Learable enhanced comparison recommendation method capable of adaptively fusing multiple views

Abstract

The invention discloses a learning enhancement comparison recommendation method capable of adaptively fusing multiple views, and belongs to the technical field of recommendation systems. The method comprises the following steps: constructing a high-order collaborative neighbor graph based on user-project interaction data, and dynamically fusing multi-source information by using a self-adaptive non-parametric fusion strategy; adopting a multi-head attention network to adaptively generate a node-level enhanced view; and performing multi-view comparative learning on the original view, the collaborative view and the enhanced view, and jointly optimizing the comparative loss and the recommended loss. Experimental results show that the method can effectively improve the performance of a recommendation system in sparse and noise environments, and has high recommendation accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a learnable enhanced contrast recommendation method for adaptively fusing multiple views. Background Art

[0002] Existing recommendation systems primarily rely on collaborative filtering algorithms, modeling user preferences through user-item interaction history. In recent years, graph neural network (GNN) technology has been introduced into recommendation systems to capture high-order connections between users and items, effectively improving recommendation performance. However, traditional graph neural network models still face challenges such as data sparsity and noise interference. Graph contrastive learning, as a self-supervised learning method, enhances representation capabilities by comparing multiple views, improving model robustness. However, most existing solutions fail to fully utilize high-order information and easily destroy the original graph structure during the view enhancement process, resulting in limited recommendation accuracy. Therefore, there is an urgent need to propose a recommendation method that can dynamically integrate high-order information from multiple views and has adaptive noise suppression capabilities. Summary of the Invention

[0003] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a learnable enhanced contrast recommendation method for adaptive multi-view fusion, which is characterized by comprising:

[0004] S1, build a multi-view graph structure:

[0005] Based on user-item interaction data, a user-item interaction graph and a collaborative neighbor graph are constructed. The collaborative neighbor graph includes a collaborative social graph and a collaborative knowledge graph. Collaborative edges are generated by counting the high-order similarities in the interaction history between users and items, and collaborative high-order neighbor relationships are filtered according to thresholds and edge number limits.

[0006] S2, Multi-graph Embedding and Information Propagation:

[0007] Based on the user-item interaction graph and the collaborative neighborhood graph, a graph neural network is used to perform multi-layer information propagation to obtain the embedded representation of each node in different views;

[0008] S3, adaptive multi-view fusion:

[0009] For each layer of node embedding, a non-parametric fusion strategy is used to dynamically adjust the contribution of different view embeddings based on inter-layer attenuation, node activity, and redundancy similarity between views to obtain the fused node representation.

[0010] S4, learnable enhanced embedding generation:

[0011] Using a multi-head graph attention network, we perform adaptive attention weighted aggregation on nodes, generate learnable enhanced embedding representations, and aggregate the results of different heads;

[0012] S5, Joint contrastive learning and recommendation task optimization:

[0013] Multi-view comparative learning is performed on the original view, collaborative neighbor views, and enhanced views. The fused view is used as an anchor to align with other views. The recommendation loss and contrast loss are jointly optimized to improve the discriminability of node representation and model robustness.

[0014] S6, final recommendation and output:

[0015] For target users and candidate items, preference scores are calculated based on the final fusion embedding using methods such as inner product to output personalized recommendation results.

[0016] In a preferred embodiment of the present invention, constructing a multi-view graph structure includes:

[0017] definition Represents a collection of users, Represents a collection of items;

[0018] Where M and J represent the number of users and projects respectively;

[0019] The observed implicit feedback matrix is ​​expressed as when When , it means that user u interacts with item i; otherwise use Represents a user-item interaction graph;

[0020] represents the collaborative neighborhood graph, which contains the collaborative social graph and collaborative knowledge graphs It can be expressed as the following formula:

[0021]

[0022] in, is the interaction adjacency matrix, is the transpose of the interaction adjacency matrix, T represents the transpose of the matrix;

[0023] It is a collaborative social graph, which represents the interaction between users;

[0024] A collaborative knowledge graph that represents the interaction between projects;

[0025] In the recommendation system based on graph neural network, the user-item interaction graph is usually constructed as a bipartite graph;

[0026] Right now

[0027] in, It is a set of nodes, including project nodes and user nodes;

[0028] represents an edge set;

[0029] The initial embedding vector for each user u and item i is represents a d-dimensional vector space;

[0030] The embedding sets of users and items are represented as and Where d represents the embedding dimension;

[0031] Constructing feature graph for multi-head GAT based on user-item interaction graph in, represents the feature matrix; q u and q i are the feature vectors of users and items used by GAT respectively;

[0032] Therefore, the probability of interaction between user u and item i can be expressed as follows:

[0033]

[0034] Where, represents the probability of user u interacting with item i;

[0035] e u The final embedding representation of the user, e i The final embedding representation of the represented item;

[0036] The transpose of the final embedding representation representing the user;

[0037] The collaboration similarity is calculated using the Jaccard similarity coefficient based on the common interaction history of user i and user j as shown in the formula:

[0038]

[0039] Among them, sim i,j Jaccard similarity coefficient between node i and node j;

[0040] It's a picture The first-order neighbor set of node i;

[0041] It's a picture The first-order neighbor set of node j;

[0042] ∩ and ∪ represent the intersection and union of sets;

[0043] To simplify the calculation and remove redundant high-order information, the high-order information interaction diagram It can be defined as follows:

[0044]

[0045] in, is the adjacency relationship between node i and node j in the high-order information interaction graph;

[0046] η represents the collaboration similarity threshold between users and projects. The collaboration similarity of similar users should be greater than or equal to this threshold.

[0047] K represents the top K items recommended for the user, that is, the maximum number of high-order collaborative edges retained by each node.

[0048] In a preferred embodiment of the present invention, the multi-graph embedding and information propagation includes:

[0049] For each node i, at layer l, The embedding learned in Correspondingly, from The learned embeddings are available Represented as, then the user-item interaction graph is fused at layer l and collaboration diagrams The embedding representation of Embedded representation after fusion at layer l will be passed to the next graph neural network layer; thus, the information propagation process of AMLECR can be expressed as follows, in which the nonlinear activation and feature change process in the propagation function are omitted:

[0050]

[0051] in, and It's a picture and pictures The first-order neighbor set of node i;

[0052] It's a picture The first-order neighbor set of node j;

[0053] and In the figure The number of first-order neighbors of node i and node j;

[0054] Indicated in the figure The number of first-order neighbors of node i;

[0055] Is node i in layer l, based on the graph Embedding after propagation;

[0056] Is node i in layer l, based on the graph Embedding after propagation;

[0057] is the fused embedding representation of node j at layer l-1;

[0058] Indicates that In the summation, all nodes j directly connected to node i are summed;

[0059] Indicates that The sum of all nodes j directly connected to node i is calculated.

[0060] In a preferred embodiment of the present invention, the adaptive multi-view fusion includes:

[0061] To fully capture the user-item interaction graph and high-level collaboration infographic The relationship between the two is carefully designed. An adaptive non-parametric fusion strategy is and The information is used to get the final embedding representation of users and items To more accurately grasp the contribution of auxiliary user-user relationships;

[0062] The designed adaptive multi-view fusion strategy can be specifically expressed as:

[0063]

[0064] in, is the fused embedding representation;

[0065] It's a picture The first-order neighbor set of the middle node v;

[0066] and It's a picture The number of first-order neighbors of node i and node v;

[0067] β i is the high-order information weight of the i-th node, l is the number of network layers;

[0068] sim(*) represents the cosine similarity function; γ is used to control Hyperparameters of weights;

[0069] d irepresents the normalized user activity, where the logarithm is taken to mitigate the large differences in user activity;

[0070] is the total number of nodes.

[0071] In a preferred embodiment of the present invention, the learnable enhanced embedding generation includes:

[0072] Graph Attention Network, a neural architecture executed on graph data structures, can effectively stabilize the training process of the self-attention layer;

[0073] according to Generate adaptive attention correlation coefficients using multi-head GAT The learnable enhanced embedding representation of , where N represents the number of hidden attention heads;

[0074] Considering the self-attention mechanism of each head as a single-layer feedforward neural network, the adaptive attention coefficient between user u and item i in each head can be specifically expressed as:

[0075]

[0076] in, represents the adaptive attention weight coefficient of user u and item i under the nth attention head;

[0077] exp(*) represents the exponential function with the natural base e as the base, and LeakyReLU(*) is the LeakyReLu activation function;

[0078] W (n) is the weight matrix of the nth head;

[0079] q u and q i are the feature vectors of users and items used by GAT respectively;

[0080] q p is the feature vector of node p;

[0081] Represents the learnable weight vector of the nth head;

[0082] || represents cascade operation;

[0083] N u represents the set of neighbor nodes of user u;

[0084] p is the neighbor node index;

[0085] p∈N u ∪{u} means traversing all neighbor nodes including user u itself;

[0086] To simplify the calculation, when there is no interaction between user u and item i,

[0087] Finally, the adaptive attention coefficient calculated in each head is aggregated with the adjacent features by formula:

[0088]

[0089] Where, represents the enhanced representation of user u after passing the nth attention head;

[0090] ELU(*) represents the ELU activation function;

[0091] The enhanced representation of item i corresponding to the above user u, that is, the enhanced representation of item i after the nth attention head is Averaging operations across multiple heads:

[0092]

[0093] in, represents the enhanced representation of user u at layer l;

[0094] represents the enhanced representation of item i at layer l;

[0095] represents the enhanced representation of item i after passing through the nth attention head;

[0096] N represents the number of hidden attention heads;

[0097] l represents the number of network layers.

[0098] In a preferred embodiment of the present invention, the combined contrastive learning and recommendation task optimization includes:

[0099] After iterative propagation of a total of L layers of graph neural networks, the initial embedding vector and the embedding representation of all layers are averaged and pooled to obtain the final embedding representation:

[0100]

[0101] in, Represents the embedding vector of user u at layer l;

[0102] is the embedding vector of item i at layer l;

[0103] Based on the final embedding representation of user u and item i, the possibility of interaction between target user u and candidate item i is predicted by inner product:

[0104]

[0105] in, represents the probability of user u interacting with item i;

[0106] is the transpose of the final embedding vector of user u, e i is the final embedding vector of item i

[0107] L represents the total number of layers of iterative propagation of the neural network;

[0108] The Bayesian personalized ranking loss, a ranking objective function carefully designed for recommendation tasks, is used to directly capture interaction information. The goal is to encourage the model to predict the preference probability for items with which the user has interacted, which is much higher than the preference probability predicted for items with which the user has not interacted. It can be specifically expressed as:

[0109]

[0110] in, is the BPR recommendation loss;

[0111] represents the positive sample, represents negative samples;

[0112] u is the user index, i+ is the index of the positive sample of the item, and i- is the index of the negative sample of the item;

[0113] σ(*) is the Sigmoid activation function;

[0114] User-Item Interaction Diagram Collaborative Neighborhood Graph The final representation of the user or item is fused as shown below:

[0115]

[0116] in, Indicates that node v is in the interaction graph The final embedding vector in ;

[0117] Indicates that node v is in the collaborative neighbor graph The final embedding vector in ;

[0118] Indicates that node v is at layer l, based on the interaction graph Embedding vector after propagation;

[0119] Indicates that node v is at layer l, based on the collaborative neighbor graph Embedding vector after propagation;

[0120] The fusion view can be regarded as an enhancement of the user-item interaction graph and the collaborative neighborhood view, and contains richer information. Therefore, compared with the direct comparison of the interaction view and the collaborative neighborhood view, we choose to compare them with the fusion view separately as follows:

[0121]

[0122] is the contrast loss of the interaction graph view;

[0123] is the contrastive loss of the collaborative neighborhood graph view;

[0124] is the multi-view weighted combination contrast loss;

[0125] represents the transpose of the final embedding vector of node v;

[0126] and Respectively represent node j and node v in the collaborative neighbor graph The final embedding vector in ;

[0127] and Respectively represent node j and node v in the interaction graph The final embedding vector in ;

[0128] exp(*) represents the exponential function with the natural base e as the base;

[0129] is a sampling set;

[0130] τ is the smoothing coefficient, which is used to adjust the smoothness of the softmax distribution;

[0131] α is a weighting coefficient used to adjust the importance of the two view losses;

[0132] In addition, in order to avoid losing important original information, GAT is used to make up for the limitations of GCN and convert the original view e u With enhanced view h u The contrastive learning is performed as shown below, aiming to optimize the representation of downstream tasks. Thanks to the characteristics of self-supervised learning, the generalization ability of the model is also improved:

[0133]

[0134] is the user enhancement contrast loss, which is used to measure the original view e u With enhanced view hu similarity;

[0135] is the item enhancement contrast loss, which is used to measure the original view e i With enhanced view h i similarity;

[0136] is the total enhanced view contrast loss;

[0137] h u′ represents the embedding representation vector of other users except user u in the augmented view;

[0138] h i′ represents the embedding representation vector of other items except item i in the enhanced view;

[0139] u′ and i′ are used as negative users and negative items in the contrastive loss, respectively;

[0140] and are collections of users and projects respectively;

[0141] Finally, the contrast loss and recommendation loss are combined through joint learning, and the final loss calculation is as follows:

[0142]

[0143] in, is the total loss;

[0144] is the multi-view weighted combination contrast loss;

[0145] is the enhanced view contrast loss;

[0146] represents the Frobenius regularization term with parameter Θ;

[0147] λ1 and λ2 are designed to control the weight of the contrast loss, and λ3 is used to control the weight of the regularization term.

[0148] The present invention also discloses a computer system, comprising:

[0149] processor;

[0150] a memory for storing processor-executable instructions;

[0151] The processor is configured to implement the adaptive multi-view fusion learnable enhanced contrast recommendation method when executing the executable instructions.

[0152] The present invention also discloses a computer-readable storage medium, comprising:

[0153] a memory having a computer program stored thereon;

[0154] A processor is used to execute the program in the memory to implement the adaptive multi-view fusion learnable enhanced contrast recommendation method.

[0155] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0156] By introducing high-order neighborhood views such as collaborative social graphs and collaborative knowledge graphs, it is possible to explicitly model complex collaborative relationships between users (or projects). Compared with existing methods that only use user-project interaction information, it can capture richer semantic structure features and improve the expressiveness and generalization performance of the recommendation model.

[0157] The present invention adopts a fusion strategy based on node activity, inter-layer attenuation and inter-view redundant similarity, which can dynamically adjust the contribution ratio of different view information to node representation, effectively suppress the interference of irrelevant or redundant high-order neighbor information, and improve the robustness of the model and the recommendation accuracy.

[0158] By adaptively aggregating various neighborhood information through the Multi-head Graph Attention Network (Multi-head GAT), we can enhance the node embedding representation in a fine-grained manner, assign higher weights to important neighbors, improve the discriminability of node representation, and further enhance the recommendation performance.

[0159] This paper jointly optimizes multi-view contrastive learning and recommendation ranking tasks. By aligning the original view, collaborative neighbor view and enhanced view, the node representation has stronger discriminative ability, while improving the model's adaptability to noise and sparse data, ensuring the stability of the recommendation results.

[0160] The method of the present invention has a flexible structure and can be combined with a variety of graph neural network models. It is suitable for various large-scale recommendation scenarios. It can achieve excellent recommendation effects under different data sets and parameter settings, and has broad industrial and commercial application prospects.

[0161] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0162] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0163] Figure 1 It is a schematic flow chart of the method of the present invention.

[0164] Figure 2 Schematic diagram of the ablation experiment analysis of the present invention.

[0165] Figure 3 2 is a comparative schematic diagram of the AMLECR sparsity experimental analysis of the present invention.

[0166] Figure 4 It is a schematic diagram of the comparison of the AMLECR noise experimental analysis of the present invention.

[0167] Figure 5 2 is a schematic diagram of the hyperparameter impact analysis of the present invention. DETAILED DESCRIPTION

[0168] The following describes embodiments of the present invention in detail. Examples of the embodiments 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 intended only to explain the present invention and are not to be construed as limiting the present invention.

[0169] like Figure 1 As shown, the present invention discloses a learnable enhanced contrast recommendation method for adaptive fusion of multiple views, comprising:

[0170] 1. Preprocessing

[0171] Following the recommendation system convention, define Represents a collection of users, Represents a collection of items;

[0172] Where M and J represent the number of users and projects respectively;

[0173] The observed implicit feedback matrix is ​​expressed as when When , it means that user u interacts with item i; otherwise

[0174] use Represents a user-item interaction graph;

[0175] represents the collaborative neighborhood graph, which contains the collaborative social graph and collaborative knowledge graphs It can be expressed as the following formula:

[0176]

[0177] in, is the interaction adjacency matrix, is the transpose of the interaction adjacency matrix, T represents the transpose of the matrix;

[0178] It is a collaborative social graph, which represents the interaction between users;

[0179] A collaborative knowledge graph that represents the interaction between projects;

[0180] In the recommendation system based on graph neural network, the user-item interaction graph is usually constructed as a bipartite graph;

[0181] Right now

[0182] in, It is a set of nodes, including project nodes and user nodes;

[0183] represents an edge set;

[0184] The initial embedding vector for each user u and item i is represents a d-dimensional vector space;

[0185] The embedding sets of users and items are represented as and Where d represents the embedding dimension;

[0186] Constructing feature graph for multi-head GAT based on user-item interaction graph in, represents the feature matrix;

[0187] q u and q i are the feature vectors of users and items used by GAT respectively;

[0188] Therefore, the probability of interaction between user u and item i can be expressed as follows:

[0189]

[0190] Where, represents the probability of user u interacting with item i;

[0191] e u The final embedding representation of the user, e i The final embedding representation of the represented item;

[0192] Represents the transpose of the final embedding representation of the user.

[0193] 2. Collaborative Neighborhood Graph

[0194] Users with similar preferences are defined as collaboratively similar users, and items that have interacted with the same user are defined as collaboratively similar items. The original user-item interaction graph is used for explicit modeling. The high-level information of the project is used to build a collaborative relationship diagram (i.e., user-user diagram and project-project diagram). ), the collaborative similarity is calculated using the Jaccard similarity coefficient based on the common interaction history of user (item) i and user (item) j as shown in Eq.

[0195]

[0196] Among them, sim i,j Jaccard similarity coefficient between node i and node j;

[0197] It's a picture The first-order neighbor set of node i;

[0198] It's a picture The first-order neighbor set of node j;

[0199] ∩ and ∪ represent the intersection and union of sets;

[0200] To simplify the calculation and remove redundant high-order information, the high-order information interaction diagram It can be defined as follows:

[0201]

[0202] in, is the adjacency relationship between node i and node j in the high-order information interaction graph;

[0203] η represents the collaboration similarity threshold between users and projects. The collaboration similarity of similar users should be greater than or equal to this threshold.

[0204] K represents the top K items recommended for the user, that is, the maximum number of high-order collaborative edges retained by each node.

[0205] 3. Multi-image fusion embedding expression

[0206] We use the classic LightGCN as a graph data information extractor for the user-item interaction graph and collaboration neighborhood graph, explicitly encoding the user and item propagation process to achieve high-order connections.

[0207] 3.1 Information Dissemination

[0208] For each node i, at layer l, The embedding learned in Correspondingly, from The learned embeddings are available Represented as, then the user-item interaction graph is fused at layer l and collaboration diagrams The embedding representation of Embedded representation after fusion at layer l will be passed to the next graph neural network layer; thus, the information propagation process of AMLECR can be expressed as follows, in which the nonlinear activation and feature change process in the propagation function are omitted:

[0209]

[0210] in, and It's a picture and pictures The first-order neighbor set of node i;

[0211] It's a picture The first-order neighbor set of node j;

[0212] and In the figure The number of first-order neighbors of node i and node j;

[0213] Indicated in the figure The number of first-order neighbors of node i;

[0214] Is node i in layer l, based on the graph Embedding after propagation;

[0215] Is node i in layer l, based on the graph Embedding after propagation;

[0216] is the fused embedding representation of node j at layer l-1;

[0217] Indicates that In the summation, all nodes j directly connected to node i are summed;

[0218] Indicates that The sum of all nodes j directly connected to node i is calculated.

[0219] 3.2 Adaptive High-Order Information Fusion

[0220] To fully capture the user-item interaction graph and high-level collaboration infographic The relationship between the two is carefully designed. An adaptive non-parametric fusion strategy is and The information of the user (item) is used to obtain the final embedding representation In order to more accurately grasp the contribution of auxiliary user-user relationships, the designed adaptive non-parametric fusion strategy takes into full consideration the following three key aspects:

[0221] Inter-layer attenuation constraints: In the collaborative neighbor graph As the number of layers in the graph neural network increases, the rapid expansion of the neighborhood receptive field may make the added high-order information useless, or even hinder the model performance. Therefore, the weight of the number of layers in the graph convolutional network is It should be negatively correlated with the number of layers to effectively eliminate useless high-order information.

[0222] Activity perception mechanism: In the user-item original interaction graph If the user is highly active, The information provided by the first-order neighbors in is sufficient, and usually there is no need to consider higher-order information to avoid information overload of highly active nodes.

[0223] Redundancy suppression: If the representation in the user-item interaction graph Representation in the collaborative infographic With a high similarity, it is determined that the high-order information is redundant at this time. Adding redundant information may reduce performance, so there is no need to incorporate high-order information.

[0224] Based on the above conditions, the designed adaptive multi-view fusion strategy can be specifically expressed as:

[0225]

[0226] in, is the fused embedding representation;

[0227] It's a picture The first-order neighbor set of the middle node v;

[0228] and It's a picture The number of first-order neighbors of node i and node v;

[0229] β i is the high-order information weight of the i-th node, l is the number of network layers;

[0230] sim(*) represents the cosine similarity function; γ is used to control Hyperparameters of weights;

[0231] d i represents the normalized user activity, where the logarithm is taken to mitigate the large differences in user activity;

[0232] is the total number of nodes.

[0233] 3.3 Data Augmentation

[0234] Graph Attention Network, a neural architecture executed on graph data structures, can effectively stabilize the training process of the self-attention layer;

[0235] according to Generate adaptive attention correlation coefficients using multi-head GAT The learnable enhanced embedding representation of , where N represents the number of hidden attention heads;

[0236] Considering the self-attention mechanism of each head as a single-layer feedforward neural network, the adaptive attention coefficient between user u and item i in each head can be specifically expressed as:

[0237]

[0238] in, represents the adaptive attention weight coefficient of user u and item i under the nth attention head;

[0239] exp(*) represents the exponential function with the natural base e as the base, and LeakyReLU(*) is the LeakyReLu activation function;

[0240] W (n) is the weight matrix of the nth head;

[0241] q u and q i are the feature vectors of users and items used by GAT respectively;

[0242] q p is the feature vector of node p;

[0243] Represents the learnable weight vector of the nth head;

[0244] || represents cascade operation;

[0245] N u represents the set of neighbor nodes of user u;

[0246] p is the neighbor node index;

[0247] p∈N u ∪{u} means traversing all neighbor nodes including user u itself;

[0248] To simplify the calculation, when there is no interaction between user u and item i,

[0249] Finally, the adaptive attention coefficient calculated in each head is aggregated with the adjacent features by formula:

[0250]

[0251] Where, represents the enhanced representation of user u after passing the nth attention head;

[0252] ELU(*) represents the ELU activation function;

[0253] The enhanced representation of item i corresponding to the above user u, that is, the enhanced representation of item i after the nth attention head is Averaging operations across multiple heads:

[0254]

[0255] in, represents the enhanced representation of user u at layer l;

[0256] represents the enhanced representation of item i at layer l;

[0257] represents the enhanced representation of item i after passing through the nth attention head;

[0258] N represents the number of hidden attention heads;

[0259] l represents the number of network layers.

[0260] 3.4 Layer Combination and Model Prediction

[0261] After iterative propagation of a total of L layers of graph neural networks, the initial embedding vector and the embedding representation of all layers are averaged and pooled to obtain the final embedding representation:

[0262]

[0263] in, Represents the embedding vector of user u at layer l;

[0264] is the embedding vector of item i at layer l;

[0265] Based on the final embedding representation of user u and item i, the possibility of interaction between target user u and candidate item i is predicted by inner product:

[0266]

[0267] in, represents the probability of user u interacting with item i;

[0268] is the transpose of the final embedding vector of user u, e i is the final embedding vector of item i

[0269] L represents the total number of layers of the neural network iterative propagation.

[0270] The Bayesian Personalized Ranking Loss (BPR) is a ranking objective function designed for recommendation tasks to directly capture interaction information. The goal is to encourage the model to predict the preference probability for items with which the user has interacted (positive samples) much higher than the preference probability for items with which the user has not interacted (negative samples). It can be specifically expressed as:

[0271]

[0272] in, is the BPR recommendation loss;

[0273] represents the positive sample, represents negative samples;

[0274] u is the user index, i + The index of the positive sample of the item, i - is the index of the negative sample of the item;

[0275] σ(*) is the Sigmoid activation function.

[0276] 4. Multi-view fusion contrast learning

[0277] User-Item Interaction Diagram Collaborative Neighborhood Graph The final representation of the user or item is fused as shown below:

[0278]

[0279] in, Indicates that node v is in the interaction graph The final embedding vector in ;

[0280] Indicates that node v is in the collaborative neighbor graph The final embedding vector in ;

[0281] Indicates that node v is at layer l, based on the interaction graph Embedding vector after propagation;

[0282] Indicates that node v is at layer l, based on the collaborative neighbor graph Embedding vector after propagation;

[0283] The fusion view can be regarded as an enhancement of the user-item interaction graph and the collaborative neighborhood view, and contains richer information. Therefore, compared with the direct comparison of the interaction view and the collaborative neighborhood view, we choose to compare them with the fusion view separately as follows:

[0284]

[0285] is the contrast loss of the interaction graph view;

[0286] is the contrastive loss of the collaborative neighborhood graph view;

[0287] is the multi-view weighted combination contrast loss;

[0288] represents the transpose of the final embedding vector of node v;

[0289] and Respectively represent node j and node v in the collaborative neighbor graph The final embedding vector in ;

[0290] and Respectively represent node j and node v in the interaction graph The final embedding vector in ;

[0291] exp(*) represents the exponential function with the natural base e as the base;

[0292] is a sampling set;

[0293] τ is the smoothing coefficient, which is used to adjust the smoothness of the softmax distribution;

[0294] α is a weighting coefficient used to adjust the importance of the two view losses;

[0295] In addition, in order to avoid losing important original information, GAT is used to make up for the limitations of GCN and convert the original view e u With enhanced view h u The contrastive learning is performed as shown below, aiming to optimize the representation of downstream tasks. Thanks to the characteristics of self-supervised learning, the generalization ability of the model is also improved:

[0296]

[0297] is the user enhancement contrast loss, which is used to measure the original view e u With enhanced view h usimilarity;

[0298] is the item enhancement contrast loss, which is used to measure the original view e i With enhanced view h i similarity;

[0299] is the total enhanced view contrast loss;

[0300] h u′ represents the embedding representation vector of other users except user u in the augmented view;

[0301] h i′ represents the embedding representation vector of other items except item i in the enhanced view;

[0302] u′ and i′ are used as negative users and negative items in the contrastive loss, respectively;

[0303] and are collections of users and projects respectively;

[0304] Finally, the contrast loss and recommendation loss are combined through joint learning, and the final loss calculation is as follows:

[0305]

[0306] in, is the total loss;

[0307] is the multi-view weighted combination contrast loss;

[0308] is the enhanced view contrast loss;

[0309] represents the Frobenius regularization term with parameter Θ;

[0310] λ1 and λ2 are designed to control the weight of the contrast loss, and λ3 is used to control the weight of the regularization term.

[0311] 4.1 Training Process

[0312] The AMLECR training process is shown in ALGORITHM 1:

[0313]

[0314]

[0315] 5. Experiment

[0316] Table 1 shows a comparative analysis of the model with eight baselines on four public datasets: Yelp, Amazon-Books, Gowalla, and Alibaba-iFashion. The following questions are answered to fully verify the superiority and feasibility of the model:

[0317] (1) RQ1: How does the AMLECR model perform compared to some current advanced models?

[0318] (2) RQ2: Do the proposed key components have a positive effect on improving model performance?

[0319] (3) RQ3: How robust is the proposed self-supervised learning method compared with other methods?

[0320] (4) RQ4: To what extent will different hyperparameter settings affect the model?

[0321] Table 1 Experimental results

[0322]

[0323]

[0324] 5.1 Dataset

[0325] To evaluate the performance of the AMLECR model, we selected four representative public datasets: Yelp, Amazon-Books, Gowalla, and Alibaba-iFashion. Yelp: A real-world commercial dataset provided by Yelp, containing basic information such as user ID, registration time, and nickname; Amazon-Books: Derived from the Amazon book sales platform, it is widely used in various fields, including recommendation systems, user behavior analysis, and book market research; Gowalla: Records user check-ins at various locations and is often used to study location-based social behavior and recommendations; Alibaba-iFashion: Contains a large amount of product information and user-product interaction data, providing rich data support for recommendation algorithm research. These datasets vary in scope, scale, and density. Table 2 provides detailed information on the relevant datasets.

[0326] Table 2. Detailed information of the dataset

[0327]

[0328] 5.2 Evaluation Metrics

[0329] The top-p recommendation performance of the AMLECR method is evaluated using two key evaluation metrics commonly used in recommendation systems: recall (Recall@P) and normalized discounted cumulative gain (NDCG@P). To be consistent with most baselines, the values ​​of p are set to 10, 20, and 50. Unless otherwise specified, p defaults to 10. The specific calculation methods for recall (Recall) and normalized discounted cumulative gain (NDCG) are as follows:

[0330] 1) Recall: used to evaluate the true items in the test set

[0331] Whether the item exists in the first p lists retrieved.

[0332] Where R(u) represents the p items recommended for user u, and T(u) represents the set of items that user u likes on the test set;

[0333]

[0334] Recall is the recall rate indicator;

[0335] U is the complete set of users in the experiment;

[0336] T(U) is the set of all items liked by all users in the test set;

[0337] u∈U means traversing the entire set of users;

[0338] 2) CG (Cumulative Gain): The gains of all items i in the list recommended to the user are added up. The result represents the total number of items that the user likes in the list, which only reflects the quantitative meaning.

[0339]

[0340] CG P is the cumulative gain, which represents the total relevance score of the top K recommended items;

[0341] rel i is the relevance of the i-th recommended item;

[0342] 3) DCG (Discounted Cumulative Gain): Taking into account the position factor, for each rel i A depreciation value has been added, which is closely related to the position of the item in the list. The earlier the item is in the list, the smaller the corresponding depreciation value; conversely, the later the position, the greater the depreciation value.

[0343]

[0344] DCG p represents the discounted cumulative gain, taking into account the impact of the recommended item ranking on relevance;

[0345] log2(i+1) is the loss coefficient;

[0346] 4) NDCG (Normalized Discounted Cumulative Gain): This evaluates the position of real items in the top p lists, taking into account both the item's relevance and its ranking within the list. NDCG is a normalized version of DCG using IDCG, measuring how close the current result is to the ideal result. IDCG is a special case of DCG.

[0347]

[0348] Among them, IDCG p is the ideal discounted cumulative gain, which represents the maximum gain under the ideal ranking of the first p items;

[0349] REL represents the set of truly relevant items, and |REL| represents the number of relevant items;

[0350] NDCG p is the normalized discounted cumulative gain;

[0351] 5.3 Hyperparameter Setting

[0352] During the collaborative neighborhood graph construction phase, the filter threshold K was set to 10 for the Yelp dataset and 5 for the other datasets. Furthermore, the filter threshold η was set to 0.8 for all datasets. To ensure fair comparison across all baselines, we used Adam as the optimizer, the batch size was set to 4096, and the embedding size d was set to 64. In GCN, the number of graph neural network layers was selected from {1, 2, 3, 4}, and in the multi-head GAT, the number of hidden attention heads K was selected from {2, 4, 6, 8}. The experiment was repeated five times, and the average results are reported.

[0353] 5.4 Baseline Comparison

[0354] AMLECR is comprehensively compared with the following eight methods in terms of Recall@K and NDCG@K to evaluate the performance of AMLECR, as shown below. Table 1 shows the experimental results.

[0355] NGCF: This method incorporates collaborative signals into the interaction modeling process based on graph neural networks (GNNs) to capture higher-order connection patterns.

[0356] LightGCN: This is a lightweight GCN model that only retains the most critical component, neighborhood aggregation, which significantly improves the performance of the model.

[0357] DGCF: This method focuses on the user-item relationship under the fine-grained user intent. By modeling the interaction intent, iteratively optimizes the intent perception graph and representation, promotes intent independence, and achieves effective information extraction.

[0358] HCCF: This method proposes a self-supervised recommendation framework, which uses a hypergraph-enhanced cross-view comparative learning architecture to capture local and global collaborative relationships and improve the representation quality of the recommendation system.

[0359] NCL: This method incorporates potential neighbors into the comparison pairs, introduces neighbors from structural and semantic spaces, and develops a novel structural comparison objective by optimizing a specific objective function.

[0360] SGL: Uses self-supervised learning for graph structure data enhancement, generating different comparative views through methods such as node discarding, edge discarding, and random walks, thereby effectively achieving data enhancement.

[0361] AdaMCL: This method effectively utilizes high-order information and noise problems through a multi-view fusion contrastive learning paradigm and a layer-level contrastive learning task.

[0362] SimGCL: This method implements a simple contrastive approach by adding uniform noise in the embedding space to create contrastive views.

[0363] 5.5 Performance Comparison

[0364] 5.5.1 Overall Performance Analysis

[0365] Table 1 shows a comparative analysis of the performance of AMLECR against eight cutting-edge collaborative filtering models in terms of Recall@P (P=10, 20, 50) and NDCG@P. AMLECR outperforms the baselines overall, demonstrating the effectiveness of applying high-order relationships in collaborative neighbor graphs and the superiority of multi-head GAT for adaptively generating node-level enhanced representations. This approach preserves the original topological structure while overcoming the drawback of the GCN model's exclusive reliance on the graph-level adjacency matrix.

[0366] Among all graph neural network-based collaborative filtering models, LightGCN achieved the best baseline performance in most datasets due to its lightweight architecture, while the tangle representation learning method DGCF performed slightly worse than LightGCN, especially on sparse datasets. Most self-supervised learning methods (including SGL, HCCF, and NCL) showed consistent superiority, indicating that contrastive learning is beneficial for improving recommendation performance. However, SimGCL, also a self-supervised method, performed worse than SGL and NCL, mainly because perturbations in the representation space may destroy important supervised interaction labels. In contrast, NCL achieved better performance by leveraging semantic and structural neighbors.

[0367] In summary, AMLECR achieved the best performance on all datasets, which fully validates the efficiency of adaptive multi-view fusion contrastive learning. The reasons can be attributed to the following two points: First, explicit modeling of high-level user-item information enriches the embedding representations of users and items, making the generated representations more refined; second, the multi-head GAT can adaptively generate node-level enhanced representations, which not only preserves the important original topological structure, but also makes up for the defect of the GCN model that relies entirely on the graph-level adjacency matrix and can dynamically learn neighbor weights.

[0368] 5.5.2 Ablation Analysis

[0369] To verify the effectiveness of the proposed key components, three ablation variant experiments are constructed: i) removing the high-order neighborhood module (HNM) for explicit modeling; ii) removing the learnable augmented representation module (LARM); and iii) non-adaptive enhanced fusion strategy (NAEF).

[0370] Compared with the AMLECR method, the variant experiments were retrained on two datasets: Amazon-books and Gowalla. Figure 2 Detailed information and performance results are listed, where (a) is a schematic diagram of the display on Amazon-books; (b) is a schematic diagram of the display on Gowalla.

[0371] from Figure 2 You can find:

[0372] (1) After removing the high-order collaborative domain module, the model performance dropped significantly, which shows that applying high-order neighborhood information is very important for collaborative filtering;

[0373] (2) When graph convolutional networks aggregate neighboring nodes, their method is closely related to the graph structure, which to some extent restricts the generalization performance of the model. In contrast, the multi-head attention network (GAT) can obtain learnable dynamic weights independent of the graph structure based solely on node features. In view of this, the effective integration of GAT can enhance the overall expressive power of the collaborative filtering paradigm;

[0374] (3) Removing any contrastive learning task module will harm the performance, which fully demonstrates that contrastive learning is beneficial to enriching user-item information. In addition, the adaptive fusion strategy can effectively identify and filter redundant information and improve the accuracy of the model.

[0375] 5.5.3 Deeper Analysis

[0376] Robustness to sparsity. Data sparsity is a common problem in recommendation systems. Therefore, whether it can effectively combat data sparsity is an important indicator for measuring recommendation performance. To further verify that the proposed method can effectively alleviate the data sparsity problem, the Recall@10 indicator is used to compare AMLECR with the current advanced LightGCN, NCL and AdaMCL on the Gowalla and Amazon datasets to illustrate the robustness of the model to data sparsity. For reference, all users are divided into five groups according to the number of user-item interactions, and the total number of interactions in each group is kept unchanged. Figure 3 The comparison results of the four experimental performances are shown, among which (a) is the display diagram of Recall@10 on Amazon-books; (b) is the display diagram of NDCG@10 on Amazon-books; (c) is the display diagram of NDCG@10 on Gowalla; and (d) is the display diagram of Recall@10 on Gowall. Figure 3 It can be found that AMLECR outperforms LightGCN, NCL and AdaMCL all the time, which shows the potential of AMLECR in solving the problem of data sparsity. In addition, as the number of user-item interactions decreases, the performance of AMLECR increases.

[0377] This indicates that the generated self-supervisory signal is robust in alleviating data sparsity and achieves high-quality recommendations.

[0378] Regarding the robustness to noise, in order to test the robustness of AMLECR to the noisy data that is widely present in the user-item interaction graph in the real dataset, negative interactions are added to the original training set with a certain probability (e.g., 5%, 10%, 15%, 20%). Figure 4The Recall@10 performance metrics for the corresponding contaminated training sets on the Gowalla and Amazon datasets are shown, with (a) showing a schematic diagram for Amazon-books and (b) showing a schematic diagram for Gowalla. Clearly, noisy data interferes with the encoder's construction of a user preference model, negatively impacting recommendation results. The more noise injected, the more pronounced the performance degradation. However, compared to other comparison methods, AMLECR consistently outperforms, demonstrating its ability to more effectively discern noisy data in sparse interactions, ensuring stable performance and strong noise resistance. This is because AMLECR leverages multi-head attention to adaptively assign weights to adjacent nodes. This not only effectively identifies noisy information but also uses weights as collaborative supervisory signals to normalize the aggregation process of the graph neural network. Overall, the experimental results demonstrate that AMLECR can effectively handle noise in interaction data and exhibits good robustness even in the presence of high noise levels.

[0379] 5.5.4 Hyperparameter Analysis

[0380] In this section, we explore the impact of the key hyperparameters of MLECR (such as the number of hidden attention heads N, γ, and α) on the experimental performance on the Amazon-books and Gowalla datasets. Figure 5 (a) to (c) are schematic diagrams of the display on Gowalla, (d) to (f) are schematic diagrams of the display on Amazon-books, and Recall@10 is used as the evaluation indicator.

[0381] (1) The influence of γ. γ is used to regulate the weight of high-order information. The value of γ is set in the range of [0, 0.2, 0.4, 0.6, 0.8, 1.0, 1.2] for experiments. The experimental results are as follows: Figure 5 As shown in the figure, we can observe that when the value of γ is small, the model can only capture limited high-order information, resulting in poor performance. When the value of γ is around 0.8, AMLECR can achieve relatively ideal performance, which shows that γ is of great significance in improving experimental performance.

[0382] (2) The influence of α. α is defined to balance the information from the user-item interaction graph and the collaborative neighbor graph. The value of α is set in the range of [0.0, 0.05, 1.0, 1.5, 2.0] and the experimental results are shown in Figure 5 From the figure, we can observe that the performance of AMLECR shows strong stability to the change of α within a certain range, but if the value of α is too large, the model performance will deteriorate.

[0383] (3) The influence of N. N independent attention mechanisms are responsible for implementing feature transformation, and then averaging these features. Figure 5 It can be seen that increasing the value of N gradually improves model performance. This is because by aggregating features from different attention heads, richer information can be obtained. A model with a smaller value of N generates information-poor and discriminative augmented views. Therefore, the multi-head attention mechanism plays an important role in the model. However, a larger value of N significantly increases the complexity of the model. Therefore, while ensuring performance, an appropriate value of N should be set to improve model efficiency.

[0384] While 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 invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A learnable enhanced contrast recommendation method for adaptive fusion of multiple views, characterized by: include: S1, build a multi-view graph structure: Based on user-item interaction data, a user-item interaction graph and a collaborative neighbor graph are constructed. The collaborative neighbor graph includes a collaborative social graph and a collaborative knowledge graph. Collaborative edges are generated by counting the high-order similarities in the interaction history between users and items, and collaborative high-order neighbor relationships are filtered according to thresholds and edge number limits. S2, Multi-graph Embedding and Information Propagation: Based on the user-item interaction graph and the collaborative neighborhood graph, a graph neural network is used to perform multi-layer information propagation to obtain the embedded representation of each node in different views; S3, adaptive multi-view fusion: For each layer of node embedding, a non-parametric fusion strategy is used to dynamically adjust the contribution of different view embeddings based on inter-layer attenuation, node activity, and redundancy similarity between views to obtain the fused node representation. S4, learnable enhanced embedding generation: Using a multi-head graph attention network, we perform adaptive attention weighted aggregation on nodes, generate learnable enhanced embedding representations, and aggregate the results of different heads; S5, Joint contrastive learning and recommendation task optimization: Multi-view comparative learning is performed on the original view, collaborative neighbor views, and enhanced views. The fused view is used as an anchor to align with other views. The recommendation loss and contrast loss are jointly optimized to improve the discriminability of node representation and model robustness. S6, final recommendation and output: For target users and candidate items, preference scores are calculated based on the final fusion embedding using methods such as inner product to output personalized recommendation results.

2. The adaptive multi-view fusion learning enhanced contrast recommendation method according to claim 1, characterized in that: The constructing of the multi-view graph structure includes: definition Represents a collection of users, Represents a collection of items; Where M and J represent the number of users and projects respectively; The observed implicit feedback matrix is ​​expressed as when When , it means that user u interacts with item i; otherwise use Represents a user-item interaction graph; represents the collaborative neighborhood graph, which contains the collaborative social graph and collaborative knowledge graphs It can be expressed as the following formula: in, is the interaction adjacency matrix, is the transpose of the interaction adjacency matrix, T represents the transpose of the matrix; It is a collaborative social graph, which represents the interaction between users; A collaborative knowledge graph that represents the interaction between projects; In the recommendation system based on graph neural network, the user-item interaction graph is usually constructed as a bipartite graph; Right now in, It is a set of nodes, including project nodes and user nodes; represents an edge set; The initial embedding vector for each user u and item i is represents a d-dimensional vector space; The embedding sets of users and items are represented as and Where d represents the embedding dimension; Constructing feature graph for multi-head GAT based on user-item interaction graph in, represents the feature matrix; q u and q i are the feature vectors of users and items used by GAT respectively; Therefore, the probability of interaction between user u and item i can be expressed as follows: Where, represents the probability of user u interacting with item i; e u The final embedding representation of the user, e i The final embedding representation of the represented item; The transpose of the final embedding representation representing the user; The collaboration similarity is calculated using the Jaccard similarity coefficient based on the common interaction history of user i and user j as shown in the formula: Among them, sim i,j Jaccard similarity coefficient between node i and node j; It's a picture The first-order neighbor set of node i; It's a picture The first-order neighbor set of node j; ∩ and ∪ represent the intersection and union of sets; To simplify the calculation and remove redundant high-order information, the high-order information interaction diagram It can be defined as follows: in, is the adjacency relationship between node i and node j in the high-order information interaction graph; η represents the collaboration similarity threshold between users and projects. The collaboration similarity of similar users should be greater than or equal to this threshold. K represents the top K items recommended for the user, that is, the maximum number of high-order collaborative edges retained by each node.

3. The adaptive multi-view fusion learning enhanced contrast recommendation method according to claim 1, characterized in that: The multi-graph embedding and information propagation include: For each node i, at layer l, The embedding learned in Correspondingly, from The learned embeddings are available Represented as, then the user-item interaction graph is fused at layer l and collaboration diagrams The embedding representation of Embedded representation after fusion at layer l will be passed to the next graph neural network layer; thus, the information propagation process of AMLECR can be expressed as follows, in which the nonlinear activation and feature change process in the propagation function are omitted: in, and It's a picture and pictures The first-order neighbor set of node i; It's a picture The first-order neighbor set of node j; and In the figure The number of first-order neighbors of node i and node j; Indicated in the figure The number of first-order neighbors of node i; Is node i in layer l, based on the graph Embedding after propagation; Is node i in layer l, based on the graph Embedding after propagation; is the fused embedding representation of node j at layer l-1; Indicates that In the summation, all nodes j directly connected to node i are summed; Indicates that The sum of all nodes j directly connected to node i is calculated.

4. The adaptive multi-view fusion learning enhanced contrast recommendation method according to claim 1, characterized in that: The adaptive multi-view fusion includes: To fully capture the user-item interaction graph and high-level collaboration infographic The relationship between the two is carefully designed. An adaptive non-parametric fusion strategy is and The information is used to get the final embedding representation of users and items To more accurately grasp the contribution of auxiliary user-user relationships; The designed adaptive multi-view fusion strategy can be specifically expressed as: in, is the fused embedding representation; It's a picture The first-order neighbor set of the middle node v; and It's a picture The number of first-order neighbors of node i and node v; β i is the high-order information weight of the i-th node, l is the number of network layers; sim(*) represents the cosine similarity function; γ is used to control Hyperparameters of weights; d i represents the normalized user activity, where the logarithm is taken to mitigate the large differences in user activity; is the total number of nodes.

5. The adaptive multi-view fusion learning enhanced contrast recommendation method according to claim 2, characterized in that: The learnable enhanced embedding generation includes: Graph Attention Network, a neural architecture executed on graph data structures, can effectively stabilize the training process of the self-attention layer; according to Generate adaptive attention correlation coefficients using multi-head GAT The learnable enhanced embedding representation of , where N represents the number of hidden attention heads; Considering the self-attention mechanism of each head as a single-layer feedforward neural network, the adaptive attention coefficient between user u and item i in each head can be specifically expressed as: in, represents the adaptive attention weight coefficient of user u and item i under the nth attention head; exp(*) represents the exponential function with the natural base e as the base, and LeakyReLU(*) is the LeakyReLu activation function; W (n) is the weight matrix of the nth head; q u and q i are the feature vectors of users and items used by GAT respectively; q p is the feature vector of node p; Represents the learnable weight vector of the nth head; || represents cascade operation; N u represents the set of neighbor nodes of user u; p is the neighbor node index; p∈N u ∪{u} means traversing all neighbor nodes including user u itself; To simplify the calculation, when there is no interaction between user u and item i, Finally, the adaptive attention coefficient calculated in each head is aggregated with the adjacent features by formula: Where, represents the enhanced representation of user u after passing the nth attention head; ELU(*) represents the ELU activation function; The enhanced representation of item i corresponding to the above user u, that is, the enhanced representation of item i after the nth attention head is Averaging operations across multiple heads: in, represents the enhanced representation of user u at layer l; represents the enhanced representation of item i at layer l; represents the enhanced representation of item i after passing through the nth attention head; N represents the number of hidden attention heads; l represents the number of network layers.

6. The adaptive multi-view fusion learning enhanced contrast recommendation method according to claim 1, characterized in that: The joint contrastive learning and recommendation task optimization includes: After iterative propagation of a total of L layers of graph neural networks, the initial embedding vector and the embedding representation of all layers are averaged and pooled to obtain the final embedding representation: in, Represents the embedding vector of user u at layer l; r i (l) is the embedding vector of item i at layer l; Based on the final embedding representation of user u and item i, the possibility of interaction between target user u and candidate item i is predicted by inner product: in, represents the probability of user u interacting with item i; is the transpose of the final embedding vector of user u, e i is the final embedding vector of item i L represents the total number of layers of iterative propagation of the neural network; The Bayesian personalized ranking loss, a ranking objective function carefully designed for recommendation tasks, is used to directly capture interaction information. The goal is to encourage the model to predict the preference probability for items with which the user has interacted, which is much higher than the preference probability predicted for items with which the user has not interacted. It can be specifically expressed as: in, is the BPR recommendation loss; represents the positive sample, represents negative samples; u is the user index, i + The index of the positive sample of the item, i - is the index of the negative sample of the item; σ(*) is the Sigmoid activation function; User-Item Interaction Diagram Collaborative Neighborhood Graph The final representation of the user or item is fused as shown below: in, Indicates that node v is in the interaction graph The final embedding vector in ; Indicates that node v is in the collaborative neighbor graph The final embedding vector in ; Indicates that node v is at layer l, based on the interaction graph Embedding vector after propagation; Indicates that node v is at layer l, based on the collaborative neighbor graph Embedding vector after propagation; The fusion view can be regarded as an enhancement of the user-item interaction graph and the collaborative neighborhood view, and contains richer information. Therefore, compared with the direct comparison of the interaction view and the collaborative neighborhood view, we choose to compare them with the fusion view separately as follows: is the contrast loss of the interaction graph view; is the contrastive loss of the collaborative neighborhood graph view; is the multi-view weighted combination contrast loss; represents the transpose of the final embedding vector of node v; and Respectively represent node j and node v in the collaborative neighbor graph The final embedding vector in ; and Respectively represent node j and node v in the interaction graph The final embedding vector in ; exp(*) represents the exponential function with the natural base e as the base; is a sampling set; τ is the smoothing coefficient, which is used to adjust the smoothness of the softmax distribution; α is a weighting coefficient used to adjust the importance of the two view losses; In addition, in order to avoid losing important original information, GAT is used to make up for the limitations of GCN and convert the original view e u With enhanced view h u The contrastive learning is performed as shown below, aiming to optimize the representation of downstream tasks. Thanks to the characteristics of self-supervised learning, the generalization ability of the model is also improved: is the user enhancement contrast loss, which is used to measure the original view e u With enhanced view h u similarity; is the item enhancement contrast loss, which is used to measure the original view e i With enhanced view h i similarity; is the total enhanced view contrast loss; h u′ represents the embedding representation vector of other users except user u in the augmented view; h i′ represents the embedding representation vector of other items except item i in the enhanced view; u′ and i′ are used as negative users and negative items in the contrastive loss, respectively; and are collections of users and projects respectively; Finally, the contrast loss and recommendation loss are combined through joint learning, and the final loss calculation is as follows: in, is the total loss; is the multi-view weighted combination contrast loss; is the enhanced view contrast loss; represents the Frobenius regularization term with parameter Θ; λ1 and λ2 are designed to control the weight of the contrast loss, and λ3 is used to control the weight of the regularization term.

7. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to implement the adaptive fusion multi-view learnable enhanced contrast recommendation method according to any one of claims 1 to 6 when executing the executable instructions.

8. A computer-readable storage medium, characterized in that include: a memory having a computer program stored thereon; A processor is configured to execute the program in the memory to implement the adaptive fusion multi-view learnable enhanced contrast recommendation method according to any one of claims 1 to 6.

Images