Intelligent recommendation verification method fusing hypergraph and context information

Abstract

The invention provides an intelligent recommendation verification method fusing a hypergraph and context information, which comprises the following steps: S1, extracting prejudice article characteristics, article popularity characteristics and article semantic characteristics from an interaction graph of user articles, and combining the prejudice article characteristics, the article popularity characteristics and the article semantic characteristics of individuals and neighbor nodes to obtain an article recommendation result; coding the representation of the user and the article by adopting hypergraph convolution to obtain a coded embedding; s2, performing hypergraph convolution enhancement on the encoded embedding by adopting hypergraph convolution to obtain a user embedding representation and an object embedding representation which are subjected to hypergraph convolution optimization; and then scores of the interactive items recommended by the user are calculated, and the scores are arranged in a descending order to generate a recommendation list of related candidate items. According to the method, user article embedding representation is optimized through a self-supervised signal of hard negative supervised contrast learning, and the method is combined with a multi-label cross entropy loss optimization strategy, so that the learning of model perception on user preferences and the distinction in an embedding space are enhanced, and context features are enriched.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation, and particularly to an intelligent recommendation verification method integrating hypergraph and context information. Background Art

[0002] With the problem of information overload in online applications brought about by the development of the Internet, high-quality personalized recommendation systems have become an essential key role in daily life. Recommendation systems generally improve the quality of recommendation results by mining historical behavior data of user-item interactions. Most of them use graph convolutional neural networks to enhance the embedded representation of user-item interactions and capture collaborative signals for recommendation. In recommendation systems based on deep learning, using neural networks to learn can effectively represent complex user behaviors and item features, and explore user interests to provide accurate recommendations. Graph representation learning can more effectively capture multi-hop connections and adjacent node information between users and items, and learn the representations of users and items through graph neural networks.

[0003] When most recommendation methods improve performance, it is often difficult to comprehensively learn the feature representations in the context, resulting in poor generalization ability. Therefore, in many cases, graph neural networks play an important role in recommendation systems. However, how to introduce other effective methods to balance the contribution of graph convolutional neural networks, especially when dealing with user interest bias and popularity bias problems, is particularly crucial. Only by better learning the feature representations in the context can the influence of these biases be effectively reduced, thereby improving the performance and accuracy of the recommendation system. Summary of the Invention

[0004] The present invention aims to at least solve the technical problems existing in the prior art, and particularly innovatively proposes an intelligent recommendation verification method integrating hypergraph and context information.

[0005] To achieve the above object of the present invention, the present invention provides an intelligent recommendation verification method integrating hypergraph and context information, including the following steps:

[0006] S1, extracting biased item features, item popularity features, and item semantic features from the interaction graph of users and items, combining the biased item features, item popularity features, and item semantic features of an individual and its neighbor nodes, and then using hypergraph convolution to encode the representations of users and items to obtain encoded embeddings, where the encoded embeddings include the embedded representations of users and items;

[0007] S2. Enhance the encoded embeddings through hypergraph convolution to enhance signal propagation. Hypergraph convolution can effectively select the corresponding rows in hyperedges, effectively learn the context information in hyperedges, and reduce computational complexity. Obtain the user embedding representation and item embedding representation optimized by hypergraph convolution. Then, calculate the scores of the items recommended for the user, and sort the scores in descending order to generate a recommended list of relevant candidate items.

[0008] Preferably, S1 includes the following steps:

[0009] S1-1. Convert the user-item interaction graph information into an embedding representation of user-item interaction, and derive the historical interaction sequence H of user u and item i u,i , determine the user u interacting with the current item i, and then obtain the item embedding representation y through random sampling with a uniform distribution, where the embedding values of the item are randomly generated within the uniform distribution range;

[0010] S1-2. In the historical behavior sequence of user-item interaction, according to the items interacting with the current user, integrate the embedding vectors of all interacting items to obtain the user embedding representation x;

[0011] S1-3. Integrate the embedding representations of the items interacting with the user. Based on the item bias characteristics, item popularity characteristics, and item semantic information during the user interaction, separate the user bias in the item embedding representation, fit the biased interaction, and perform self-attention mechanism enhancement to obtain the final user embedding feature representation, and obtain the event sequence representation (x, y) of the user-item interaction characteristics.

[0012] Preferably, the calculation formulas for the user embedding representation and the item embedding representation are as follows:

[0013]

[0014] where x is the user embedding representation;

[0015] y is the item embedding representation;

[0016] y i is the embedding representation of the i-th item, and is the item interacting with user u;

[0017] H u,i is the historical interaction sequence of user u and item i;

[0018] r u is the embedding vector of the user individual bias of the original user-item interaction graph;

[0019] ||r u || is the modulus length of r u ;

[0020] sign() is the sign function;

[0021] ⊙ is the Hadamard product symbol;

[0022] Hgc(·) is the hypergraph convolution, which multiplies the normalized user embedding and item embedding, combines the information of both, and takes the updated embedding representation as the input through the encoder, fuses the user embedding and the adjacency matrix, and generates the final user embedding representation;

[0023] is the adjacency matrix;

[0024] Encoder() is the hypergraph convolution encoding representation.

[0025] Preferably, the hypergraph convolution enhancement includes:

[0026] S2-1, extracting feature information through multi-layer convolutional propagation to enhance the user embedding representation:

[0027]

[0028] where M l is the output after enhancement by the l-th convolutional layer;

[0029] is the ELU activation function;

[0030] D -1 / 2 BD -1 / 2 is the symmetric normalized adjacency matrix;

[0031] B is the bipartite graph adjacency matrix;

[0032] D represents the diagonal matrix;

[0033] W 2 is the parameter weight matrix of convolutional propagation;

[0034] S2-2, modeling the context representation of the user using hypergraph convolution, learning the context representation by the user hyper-encoding in the hypergraph, enhancing the signal propagation in the convolutional process, adjusting the critical matrix and feature matrix according to the introduced hypergraph information, and selecting the specific rows and columns corresponding to the user;

[0035]

[0036] where ψ u is the hyperedge of user u;

[0037] T is the transpose symbol;

[0038] Norm is the L2 regularization calculation;

[0039] d is the dimension in the self-attention mechanism;

[0040] is the ELU activation function;

[0041] Q, K, and V respectively represent the query, key, and value matrices in the attention mechanism;

[0042] W Q and W K are the corresponding parameter weight matrices;

[0043] Φ is a hyperparameter that controls the degree of attention;

[0044] S2-3, aggregating multiple subspace information through the multi-head attention mechanism:

[0045]

[0046] where, W 3 is the weight of each head in the multi-head attention mechanism;

[0047] k is the value of the number of attention heads;

[0048] n is the total number of attention heads;

[0049] represents the result after the input embedding at the l-th layer is processed by different numbers of attention heads;

[0050] S2-4, obtaining the user embedding representation and item embedding representation optimized by hypergraph convolution through average pooling and normalization;

[0051]

[0052]

[0053] where, x is the user embedding representation optimized by hypergraph convolution;

[0054] M i the previously processed embedding;

[0055] i ∈ H u,i means that the selection of the embedding belongs to the user-item interaction sequence;

[0056] || is the absolute value symbol, taking the number of the sequence;

[0058] H u,i represents the historical interaction sequence between user u and item i;

[0059] y is the item embedding representation optimized by hypergraph convolution;

[0060] y iis the embedded representation of the i-th item, and is the item interacting with user u;

[0061] sign() is the sign function;

[0062] ⊙ is the Hadamard product symbol;

[0063] r u is the embedded vector of the user individual bias of the original user-item interaction graph;

[0064] ||r u || is the modulus length of r u .

[0065] Preferably, the loss function enhanced by hypergraph convolution is

[0066]

[0067] where is the user set;

[0068] is the true label vector of user u's interaction with item 1;

[0069] is the true label vector of user u's interaction with item N;

[0070] α u,1 ......α u,N is the predicted score of user u for items 1 to N;

[0071] σ is the sigmoid activation function.

[0072] Preferably, it further includes contrastive self-supervised optimization: adopting the contrastive self-supervised learning method, performing cross-view contrastive learning enhancement on the multi-feature embedded representation of user-item interaction, learning the personalized features of different nodes to prevent bias; and adopting the hard negative supervision contrastive learning loss optimization strategy, selecting samples with different but similar labels in the current range as negative samples.

[0073] Preferably, the contrastive self-supervised optimization includes:

[0074] First, a sample selection strategy based on random probability is designed to calculate the distribution probability (0 to 1) of positive and negative samples in the sample set. If the sample label belongs to the current sample set then it is selected as a positive sample; otherwise, it is regarded as a negative sample; and multiple different negative sample selection structures are set in the same input sample space to assist in the selection of difficult negative samples in the contrastive learning framework;

[0075]

[0076] Among them, p(z + ) is the probability distribution of positive samples under different conditions;

[0077] z + and z - are positive samples and negative samples respectively;

[0078] z is a sample;

[0079] p U (z - ) is the probability distribution of negative samples under different conditions in the sample set;

[0080] represents the probability distribution of selecting positive samples;

[0081] represents the probability distribution of selecting negative samples;

[0082] represent the positive sample set and the negative sample set respectively;

[0083] Under the current negative sample selection structure, the work of selecting hard negative samples is carried out: after determining the anchor point, calculate the cosine similarity between the anchor point and the sample according to the sample space set, construct the hard negative sample set, and obtain the negative probability distribution p HU (z - ):

[0084]

[0085] Among them, is the negative sample expectation within the currently defined range, i.e., Ψ HU ;

[0086] Ψ HU is the specific range of current conditional sample selection;

[0087] exp() is the exponential operation with the natural base;

[0088] w is the sample label;

[0089] is the sample label set;

[0090] d(z, ξ) is the similarity between the currently selected node (sample z) and the anchor point;

[0091] is the threshold for controlling the sampling hardness;

[0092] Under the sample space of combining the same anchor point, the method of selecting samples that are similar to the anchor point but of different types is used to construct the hard negative sample set and obtain the corresponding negative probability distribution:

[0093]

[0094] Among them, p S (z - ) is the negative probability distribution for negative sample selection under conditions;

[0095] The label information w of the sample z and the label information w of the anchor point ξ are comprehensively considered; finally, the above two sets of hard negative samples are combined to construct a hard negative sample selection set based on the contrast learning framework; this set selects negative samples within a specified range, satisfying two conditions: being similar to the anchor point but having different label types, and satisfying the selection conditions within the specified selection range

[0096]

[0097] Among them, p HS (z - ) is the negative probability distribution of hard negative samples that meet the conditions;

[0098] Ψ HS 、Ψ S 、Ψ HU respectively represent different limited ranges in the sample selection process.

[0099] Preferably, the contrast loss function of the hard negative samples is:

[0100]

[0101] Among them, d(·,·) represents calculating the similarity;

[0102] n is the number of negative samples;

[0103] k is the negative sample index;

[0104] represents the hard negative samples belonging to the current limited range;

[0105] In addition, in order to effectively balance the user's bias characteristics of items, the user's current interaction item characteristics, and the item characteristics that the user may interact with in the future, and prevent the problem of uneven node embedding. By changing the weighting scheme to balance the gradient and reducing the recommendation that is affected by popular items on the user's true preferences, at this time, the loss function is expressed as follows:

[0106]

[0107] Among them, λ is a hyperparameter for balancing the weight of the item embedding matrix interacting with the user;

[0108] λ 1 For use in the embedding matrix of items with which the user has interacted;

[0109] ⊙ is the Hadamard product symbol;

[0110] η is a weight vector used to weight the losses of different items;

[0111] Θ is a regularization coefficient that controls the regularization weight;

[0112] Is the previous loss function;

[0113] Is the set of users;

[0114] H u,i Is the historical interaction sequence between user u and item i.

[0115] Preferably, the formula for calculating the score of the recommended interaction item for the user is as follows:

[0116] β = <x||μ·r u ,y||μ·r i > (14)

[0117] Among them, β represents the score of the recommended interaction item for the user;

[0118] <·,·> is the inner product operation;

[0119] || is the concatenation operation;

[0120] <x||μ·r u ,y||μ·r i > When calculating, first perform concatenation respectively and then perform the inner product operation;

[0121] x and y represent the user embedding representation and the item embedding representation respectively;

[0122] μ represents a hyperparameter that adjusts the influence of popularity bias;

[0123] r u and r i respectively represent the bias vectors of the user and the item.

[0124] Preferably, it also includes verifying the results of the generated recommended list of relevant candidate items by calculating recall rate, precision rate, cumulative gain, discounted cumulative gain, and normalized discounted cumulative gain.

[0125] In summary, due to the adoption of the above technical solutions, the present invention optimizes the user-item embedding representation through the self-supervised signal of hard negative supervised contrastive learning and combines it with the multi-label cross-entropy loss optimization strategy, enhancing the model's perception of user preference learning and discriminability in the embedding space, and enriching the context features. Specifically, the beneficial effects are as follows: In order to effectively alleviate the problems of user interest bias and popularity bias, high-quality graph collaborative interaction information is obtained through contrastive self-supervised learning, and self-supervised signals are introduced to enhance the model performance, so as to better learn the context feature information. This method not only optimizes the modeling of user individual bias but also effectively alleviates the problem of exacerbating popularity bias due to excessive noise in the interaction information. In addition, combining the self-supervised signal of hard negative sample supervised contrastive learning to optimize the embedding representations of users and items and combining it with the multi-label cross-entropy loss optimization strategy further enhances the model's perception ability of user preference, improves the discriminability of the embedding space, and thus enriches the learning of context features.

[0126] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0127] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the description of the embodiments in conjunction with the following drawings, wherein:

[0128] Figure 1 is the DDCLRec framework diagram of the present invention.

[0129] Figure 2 is a schematic diagram of the performance comparison of DDCLRec, EEDN, and CaDRec in terms of Precision@k, Recall@k, and NDCG@k on four datasets such as ML-1M and Yelp2018 in the sparse user-item interaction graph scenario; Figure 2 (a) is ML-1M, Figure 2 (b) is Yelp2018, Figure 2 (c) Yelp2018, Figure 2 (c) Douban-book, Figure 2 (d) is Foursquare.

[0130] Figure 3 is a schematic diagram of the performance comparison of DDCLRec, EEDN, and CaDRec in terms of Precision@k, Recall@k, and NDCG@k on four real datasets such as ML-1M and Yelp2018 in the random embedding vector scenario; Figure 3 (a) is Precision@5, Figure 3 (b) is Precision@10,Figure 3 (c) is Precision@20, Figure 3 (d) is Recall@5, Figure 3 (e) is Recall@10, Figure 3 (f) is Recall@20, Figure 3 (g) is NDCG@5, Figure 3 (h) is NDCG@10, Figure 3 (i) is NDCG@20.

[0131] Figure 4 It is a schematic diagram analyzing the change of the Precision@k index of DDCLRec compared with other models on four real datasets such as ML-1M and Yelp2018. Figure 4 (a) is ML-1M, Figure 4 (b) is Yelp2018, Figure 4 (c) is Douban-book, Figure 4 (d) is Foursquare.

[0132] Figure 5 It is to analyze λ 1 and τ, and λ 2 and τ on the impact of the Recall@k and NDCG@k indexes of DDCLRec on the dataset ML-1M. Figure 5 (a) is λ 1 and τ on Recall@10 on ML-1M, Figure 5 (b) is λ 1 and τ on Recall@20 on ML-1M, Figure 5 (c) is λ 1 and τ on NDCG@10 on ML-1M, Figure 5 (d) is λ 1 and τ on NDCG@20 on ML-1M, Figure 5 (e) is λ 2 and τ on Recall@10 on ML-1M, Figure 5 (f) is λ 2 and τ on Recall@20 on ML-1M, Figure 5 (g) is λ 2 and τ on NDCG@10 on ML-1M, Figure 5 (h) is λ 2 and τ on NDCG@20 on ML-1M. Specific implementation manner

[0133] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0134] Compared with traditional recommendation methods, the recommendation system based on graph neural networks can better capture complex user-item interaction relationships. By using the rich information of the interaction graph structure to model the complex interaction relationships between users and items, optimizing the embedded encoding representations of users and items, effectively dealing with the data sparsity problem, and enhancing the interpretability of the recommendation system. However, overusing graph neural networks will also be accompanied by problems such as high computational complexity, data sparsity, and interpretability, and there are potential challenges such as constructing high-quality interaction graphs.

[0135] The introduction of contrastive learning has greatly increased the inclusiveness of recommendation methods and is relatively general for any task. It reduces the dependence on labeled data, improves the representation ability of the model, and enhances the robustness of the model. It has the advantages of lightweight model and flexible design, reduces cost consumption, and improves the generalization ability of the model. However, contrastive learning depends on the selection of negative samples, and improper selection strategies will affect the model's learning features, and there are relatively high requirements for the selection of negative samples.

[0136] Self-supervised learning, as an effective paradigm for recommendation enhancement, can train a large amount of unlabeled data and improve recommendation performance. The input data itself is the supervision, which is beneficial to various types of downstream tasks.

[0137] In recent years, with the continuous improvement of contrastive self-supervised recommendation methods, new recommendation methods are no longer limited to recommendation accuracy, but better promote the diversity of recommendations. Hypergraph convolutional networks can express richer relationship information through hyperedges, effectively process multi-type data relationships, and capture high-order relationships between nodes.

[0138] To better solve the problem of over-smoothing of node embeddings in the recommendation system, by combining the advantages of diffusion models and contrastive learning, selectively retaining more valuable embedded representation information, improving recommendation performance while promoting the diversity of recommendations, a contrastive learning debiasing recommendation method based on diffusion models, DDCLRec (Diffusion-based debiased contrastive learning for recommendation), is proposed.

[0139] The user-item interaction process is affected by many factors such as popularity and exposure, and the recommendation effect is affected by the differences in rating standards of different users. Some users will give high scores, some will give low scores, and some will give medium scores. This difference will interfere with the accurate capture of the true semantics of the item. At the same time, different users have different focuses on specific dimensions, which will widen the semantic distribution of items and expand the differences. In actual recommendation scenarios, user interest bias will also fluctuate due to factors such as the accumulation effect of historical behaviors and short-term interest hotspots, which will increase the interference with the true semantics of items and put forward higher requirements for the robustness of recommendation methods. However, the user interest bias factor can be used as a learnable perturbation, which can serve as an important basis for the change in the semantic distribution of items, and make corresponding strategies based on this change, which is more conducive to improving the robustness of recommendation methods. Under the interaction of hypergraph convolution enhancement and diffusion model simulation, we can learn important features more effectively, reconsider the generation of user-item embedding representations, balance user interest biases in the recommendation process, update item embeddings by balancing gradients through different weighting schemes, and use regularization to further balance embedding updates, prevent popular items from occupying too large a weight in the embedding space, and tap into users' true preferences.

[0140] The key symbols used in the present invention are shown in Table 1.

[0141] Table 1 Mathematical symbols

[0142]

[0143] The task is formulated as follows: Input: user-item interaction graph, Output: Use the denoised user-item interaction embedding representation to generate a recommendation list containing relevant candidate items for the user, and predict the items that the user ultimately wants to interact with based on the user's preferences.

[0144] For the user-item interaction graph, the biased item features of individuals and neighboring nodes, item popularity features and item semantic features are combined to obtain a multi-feature interaction representation of users to items. Hypergraph convolution is used to encode the representations of users and items to obtain the enhanced embedding representation of user-item interaction. The diffusion model is introduced to incorporate a variety of semantic information and bias features to enhance the embedding representation of users and items and model user preferences. Secondly, hypergraph convolution is used to enhance the signal propagation process and effectively learn the contextual information in the hyperedge. Hypergraph convolution is combined with the self-attention mechanism to assist information diffusion and aggregate the semantic information of neighboring nodes. Finally, through the contrastive self-supervised learning method, the multi-feature embedding representation of user-item interaction is enhanced by cross-view contrastive learning. The hard negative supervision contrastive learning loss optimization strategy is adopted, and the multi-label cross entropy loss and NCE loss are combined as the objective function to better utilize contextual information. The multi-label cross entropy loss corresponds to The NCE loss is a commonly used loss function in contrastive learning and is included in it. It corresponds to the contrastive optimization loss with hard negative sample sampling.

[0145] The overall framework of DDCLRec proposed is as Figure 1 shown: For the user-item interaction graph, combining the biased item features of individuals and neighbor nodes, item popularity features, and item semantic features, hypergraph convolution is used to encode the representations of users and items. Among them, three different token information, personal bias information, semantic information, and popular item information, are considered and injected as learnable perturbations into the hypergraph convolution operation to decouple the popularity bias and user individual bias in user-item interactions and generate de-biased user and item representations. A diffusion model is introduced to integrate multiple semantic information and bias features to enhance the embedding representations of users and items. The diffusion process is first a forward process, gradually adding Gaussian noise to disrupt the interaction history of users. In this process, a noise limit is set to retain user personalized information. In the reverse process, the original interaction data is gradually recovered from the interaction history damaged by noise through a parameterized neural network, and the interaction probability is recovered for ranking and recommending un-interacted items.

[0146] Secondly, use hypergraph convolution to enhance the signal propagation process and effectively learn the context information in hyperedges. Use slicing operations to enable hypergraph convolution to effectively select the corresponding rows in hyperedges and reduce the computational complexity. Inject the attention mechanism as a perturbation into the convolution process to select effective neighbor nodes, assist in information diffusion, and aggregate the semantic information of adjacent nodes. The multi-head hypergraph convolutional network layer learns features in multiple subspaces to generate diverse representations, and finally obtains the user representation through average pooling and normalization to learn context information.

[0147] Finally, through the contrastive self-supervised learning method, cross-view contrastive learning is enhanced for the multi-feature embedding representations of user-item interactions to learn the personalized features of different nodes and prevent bias. Adopt the hard negative supervised contrastive learning loss optimization strategy, preferentially select samples with different but similar labels within the current range as negative samples, and combine the multi-label cross-entropy loss and the NCE loss as the objective function to learn the features of real data, distinguish real data from noise data, and improve the generalization ability and robustness of the model.

[0148] 1. Multi-feature Interaction Representation Learning

[0149] 1.1 Extract the multi-feature embedding representations of user-item interactions

[0150] According to the interaction information of different users on the interaction graph, the obtained information is converted into entity embedding representations and relationship embedding representations. The bias item features, item popularity features, and item semantic features of user nodes and neighbors on the graph are embedded to construct a multi-feature interaction embedding representation. The user-item interaction graph information is converted into an embedding representation of user-item interaction, and the historical interaction sequence H of user u and item i is derived. u,i Determine which users u interact with the current item i. The embedding representation y of the item is obtained through random sampling from a uniform distribution, where the embedding values of the item are randomly generated within the range of the uniform distribution. In the historical behavior sequence of the current user-item interaction, according to the items that interact with the current user, by integrating the embedding vectors of all interacting items, the user embedding representation x is obtained. Integrating the embedding representations of the items that interact with the current user is to separate the user bias in the item embedding representation according to the item bias feature, item popularity feature, and item semantic information during the user interaction process, fit the biased interaction, and perform self-attention mechanism enhancement to obtain the final user embedding feature representation, and obtain the event sequence representation (x, y) that enriches the user-item interaction features.

[0151]

[0152] Among them, x is the user embedding representation;

[0153] y is the embedding representation of the item;

[0154] y i is the embedding representation of the i-th item, and is the item that interacts with user u;

[0155] H u,i is the historical interaction sequence of user u and item i;

[0156] ||r u || is the modulus length, This calculation is a regularization calculation;

[0157] sign() is, for example, sign(x). If x is positive, output 1; if x is negative, output -1; if x is 0, output 0. The purpose is to extract the sign information of x therein and perform element-wise multiplication with the subsequent content.

[0158] ⊙ is element-wise multiplication, Hadamard product. The corresponding elements in two vectors or matrices are multiplied, and the result is a new vector or matrix, where each element is the product of the corresponding elements of the original two vectors or matrices.

[0159] Hgc(·) multiplies the normalized user embedding and the item embedding, combines the two pieces of information, and takes the updated embedding representation as the input through the encoder, fuses the user embedding and the adjacency matrix, and generates the final user embedding representation.

[0160] r u The embedded vector of the user individual bias for the original user-item interaction graph;

[0161] Is the adjacency matrix;

[0162] Encoder() is the hypergraph convolutional encoding representation injecting the self-attention mechanism.

[0163] To better model user preferences, semantic information is introduced to separate item biases. A preference evaluation is set for the current user and calculated through the inner product of the user and the item.

[0164] α = <x, y> (3)

[0165] Where x is the embedded representation of the user, y is the representation of the item interacted with, and <> is the inner product operation.

[0166] 1.2 Probability Diffusion Paradigm of Interaction

[0167] For the event sequence representation (x, y) containing rich user-item interaction features obtained, a diffusion model is introduced to enhance user preference modeling and reduce the influence of irrelevant features in the recommendation process. Unify the collaborative signals and feature information of user-item interactions. By disrupting the original user-item interactions, iterative learning is used to restore the initial state through probability diffusion, and iterative denoising training integrates the information into the user-item interaction embedded representation to mitigate the negative impact of noise features. Introduce a diffusion process to the user-item interaction graph. First, gradually introduce Gaussian noise to disrupt the original user-item interaction graph, gradually disrupting the interactions between users and items to simulate the negative impact of noise features. Second, in the reverse process, focus on learning and denoising the damaged graph connection structure, aiming to gradually refine the damaged interaction information to restore the original interactions between users and items.

[0168] In the forward diffusion process, define the interaction z between a currently selected user and a set of items u =(x, y) u =[(x, y 0 ),(x, y 1 ),…,(x, y I )], initialize the diffusion process γ 0 =z u .

[0169] By gradually introducing Gaussian noise at step S, parameterize the whole process with s as the index, gradually introduce noise to the user-item interaction, and evolve the original interaction into a noisy state.

[0170]

[0171] Among them, δ(γ s |γ s-1 ) destroys the user's interaction history by gradually adding Gaussian noise in the forward process.

[0172] is a Gaussian distribution;

[0173] s ∈ {0, …, S} is each time step, and the current state γ s-1 is transferred to the state γ s by adding noise. Here, the noise follows a Gaussian distribution with a mean of and a variance of μ s ·I.

[0174] Two hyperparameters and are introduced to control the noise at the current state through a linear noise scheduler, and noise is gradually added to mimic the randomness in user-item interactions.

[0175]

[0176] Among them, γ s is the state at the current step in the diffusion;

[0177] γ 0 is the state at the initial step;

[0178] is the scaling factor, representing the intensity of the added noise;

[0179] controls the scale of the Gaussian noise added at each step (i.e., the s-step), and its value range is (0, 1);

[0180] ε is the standard Gaussian noise, following a Gaussian distribution with a mean of 0 and a covariance matrix of the identity matrix I

[0181] Performs a linear noise schedule for the current state.

[0182]

[0183] Among them, controls the scale of the Gaussian noise added at each step (i.e., the s-step), and its value range is (0, 1);

[0184] The hyperparameter c ∈ [0, 1] controls the noise scale, and the two hyperparameters η max and η min represent the upper and lower bounds of the added noise.

[0185] During the reverse diffusion process, the current noisy state is eliminated and the initial state is restored, enabling the diffusion model to effectively capture subtle changes in complex generation processes. Starting from the current state γ s The denoising transition steps gradually restore user-item interactions.

[0186]

[0187] where p θ (γ s-1 |γ s ) represents the reverse process in diffusion, which predicts the conditional probability distribution of the previous state given the current state. χ θ (γ s , s) is the mean and Σ θ (γ s , s) is the covariance matrix. These two are the mean and covariance predicted by a parameterized neural network based on the current state and time step.

[0188] represents a Gaussian distribution, and the content in the brackets specifically demonstrates the gradual addition of noise in different states and finally conforming to the Gaussian distribution.

[0189] γ s-1 represents the state at the previous time step;

[0190] χ θ (γ s , s) and Σ θ (γ s , s) are the mean and covariance representing the predicted Gaussian distribution, respectively, generated by two neural networks with learnable parameters.

[0191] 1.3 Diffusion Optimization

[0192] To effectively guide the learning of θ in the reverse graph diffusion process, maximize the evidence lower bound (ELBO) of the negative log-likelihood of user interactions. This is achieved by optimizing the log-likelihood function of the latent variable model, maximizing the log-likelihood of the result to enhance the model's generative ability and maintain the model's stability and diversity. Define which reflects the model training objective at different time steps s during the diffusion model optimization process. It is optimized at different stages according to the changes in different steps to effectively obtain necessary information from the noise.

[0193]

[0194] where is the KL divergence between the denoising distribution predicted by the model and the true posterior distribution;

[0195] is the reconstruction term, which represents the ability of the model to reconstruct the original data in a given state, and calculates the expected value of the log-likelihood of the initial state under the condition of the current time step;

[0196] is the denoising matching term, which elaborates on the denoising overshoot step and describes how to gradually restore the user's interaction history during the reverse process. It measures the KL divergence between the denoising distribution predicted by the model and the true posterior distribution at each time step;

[0197] logp(γ 0 ) represents the log-likelihood of the original user interaction data and measures the ability of the model to restore the original user interaction data during the reverse process;

[0198] p θ (γ 0 |γ 1 ) is the reverse process in diffusion, which predicts the conditional probability distribution of the previous state given the current state.

[0199] D KL (δ(γ s-1 |γ s ,γ 0 )||p θ (γ s-1 |γ s )) measures the difference between the true posterior distribution δ(γ s-1 |γ s ,γ 0 ) and the posterior distribution p θ (γ s-1 |γ s ) predicted by the model;

[0200] δ(γ s-1 |γ s ,γ 0 ) represents the true posterior distribution, which refers to the true distribution of the previous state γ s given the current state γ 0 and the initial state γ s-1 during the diffusion process, and is gradually defined by adding Gaussian noise;

[0201] γ 0 represents the initial state;

[0202] p θ (γ s-1 |γ s ) is the reverse process in diffusion, which predicts the conditional probability distribution of the previous state given the current state.

[0203] When s = 0,

[0204] s is the time step;

[0205] γ s is the state at the current step.

[0206] In the diffusion process, the Gaussian distribution reflects the reverse update of the whole process, ensuring that the reverse sampling in each step can be completely restored to the state before the noise is introduced, and finding the complete user-item interaction information from the noise.

[0207] By adjusting the parameters therein, the accuracy of the reverse update is changed. The mean value is used to calculate and obtain the current diffusion state and the initial diffusion state, ensuring the effectiveness of its reverse recovery while ensuring the data diversity characteristics.

[0208]

[0209]

[0210] where ω(γ s , γ 0 , s) is the mean expression for each step in the calculation of the reverse process, describing how to calculate the mean of the next step given the current noise state, the original state, and the time s.

[0211] ω θ (γ s , s) is an elaboration of the transition process for facilitating the understanding of the calculation of the reverse process. The final change is to push Equation 11 closer to Equation 10;

[0212] is the noise ratio parameter at time step s, which controls the amount of noise added in each step;

[0213] is the cumulative noise ratio parameter, representing the cumulative noise ratio from the initial state to the current time step s;

[0214] s is the time step;

[0215] ω is the mean of the predicted state at the previous time step given the current state γ s and the initial state.

[0216] is the prediction of the initial state γ s based on the existing state γ 0 . Receiving the embedding of the current state and the time step as input, it outputs the prediction of the initial state γ 0 .

[0217] Calculate the reverse distribution probability from time step s to s - 1 in the current diffusion state. Optimize the KL divergence to measure the approximate distribution of the reverse probability distribution, and use Bayes' rule to repair the expression of the reverse probability distribution.

[0218]

[0219] Among them, δ(γ s-1 ∣γ s ,γ 0 ) represents how to estimate the conditional probability distribution of the previous state given the current state and the initial state.

[0220] ∝ means proportional, that is, the probability density on the right is proportional to the conditional probability on the left.

[0221] represents the Gaussian distribution (normal distribution).

[0222] ω(γ s ,γ 0 , s) represents the mean expression for each step in the reverse process, describing how to calculate the mean of the next step given the current noise state, the original state, and time s.

[0223] I represents the identity matrix.

[0224] Variance Add randomness to the sampling process to ensure the diversity of the generated data.

[0225] When the initial state is transformed into the state at a certain time step, it is for the model to learn to capture complex dynamic processes. To effectively improve the training efficiency, prevent the instability caused by introducing noise, and avoid the model falling into local minima and affecting the final training effect. Simplify the state change in the reverse process to ensure the training efficiency and stability, and instantiate through a multi-layer perceptron (MLP) And based on γ at step s s State to predict the initial state γ 0 , define the loss function for each time step.

[0226]

[0227] Among them, ‖‖ 2 Is the two-norm;

[0228] Is the prediction of the initial state γ s Based on the existing state γ 0 . Receive the embeddings of the current state and time step as input, and output the initial state γ 0 .

[0229] 1.4 Debiasing Optimization

[0230] In order to better adjust the embedding representations of users and items during training, introducing popularity bias can effectively improve the learning ability of the model. During training, the popularity biases of users and items are added, but only the representations of users and items are considered in the test phase. By combining the popularity feature information and interaction item bias information of the interaction items in the user context, the embedding representation of the user is enriched, and the score for recommending interaction items to the user is set.

[0231] β = <x||μ·r u ,y||μ·r i > (14)

[0232] Among them, β represents the score for recommending interaction items to the user;

[0233] <·,·> is the inner product operation;

[0234] || is the concatenation operation;

[0235] <x||μ·r u ,y||μ·r i > When calculating, first perform concatenation respectively and then perform the inner product operation;

[0236] x and y represent the user embedding representation and the item embedding representation respectively;

[0237] μ represents the hyperparameter that adjusts the influence of popularity bias; changes the influence level of popularity.

[0238] r u and r i respectively represent the bias vectors of the user and the item.

[0239] According to the historical behavior data of the user's past interactions and the relevance of the user-item interactions in the interaction graph, the embeddings of the items interacted with the user are obtained

[0240]

[0241] Among them, is the embedding matrix of the items interacted with by the user in the past;

[0242] is the embedding matrix of the items that the user may interact with in the future.

[0243] The score for the items recommended to the user is obtained through the inner product of the user embedding representation that only contains the items interacted with by the user in the past and the representation of the items that the user may interact with, and the cross-entropy loss function is calculated.

[0244]

[0245] Among them, W 1T is the transpose of W 1 , where W 1 is the parameter weight matrix for calculating the cross - entropy loss function;

[0246] σ is the activation function.

[0247] Based on the cross - entropy loss function, by combining parameter weights and regularization, the loss gradient is adjusted so that it can learn more accurate user - item interaction relationships that are not biased towards popular items or user personal preferences, reducing the risk of one - sided prediction of items interacted with users caused by personal bias. Set a gradient update parameter weight matrix Z. Through partial derivative operations, use specific optimization algorithms such as gradient descent to adjust the parameters, minimize the overall loss function to improve the model recommendation accuracy, and make the model reduce popularity bias while adapting to data characteristics. The parameter matrix adjustment process is divided into three categories: one is the existing user - item interaction in the current interaction scenario, the second is the possible user - item interaction in the future interaction scenario, and the third is the non - interaction situation. Introduce the embedding matrix of items that may interact with users in the future to correct possible bias situations in the future. Compare and analyze the influence of bias in different interaction scenarios through the hyperparameter weight matrix Z, and adjust the parameter matrix Z through gradient calculation to reduce the influence brought by bias error, making the model have better robustness.

[0248]

[0249] 2. Context Representation Contrast Enhancement

[0250] 2.1 Hypergraph Convolution Enhancement

[0251] Use the attention mechanism as a perturbation injection into the hypergraph convolution operation, learn features in multiple sub - spaces, effectively select neighbor nodes, and consider context and sequence context during propagation.

[0252] Enhance the user embedding in the way of graph convolution. First, extract feature information through multi - layer convolution propagation to enhance the user embedding representation.

[0253]

[0254] Among them, is the output after enhancement by the l - th convolution layer;

[0255] is the ELU activation function;

[0256] D -1 / 2 BD -1 / 2 calculates the symmetric normalized adjacency matrix;

[0257] is the bipartite graph adjacency matrix. If the same user and item i are in Bf and item i g interacts, then the b in the matrix f,g position is 1, otherwise 0.

[0258] represents a diagonal matrix.

[0259] is the parameter weight matrix for convolutional propagation.

[0260] Model the context representation of users using hypergraph convolution. Enhance the signal propagation in the convolution process by learning the context representation from the user hyperedges in the hypergraph. Adjust the adjacency matrix and the feature matrix according to the introduced hypergraph information, and specific rows and columns corresponding to the user can be selected. Combine hypergraph convolution with the attention mechanism to enhance the ability to select effective neighbors, improve the information diffusion in the hypergraph neural network, and consider the multi-layer implications of structure and order to promote different context-related representations.

[0261]

[0262] where, ψ u is the hyperedge of user u;

[0263] T is the transpose symbol;

[0264] Norm is the L2 regularization calculation;

[0265] d is the dimension in the self-attention mechanism, which is the dimension of the query matrix Q and the key matrix K and is used for the calculation of the self-attention mechanism.

[0266] is the ELU activation function;

[0267] Q, K, and V respectively represent the query, key, and value matrices in the attention mechanism;

[0268] W Q and W K are the corresponding parameter weight matrices;

[0269] Φ is the hyperparameter that controls the attention degree.

[0270] Then aggregate the information of multiple subspaces through the multi-head attention mechanism.

[0271]

[0272] where, W 3 is the weight of each head in the multi-head attention mechanism;

[0273] k is the value of the number of attention heads;

[0274] n is the total number of attention heads;

[0275] represents the result after processing the input embedding at the l-th layer through different numbers of attention heads;

[0276] Then, through average pooling and normalization, the user embedding representation optimized by hypergraph convolution is obtained.

[0277]

[0278] M i previously processed embeddings;

[0279] i ∈ H u,i is the selection of the embedding belonging to the user-item interaction sequence;

[0280] || is the absolute value symbol;

[0281] H u,i represents the historical interaction sequence between user u and item i;

[0282] The hypergraph convolution enhancement process is optimized through a multi-label cross-entropy loss function.

[0283]

[0284] Among them, is the true label vector of the current user-item interaction;

[0285] is the true label vector of the user-item interaction; and when the user has accessed the item, each dimension of this vector is equal to 1; otherwise, it is 0.

[0286] α u,1 α u,N are the predicted scores of the model for user u on items 1 to n respectively, and each represents the predicted score of the user for the item.

[0287] σ is the sigmoid activation function.

[0288] 2.2 Comparative self-supervised optimization

[0289] In a self-supervised recommendation system, introducing a contrastive learning framework and combining a hard negative sampling strategy can significantly improve the learning performance. A self-supervised recommendation method based on hard negative sampling is adopted. This method fully utilizes label information and the selection of hard negative samples during the training process to optimize the learning process of the model. First, a sample selection strategy based on random probability is designed to calculate the distribution probability (0 - 1) of positive and negative samples in the sample set. In this strategy, if the sample label belongs to the current sample set It is selected as a positive sample; otherwise, it is regarded as a negative sample. Further, multiple different negative sample selection structures are set within the same input sample space to assist in the selection of hard negative samples in the contrastive learning framework. The similarity between samples is calculated through an auxiliary function, and the final loss function is inferred based on this, thereby optimizing the learning process of the model. This process is similar to unsupervised contrastive learning, where negative samples are randomly selected throughout the input space.

[0290]

[0291] Among them, p(z + ) is the probability distribution of positive samples under different conditions;

[0292] z + and z - are positive samples and negative samples respectively;

[0293] z is a sample;

[0294] p U (z - ) is the probability distribution of negative samples under different conditions in the sample set;

[0295] represents the probability distribution of selecting positive samples;

[0296] represents the probability distribution of selecting negative samples;

[0297] represent the positive sample set and the negative sample set respectively;

[0298] Under the current negative sample selection structure, the work of selecting hard negative samples is carried out. After selecting the anchor point, the cosine similarity between the anchor point and the samples is calculated according to the sample space set, a hard negative sample set is constructed, and the negative probability distribution p HU (z - ) of hard negative sample selection is obtained.

[0299]

[0300] Among them, is the negative sample expectation of the currently selected range, that is, Ψ HU .

[0301] Ψ HU is the specific range of current conditional sample selection.

[0302] exp() is the exponential operation with the natural base;

[0303] w is the sample label;

[0304] is the set of sample labels;

[0305] d(z, ξ) is the similarity between the currently selected node (sample z) and the anchor point;

[0306] is the threshold for controlling the sampling hardness.

[0307] Under the sample space that combines the same anchor points, a method of selecting samples that are similar to the anchor points but have different types is used to construct a hard negative sample set and obtain the corresponding negative probability distribution.

[0308]

[0309] where p S (z - ) is the negative probability distribution of negative sample selection under the condition.

[0310] The label information w of the sample z and the label information w of the anchor point ξ are comprehensively considered. Finally, the above two hard negative sample sets are combined to construct a hard negative sample selection set based on the contrast learning framework. This set selects negative samples within a specified range, satisfying two conditions: being similar to the anchor point but having different label types, and satisfying the selection conditions within the specified selection range

[0311]

[0312] where p HS (z - ) is the negative probability distribution of hard negative samples that meet the conditions;

[0313] Ψ HS 、Ψ S 、Ψ HU respectively represent different limited ranges in the sample selection process.

[0314] Based on the above preparation for hard negative sample selection, a hard negative sample contrast loss function is established, allowing the sample selection set that meets the above two requirements to be included in the calculation, maximizing the similarity between the anchor point and the positive sample, while minimizing the similarity between the anchor point and the negative sample, pulling similar samples closer and pushing dissimilar samples farther away in the current sample space.

[0315]

[0316] where d(·,·) represents calculating the similarity;

[0317] n is the number of negative samples;

[0318] k is the negative sample index;

[0319] Indicates difficult negative samples within the current defined range;

[0320] To effectively balance the user's bias features of items, the user's current interaction item features, and the item features that the user may interact with in the future, and prevent the problem of uneven node embedding. By changing the weighting scheme to balance the gradient, reduce the recommendation that is affected by popular items and influences the user's true preferences.

[0321]

[0322] Among them, λ is a hyperparameter that balances the weights of the item embedding matrix for user interactions;

[0323] λ 1 Is used for the item embedding matrix that the user has interacted with;

[0324] ⊙ is the element-wise multiplication, Hadamard product. The corresponding elements in two vectors or matrices are multiplied, and the result is a new vector or matrix, where each element is the product of the corresponding elements of the original two vectors or matrices.

[0325] η is the weight vector used to weight the losses of different items;

[0326] Θ is the regularization coefficient that controls the regularization weight;

[0327] Is the previous loss function;

[0328] Is the user set;

[0329] H u,i Is the historical interaction sequence between user u and item i;

[0330] λ is used for the item embedding matrix that the user may interact with in the future 2 .

[0331]

[0332] 3. Experimental Analysis

[0333] By comparing with various latest recommendation methods and conducting experiments on different datasets to evaluate the performance of DDCLRec, the goal is to answer the following research questions:

[0334] ·RQ1: How does DDCLRec perform when competing with different types of recommendation methods?

[0335] ·RQ2: How does DDCLRec perform in alleviating the data sparsity problem?

[0336] ·RQ3: How effective is DDCLRec in alleviating noise problems?

[0337] ·RQ4: How robust is the recommendation of DDCLRec?

[0338] ·RQ5: What contributions do different key modules in DDCLRec make to the overall performance?

[0339] ·RQ6: What is the impact of the hyperparameters set in DDCLRec on the changes in the performance metrics Recall and NDCG?

[0340] 3.1 Dataset

[0341] To evaluate the effectiveness of DDCLRec, experiments were conducted on four public datasets collected from different real - life platforms to reflect real - world scenarios: ML - 1M, Yelp2018, Douban - book, Foursquare, which differ in terms of size and sparsity. The dataset statistics are shown in Table 3.

[0342] Table 3 Experimental dataset statistics

[0343]

[0344] 3.2 Evaluation Metrics

[0345] For performance evaluation, three representative metrics: Precision@k, Recall@k, and NDCG@k are used to measure the performance of the proposed DDCLRec. For the top - k recommendations, the value of k is set to 5, 10, 20. The DDCLRec implementation method is realized using PyTorch, and its performance is compared with various baseline methods using official or third - party code. The calculation formulas for Precision, Recall, and NDCG are as follows:

[0346] 1) Recall (recall rate) reflects how much of the information that users are interested in is perceived by us. R(u) represents the top - k recommendation list made for the user based on the user's behavior on the training set;

[0347] T(u) represents the set of items that the user actually selects after the system recommends items to the user.

[0348]

[0349] 2) Precision (precision rate), R(u) represents the top - k recommendation list made for the user based on the user's behavior on the training set; T(u) represents the set of items that the user actually selects after the system recommends items to the user.

[0350]

[0351] 3) CG (Cumulative Gain, cumulative gain), add up each recommendation result (relevance score) in the recommendation list; however, it is possible that in a list, high-score items are ranked lower and low-score items are ranked higher.

[0352]

[0353] 4) DCG (Discounted Cumulative Gain, discounted cumulative gain), on the basis of CG, introduce the position influence factor and "discount" the recommendation effect of the recommendation results ranked later.

[0354]

[0355] 5) NDCG (Normalized Discounted Cumulative Gain, normalized discounted cumulative gain), the recommendation system needs to conduct an overall evaluation of the recommendation lists of all users in the entire test set, and normalize the evaluation scores of the recommendation lists of different users. IDCG refers to the best recommendation result list returned by the recommendation system for a certain user.

[0356]

[0357] 3.3 Baseline methods

[0358] Compare the performance evaluations of DDCLRec with 16 existing popular recommendation system methods such as LightGCN and AutoCF on Recall@5, @10, @20 and NDCG@5, @10, @20 on four datasets including ML-1M and Yelp2018.

[0359] Graph Convolutional Neural Network:

[0360] · LightGCN: Use a lightweight convolutional graph encoder for better representation learning and model training.

[0361] · LCFN: A triple training framework based on self-supervised learning, integrating users' social information, enhancing the learning of multi-view encoders, and using self-supervised signals generated by other users to iteratively optimize the representation learning in the recommendation system.

[0362] Graph Contrastive Learning:

[0363] · AutoCF: Learn generative self-supervised learning for automatic data augmentation to improve the representation ability of the collaborative filtering model.

[0364] · HCCF: A hypergraph-enhanced cross-view contrastive learning structure to jointly capture local and global collaborative relationships.

[0365] · NCL: Incorporate potential neighbors in the structural and semantic spaces into the contrast pairs to enhance the performance of graph collaborative filtering.

[0366] · SGL: Generate multiple views and maximize the similarity between different views of the same node to enhance representation learning.

[0367] · XSimGCL: Adopt noise-based embedding to enhance the generated views.

[0368] · LightGCL: A simplified graph contrast learning paradigm that performs contrast enhancement by leveraging singular value decomposition.

[0369] Debiased Deep Learning:

[0370] · DICE: Causal inference separates users' interests and herd behaviors to learn independent representations.

[0371] · InvCF: Discover decoupled representations that are not affected by changes in the popularity distribution, reflecting latent preferences and popularity semantics.

[0372] · STaTRL: Transformers capture long-range dependencies in users' check-in sequences.

[0373] Diffusion Models:

[0374] · DiffRec: The denoising process learns the generative process of user interactions.

[0375] Mitigating Over-Smoothing:

[0376] · IMP-GCN: An interest-aware message-passing graph convolutional network.

[0377] · GDE: A simple and effective graph denoising encoder.

[0378] · EEDN: An enhanced encoder-decoder network that combines hybrid hypergraph convolution to enhance the aggregation in the graph convolution step.

[0379] · CaDRec: Introduce hypergraph convolution operations with structural and sequential contexts to select effective neighbors and mitigate over-smoothing.

[0380] 3.4 Comparative Analysis Experiment with Recommendation System Baseline Methods (RQ1)

[0381] Table 4 shows the performance evaluations of DDCLRec compared with 16 baseline methods such as LightGCN and AutoCF on four real-world datasets, ML-1M, Yelp2018, Douban-book, and Foursquare, in terms of Recall@k and NDCG@k, where the k values are set to 5, 10, and 20.

[0382] Table 4 presents the best performance comparison of DDCLRec proposed with 16 baseline methods in terms of Recall@k and NDCG@k on four datasets such as ML-1M and Yelp2018.

[0383]

[0384]

[0385] The experimental results in Table 4 show that DDCLRec outperforms other baselines in all cases, which verifies the effectiveness of introducing the diffusion model to enhance the embedding representation and the hard negative supervision contrast learning optimization strategy. The diversity of the evaluation datasets varies due to the sparsity of the interaction graph, the characteristics of the knowledge graph, and the recommendation scenarios. The research results demonstrate the generality and flexibility of DDCLRec. Overall, the progress of DDCLRec can be attributed to two aspects:

[0386] 1) Using the hypergraph convolutional network to encode the embedding representation of users and items to capture the complex multi-sided relationship structure. Introducing the diffusion model to add random noise, retaining and enhancing the diversity of the original data, making the embedding representation more differentiated. Further improving the modeling ability of user preferences and promoting the derivation of unbiased recommendations.

[0387] 2) Optimizing the embedding representation of users and items through the self-supervised signal of hard negative sample contrast learning and combining it with the multi-label cross-entropy loss optimization rate, enhancing the model's perception ability of user preferences and the discriminability in the embedding space, and enriching the context features.

[0388] It can be seen from the content of Table 4 that DDCLRec achieved better performance than most recommendation methods on four real datasets such as ML-1M and Yelp2018. For example, for the dataset ML-1M, DDCLRec improved by approximately 11.2% in Recall@20 and approximately 7% in NDCG@20 compared to LightGCN. Compared with AutoCF, it improved by approximately 27.5% in Recall@20 and approximately 22.4% in NDCG@20. For the dataset Douban-book, DDCLRec improved by approximately 42.4% in Recall@20 and approximately 56.3% in NDCG@20 compared to LightGCN. Compared with AutoCF, it improved by approximately 54.2% in Recall@20 and approximately 57.9% in NDCG@20. Similarly, good performance was also achieved on the other two datasets. This confirmed the effectiveness of combining hypergraph convolutional encoding with diffusion models to address issues such as popularity bias in recommendation systems. Existing recommendation methods mainly use the items associated in the current user's historical behavior data for collaborative filtering calculation of recommendations, often ignoring the correlation information of many edge sets. However, the advantages of DDCLRec are not obvious on some datasets, indicating that the noise introduced by the diffusion model affects the analysis of user individual biases, resulting in the distortion of user-item interaction data, and the selection strategy of difficult negative samples may not be perfect enough, leading to user interest bias. There is still much room for improvement in these aspects of DDCLRec.

[0389] 3.5 Experimental Analysis of Alleviating Data Sparsity (RQ2)

[0390] To explore the robustness of the model in data-sparse scenarios, a sparsity experiment was conducted by randomly deleting interaction edges from the user-item interaction graph at a ratio of 10%. By reducing the edge information of the user-item interaction graph in the current scenario by 10%, in the absence of sufficient interaction information, the performance of DDCLRec and other different recommendation methods in this scenario was evaluated, with particular attention paid to the performance of handling the problem of user data sparsity in the absence of sufficient interaction information as Figure 2 shown. The Precision@k, Recall@k, and NDCG@k metrics were used, where the k value was set to 10 and 20.

[0391] In Figure 2In the scenario of the medium-sparse user-item interaction information graph, DDCLRec has certain advantages over EEDN and CaDRec in Precision@10, @20, Recall@10, @20, and NDCG@10, @20. The experimental results prove the superiority of DDCLRec in alleviating the data sparsity problem. The advantage of DDCLRec in alleviating data sparsity is more prominent on the ML-1M dataset, but the advantage is weaker on Yelp2018. The possible reason is that Yelp2018 contains more complex node and interaction edge information, and a certain proportion of user-item interaction edge information is reduced, resulting in a greater impact of user individual bias. There are fewer recommendation bases, causing bias problems, indicating that DDCLRec is not very excellent in dealing with complex network node mining and training with a large amount of unlabeled data, and there is still much room for improvement. At the same time, it is also observed that the lower the integrity of the interaction graph, the lower the performance. Slightly reducing the integrity of the interaction graph is also beneficial to alleviating the overfitting problem and improving the recommendation performance, indicating that reasonably selecting the proportion of interaction graph interaction information and the learning rate during the training process is more conducive to improving the generalization ability of the model.

[0392] 3.6 Mitigation of Noise Analysis Experiment (RQ3)

[0393] Noise in recommendations may distort user-item interaction data, leading to problems such as recommendation bias and popularity bias. A high noise level may reduce the signal-to-noise ratio in the user-item interaction graph, reducing the recommendation quality and recommendation stability. After encoding and representing user-item interactions using hypergraph convolution, a random scaling tensor with a 10% ratio is added to the output of the encoder to introduce noise and incorporate it into the recommendation basis. The robustness of the recommendation method is verified through performance evaluation as Figure 3 shown. The Precision@k, Recall@k, and NDCG@k metrics are used, where the k value is set to 5, 10, and 20.

[0394] In Figure 3Among them, the abscissa represents different data sets, and the ordinate represents Precision@5, @10, @20, Recall@5, @10, @20, and NDCG@5, @10, @20. For the existing set of embedding representations, adding a certain proportion of randomly scaled tensors to it will cause bias in recommendations when incorporating the basis for recommendations. In the current noise scenario, DDCLRec has advantages on four data sets such as ML-1M and Yelp2018, demonstrating the effectiveness of DDCLRec in alleviating noise perturbation. It shows the necessity of combining the hypergraph convolutional encoding and diffusion model of DDCLRec, which can effectively alleviate the problem of user interest bias in the noise perturbation scenario. Among them, the advantage on the ML-1M data set is more obvious, while the advantage on the Foursquare data set is weaker. This may be because there are more user nodes and fewer item nodes in the Foursquare data set, and there is less node interaction information. Adding noise in the process of storing hyperedge information in complex hypergraph convolution is more likely to cause user interest bias, resulting in the true preferences of users being covered by popular items, and it is difficult to discover the true interests of users. It shows that in terms of alleviating noise perturbation, there is still a large room for improvement in the encoding method and embedding vector calculation method of DDCLRec.

[0395] 3.7 Robustness Analysis Experiment (RQ4)

[0396] In the evaluation of recommendation systems, Precision is an indicator used to measure the accuracy of recommendation results, defined as the proportion of truly relevant items among the recommended items. A high Precision indicates that the proportion of relevant items in the recommendation results of the recommendation system is relatively high, indicating a high accuracy of the recommendation system. It generally decreases as the value of K in Top-K increases. The stability of its change process can be used to judge the effectiveness of the recommendation strategy. Set K to (1, 5, 10, 20) and conduct multiple experiments on four data sets such as ML-1M and Yelp2018 to analyze the change of Precision of different models as Figure 4 shown.

[0397] In Figure 4It is observed that on four real-world datasets such as ML-1M and Yelp2018, as the value of K in Top-K increases, the Precision of different models gradually decreases. In the initial stage when K is small, the Precision of CaDRec is relatively high in some datasets. However, as K increases, compared with EEDN and CaDRec, the change in the Precision of DDCLRec is more stable, especially on the Yelp2018 dataset and the Foursquare dataset. This shows that DDCLRec can still maintain high accuracy in a long recommendation list. Even in a long recommendation list, it can effectively mine the true interests of users. It proves that DDCLRec can maintain good recommendation quality and robustness when facing more recommendation options. However, the advantage of DDCLRec on the ML-1M dataset is not obvious. Compared with other datasets, ML-1M has fewer nodes and interaction information, indicating that there is still much room for improvement in the strategy of effectively mining user interest preferences during the propagation process of hypergraph convolution and multi-layer graph neural networks.

[0398] 3.8 Ablation Experiment (RQ5)

[0399] DDCLRec has an advantage over other baseline methods in terms of the two performance metrics of Recall@k and NDCG@k. The superior results obtained by DDCLRec can be attributed to the following factors:

[0400] The hypergraph convolutional network is used to encode and embed users and items to capture complex multiple relationships. The diffusion model is introduced to preserve and enhance the diversity of the original data, distinguish the embedding representations, strengthen user preference modeling, and achieve unbiased recommendation. In addition, the hard negative supervised contrast learning loss optimization and the multi-label cross-entropy loss optimization strategies are combined to optimize the embedding representations, enrich the context features, and enhance the model's ability to identify user preferences in the embedding space.

[0401] To study the effectiveness of the key components of DDCLRec, an ablation study is conducted on three model variants, as shown in Table 5. The following are the two method variants of DDCLRec:

[0402] 1) DR: Only use the hypergraph convolution to encode the embedding representation and perform multi-layer graph neural network propagation enhancement. Do not introduce the diffusion model and abandon the hard negative contrast learning loss optimization strategy.

[0403] 2) DDR: Use the hypergraph convolution to encode the embedding representation, introduce the diffusion model, and perform multi-layer graph neural network propagation enhancement. Abandon the hard negative contrast learning loss optimization strategy.

[0404] 3) DCLR: Use hypergraph convolutional coding embedding representation to perform multi-layer graph neural network propagation enhancement, and adopt the strategy of discarding hard negative contrast learning loss optimization. No diffusion model is introduced.

[0405] The comparative analysis experiments of Recall@k and NDCG@k indicators of DDCLRec and three method variants on four real datasets such as ML-1M are shown in Table 5, where the value of k is set to 5, 10, and 20.

[0406] Table 5 Evaluation results of Recall@k and NDCG@k indicators of DDCLRec and three method variants on four real datasets

[0407]

[0408] In Table 5, DDCLRec has superior performance in Recall@5, @10, @20 and NDCG@5, @10, @20 indicators of four datasets including ML-1M and Yelp2018. This shows the necessity of combining hypergraph convolutional coding and diffusion model, verifies the importance of combining hard negative contrast learning loss optimization strategy with multi-label cross entropy loss optimization strategy for user preference evaluation, and proves that DDCLRec is reasonable and effective in solving problems such as popularity bias in recommendation systems.

[0409] 3.9 Hyperparameter Analysis Experiment (RQ6)

[0410] DDCLRec has three important hyperparameters. The important weight parameter λ in the loss optimization process 1 , 2 and hard negative sampling depth τ. Study λ 1 and different combinations of τ and λ 2 The influence of different combinations of and τ on the recommendation performance is shown in (10 -3 ,10 -2 ,10 -1 ), (0.35,0.4,0.46,0.5,0.55) and (0.3,0.35,0.41,0.45,0.5) in the range of τ and λ 1 and λ 2 Taking ML-1M data as an example, we explore the influence of hyperparameters on the Recall@k and NDCG@k indicators of DDCLRec. Figure 5 As shown in the figure, the value of k is set to 10 and 20.

[0411] exist Figure 5 (a) and Figure 5 In (b), it is found that when τ is 10 -1 ,λ 1When the value is 0.5, the performance of DDCLRec is optimal. By changing the values of the two parameters, a significant performance drop is observed. In Figure 5 (c) and Figure 5 (d), it is found that when the value of τ is 10 -2 , and the value of λ 2 is 0.41, the performance of DDCLRec is optimal. According to the experimental results, it is found that DDCLRec is more sensitive to τ. When λ 1 is appropriately increased, the convergence is faster and it helps to improve the model performance. However, too high a value of λ 2 will affect the performance, indicating that the participation and change of τ can greatly improve the recommendation performance, but comprehensive excellent performance also requires λ 1 and λ 2 to balance. It proves the necessity of hard negative contrast learning for optimizing the recommendation method. Therefore, according to the consideration of weight parameters in different ranges, an effective weighting scheme is synthesized, and τ, λ 1 and λ 2 are set to 10 -1 , 0.5, 0.41 respectively, and τ, λ 1 and λ 2 are set to 10 -2 , 0.46, 0.41 respectively, which can make DDCLRec achieve better performance.

[0412] In summary, the present invention proposes a contrastive learning debiasing recommendation method DDCLRec based on a diffusion model. First, a hypergraph convolutional network is used to capture the multi-sided complex relationships between users and items by combining context information to encode user-item interactions, effectively transmitting valuable information into the modeling process of user-item interactions. At the same time, a diffusion model is introduced to preserve data diversity, differentiate the embedding representations, enhance user preference modeling, and derive debiased representations. Second, different combinations of regularization and weighting schemes are studied, and the self-supervised signals of hard negative supervised contrastive learning are used to enhance the embedding representations of users and items, further improving the ability of this method in user preference learning, as well as its performance in user preference perception and embedding space discriminability, and enriching the understanding of context features by this method. To verify the effectiveness and feasibility of this method, multiple comparative analysis experiments are carried out on four real-world datasets such as ML-1M and Yelp2018, and DDCLRec is compared with 16 existing popular recommendation methods. The experimental results show that: in terms of the performance evaluation of the recommendation system, DDCLRec proposed on the ML-1M dataset is superior to the latest CaDRec method by 0.8% - 1.4% in the Recall@k metric and 0.5% - 1.1% in the NDCG@k metric, demonstrating the accuracy and superiority of DDCLRec; in the scenario of sparse user-item interaction graphs, DDCLRec has certain advantages in Precision@k, Recall@k, and NDCG@k on four real-world datasets such as ML-1M and Yelp2018, proving that DDCLRec can better recommend items for users in dealing with sparsity problems; in the scenario with noise, DDCLRec outperforms the existing popular recommendation methods EEDN and CaDRec in Precision@k, Recall@k, and NDCG@k on four real-world datasets such as ML-1M and Yelp2018; in terms of recommendation stability, as the recommended item list increases, DDCLRec performs more stably, proving that DDCLRec has more stable and effective method performance in the scenario of a large number of item recommendations; in ablation analysis and hyperparameter sensitivity, the effects of different hyperparameters on DDCLRec in Recall@k and NDCG@k are explored, proving the necessity of introducing a diffusion model in DDCLRec to enhance the hypergraph convolutional encoded embedding representations and the importance of the hard negative supervised contrastive learning loss optimization strategy. The above experiments fully demonstrate the advantages of DDCLRec in solving the key problems faced by the recommendation system, highlighting the effectiveness of DDCLRec.

[0413] However, DDCLRec still has a lot of room for improvement, especially in dealing with long-tail items and user bias. Effectively distinguishing between users' real preference items and popular items is crucial to alleviating user interest bias. In addition, the application of contrastive learning in recommendation systems can improve the accuracy and efficiency of recommendation systems, which requires combining user interaction modeling and contrastive embedding enhancement. In terms of scalability, DDCLRec is not much different from other models. Although DDCLRec is also based on a lightweight framework for recommendation, it also faces increased memory requirements and time when the data set is larger, and its advantage in scalability is not obvious. Future recommendation research should not only focus on accuracy and diversity, but also take scalability as a key feature. In the context of the era of massive information overload on the Internet, recommendation methods are no longer limited to the accuracy of personalized recommendations, but also focus on the scalability of recommendation methods, effectively saving time and space resources, and are more conducive to the development of recommendation system research. Therefore, developing lightweight recommendation methods and exploring effective recommendation methods to remove bias to tap users' real preferences are crucial for the high-quality development of recommendation systems.

[0414] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. An intelligent recommendation verification method integrating hypergraph and context information, characterized in that: The following steps are involved: S1, extract biased item features, item popularity features and item semantic features from the user-item interaction graph, combine the biased item features, item popularity features and item semantic features of the individual and neighboring nodes, and then use hypergraph convolution to encode the representations of the user and item to obtain the encoded embedding, which includes the embedded representation of the user and the embedded representation of the item; S2, the encoded embedding is enhanced by hypergraph convolution to obtain the user embedding representation and item embedding representation optimized by hypergraph convolution; then the scores of the user recommended interactive items are calculated, and the scores are arranged in descending order to generate a recommendation list of related candidate items.

2. According to claim 1, the intelligent recommendation verification method integrating hypergraph and context information is characterized in that: S1 includes the following steps: S1-1, convert the user-item interaction graph information into an embedded representation of the user-item interaction, and derive the historical interaction sequence H between user u and item i u,i , determine the user u who interacts with the current item i, and then obtain the item's embedding representation y through uniformly distributed random sampling, where the item's embedding value is randomly generated within the uniform distribution range; S1-2, in the historical behavior sequence of user-item interaction, according to the items interacted with the current user, the embedding vectors of all the interacted items are integrated to obtain the user embedding representation x; S1-3, the embedded representation of items that interact with users is integrated. It fits biased interactions based on item bias features, item popularity features, and item semantic information during the user interaction process to obtain the final user embedded feature representation and the event sequence representation (x, y) of the user-item interaction features.

3. According to claim 2, the intelligent recommendation verification method integrating hypergraph and context information is characterized in that: The calculation formulas for the embedding representation of the user and the embedding representation of the item are as follows: Among them, x is the user embedding representation; y is the embedding representation of the item; y i is the embedding representation of the i-th item, and is the item that user u interacts with; H u,i is the historical interaction sequence between user u and item i; r u is the embedding vector of individual user biases in the original user-item interaction graph; ||r u || is r u The module length; sign() is the sign function; ⊙ is the Hadamard product symbol; Hgc(·) is the hypergraph convolution; is the adjacency matrix; Encoder() is a hypergraph convolutional coding representation.

4. According to claim 1, the intelligent recommendation verification method integrating hypergraph and context information is characterized in that: The hypergraph convolution enhancement includes: S2-1, extract feature information through multi-layer convolution propagation to enhance user embedding representation: Among them, M l is the enhanced output of the lth convolutional layer; is the ELU activation function; D -1 / 2 BD -1 / 2 is the symmetric normalized adjacency matrix; B is the bipartite graph adjacency matrix; D represents a diagonal matrix; W2 is the parameter weight matrix of convolution propagation; S2-2, using hypergraph convolution to model the contextual representation of users, by learning the contextual representation from the user hypergraph in the hypergraph, enhancing the signal propagation in the convolution process, adjusting the critical matrix and feature matrix according to the introduced hypergraph information, and selecting the specific rows and columns corresponding to the user; Among them, ψ u is the hyperedge of user u; T is the transpose symbol; Norm is the L2 regularization calculation; d is the dimension in the self-attention mechanism; is the ELU activation function; Q, K, and V represent the query, key, and value matrices in the attention mechanism, respectively; W Q , W K is the corresponding parameter weight matrix; Φ is a hyperparameter that controls the degree of attention; S2-3, aggregate multiple subspace information through multi-head attention mechanism: Among them, W3 is the weight of each head in the multi-head attention mechanism; k is the value of the number of attention heads; n is the total number of attention heads; It represents the result of input embedding layer l after being processed by different numbers of attention heads; S2-4, after average pooling and normalization, the user embedding representation and item embedding representation optimized by hypergraph convolution are obtained; Among them, x is the user embedding representation optimized by hypergraph convolution; M i Embeddings that have been processed previously; i∈H u,i The embedded selection belongs to the user-item interaction sequence; || is the absolute value symbol; H u,i Represents the historical interaction sequence between user u and item i; y is the embedding representation of the item optimized by hypergraph convolution; y i is the embedding representation of the i-th item, and is the item that user u interacts with; sign() is the sign function; ⊙ is the Hadamard product symbol; r u is the embedding vector of individual user biases in the original user-item interaction graph; ||r u || is r u The mold length.

5. The intelligent recommendation verification method integrating hypergraph and context information according to claim 1, characterized in that: The loss function of hypergraph convolution enhancement is in, For user collection; is the true label vector of user u’s interaction with item 1; is the true label vector of user u’s interaction with item N; α u,1 ......α u,N is the predicted rating of user u for items 1 to N; σ is the sigmoid activation function.

6. The intelligent recommendation verification method integrating hypergraph and context information according to claim 1, characterized in that: It also includes contrastive self-supervised optimization: using contrastive self-supervised learning methods to enhance cross-view contrastive learning of multi-feature embedding representations of user-item interactions, and learning personalized features of different nodes to prevent bias; A hard negative supervised contrastive learning loss optimization strategy is adopted to select samples with different but similar labels in the current range as negative samples.

7. The intelligent recommendation verification method integrating hypergraph and context information according to claim 6, characterized in that: The contrastive self-supervised optimization includes: First, a sample selection strategy based on random probability is designed to calculate the distribution probability (0-1) of positive and negative samples in the sample set. If the sample label belongs to the current sample set, It is selected as a positive sample; otherwise, it is regarded as a negative sample; and it is in the same input sample space; Among them, p(z + ) is the probability distribution of positive samples under different conditions; z + 、z - They are positive samples and negative samples respectively; z is the sample; p U (z - ) is the probability distribution of negative samples under different conditions in the sample set; Represents the probability distribution of selecting positive samples; Represents the probability distribution of selecting negative samples; Represent the positive sample set and the negative sample set respectively; Under the current negative sample selection structure, the selection of difficult negative samples is performed: after the anchor point is determined, the cosine similarity between the anchor point and the sample is calculated according to the sample space set, the difficult negative sample set is constructed, and the negative probability distribution p of the difficult negative sample selection is obtained. HU (z - ): in, is to select the current limited range, that is, HU Negative sample expectation; Ψ HU It is the specific range of the current condition sample selection; exp() is an exponential operation with a natural base as the base; w is the sample label; is a set of sample labels; d(z,ξ) is the similarity between the currently selected node (sample z) and the anchor point; To control the threshold of sampling hardness; In the sample space combined with the same anchor point, the method of selecting samples similar to the anchor point but of different types is used to construct a set of difficult negative samples and obtain the corresponding negative probability distribution: Among them, p S (z - ) is in Negative probability distribution of negative sample selection under conditions; The label information of the sample w z and the label information w of the anchor point ξ are comprehensively considered; finally, the above two difficult negative sample sets are combined to construct a difficult negative sample selection set based on the contrastive learning framework; this set selects negative samples within the specified range and meets two conditions: similar to the anchor point but with different label types, and meets the selection conditions within the specified selection range. Among them, p HS (z - ) is the negative probability distribution of difficult negative samples that meet the conditions; Ψ HS , S , HU They represent different limiting ranges in the sample selection process.

8. The intelligent recommendation verification method integrating hypergraph and context information according to claim 7, characterized in that: The contrast loss function of the difficult negative sample is: Among them, d(·,·) represents the similarity; n is the number of negative samples; k is the negative sample index; Indicates difficult negative samples that fall within the current limited range; In addition, by changing the weighting scheme to balance the gradient, the recommendation of popular items that affect the user's true preference is reduced. At this time, the loss function is expressed as follows: Among them, λ is a hyperparameter that balances the weight of the item embedding matrix that interacts with the user; λ1 is used for the embedding matrix of items that the user has interacted with; ⊙ is the Hadamard product symbol; η is a weight vector used to weight the losses of different items; Θ is the regularization coefficient that controls the regularization weight; is the previous loss function; For user collection; H u,i is the historical interaction sequence between user u and item i.

9. The intelligent recommendation verification method integrating hypergraph and context information according to claim 1, characterized in that: The formula for calculating the score of user-recommended interactive items is as follows: β= <x||μ·r u ,y||μ·r i > (14) Among them, β represents the score of the interactive items recommended by the user; <·,·> is the inner product operation; || is the concatenation operation; <x||μ·r u ,y||μ·r i >When calculating, first concatenate each part and then perform inner product operation; x and y represent the user embedding representation and item embedding representation respectively; μ represents a hyperparameter that adjusts the influence of popularity bias; r u and r i denote the bias vectors of users and items respectively.

10. The intelligent recommendation verification method integrating hypergraph and context information according to claim 1, characterized in that: It also includes verifying the results of the generated recommendation list of related candidate items by calculating the recall rate, precision rate, cumulative gain, discounted cumulative gain, and normalized discounted cumulative gain.

Images