation vectors of users and items in the l-th layer of graph neural network respectively; The preference scores of users and items are calculated as follows: in, Represents user u i and project v j Embedding representation of ; T is the transposition symbol; Y i,j represents the interaction score between user i and item j; Y represents the final interaction score between user i and item j; E (u) represents the final embedding representation of user i and item j.

2. The adaptive disentanglement comparison recommendation method according to claim 1, characterized in that: The neighborhood aggregation of the user-item interaction graph includes: Where σ(·) represents the LeakyReLU activation function; Represents the central node u i and v j The embedded representation obtained by aggregating the neighbor nodes of Represent the user-item interaction matrix The i-th row and interaction matrix The jth column in ; E (u) 、E (v) Represent the initialization embedding representation of users and items respectively; the initialization results obtained by using the xavier function on users and items.

3. The adaptive disentanglement comparison recommendation method according to claim 2, characterized in that: is the normalized adjacency matrix derived from the user-item interaction matrix: D in the formula (u) and D (v) Represents the angle matrices of users and items respectively; represents the user-item adjacency matrix; and Represents user u i and project v j The number of neighbor nodes set; || indicates normalization operation; Represents the element in the i-th row and j-th column of the adjacency matrix of the user-item interaction graph.

4. The adaptive disentanglement comparison recommendation method according to claim 1, characterized in that: Mask subgraphs generated using user-item interaction matrix and cosine similarity using intent vectors Get the relationship matrix include: in, represents the user-item interaction matrix; Represents a relationship matrix; Represents the mask sub-image, and the solution formula is as follows: in, Represents the mask sub-graph For each value in ; || ||2 represents the two-norm; for The transpose of denote the disentangled representations of users and items, respectively.

5. The adaptive disentanglement comparison recommendation method according to claim 4, characterized in that: In the solution process, the relationship matrix Using learnable data augmentation: Use the node degree pair relationship matrix Perform normalization operation, expressed as: in, Represents the relationship matrix after normalization operation The element value of the i-th row and j-th column; Representing the relationship matrix The element at row i and column j; U represents the size of the user; I represents the size of the project; T is the transpose symbol.

6. The adaptive disentanglement comparison recommendation method according to claim 1, characterized in that: S4, generates local collaborative views and global collaborative views with adaptive enhancement, including: First, the normalized learning relationship matrix Applied to node messages to achieve learnable message propagation, the graph structure is perturbed to generate contrastive learning views with adaptive enhancement. This enhancement with adaptive mask is expressed as: in, They respectively represent the graph perturbation operation on the representation embedding vector of layer l; Represent the representation embedding vectors of users and items obtained by traversing the graph neural network at layer l; represents the relationship matrix obtained after normalization operation, In formula (14) The following formula (6) is used to obtain the enhanced representation of the local level of users and items: in, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l; The representation embedding vector obtained by passing the aggregated information of the user and its adjacent target nodes through the l-1 layer graph neural network; The representation embedding vector obtained by passing the aggregated information of the project and its adjacent target nodes through the l-1 layer graph neural network; In formula (14) The following formula (9) is used to obtain the global level enhanced representation of users and items: In the formula, Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l; They represent the first l-1 layers from the central node u i and v j The embedded representation obtained by aggregating the neighbor nodes of and They represent the users after the first l-1 layers of the graph neural network are stacked. and Projects Intent-aware representation, They respectively represent the disentangled representations of users and items obtained after the l-1 layer of graph neural network.

7. The adaptive disentanglement comparison recommendation method according to claim 6, characterized in that: Obtained by the following formula: in, The disentangled representations of users and items respectively; Indicates that in the lth layer of the graph neural network, after passing through the encoder, in the intention c u Below, the user’s representation embedding vector; It means that when the l-th layer of the graph neural network is passed through the encoder, the intention c v Below, the representation embedding vector of the item; c u 、c v Represent the user and intention representation vector respectively; They represent the corresponding i-th user and intention representation vector respectively; and Represents the representation vector of users and items at level l; Represents user u i and each intent prototype c u relevance; Indicates that the user v j and each intent prototype c v of relevance.

8. The adaptive disentanglement comparison recommendation method according to claim 7, characterized in that: Correlation k Use intent relevance score Calculate and use formula (8) to deduce: in, Represents the embedding representation vector of users and items at the lth layer; for The transpose of represents the embedded representation vector of the user at the l-1 layer; They all represent the representation vectors of the user's i-th and i′-th intentions, where i′ represents the index, which means it includes the traversal operation; for The transpose of Represents the embedding representation vector of the item at level l-1; They represent the representation vectors of the i-th and i′-th intentions of the item respectively, where i′ represents the index, which means that the traversal operation is included.

9. The adaptive disentanglement comparison recommendation method according to claim 1, characterized in that: The weight is obtained by the following formula: In the formula, δ (l) Represents the attention weight assigned to the lth layer; L represents the number of layers of the graph convolutional network; ⊙ is the dot product operation; E (l) 、E (l′) Both represent the lth layer and the lth layer. ′ Feature vectors of layer graph neural networks; W Att are the parameters used for attention calculation.

10. The adaptive disentanglement comparison recommendation method according to claim 1, characterized in that: The negative samples used in training are obtained through the rating function: Among them, r m Represents candidate negative sample i m Rating value of || is the absolute value symbol; represents the transpose of the embedding representation of user u; Represent the embedding representation of candidate negative samples; Represents the embedding representation of the positive sample; α, β, p are predefined hyperparameters; After calculating the ratings of all candidate negative samples, we get a rating set exist Select r m The candidate negative sample i with the smallest value m as the final negative sample.

Images