Graph contrast learning recommendation method for self-adaptive intention perception enhancement

Abstract

The invention provides a self-adaptive intent perception enhanced graph contrast learning recommendation method, which comprises the following steps of: constructing a user-article bipartite graph, carrying out multi-layer embedded coding on a user and an article by adopting a graph neural network, and obtaining a multi-intention embedded representation of Gaussian distribution through a variational auto-encoder in combination with a multi-intention hypothesis; noise disturbance is adaptively added to intention embedding so as to enhance feature robustness, and the method respectively implements comparative learning of isomorphic and heterogeneous nodes in a node interaction space domain and an intention perception domain, so that the problems of data sparsity and intention entanglement are effectively relieved; and finally, carrying out joint optimization on recommendation task loss, KL divergence loss and double-domain comparison loss, and realizing accurate modeling and recommendation of the personalized preference of the user. Experimental results show that the method is superior to a mainstream recommendation system on a plurality of real data sets, and has strong generalization ability and robustness.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a graph comparative learning recommendation method with adaptive intent perception enhancement. Background Art

[0002] Recommender systems, as an important technical means to address information overload and meet personalized needs, rely on accurately capturing user preferences. Traditional collaborative filtering methods rely on user-item interaction data, but in practical applications, data sparsity is common, hindering recommendation effectiveness. Graph neural networks (GNNs) have made significant progress in mining high-order collaborative relationships, improving the expressive power of embedded features through graph structure modeling. However, existing recommendation methods based on graph contrastive learning (GCL), while incorporating self-supervisory signals to alleviate the sparsity problem, generally suffer from insufficient modeling of potential multiple intents between users and items, high levels of intent entanglement, and unstable data augmentation. These shortcomings limit the recommender system's ability to capture fine-grained preferences and complex behavioral motivations. Therefore, a new graph contrastive learning recommendation method that can adaptively perceive multiple user intents and effectively enhance the robustness of embeddings is urgently needed to further improve the accuracy and interpretability of recommendations. 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 an adaptive intent-aware enhanced graph comparison learning recommendation method, comprising:

[0004] S1: Construct a bipartite graph based on user-item interactions and use a graph neural network to perform multi-level embedding feature encoding on the bipartite graph to obtain embedded representations of users and items.

[0005] S2, inputs the embedded representation into the variational autoencoder to obtain Gaussian distribution embedding representations of multiple potential intents, and obtains multi-intent embedding through sampling;

[0006] S3, adaptively adding noise perturbation to the multi-intent embedding according to its norm to obtain an enhanced intent embedding representation;

[0007] S4 implements contrastive learning in the node interaction space domain and the intention perception domain respectively:

[0008] In the node interaction space domain, comparative learning is performed on the embedding of homogeneous and heterogeneous nodes;

[0009] In the intent-aware domain, isomorphic node comparative learning is performed on the enhanced multi-intent embedding;

[0010] S5, comprehensively considers the recommendation task loss, KL divergence loss, and the above-mentioned dual-domain contrastive learning loss, jointly optimizes the training model parameters, and outputs the recommendation results.

[0011] In a preferred embodiment of the present invention, step S1 includes:

[0012] In the recommendation scenario, the collaborative filtering paradigm recommendation system uses graph structure relationships to model the relationship between users and items;

[0013] It contains a set of user sets

[0014] and a set of items

[0015] Among them, u K Represents a user set The Kth user in , where K is the total number of users;

[0016] i M Represents an item set The Mth item in , where M is the total number of items;

[0017] The potential interaction relationship between users and items is represented by the interaction matrix Sure, Represents the vector space of the matrix, K×M is the dimension;

[0018] The user-item interaction relationship can be abstracted into a bipartite graph structure, namely:

[0019]

[0020] Among them, ε ui =ε iu =R ui ;

[0021]

[0022] Among them, ε ui =ε iu =R ui ;

[0023] It is a bipartite graph consisting of users and items;

[0024] is a collection of nodes;

[0025] ε is the set of edges in the bipartite graph;

[0026] ε ui It is the edge between users and items in the bipartite graph, that is, the connection relationship;

[0027] ε iu is the edge between items and users in the bipartite graph;

[0028] R uiIt is the interaction value between the user and the item, which can be 0 or 1;

[0029] When ε ui =ε iu =R ui = 0, it means there is no connection relationship. ui =ε iu =R ui =1, indicating that there is a connection relationship;

[0030] Predicting the user-item interaction bipartite graph The probability scores above are expressed as follows:

[0031]

[0032] in, Indicates that the initial embedding representation matrix E is known (0) Under the conditions of and interaction matrix R, the predicted interaction matrix The conditional probability of

[0033] It is represented by the user embedding representation e u , item embedding representation e i Condition, user-item bipartite graph interaction probability score;

[0034] is the interaction value between the user and the item predicted by the model, and its value is in the interval [0,1];

[0035] represents the initial embedding representation matrix;

[0036] Represents the user's initial embedding representation matrix;

[0037] The initial embedding representation matrix representing the item;

[0038] The factors that need to be added to the intention guidance at the interaction probability level are:

[0039]

[0040] in, Indicates the initial embedding representation matrix E (0) , user-item potential interaction matrix R and intention representation C, predict the interaction matrix The conditional probability of

[0041] e u represents the user embedding representation; e i Representation of item embedding representation;

[0042] represents the intention representation of user and item nodes, and S represents the set of intentions;

[0043] × represents multiplication operation;

[0044] Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the user set Conditional probability score of user embedding representation;

[0045] Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the item set Conditional probability scores of item embedding representations under the given condition;

[0046] and They represent the user's intention embedding representation and the item's intention embedding representation respectively.

[0047] In a preferred embodiment of the present invention, step S2 includes:

[0048] Construct a subgraph representation embedding information propagation layer to aggregate comprehensive subgraph structure information:

[0049]

[0050] in, is the aggregate representation of the user's neighboring nodes to the target node, K is the total number of users, and d is the number of features of the embedding vector;

[0051] It is the aggregate representation of the item’s neighboring nodes to the target node, and M is the total number of items;

[0052] Represents the user-item interaction graph;

[0053] The normalized adjacency matrix representing the normalized user-item interaction graph;

[0054] and Diagonal matrices for users and items respectively;

[0055] and are the normalized diagonal matrices of users and items respectively;

[0056] and Represent the embedding matrices of users and items respectively;

[0057] The residual connection structure is applied to the aggregation layer to prevent over-smoothing problems. In order to aggregate high-order collaborative filtering signals, GCN information propagation is performed at different graph structure layers:

[0058]

[0059] in, represents the user embedding representation obtained after l-layer GCN encoding;

[0060] It is the user embedding representation obtained after l-1 layer GCN encoding;

[0061] It is the aggregated representation of the neighboring nodes to the target user node after being encoded by the l-1 layer GCN;

[0062] It is the item embedding representation obtained after l-layer GCN encoding;

[0063] It is the item embedding representation obtained after l-1 layer GCN encoding;

[0064] Represents the aggregated representation of the neighboring nodes to the target item node after encoding by the l-1 layer GCN;

[0065] Calculate the correlation score between the user and item embedding structures and the potential intent of users and items:

[0066]

[0067] in, The probability score representing the similarity of the user's preference for a specific intent category;

[0068] The probability score representing the preference similarity of an item to a specific intent category;

[0069] and Respectively represent the embedding representation of users and items at the lth layer, which represents the probability score of the preference similarity of items to specific intent categories; represents the user embedding representation obtained after l-layer GCN encoding;

[0070] It is the item embedding representation obtained after l-layer GCN encoding;

[0071] exp(*) is the exponential function with the natural base e as the base;

[0072] s′ is the index that traverses all intent categories;

[0073] Τ is used to adjust the smoothness of the softmax probability distribution;

[0074] is the user's embedded representation under the sth intent category;

[0075] is the user's embedded representation under the s′th intent category;

[0076] is the embedding representation of the item under the sth intent category;

[0077] is the embedding representation of the item under the s′th intent category;

[0078] τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution;

[0079] After calculating the relevance score, we can aggregate S different intent information and the global sampled subgraph to represent the user and item information as follows:

[0080]

[0081] in, and They represent the embedded representations of users and items respectively, and the embedded representations can be stacked and recorded as and

[0082] and are the embedded representations of users and items under the l-th layer of intention perception;

[0083] S represents the number of intention information;

[0084] In order to obtain the posterior predictive distribution of user preference features, the encoder is defined as a variational autoencoder, and the embedded features of users and items are expressed in Gaussian forms as follows:

[0085]

[0086] Among them, P(e u |Ε (0) ,R,C u ) is the user embedding representation e under the condition that the initial embedding, interaction matrix and user intention set are known. u The probability distribution of C u Represents the user intent set;

[0087] P(e i |Ε(0) ,R,C i ) is the item embedding representation e under the condition that the initial embedding, interaction matrix and item intention set are known i The probability distribution of C i Represents the item intention set;

[0088] Indicates that the user embeds e u At the mean μ u and the variance is Normal distribution when ;

[0089] Indicates that the item is embedded in e i At the mean μ i and the variance is Normal distribution when ;

[0090] It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each user node in the bipartite graph;

[0091] It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each node of the item in the bipartite graph;

[0092] is the approximate variance of the Gaussian distribution of each user node in the bipartite graph;

[0093] is the approximate variance of the Gaussian distribution of each node representing the item in the bipartite graph;

[0094] e u and e i Represent the embedding representations of users and items respectively;

[0095] The reparameterization method is used to avoid back propagation, which is as follows:

[0096]

[0097] Among them, ∈ u and ∈ i denote Gaussian noise variables for users and items respectively; ⊙ denotes element-wise product;

[0098] This re-parameterized adaptive fusion feature avoids manual adjustment. Combining the above formulas, the global graph structure encoding and the intent-aware graph structure synergy can be integrated:

[0099]

[0100] and They are the embedding representations of users and items under the l-1 layer of intent perception.

[0101] In a preferred embodiment of the present invention, step S3 includes:

[0102] A noise matrix G with the same shape as the intent embedding matrix and an adaptive enhanced intent embedding representation H are introduced, and the enhanced intent embedding representation is obtained by matrix addition:

[0103] H′=H+G;

[0104] H′ is the intent embedding representation after adding noise;

[0105] H is the adaptive enhanced intent embedding representation;

[0106] G is the noise matrix;

[0107] In order to adapt the injected noise and ensure that high-confidence features are not overly disturbed, the strength of the noise matrix G of each feature vector is dynamically controlled according to the L2 norm of the intent embedding vector. The noise strength is calculated as follows:

[0108]

[0109] Where Ψ is the noise intensity;

[0110] ||H||2 represents the L2 norm of the intention vector;

[0111] ω is a hyperparameter that represents the basic noise intensity;

[0112] Is a constant used to avoid the situation where the denominator is left blank;

[0113] In order to avoid the sign reversal problem caused by the addition of noise to the values ​​close to zero in the original embedding, the embedding values ​​close to zero are smoothed:

[0114]

[0115] Where sign′(·) is a sign smoothing function. When the number of elements in the intention embedding vector is less than φ, the output is 1, otherwise the standard sign function is output.

[0116] H n Represents the nth element of the intent embedding vector;

[0117] |H n | is the total number of elements in the intent embedding vector;

[0118] φ is the judgment threshold for sign adjustment;

[0119] sign(·) represents a sign function, which returns the sign of the input value;

[0120] In order to introduce randomness while keeping the direction of the noise consistent, a random noise vector is generated and normalized:

[0121]

[0122] in, represents the normalized Gaussian noise vector;

[0123] η represents the Gaussian noise vector corresponding to the intent embedding vector;

[0124] ||η||2 is the L2 norm of the Gaussian noise vector;

[0125] represents the standard Gaussian distribution;

[0126] According to H′, sign′(H n )and The noise matrix G is calculated:

[0127]

[0128] Where ⊙ represents the element-wise product.

[0129] In a preferred embodiment of the present invention, step S4 includes:

[0130] Given any node, the widely used contrast loss infoNCE is calculated as follows:

[0131]

[0132] Where f(*) is the infoNCE loss function;

[0133] x′ and x″ represent the embedding representation vectors of the same embedding representation vector x under different enhanced views;

[0134] x is the embedding representation vector;

[0135] exp(*) is the exponential function with the natural base e as the base;

[0136] cos(·,·) represents cosine similarity;

[0137] The hyperparameter τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution;

[0138] y∈β represents the sampling from the mini-batch as the negative sample node of x, β is the sampling set in the mini-batch;

[0139] By maintaining the distance between different nodes in the homogeneous node set, we can avoid the nodes from concentrating in a specific direction. The calculation method is as follows:

[0140]

[0141] in, represents the loss function for contrastive learning of isomorphic nodes;

[0142] e u and e i Represent the embedding representations of users and items respectively;

[0143] Specifically, the node interaction space domain contrast loss function is:

[0144]

[0145] in, represents the node interaction space domain loss function;

[0146] represents the contrastive learning loss function of heterogeneous nodes;

[0147] represents the expansion of adding the comparative learning loss function of homogeneous nodes and the comparative learning loss function of heterogeneous nodes, and <u,i>∈β represents the nodes sampled from the mini-batch;

[0148] The intent-aware domain contrast loss function is calculated as:

[0149]

[0150] in, is the intent-aware contrast loss function;

[0151] h′ u and h′ i These are the enhanced embedding representations of users and items respectively;

[0152] h u and h i are the embedding representations of original users and items respectively;

[0153] The Bayesian personalized ranking loss function is introduced in combination with KL divergence to optimize the graph generation process:

[0154]

[0155] in, represents the Bayesian personalized ranking loss function;

[0156] <u,p,n>∈β means accumulating each triple in the sampling set β, where u,p,n represent user index, positive sample and negative sample respectively;

[0157] σ(*) represents the Sigmoid activation function;

[0158] and Represents the probability distribution of positive sample and negative sample users respectively;

[0159] and Represent positive sample users and negative sample users respectively;

[0160] KL(*) is the Kullback-Leibler divergence loss calculation function, which is used to measure the difference between two probability distributions. || is the separator in Kullback-Leibler divergence, which is used to separate the probability distributions before and after.

[0161] represents the probability distribution of the embedding representation matrix E under the conditions of parameter sets Θ and R;

[0162] R represents the user-item interaction matrix; Θ = {E, H} represents the set of trainable model parameters;

[0163] represents the prior test probability distribution of the embedding representation matrix E;

[0164] Finally, in order to integrate intent modeling and recommendation tasks, a multi-task joint training strategy is adopted to optimize the self-supervised loss and the classic recommendation loss:

[0165]

[0166] in, It is the total loss function, which combines the Bayesian personalized ranking loss function, the interactive space domain loss function, the intent-aware contrast loss function and the regularization term;

[0167] λ1 and λ2 are adjustable weight parameters;

[0168] is the square of the Frobenius norm of the model parameter set Θ.

[0169] The present invention also discloses a computer system, comprising:

[0170] processor;

[0171] a memory for storing processor-executable instructions;

[0172] Wherein, the processor is configured to implement the adaptive intent-aware enhanced graph comparison learning recommendation method when executing the executable instructions.

[0173] The present invention also discloses a computer-readable storage medium, comprising:

[0174] a memory having a computer program stored thereon;

[0175] A processor is used to execute the program in the memory to implement the adaptive intent-aware enhanced graph comparison learning recommendation method.

[0176] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0177] By introducing an adaptive intent-aware mechanism and a noise enhancement strategy, the present invention achieves detailed modeling of multiple potential user-item intentions, effectively alleviating the problem of degraded recommendation performance caused by data sparsity and intent entanglement. The dual-domain collaborative graph contrast learning framework implements isomorphic and heterogeneous comparisons in the node interaction space and the intent-aware space, respectively, which helps to improve the diversity and expressiveness of embedded features. The adaptive noise perturbation enhancement strategy enhances the robustness of embedding and improves the generalization performance of the model on different user activity groups and different data sets. Experiments show that the present invention significantly outperforms existing mainstream recommendation methods on multiple public data sets, and has higher accuracy, robustness and practical application value.

[0178] 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

[0179] 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:

[0180] Figure 1 It is a schematic block diagram of the process of the present invention.

[0181] Figure 2 This is a schematic diagram comparing the performance of different model variants of the present invention on three data sets.

[0182] Figure 3 Schematic diagram of the performance comparison of the present invention on three data sets.

[0183] Figure 4 2 is a schematic diagram comparing the data sparsity performance of user groups with different interactivity levels according to the present invention.

[0184] Figure 5 Schematic diagram of the comparison of the sensitivity of the learning weight hyperparameters in the present invention.

[0185] Figure 6 It is a schematic diagram of the sensitivity analysis of the intended quantity hyperparameters of the present invention.

[0186] Figure 7 It is a schematic diagram of the potential intention prototype distribution for comparative analysis of different numbers of intentions in the present invention. DETAILED DESCRIPTION

[0187] 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.

[0188] 1. Introduction

[0189] 1.1 Background

[0190] Recommender systems have become a primary means of addressing information overload and personalization needs in our daily lives. Their core lies in their ability to accurately capture user preferences. Technically, collaborative filtering (CF), the cornerstone of recommender systems, can be defined as inferring personalized user preferences from historical interaction data and predicting user behavior. Current recommendation methods within the CF paradigm focus on extracting high-quality user-item embedding representations. Graph neural networks (GNNs), such as NGCF, LightGCN, and GCCF, have achieved significant success in modeling the high-order connectivity of user-item interactions. These recommendation methods all aggregate collaborative information by encoding the user-item interaction structure through information passing mechanisms, capturing high-level user or item similarities in a latent embedding space. Specifically, LightGCN has become a standard approach for graph-based recommender systems. However, with the development of graph contrastive learning (GCL) in recommender systems, GCL has become a mainstream approach in recommender systems.

[0191] Since GNN recommendation systems rely on user-item interactions as important labels for supervising recommendation models, real-world recommendation system datasets are often highly sparse. Therefore, recent research has combined GNN with contrastive learning to address this issue, using self-supervisory signals to address data sparsity. This allows recommendation systems to extract meaningful representations from unlabeled data, allowing for more accurate recommendations and predictions even with sparse data. By performing data augmentation operations on the graph structure to generate contrastive views, the consistency between the established contrastive enhanced views is maximized to learn invariant user and item embedding representations, thereby enhancing the similarity of positive sample pairs and pushing away the distance between negative sample pairs in the embedding space. In recent years, research on graph contrastive learning recommendation systems has mainly focused on: graph topology enhancement, graph feature-level enhancement, and local-global enhancement. Specifically, SGL randomly discards nodes and edges at the graph structure level to generate comparative views. It is an enhancement method based on the graph topology structure; SimGCL and XSimGCL compare the hidden feature representation obtained by embedding them with the random noise obtained for graph enhancement. It is a data enhancement method from the feature-level structure of the graph; HCCF and NCL use different information aggregators to generate global node embedding representations and compare them with local node representations. It is a data enhancement method from the global to local features of the graph structure.

[0192] The application of GCL to recommender systems has achieved some success, but this work still struggles to provide detailed explanations of user behavior. Human behavior isn't triggered unconsciously; rather, all user actions are driven by multiple factors and influenced by social phenomena such as popularity and the herd mentality. Therefore, it's crucial to uncover the underlying factors behind user-item interactions. Pioneering work has explored the underlying intentions behind user item selection, shedding light on the intentions behind recommendation systems. Inspired by disentangled GCNs, research on intention disentanglement has also been introduced into recommender systems to enhance model interpretability. Specifically, DGCF aims to iteratively optimize intent-aware interaction graphs and representations, encouraging the independence of different intents to achieve a separate representation of fine-grained user-item relationships and effectively extract relevant information for each intent. KGIN models each user intent as an attentional combination of KG relationships and designs a novel GNN information aggregation scheme to integrate remotely connected relationship paths, thereby identifying user-item relationships at a fine-grained level and improving the performance and interpretability of recommendation systems. DCCF also considers the underlying intent of items when modeling intent. Existing research has shown that user behavior is not isolated and may be influenced by a single individual or made in accordance with the crowd, an effect known in social psychology as the herd effect. Similar to users, items also experience a popularity effect. The widespread popularity of an item among a user group reflects the stickiness of the user behind the item, and thus reflects the importance of the item to that user group.

[0193] 1.2 Research Motivation

[0194] Although the research has achieved some results, the GCL-based recommendation system still has two key shortcomings. The motivations for this research are as follows:

[0195] In today's real-life recommendation systems, user-item interactions are often not driven by a single intent, but rather by the synergistic influence of multiple intents. Previous research has often overlooked the fact that the factors in the latent space behind user-item interactions are highly entangled, thus affecting user representations. Entanglement refers to the fact that, in GCL recommendation methods, user-item interactions are often driven by multiple different factors, and these factors (i.e., the user's latent intent or preferences) are often not clearly distinguished but rather mixed and encoded in the representation learned by the model. In this case, even the so-called "invariant representation" obtained by the GCL method contains information from different intents, making it difficult to capture the finer-grained latent intent between user and item interactions, thereby limiting the recommendation system's ability to better capture more realistic user preferences and intent-aware preferences.

[0196] Existing GCL-based recommendation methods struggle to provide stable self-supervised learning (SSL) signals, making contrastive learning difficult to generalize to diverse user-item interaction structures. Traditional random data augmentation, such as random dropout of nodes or edges, is believed to produce high-quality, invariant representations. However, studies have found that this approach is limited in effectiveness and instability, making it difficult to maintain the original semantic relationships. Specifically, randomly dropping edges or nodes can disrupt important community structures and potentially perturb the original user-item interaction features with non-positive correlations.

[0197] 1.3 Research Contributions

[0198] Based on the limitations and challenges of existing recommendation systems, this patent proposes a new adaptive intent-guided graph comparative learning recommendation method (AIARec). Specifically, the model encodes the user-item interaction structure through iterative GCN, and defines the obtained feature distribution as the preference distribution of users and items, which can be abstracted as the potential intention relationship between users and items. The purpose of this work is to better explore the potential relationship behind users and items to capture higher-granularity spatial intent vectors. In order to achieve the goal, an adaptive node feature enhancement method is proposed, which combines the GCL self-supervision signal to avoid high entanglement of intent, and compares the view combined with global intent perception with the view obtained by the adaptive enhancement method to improve the stability of the self-supervision signal and the model's ability to alleviate data sparsity problems. A dual-domain latent space comparison method is introduced to take into account the relative comparison of isomorphic nodes to prevent the Matthew effect of popularity bias problem. The node representation learned in the embedding layer is concentrated in one direction, causing the model to collapse.

[0199] In summary, the main contributions of this study are as follows:

[0200] In this study, we investigate the GCL recommendation system based on challenging problems in real-world scenarios, decompose the potential intention relationship between user-item interactions, and propose a new recommendation method, AIARec, to explore more fine-grained potential intention relationships between users and items, thereby improving the representation ability and interpretability of the model.

[0201] An adaptive feature-level noise perturbation data enhancement method is proposed to ensure the feature distribution of the data to effectively improve the problem of excessive noise perturbation; it is used to enhance the intention feature of the node in the potential intention space of users and items to enhance the contrastive self-supervision signal in the feature space.

[0202] A dual-domain-aware graph comparative learning framework was constructed to optimize the consistency and uniformity of node representations. Extensive experimental results, conducted on three public datasets, show that the proposed AIARec recommendation method outperforms 16 existing recommendation methods and effectively taps into users' personalized preferences. Supplementary experiments and ablation studies of various components were also conducted to analyze the rationality of the recommendation method design.

[0203] 2. Related Work

[0204] GNNs have demonstrated outstanding performance in the field of leveraging graph representation learning to solve downstream tasks. Recommender systems have also been inspired by GNNs. Researchers have constructed graph neural network architectures based on the CF paradigm for recommender systems, using powerful GNN models to capture high-order connectivity information in the user-item interaction graph. STGCN, a novel STacked and Reconstructed Graph Convolutional Networks (STGCN) architecture designed for cold-start recommender systems, is used to learn node representations to improve recommender performance. Later, researchers considered transition smoothing and scalability issues, leading some recommendation methods to optimize GNN architectures based on these issues, such as GCCF, LightGCN, and HGCF. NGCF and PinSage have demonstrated the importance of high-order connections between users and items for collaborative filtering. To further optimize the message passing process within the graph convolutional neural network (GCN) framework, LightGCN lightweights GNNs for recommender systems by removing nonlinear activation functions and feature transformation layers and using a sum-based pooling operation for neighborhood aggregation. It is currently the most mainstream recommendation method within the CF paradigm.

[0205] 2.2 Recommendation Method Based on Disentanglement

[0206] Generally speaking, obtaining disentangled representations of users' latent intents in implicit feedback recommendation scenarios has been a hot topic in recent years. Some progress has been made, such as using variational autoencoders to encode high-level user intents for improved recommendations. DGCF builds on the concept of intent disentanglement and learns disentangled representations via split embedding within a graph neural network framework. DisenHAN utilizes heterogeneous graph attention mechanisms to learn disentangled embeddings. DICE learns two disentangled causal embeddings for users and items, respectively. KGIN leverages knowledge graphs to encode latent user intents, thereby improving the performance of recommendation methods. DCCF proposes decomposing users and items into factor-level representations and employing factor-level attention mechanisms to identify underlying intents. CDR designs a dynamic routing mechanism to characterize the correlations between user intents and implement denoised embeddings. Some innovative research efforts have attempted to incorporate contrastive learning into intent modeling, such as ICLRec, DiRec, and DCCF. Similar to ICLRec, sequential recommendation methods based on contrastive learning for intents obtain more fine-grained intent representations by performing K-means operations on intent representations. However, these methods have not yet explicitly defined the intent of users or items at a fine-grained level, which results in the data augmentation process for contrastive learning increasing the complexity of model training. Although these novel methods utilize contrastive learning to improve recommendation systems, they do not delve into sufficient detail to characterize the specific intent of users or items.

[0207] 2.3 Graph Comparative Learning in Recommender Systems

[0208] Table 1. Currently popular contrastive learning recommendation methods

[0209]

[0210]

[0211] In recent years, graph contrastive learning has been very successful in various scenarios of recommendation systems, including but not limited to sequential recommendation, knowledge graph-enhanced recommendation, multi-interest recommendation, and multi-behavior recommendation. A particularly relevant research trend is to enhance collaborative filtering technology based on graph neural networks (GNNs) through contrastive learning. In this field, some models have recently emerged, such as SGL, NCL, and HCCF, which have enhanced their performance by adopting contrastive learning and achieved industry-leading results.

[0212] Leading effect. For example, SGL adopts the method of random dropout operators to perform data augmentation by destroying the structure of the interaction graph. NCL focuses on aligning the representation between a single user and a semantic center node. HCCF uses hypergraphs to generate contrast signals to improve graph-based recommendation systems. Different from them, this work enhances the graph-based collaborative filtering paradigm and the generative self-supervised learning framework. Although these models have performed well in improving the accuracy of recommendation systems, there may still be room for improvement in capturing the encoding of the underlying factors behind user-item interactions, which may lead to suboptimal representations of coarse-grained user preference modeling for recommendation.

[0213] To better reflect the achievements of contrastive learning in recommendation systems, some of the most influential contrastive learning-based recommendation methods as of 2025 are summarized in Table 1. The methods are summarized based on their target tasks, published journals or conferences, publication years, and contrastive learning strategies.

[0214] Table 1 summarizes the ICMA model published in 2025. Using contrastive self-supervised learning, it integrates intent distributions learned by clustering into sequential recommendation models to maximize the consistency between sequential views and their corresponding intents. In 2024, graph contrastive learning achieved significant success across various areas of recommendation systems, including but not limited to Graph-based CF, Cross-Domain Rec, and Sequential Rec. Some recommendation methods, such as BIGCF and ICSRec, incorporate intent modeling to enhance the quality of graph representation learning. XsimGCL proposes injecting embedding noise to enhance data representations for downstream recommendation tasks. GDCCDR utilizes contrastive learning to ensure the consistency of domain-invariant features and uses mutual information to separate features from different domains, generating more refined disentangled representations to improve cross-domain recommendation performance. Looking back further, recommendation methods based on graph contrastive learning across various fields generally utilize different data augmentation strategies to improve the quality of the self-supervisory signal. Some recommendation methods, such as SGL, ICLRec, KGCL, DCCF, IOCRec, and Gformer, also combine intent modeling to enhance graph embedding representations.

[0215] 3. Implementation method

[0216] 3.1 Task Description

[0217] In the recommendation scenario, the collaborative filtering paradigm recommendation system uses graph structure relationships to model the relationship between users and items;

[0218] It contains a set of user sets

[0219] and a set of items

[0220] Among them, u K Represents a user set The Kth user in , where K is the total number of users;

[0221] i M Represents an item set The Mth item in , where M is the total number of items;

[0222] The potential interaction relationship between users and items is represented by the interaction matrix Sure, Represents the vector space of the matrix, K×M is the dimension;

[0223] The user-item interaction relationship can be abstracted into a bipartite graph structure, namely:

[0224]

[0225] Among them, ε ui =ε iu =R ui ;

[0226] It is a bipartite graph consisting of users and items;

[0227] is a collection of nodes;

[0228] ε is the set of edges in the bipartite graph;

[0229] ε ui It is the edge between users and items in the bipartite graph, that is, the connection relationship;

[0230] ε iu is the edge between items and users in the bipartite graph;

[0231] R ui It is the interaction value between the user and the item, which can be 0 or 1;

[0232] When ε ui =ε iu =R ui = 0, it means there is no connection relationship. ui =ε iu =R ui =1, indicating that there is a connection relationship;

[0233] Therefore, the recommendation task of the collaborative filtering paradigm can also be defined as a bipartite graph generation task, specifically predicting the user-item interaction bipartite graph. The probability scores above are expressed as follows:

[0234]

[0235] in, Indicates that the initial embedding representation matrix E is known (0) Under the conditions of and interaction matrix R, the predicted interaction matrix The conditional probability of

[0236] It is represented by the user embedding representation e u , item embedding representation e i Condition, user-item bipartite graph interaction probability score;

[0237] is the interaction value between the user and the item predicted by the model, and its value is in the interval [0,1];

[0238] represents the initial embedding representation matrix;

[0239] Represents the user's initial embedding representation matrix;

[0240] Represents the initial embedding representation matrix of the item.

[0241] The key to recommendation systems is to mathematically model the preference characteristics of user-item interactions. However, all user behaviors are driven by intent. Therefore, the following factors should be incorporated into the interaction probability layer to guide intent:

[0242]

[0243] in, Indicates the initial embedding representation matrix E (0) , user-item potential interaction matrix R and intention representation C, predict the interaction matrix The conditional probability of

[0244] e u represents the user embedding representation; e i Representation of item embedding representation;

[0245] represents the intention representation of user and item nodes, and S represents the set of intentions;

[0246] × represents multiplication operation;

[0247] Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the user set Conditional probability score of user embedding representation;

[0248] Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the item set Conditional probability scores of item embedding representations under the given condition;

[0249] and They represent the user's intention embedding representation and the item's intention embedding representation respectively.

[0250] The mainstream modeling method today is to abstract the intent into a learnable but unknowable shared or non-shared parameter matrix. In this section, we will elaborate on the proposed AIARec architecture and show the overall framework of the entire recommendation method. Figure 1 shown.

[0251] 3.2 AIARec model

[0252] 3.2.1 High-order graph learning

[0253] Based on powerful neural network models, numerous studies have demonstrated the effectiveness of GNNs in capturing collaborative filtering signals in state-of-the-art recommender systems, such as NGCF, LightGCN, and DGCF. Inspired by these studies, GCN-based encoding methods can effectively enrich the collaborative relationships at all levels of the user-item interaction graph. Therefore, AIARec also uses LightGCN to encode high-level collaborative relationships. This is achieved by constructing a subgraph representation embedding information propagation layer to aggregate comprehensive subgraph structural information:

[0254]

[0255] in, is the aggregate representation of the user's neighboring nodes to the target node, K is the total number of users, and d is the number of features of the embedding vector;

[0256] It is the aggregate representation of the item’s neighboring nodes to the target node, and M is the total number of items;

[0257] Represents the user-item interaction graph;

[0258] The normalized adjacency matrix representing the normalized user-item interaction graph;

[0259] and Diagonal matrices for users and items respectively;

[0260] and are the normalized diagonal matrices of users and items respectively;

[0261] and Represent the embedding matrices of users and items respectively.

[0262] The residual connection structure is applied to the aggregation layer to prevent over-smoothing. In order to aggregate high-order collaborative filtering signals, GCN information propagation is performed at different graph structure layers based on experience:

[0263]

[0264] in, represents the user embedding representation obtained after l-layer GCN encoding;

[0265] It is the user embedding representation obtained after l-1 layer GCN encoding;

[0266] It is the aggregated representation of the neighboring nodes to the target user node after being encoded by the l-1 layer GCN;

[0267] It is the item embedding representation obtained after l-layer GCN encoding;

[0268] It is the item embedding representation obtained after l-1 layer GCN encoding;

[0269] Represents the aggregated representation of the neighboring nodes to the target item node after encoding by the l-1 layer GCN.

[0270] 3.2.2 Intent-aware Graph Information Reconstruction

[0271] In a recommendation system, when users interact with items, both users and items have multiple potential intentions that influence each other. In order to capture multi-granular potential intentions, we assume that users and items have S different potential intentions in the modeling process, which can be expressed as and The latent intent vector on the item side can be defined as the contextual representation of the item. To further study the correspondence between intent and interaction, we first need to study the relationship between each user, item, and each latent intent.

[0272] Specifically, we first calculate the correlation score between the user and item embedding structures and the potential intentions of users and items:

[0273]

[0274] in, The probability score representing the similarity of the user's preference for a specific intent category;

[0275] The probability score representing the preference similarity of an item to a specific intent category;

[0276] and The embedding representations of users and items respectively represent the probability scores of the preference similarity between items and specific intent categories at the lth layer;

[0277] in, represents the user embedding representation obtained after l-layer GCN encoding;

[0278] It is the item embedding representation obtained after l-layer GCN encoding;

[0279] exp(*) is the exponential function with the natural base e as the base;

[0280] s′ is the index that traverses all intent categories;

[0281] Τ is used to adjust the smoothness of the softmax probability distribution;

[0282] is the user's embedded representation under the sth intent category;

[0283] is the user's embedded representation under the s′th intent category;

[0284] is the embedding representation of the item under the sth intent category;

[0285] is the embedding representation of the item under the s′th intent category;

[0286] τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution.

[0287] After calculating the relevance score, we can aggregate S different intent information and the global sampled subgraph to represent the user and item information as follows:

[0288]

[0289] in, and They represent the embedded representations of users and items respectively, and the embedded representations can be stacked and recorded as and

[0290] and are the embedded representations of users and items under the l-th layer of intention perception;

[0291] S represents the number of intention information;

[0292] In order to obtain the posterior predictive distribution of user preference features, the encoder is defined as a variational autoencoder (VAE), and the embedded features of users and items are represented as Gaussian forms as shown in Eq.

[0293]

[0294] Among them, P(e u |Ε (0) ,R,C u ) is the user embedding representation e under the condition that the initial embedding, interaction matrix and user intention set are known. u The probability distribution of C u Represents the user intent set;

[0295] P(e i |Ε (0) ,R,C i ) is the item embedding representation e under the condition that the initial embedding, interaction matrix and item intention set are known i The probability distribution of C i Represents the item intention set;

[0296] Indicates that the user embeds e u At the mean μ u and the variance is Normal distribution when ;

[0297] Indicates that the item is embedded in e i At the mean μ i and the variance is Normal distribution when ;

[0298] It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each user node in the bipartite graph;

[0299] It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each node of the item in the bipartite graph;

[0300] is the approximate variance of the Gaussian distribution of each user node in the bipartite graph;

[0301] is the approximate variance of the Gaussian distribution of each node representing the item in the bipartite graph;

[0302] e u and ei Represent the embedding representations of users and items respectively.

[0303] Therefore, the learnable intent embedding representation can be represented as an approximate Gaussian distribution. The mean and variance sampling of the posterior prediction distribution of given users and items are expressed as and However, considering the non-differentiable nature of the sampling process, a reparameterization method is used to avoid backpropagation, as shown below:

[0304]

[0305] Among them, ∈ u and ∈ i denote Gaussian noise variables for users and items respectively; ⊙ denotes element-wise product;

[0306] This re-parameterized adaptive fusion feature avoids manual adjustment. Combining the above formulas, the global graph structure encoding and the intention-aware graph structure synergy can be integrated:

[0307]

[0308] and They are the embedded representations of users and items under the l-1 layer of intention perception;

[0309] Reconstructing the learned embedding representation using intent-aware representations can effectively uncover the underlying factors behind user-item interactions.

[0310] 3.2.3 Adaptive Intent Enhancement

[0311] Recent research has found that data augmentation is effective even on highly sparse graph structures. Because directly manipulating graph structures is intractable and time-consuming, research has focused on the latent feature space. One example involves reconstructing the latent feature space by adding an imperceptible fixed noise perturbation to the image. This approach can be applied to recommendation systems, where a fixed perturbation is applied to the latent feature space of the graph structure. However, if the noise intensity is not set appropriately, some embeddings may be over-perturbed while others are insufficiently perturbed, potentially leading to model overfitting. To address this, random noise is added as a new data augmentation method. While this method improves overfitting, it can also lead to inconsistent noise intensity, causing the model to learn incorrect information. To address this issue, a new adaptive data augmentation method is designed to enhance user and item intent-aware embedding representations for graph comparative learning, achieving the goal of disentanglement.

[0312] Data augmentation is achieved by adding adaptive noise data, which not only preserves the main semantic information of the embedding but also enhances the robustness and generalization ability of the model through appropriate noise perturbations. In order to adaptively enhance the intent embedding representation H (for convenience, the intent embedding representations of users and items are unified into H), a noise matrix G with the same shape as the intent embedding matrix is ​​introduced, and the enhanced intent embedding is obtained by matrix addition:

[0313] H′=H+G;

[0314] H′ is the intent embedding representation after adding noise;

[0315] H is the adaptive enhanced intent embedding representation;

[0316] G is the noise matrix;

[0317] In order to adapt the injected noise and ensure that high-confidence features are not overly disturbed, the strength of the noise matrix G of each feature vector is dynamically controlled according to the L2 norm of the intent embedding vector. The noise strength is calculated as follows:

[0318]

[0319] Where Ψ is the noise intensity;

[0320] ||H||2 represents the L2 norm of the intention vector;

[0321] ω is a hyperparameter that represents the basic noise intensity;

[0322] Is a constant used to avoid the situation where the denominator is left blank;

[0323] In order to avoid the sign reversal problem caused by the addition of noise to the values ​​close to zero in the original embedding, the embedding values ​​close to zero are smoothed:

[0324]

[0325] Where sign′(·) is a sign smoothing function. When the number of elements in the intention embedding vector is less than φ, the output is 1, otherwise the standard sign function is output.

[0326] H n Represents the nth element of the intent embedding vector;

[0327] |H n | is the total number of elements in the intent embedding vector;

[0328] φ is the judgment threshold for sign adjustment;

[0329] sign(·) represents a sign function that returns the sign of the input value.

[0330] In order to introduce randomness while keeping the direction of the noise consistent, a random noise vector is generated and normalized:

[0331]

[0332] in, represents the normalized Gaussian noise vector;

[0333] η represents the Gaussian noise vector corresponding to the intent embedding vector;

[0334] ||η||2 is the L2 norm of the Gaussian noise vector;

[0335] represents the standard Gaussian distribution;

[0336] According to H′, sign′(H n )and The noise matrix G is calculated:

[0337]

[0338] Where ⊙ represents the element-wise product.

[0339] 3.2.4 Dual-domain collaborative detangling graph contrastive learning

[0340] Inspired by advances in contrastive learning, we have explored contrastive enhancement and intent disentanglement techniques to address the data sparsity challenge in recommender systems. Recent research has shown that providing additional supervisory signals through the self-supervised learning (SSL) paradigm can help alleviate data sparsity and significantly improve recommendation tasks. Contrastive learning uses data augmentation methods to construct a pair of views of anchor nodes, maximizing the similarity between related samples and minimizing the similarity between unrelated samples. This is a common strategy in SSL.

[0341] Given any node, the widely used contrast loss infoNCE is calculated as follows:

[0342]

[0343] Where f(*) is the infoNCE loss function;

[0344] x′ and x″ represent the embedding representation vectors of the same embedding representation vector x under different enhanced views;

[0345] x is the embedding representation vector;

[0346] exp(*) is the exponential function with the natural base e as the base;

[0347] cos(·,·) represents cosine similarity;

[0348] The hyperparameter τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution;

[0349] y∈β represents the sampling from the mini-batch as the negative sample node of x, and β is the sampling set in the mini-batch.

[0350] Research in recent years has shown that most recommendation methods rely on various data augmentation methods to generate different enhanced contrast views to perform SSL tasks, such as structural feature enhancement and semantic enhancement. However, these methods can increase the probability of information loss in the graph structure and increase training costs. Some research has also provided inspiration by treating the graph contrastive learning process as a regularization process for all nodes in the interaction graph to reduce training costs.

[0351] In an intent-aware recommendation approach, a dual-domain graph comparative learning framework is constructed. This framework uses graph comparative learning from both the node interaction space domain and the intent perception domain to solve the problem of intent disentanglement. Specifically, based on the node interaction space analysis, two graph comparative learning strategies are utilized to optimize the node interaction space domain: a homogeneous node optimization strategy and a heterogeneous node optimization strategy. The homogeneous node optimization strategy addresses the Matthew effect, which causes node representations to be concentrated in a single direction, leading to dimensionality collapse.

[0352] By maintaining the distance between different nodes in the homogeneous node set, we can avoid the nodes from concentrating in a specific direction. The calculation method is as follows:

[0353]

[0354] in, represents the loss function for contrastive learning of isomorphic nodes;

[0355] e u and e i Represent the embedding representations of users and items respectively;

[0356] The heterogeneous node optimization strategy is primarily used to ensure accurate similarity measurements between users and items, keeping positive pairs (i.e., items with which users have interacted) as close as possible while keeping negative pairs as far apart as possible. The two contrastive learning optimization strategies in the node interaction space domain work together to simplify the computation while preserving the original structure of the information without employing any data augmentation techniques. Specifically, the contrastive loss function in the node interaction space domain is:

[0357]

[0358] in, represents the node interaction space domain loss function;

[0359] represents the contrastive learning loss function of heterogeneous nodes;

[0360] represents the expansion of the sum of the comparative learning loss function of homogeneous nodes and the comparative learning loss function of heterogeneous nodes, and <u,i>∈β represents the nodes sampled from the mini-batch.

[0361] In the intent-aware domain, graph contrast learning is used to optimize the intent representation of users and items. Unlike the node interaction space domain, to prevent intent entanglement, a contrast optimization strategy for homogeneous node intent perception is used, while optimizing heterogeneous node intent perception. In addition, the adaptive intent enhancement method proposed in 3.2.3 is used to generate an enhanced intent-aware contrast view to optimize and improve the intent entanglement problem. Specifically, the contrast loss function in the intent-aware domain is calculated as follows:

[0362]

[0363] in, is the intent-aware contrast loss function;

[0364] h′ u and h′ i These are the enhanced embedding representations of users and items respectively;

[0365] h u and h i are the embedding representations of original users and items, respectively.

[0366] 3.2.5 Model Optimization

[0367] As we can see from the task description in 3.1, the graph-based recommendation task can be viewed as a graph generation problem, predicting the likelihood of interaction edges. Given a user-item interaction graph and interaction matrix, we assume that the data for any node originates from the graph generation process. During model training, we learn the variance and mean of the user and item approximate tests. We then empirically introduce the Bayesian Personalized Ranking (BPR) loss function in combination with the KL divergence to optimize the graph generation process:

[0368]

[0369] in, represents the Bayesian personalized ranking loss function;

[0370] <u,p,n>∈β means accumulating each triple in the sampling set β, where u,p,n represent user index, positive sample and negative sample respectively;

[0371] σ(*) represents the Sigmoid activation function;

[0372] and Represents the probability distribution of positive sample and negative sample users respectively;

[0373] and Represent positive sample users and negative sample users respectively;

[0374] KL(*) is the Kullback-Leibler divergence loss calculation function, which is used to measure the difference between two probability distributions. || is the separator in Kullback-Leibler divergence, which is used to separate the probability distributions before and after.

[0375] represents the probability distribution of the embedding representation matrix E under the conditions of parameter sets Θ and R;

[0376] R represents the user-item interaction matrix; Θ = {E, H} represents the set of trainable model parameters;

[0377] represents the prior test probability distribution of the embedding representation matrix E;

[0378] Finally, in order to integrate intent modeling and recommendation tasks, a multi-task joint training strategy is adopted to optimize the self-supervised loss and the classic recommendation loss:

[0379]

[0380] in, It is the total loss function, which combines the Bayesian personalized ranking loss function, the interactive space domain loss function, the intent-aware contrast loss function and the regularization term;

[0381] λ1 and λ2 are adjustable weight parameters;

[0382] is the square of the Frobenius norm of the model parameter set Θ;

[0383] To analyze the time complexity of the proposed recommendation method, the main sources of the AIA model's time complexity were divided into graph structure encoding, intent mining modeling, and intent enhancement methods. Specifically, due to the use of LightGCN as the basic graph encoding structure, the time complexity of this process is O(2|ε|L×d), where ε represents the number of edges in the user-item interaction graph, L represents the number of layers in the graph neural network, and d represents the dimension of the node embedding; the time complexity of intent mining modeling is O(|S|×|K+M|×d), where K+M represents the number of nodes in the user-item interaction graph; and the time complexity of the intent enhancement method is O(B×d), where B represents the batch size. Therefore, the total time complexity of the AIARec recommendation method during training is O(2|ε|L×d+|S|×|K+M|×d+B×d).

[0384] 4. Experimental Analysis

[0385] To evaluate the performance and effectiveness of AIARec, we compared it with various mainstream recommendation methods on three different real-world datasets. We designed the experiments to verify the following research questions:

[0386] Question 1: How does the proposed recommendation method perform compared with the current mainstream recommendation methods?

[0387] Question 2: Do the designed key components help improve the performance of the proposed recommendation method on different datasets?

[0388] Question 3: How does the proposed recommendation method perform on data sparsity issues compared to baseline methods?

[0389] Question 4: How do key hyperparameters affect the performance of the proposed recommendation method?

[0390] 4.1 Dataset Evaluation

[0391] To evaluate the performance of AIARec, we selected three public datasets from different real-world web platforms to evaluate the performance of different recommendation tasks for different locations, venues, and products, which vary in scale, domain, and sparsity.

[0392] The datasets include Gowalla, Amazon-book, and Tmall. Gowalla: This dataset is collected from the Gowalla platform and records user check-ins at different locations based on their movement trajectories. Amazon-book: This dataset includes user ratings of products in the book category on Amazon. Tmall: This dataset contains customer purchase behavior from the online retailer Tmall. The statistical results of these datasets are shown in Table 2.

[0393] Table 2 Statistical information of experimental data set

[0394]

[0395] 4.2 Evaluation indicators

[0396] In order to verify the effectiveness of the AIARec recommendation method and to ensure the fairness and consistency of the comparative analysis experiments, the same processing method as the existing work is adopted. Specifically, all explicit feedback is forced to be converted to implicit feedback (i.e., the ratings are only 0 and 1). The items that the user interacts with are considered positive samples, and other items are considered as negative samples of the user. The performance of all recommendation models is measured by Recall@K and NDCG@K. By default, the value of K is set to 20 and 40 for evaluating Top-K recommendations. The calculation methods of Recall and NDGC are as follows:

[0397] Recall: Recall reflects the proportion of relevant items that the recommendation system can successfully find. R(u) represents the top-K recommendation list provided to the user based on the user's behavior on the training set; T(u) indicates the set of items that the user actually selected after the system recommended items to the user:

[0398]

[0399] Among them, Recall is the proportion of related items that the recommendation system can successfully find;

[0400] u is the index of the user in the training set, and U is the set of users in the training set;

[0401] CG (Cumulative Gain): Cumulative gain, which adds up the relevance scores of each recommendation result in the recommendation list. But there may be a list with high scores at the bottom and low scores at the top. The calculation method is:

[0402]

[0403] Among them, CG κ is the cumulative gain;

[0404] rel c is the relevance score of the cth recommendation result;

[0405] κ is the number of recommended results;

[0406] DCG (Discounted Cumulative Gain): The lost cumulative gain introduces the position influence factor based on CG, and the recommendation effect of the lower-ranked recommendation results is "discounted".

[0407]

[0408] DCG κ is the cumulative gain from losses;

[0409] NDCG (Normalized Discounted Cumulative Gain): Normalized loss cumulative gain, which evaluates the recommendation system's entire user recommendation list for the entire test set and normalizes the evaluation scores of recommendation lists for different users. IDCG is the list of the best recommendations returned by the recommendation system for a specific user.

[0410]

[0411]

[0412] Among them, IDCG κ is a list of the best recommendations returned by the recommender system for a particular user;

[0413] NDCG κ is the normalized loss cumulative gain;

[0414] |REL| is the total number of relevance tags in the recommendation result list;

[0415] 4.3 Experimental parameter settings

[0416] Following the general implementation method, AIARec was implemented in the Pytorch framework, and computer simulations were performed using a GPU QuadroRTX5000 and a CPU IntelCore i9-10980XE.

[0417] Specifically, the embedding size and batch size of the model are set to 32 and 1024 respectively. The default optimizer is the Adam optimizer, and the initialization is done by the Xavier method. For AIARec, the number of GCN layers is in the range of {1, 2, 3}, the number of intents in the intent modeling |S| is in the range of {16, 32, 64, 128, 256}, and the temperature coefficient τ is set to 0.2 based on experience. The constant The initial noise intensity is set to 1e-8, the initial noise intensity is {0.1, 5e-2, 1e-2, 5e-3, 1e-3}, and the image contrast learning weight is {0.1, 0.2, 0.3, 0.4, 0.5}.

[0418] 4.4 Comparison with Baseline Methods

[0419] The performance of AIARec is compared with 16 other state-of-the-art recommender system baseline methods on Recall@20@40 and NDCG@20@40.

[0420] Traditional collaborative filtering recommendation method:

[0421] MF: This is a matrix factorization based method that transforms users and items into vector representations in a latent space while considering their respective biases to reflect user preferences.

[0422] Generation-based methods:

[0423] Mult-VAE: This is a polynomial likelihood generative model based on variational autoencoders (VAE), which breaks through the performance limitations of traditional linear factor models in implicit feedback collaborative filtering tasks.

[0424] CVGA: This method addresses the time complexity and data sparsity problems faced by graph neural network recommendation systems in large-scale graph topologies, and proposes a collaborative variational graph recommendation model CVGA based on variational graph autoencoders. Through the information propagation aggregation mechanism, the user-item collaborative relationship is encoded as a behavioral probability distribution rather than an embedded representation.

[0425] GCN-based methods:

[0426] NGCF: This method explicitly incorporates collaborative signals into the embedding process by propagating embeddings on the user-item bipartite graph, thereby effectively capturing high-order connectivity in the user-item graph and improving recommendation results.

[0427] LightGCN: This recommendation method uses linear propagation and weighted summation to learn the embedding of users and items, which simplifies the design of GCN. It is not only easy to implement and train, but also shows significant performance improvements compared to advanced models such as NGCF under the same experimental settings.

[0428] Intent-based modeling approach:

[0429] DisenGCN: This study addresses the problem of insufficient representation robustness and interpretability caused by the entanglement of latent factors in graph neural networks. It establishes a causal relationship between latent factors and node neighborhood relationships through a dynamic neighbor routing mechanism, and realizes segmented convolution of specific factor features in the channel dimension.

[0430] DisenHAN: This research addresses the issues of latent factor entanglement and insufficient utilization of meta-path semantic information in heterogeneous graph neural networks. It decomposes high-order connections through meta-relationships and designs a disentangled embedding propagation layer to iteratively identify the main semantic aspects.

[0431] MacridVAE: This method addresses the problem of insufficient interpretability and controllability of representations caused by the entanglement of latent factors in user behavior. It models users' high-level concept preferences through intention-driven macro-disentanglement and combines it with information-theoretic-inspired micro-disentanglement regularization terms to achieve independent representation of low-order factors.

[0432] DGCF: This addresses the problem of suboptimal representation and insufficient relationship modeling caused by the diversity of user intent in collaborative filtering. It imposes independence constraints by iteratively constructing an intent-aware interaction graph through intent distribution modeling.

[0433] DICE: This addresses the problem of insufficient robustness and interpretability of representations due to the mixing of user herd mentality and real interests. It constructs a structurally decoupled representation of interest / herd mentality through the collision effect of causal inference.

[0434] DGCL: This study addresses the issues of suboptimal representation and insufficient interpretability caused by the entanglement of latent factors in self-supervised learning of graph neural networks, and extracts the latent semantic dimensions of the input graph through factorized representation.

[0435] Methods based on self-supervised learning:

[0436] SGL-ED: A graph self-supervised learning framework for user-item bipartite graph recommendation systems. It performs data augmentation on the input bipartite graph by performing node and edge dropout. The enhanced graph can be viewed as a subview of the original graph.

[0437] HCCF: By introducing a hypergraph-enhanced cross-view self-supervised contrastive learning architecture, it simultaneously captures local and global collaborative relationships, effectively solving the over-smoothing problem of deep graph structures and the scarcity of supervisory signals.

[0438] LightGCL: This is a recommendation model based on graph contrastive learning, which addresses the problem of random perturbations and heuristic enhancements destroying semantic structure and introducing noise bias. It uses singular value decomposition to perform contrast enhancement on global collaborative relationship modeling to achieve unconstrained structural optimization.

[0439] DCCF: This is a method that addresses the problems of the lack of a driving mechanism for user intent diversity and the introduction of noise by non-adaptive enhancement in graph comparison collaborative filtering models. It achieves latent intent factor distillation and noise suppression through global context decoupling representation.

[0440] BICF: This method analyzes user-item interactions from a causal perspective, introduces the concepts of individual intent and collective intent, and a graph contrast regularization method to unify nodes under the self-supervised and non-augmented paradigm.

[0441] 4.5 Comparative Analysis of Baseline Model Performance

[0442] Table 3 Comparison of different indicators of different baseline methods on different datasets

[0443]

[0444] Table 3 shows the performance evaluation of AIARec on three real-world datasets (Recall@20, NDCG@20, Recall@40, and NDCG@40) against 16 baseline methods, including BIGCF and DCCF. The following observations can be made from the data:

[0445] The evaluation results show that AIARec outperforms all other 16 baseline methods on all three datasets in both top-20 and top-40 settings, and significant performance improvements are observed.

[0446] Quantitative analysis shows that AIARec improves Recall@20 and @40 by 1.5% and 1.9%, respectively, and increases NDCG@20 and @40 by 1.1% and 1.3%, respectively, on the Gowalla dataset compared to the best baseline model, BIGCF. On the Amazon-book dataset, Recall@20 and @40 increase by 1.5% and 1.6%, respectively, while NDCG@20 and @40 increase by 1.1% and 1.3%, respectively. On the Tmall dataset, Recall@20 and @40 increase by 3.2% and 2.3%, respectively, while NDCG@20 and @40 increase by 3.6% and 2.7%, respectively. These experimental results demonstrate the rationality and universality of the proposed AIARec recommendation method. AIARec's superior performance may be attributed to its deeper semantic mining through user and item intent-aware modeling, enabling more accurate recommendation decisions. AIARec combines the node interaction domain and the intent perception domain to form a dual-domain synergistic graph comparative learning framework. By extracting disentangled self-supervisory signals and using them as additional training tasks, it effectively improves the data sparsity problem common in implicit feedback in recommendation systems. The proposed adaptive intent enhancement method improves the robustness of the self-supervisory signal by adaptively injecting noise into the intent representation, addressing the problem of excessive perturbation of the embedding caused by improper noise intensity settings.

[0447] Current SSL-based recommendation methods also use various data augmentation methods to enhance SSL signals, but these methods all learn the representation of users and items at a coarse-grained level, making it difficult to mine the potential intent-aware driving behaviors of the interaction between users and items. Most SSL recommendation methods have surpassed GNN-based recommendation methods in performance, indicating that SSL signals have brought positive effects to graph-based CF models. On this basis, it has also been confirmed that graph comparative learning can improve the intent entanglement problem in intent-aware modeling. AIARec has achieved improvements to recommendation methods based on intent disentanglement, and also confirmed the necessity of promoting the mining of potential behaviors behind the interaction between users and items through intent modeling, thereby improving the interpretability of recommendation methods.

[0448] 4.6 Ablation Experiment

[0449] In this summary, in order to further verify the performance of the recommendation method, an ablation experiment is designed to further discuss the effectiveness of each component in the AIARec recommendation method.

[0450] The advantages of AIARec components are demonstrated from four perspectives:

[0451] (1) Comparative learning of isomorphic nodes;

[0452] (2) Comparative learning of heterogeneous nodes;

[0453] (3) contrastive learning of intention perception graphs;

[0454] (4) Intent-aware data enhancement strategy.

[0455] Figure 2 The following table shows the comparison of the top-20 and top-20 NDGC and Recall metrics of four different variant models with the proposed recommendation method on three datasets: Gowalla, Amazon-book, and Tmall. The specific variants are as follows:

[0456] Homo: Remove isomorphic nodes for comparative learning;

[0457] Hetero: Removes heterogeneous node contrast learning;

[0458] Intent: Remove intent-aware contrastive learning;

[0459] IntentAug: Remove intent-aware data augmentation strategy.

[0460] Figure 2In the AIARec recommendation method, performance on all three datasets significantly degraded after removing the Homo module. Furthermore, performance also declined after removing the Hetero, Intent, and IntentAug modules. Experimental results demonstrate the effectiveness and necessity of dual-domain collaborative detangling graph contrastive learning. The Homo module mitigates the Matthew effect caused by popularity bias, which can lead to dimensionality collapse by concentrating node representations in a single direction. By maintaining the distances between different nodes, the Homo module avoids this problem and ensures diversity in the embedding space. This not only helps disperse the learned embeddings but also ensures a uniform distribution across the entire feature space, thereby improving model performance and generalization. The Hetero module ensures that users are closer to their positive interactions and farther away from negative samples (i.e., other non-interacting items). This allows the Hetero module to better capture the true relationship between users and items. The Intent module improves intent entanglement through graph contrastive learning, allowing each intent to independently express its specific meaning rather than being entangled with multiple intents. The IntentAug module adaptively adjusts the noise intensity injected into the intent embedding representation, thereby enhancing the SSL signal of model training to improve the robustness and performance of the model.

[0461] In order to further verify the effectiveness of the adaptive intent enhancement strategy in the AIARec method, two noise perturbation data enhancement strategies and three variant models of the data enhancement strategy were designed to conduct comparative analysis experiments with the AIARec recommended method. Figure 3 (a) and (b) show the comparison of top-40 and top-40 NDGC and Recall indicators of the variant model on three datasets: Gowalla, Amazon-book and Tmall.

[0462] Figure 3In the data augmentation, AIARec-FN represents the strategy of injecting fixed noise perturbations; AIARec-RN represents the strategy of injecting random noise perturbations; and AIARec-DN represents the strategy of removing random noise perturbations. Experimental results show that the AIARec method using adaptive data noise perturbation significantly outperforms the other three variants. This is likely due to the effectiveness of data augmentation achieved by adding adaptive noise data, which preserves the key semantic information of the embedding while enhancing the model's robustness and generalization capabilities through appropriate noise perturbations. Furthermore, observing the performance of AIARec-RN, it is found that on the Gowalla dataset, AIARec-RN outperforms both AIARec-FN and AIARec-DN variants, but is slightly inferior to the AIARec-DN variant on the Amazon-book and Tmall datasets. This phenomenon may be due to the instability of the random noise perturbation data augmentation strategy. The poor performance of the AIARec-FN variant model stems from the fact that fixed noise perturbations, due to excessive or insufficient noise intensity, can cause the model to learn low-quality supervisory signals and potentially destroy the structural information embedded in the original features, making it difficult for the model to learn useful representations. Therefore, experimental analysis demonstrates the necessity and effectiveness of adaptive noise perturbation data augmentation strategies.

[0463] 4.7 Data Sparsity Test

[0464] In order to further verify whether the proposed AIARec recommendation method is robust to the problem of data sparsity. Therefore, users are divided into five different groups according to the number of user interactions and the NDCG@40 and Recall@40 of each group are evaluated on three datasets: Gowalla, Amazon-book and Tmall, such as 0-10, 10-20, 20-30, 30-40, and 40-50. In addition, the same user interaction grouping method is applied to the three mainstream recommendation methods: LightGCN, DCCF and BIGCF. In order to intuitively show the performance of different recommendation methods in different user groups, six line graphs are used to present the impact of coefficient data on performance trends, such as Figure 4 shown.

[0465] Figure 4In the comparison analysis of different user interaction graphs in different data sets, the AIARec recommendation method shows relatively excellent performance compared with the three mainstream recommendation methods LightGCN, DCCF and BIGCF, providing more accurate recommendation results for active users and inactive users. After analysis, this result is attributed to the fact that the graph comparative learning in the synergy of the two domains not only helps to deeply understand the actual preferences of users, but also further constrains the distribution of node representations in the feature space, and the more fine-grained intent modeling better captures the diverse needs of users. In addition, it is combined with an adaptive intent enhancement strategy, which can help the model learn high-quality self-supervision signals, thereby effectively enhancing the robustness and generalization ability of the model.

[0466] 4.8 Hyperparameter Sensitivity Analysis

[0467] In order to study the impact of different hyperparameters on the performance of recommendation methods, the effects of changes in the number of GCN layers, changes in graph contrast learning weights λ1, and changes in the number of intents on the performance of AIARec are discussed on three datasets: Gowalla, Amazon-book, and Tmall.

[0468] Table 4. Impact of the number of GCN layers on AIARec on different datasets. R@20 and NDCG@20.

[0469]

[0470] The experimental results in Table 4 show that the AIARec recommendation method, with two GCN layers, can most effectively integrate graph structural information, thereby achieving excellent user preference processing. However, as the number of GCN layers increases, the recommendation method's performance gradually declines. Analysis shows that this is attributed to the fact that as the depth of the GCN model increases, the hierarchical level of the node neighborhood in the bipartite graph structure expands simultaneously. When the number of network layers reaches a certain threshold (the threshold obtained in experimental results is 2), the node neighborhood will cover the entire graph structure, resulting in a significant decrease in the diversity of the node's local subgraph. This phenomenon makes it difficult for nodes to retain unique structural information during feature learning, ultimately leading to a decline in the model's generalization ability and reduced recommendation performance. To explore the impact of changes in graph contrastive learning weights on the performance trends of the AIARec model, we retrained it on the Gowalla, Amazon-book, and Tmall datasets and evaluated it using Recall@40 and NDCG@40. The hyperparameters are selected in the range of {0.1, 0.2, 0.3, 0.4, 0.5}. In order to observe the change trend of the model after the hyperparameters are changed, the hyperparameter with the best performance is used as the base point, and then the performance of other hyperparameters is used to calculate the relative change rate with the optimal hyperparameter performance. The figure compares the change analysis of learning weights as shown in the figure. Figure 5 shown.

[0471] Figure 5 In the example, we observe that the AIARec recommendation method achieves optimal performance when the graph contrast weight λ1 is 0.4. This demonstrates that increasing the weight of graph contrast learning in the overall loss optimization can improve the overall performance of the model, demonstrating its importance in recommendation systems. However, overemphasizing graph contrast learning can lead to over-focusing on graph contrast learning and neglecting the primary recommendation task, resulting in decreased recommendation performance.

[0472] To study the impact of the number of potential intents on model performance, we observed changes in model performance by changing the hyperparameter S, selecting parameters from the range of {32, 64, 128, 256}. We retrained the model on the Gowalla, Amazon-book, and Tmall datasets and evaluated it using Recall@40 and NDCG@40. To easily observe the trend of model changes after hyperparameter changes, we used the same method of comparing learning weights to plot the impact of the number of intents on AIARec's performance. The sensitivity analysis of the hyperparameter number of intents is shown in the figure below. Figure 6 shown.

[0473] Figure 6 In the data, we can observe that as the number of potential intents increases, model performance gradually improves. When the number of potential intents reaches 128, model performance reaches its peak. Beyond 128, the model begins to show a downward trend, particularly with the Tmall dataset. Analysis suggests this may be due to the excessive number of intents blurring intent boundaries, thus affecting the model's ability to capture user preferences during intent modeling.

[0474] Through three sets of hyperparameter experiments, we can obtain the hyperparameter combination with the best recommendation performance: the number of GCN layers is 2; the image contrast learning loss weight is 0.4; and the number of intents is 128.

[0475] 4.9 Embedded Visual Analysis

[0476] To further analyze the problem of recommendation performance degradation caused by an excessive number of potential intents, we used the t-SNE dimensionality reduction method to reduce the dimensionality of the potential intent prototypes corresponding to each parameter in the range of {32, 64, 128, 256} to a two-dimensional space. We then used the k-means algorithm to cluster and visualize them, and drew potential user intent prototypes with different numbers of intents. Figure 7 (a) is the user node intention prototype diagram, and (b) is the item node intention prototype diagram.

[0477] Figure 7Observations show that, for both user and item latent intent prototypes, the distribution of intent prototypes becomes increasingly dispersed as the number of intents increases, until some potential intents begin to cluster together when the number of intents exceeds 128. This is because the granularity of latent intents is too coarse at low numbers, potentially leading to intent entanglement. As the number of intents increases, the intent factors become too detailed and redundant, introducing noise into the learned representation.

[0478] This patent proposes an adaptive intent-aware enhanced graph comparative learning recommendation method (AIARec). Intent-aware modeling is performed on the basis of graph representation learning, and the potential intention factors behind the user-item interaction are excavated. The bi-norm change of the feature vector during the training iteration is then embedded to control the noise intensity injected for each data enhancement, thereby realizing an adaptive data enhancement strategy. Finally, a dual-domain collaborative disentangled graph comparative learning framework is constructed, and comparative learning is performed in the node interaction space domain and the intent-aware domain to optimize the consistency problem of the node space. Finally, a comparative analysis experiment is conducted with 16 mainstream baseline methods on three real-world datasets, and the role of each module, robustness test, and the impact of each key parameter on the performance of the recommendation method are analyzed in the supplementary experiment. Finally, the experimental results prove the effectiveness of the AIARec recommendation method.

[0479] 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. An adaptive intent-aware enhanced graph contrast learning recommendation method, characterized by: include: S1: Construct a bipartite graph based on user-item interactions and use a graph neural network to perform multi-level embedding feature encoding on the bipartite graph to obtain embedded representations of users and items. S2, inputs the embedded representation into the variational autoencoder to obtain Gaussian distribution embedding representations of multiple potential intents, and obtains multi-intent embedding through sampling; S3, adaptively adding noise perturbation to the multi-intent embedding according to its norm to obtain an enhanced intent embedding representation; S4, implements contrastive learning in the node interaction space domain and the intention perception domain respectively; S5, output recommendation results.

2. The adaptive intent-aware enhanced graph contrast learning recommendation method according to claim 1, characterized in that: The step S1 comprises: In the recommendation scenario, the collaborative filtering paradigm recommendation system uses graph structure relationships to model the relationship between users and items; It contains a set of user sets and a set of items Among them, u K Represents a user set The Kth user in , where K is the total number of users; i M Represents an item set The Mth item in , where M is the total number of items; The user-item interaction relationship can be abstracted into a bipartite graph structure, namely: Among them, e ui =e iu =R ui ; Among them, e ui =e iu =R ui ; It is a bipartite graph consisting of users and items; is a collection of nodes; ε is the set of edges in the bipartite graph; ε ui It is the edge between users and items in the bipartite graph, that is, the connection relationship; ε iu is the edge between items and users in the bipartite graph; R ui It is the interaction value between the user and the item, which can be 0 or 1; When ε ui =ε iu =R ui = 0, it means there is no connection relationship. ui =ε iu =R ui =1, indicating that there is a connection relationship; Predicting the user-item interaction bipartite graph The probability scores above are expressed as follows: in, Indicates that the initial embedding representation matrix E is known (0) Under the conditions of and interaction matrix R, the predicted interaction matrix The conditional probability of It is represented by the user embedding representation e u , item embedding representation e i Condition, user-item bipartite graph interaction probability score; is the interaction value between the user and the item predicted by the model, and its value is in the interval [0,1]; represents the initial embedding representation matrix; Represents the user's initial embedding representation matrix; The initial embedding representation matrix representing the item; The factors that need to be added to the intention guidance at the interaction probability level are: in, Indicates the initial embedding representation matrix E (0) , user-item potential interaction matrix R and intention representation C, predict the interaction matrix The conditional probability of e u represents the user embedding representation; e i Representation of item embedding representation; represents the intention representation of user and item nodes, and S represents the set of intentions; × represents multiplication operation; Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the user set Conditional probability score of user embedding representation; Indicates the initial embedding representation matrix E (0) , the user-item potential interaction matrix R and the s-th intention in the item set Conditional probability scores of item embedding representations under the given condition; and They represent the user's intention embedding representation and the item's intention embedding representation respectively.

3. The adaptive intent-aware enhanced graph contrast learning recommendation method according to claim 2, characterized in that: The step S2 comprises: Construct a subgraph representation embedding information propagation layer to aggregate comprehensive subgraph structure information: in, is the aggregate representation of the user's neighboring nodes to the target node, K is the total number of users, and d is the number of features of the embedding vector; It is the aggregate representation of the item’s neighboring nodes to the target node, and M is the total number of items; Represents the user-item interaction graph; The normalized adjacency matrix representing the normalized user-item interaction graph; and Diagonal matrices for users and items respectively; and are the normalized diagonal matrices of users and items respectively; and Represent the embedding matrices of users and items respectively; The residual connection structure is applied to the aggregation layer to prevent over-smoothing problems. In order to aggregate high-order collaborative filtering signals, GCN information propagation is performed at different graph structure layers: in, represents the user embedding representation obtained after l-layer GCN encoding; It is the user embedding representation obtained after l-1 layer GCN encoding; It is the aggregated representation of the neighboring nodes to the target user node after being encoded by the l-1 layer GCN; It is the item embedding representation obtained after l-layer GCN encoding; It is the item embedding representation obtained after l-1 layer GCN encoding; Represents the aggregated representation of the neighboring nodes to the target item node after encoding by the l-1 layer GCN; Calculate the correlation score between the user and item embedding structures and the potential intent of users and items: in, The probability score representing the similarity of the user's preference for a specific intent category; The probability score representing the preference similarity of an item to a specific intent category; and The embedding representations of users and items respectively represent the probability scores of the preference similarity between items and specific intent categories at the lth layer; in, represents the user embedding representation obtained after l-layer GCN encoding; It is the item embedding representation obtained after l-layer GCN encoding; exp(*) is the exponential function with the natural base e as the base; s′ is the index that traverses all intent categories; Τ is used to adjust the smoothness of the softmax probability distribution; is the user's embedded representation under the sth intent category; is the user's embedded representation under the s′th intent category; is the embedding representation of the item under the sth intent category; is the embedding representation of the item under the s′th intent category; τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution; After calculating the relevance score, we can aggregate S different intent information and the global sampled subgraph to represent the user and item information as follows: in, and They represent the embedded representations of users and items respectively, and the embedded representations can be stacked and recorded as and and are the embedded representations of users and items under the l-th layer of intention perception; S represents the number of intention information; In order to obtain the posterior predictive distribution of user preference features, the encoder is defined as a variational autoencoder, and the embedded features of users and items are expressed in Gaussian forms as follows: Among them, P(e u |Ε (0) ,R,C u ) is the user embedding representation e under the condition that the initial embedding, interaction matrix and user intention set are known. u The probability distribution of C u Represents the user intent set; P(e i |Ε (0) ,R,C i ) is the item embedding representation e under the condition that the initial embedding, interaction matrix and item intention set are known i The probability distribution of C i Represents the item intention set; Indicates that the user embeds e u At the mean μ u and the variance is Normal distribution when ; Indicates that the item is embedded in e i At the mean μ i and the variance is Normal distribution when ; It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each user node in the bipartite graph; It is the result of operating the diagonal matrix of the approximate variance of the Gaussian distribution of each node of the item in the bipartite graph; is the approximate variance of the Gaussian distribution of each user node in the bipartite graph; is the approximate variance of the Gaussian distribution of each node representing the item in the bipartite graph; e u and e i Represent the embedding representations of users and items respectively; The reparameterization method is used to avoid back propagation, which is as follows: Among them, ∈ u and ∈ i denote Gaussian noise variables for users and items respectively; ⊙ denotes element-wise product; This re-parameterized adaptive fusion feature avoids manual adjustment. Combining the above formulas, the global graph structure encoding and the intent-aware graph structure synergy can be integrated: and They are the embedding representations of users and items under the l-1 layer of intent perception.

4. The adaptive intent-aware enhanced graph contrast learning recommendation method according to claim 3 is characterized in that: The step S3 comprises: A noise matrix G with the same shape as the intent embedding matrix and an adaptive enhanced intent embedding representation H are introduced, and the enhanced intent embedding representation is obtained by matrix addition: H′=H+G; H′ is the intent embedding representation after adding noise; H is the adaptive enhanced intent embedding representation; G is the noise matrix; In order to adapt the injected noise and ensure that high-confidence features are not overly disturbed, the strength of the noise matrix G of each feature vector is dynamically controlled according to the L2 norm of the intent embedding vector. The noise strength is calculated as follows: Where Ψ is the noise intensity; ||H||2 represents the L2 norm of the intention vector; ω is a hyperparameter that represents the basic noise intensity; Is a constant used to avoid the situation where the denominator is left blank; In order to avoid the sign reversal problem caused by the addition of noise to the values ​​close to zero in the original embedding, the embedding values ​​close to zero are smoothed: Where sign′(·) is a sign smoothing function. When the number of elements in the intention embedding vector is less than φ, the output is 1, otherwise the standard sign function is output. H n Represents the nth element of the intent embedding vector; |H n | is the total number of elements in the intent embedding vector; φ is the judgment threshold for sign adjustment; sign(·) represents a sign function, which returns the sign of the input value; In order to introduce randomness while keeping the direction of the noise consistent, a random noise vector is generated and normalized: in, represents the normalized Gaussian noise vector; η represents the Gaussian noise vector corresponding to the intent embedding vector; ||η||2 is the L2 norm of the Gaussian noise vector; represents the standard Gaussian distribution; According to H′, sign′(H n )and The noise matrix G is calculated: Where ⊙ represents the element-wise product.

5. The adaptive intent-aware enhanced graph contrast learning recommendation method according to claim 4, characterized in that: The step S4 comprises: Given any node, the widely used contrast loss infoNCE is calculated as follows: Where f(*) is the infoNCE loss function; x′ and x″ represent the embedding representation vectors of the same embedding representation vector x under different enhanced views; x is the embedding representation vector; exp(*) is the exponential function with the natural base e as the base; cos(·,·) represents cosine similarity; The hyperparameter τ∈[0,1] represents the temperature coefficient, which is used to adjust the prediction distribution; y∈β represents the sampling from the mini-batch as the negative sample node of x, β is the sampling set in the mini-batch; By maintaining the distance between different nodes in the homogeneous node set, we can avoid the nodes from concentrating in a specific direction. The calculation method is as follows: in, represents the loss function for contrastive learning of isomorphic nodes; e u and e i Represent the embedding representations of users and items respectively; Specifically, the node interaction space domain contrast loss function is: in, represents the node interaction space domain loss function; represents the contrastive learning loss function of heterogeneous nodes; represents the expansion of adding the comparative learning loss function of homogeneous nodes and the comparative learning loss function of heterogeneous nodes, and <u,i>∈β represents the nodes sampled from the mini-batch; The intent-aware domain contrast loss function is calculated as: in, is the intent-aware contrast loss function; h′ u and h′ i These are the enhanced embedding representations of users and items respectively; h u and h i are the embedding representations of original users and items respectively; The Bayesian personalized ranking loss function is introduced in combination with KL divergence to optimize the graph generation process: in, represents the Bayesian personalized ranking loss function; <u,p,n>∈β means accumulating each triple in the sampling set β, where u,p,n represent user index, positive sample and negative sample respectively; σ(*) represents the Sigmoid activation function; and Represents the probability distribution of positive sample and negative sample users respectively; and Represent positive sample users and negative sample users respectively; KL(*) is the Kullback-Leibler divergence loss calculation function, which is used to measure the difference between two probability distributions. || is the separator in Kullback-Leibler divergence, which is used to separate the probability distributions before and after. represents the probability distribution of the embedding representation matrix E under the conditions of parameter sets Θ and R; R represents the user-item interaction matrix; Θ = {E, H} represents the set of trainable model parameters; represents the prior test probability distribution of the embedding representation matrix E; Finally, in order to integrate intent modeling and recommendation tasks, a multi-task joint training strategy is adopted to optimize the self-supervised loss and the classic recommendation loss: in, It is the total loss function, which combines the Bayesian personalized ranking loss function, the interactive space domain loss function, the intent-aware contrast loss function and the regularization term; λ1 and λ2 are adjustable weight parameters; is the square of the Frobenius norm of the model parameter set Θ.

6. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement an adaptive intent-aware enhanced graph comparison learning recommendation method as described in any one of claims 1 to 5 when executing the executable instructions.

7. A computer-readable storage medium, characterized in that include: a memory having a computer program stored thereon; A processor is used to execute the program in the memory to implement the adaptive intent-aware enhanced graph contrast learning recommendation method described in any one of claims 1 to 5.

Images