Intelligent recommendation verification method fusing hypergraph and context information

Abstract

The application provides a kind of intelligent recommendation verification method of fusing supergraph and context information, comprising: S1, the bias item feature, item popularity feature and item semantic feature are extracted from the interaction graph of user item, the bias item feature, item popularity feature and item semantic feature of individual and neighbor node are combined, then the representation of user and item is encoded using supergraph convolution, to obtain the embedding after encoding;S2, the embedding after encoding is enhanced using supergraph convolution by supergraph convolution, to obtain the embedding representation of user and the embedding representation of item after supergraph convolution optimization;Then the score of user recommendation interaction item is calculated, and the recommendation list of relevant candidate item is generated by arranging the score in descending order.The self-supervised signal of hard negative supervision contrast learning is used to optimize the embedding representation of user and item, and is combined with multi-label cross-entropy loss optimization strategy, which enhances the learning of model perception to user preference and discriminability in embedding space, and enriches the context feature.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent recommendation, and in particular to an intelligent recommendation verification method fusing supergraph and context information. BACKGROUND

[0002] With the problem of information overload in online applications brought by the development of the Internet, high-quality personalized recommendation systems have become an indispensable key role in daily life. The recommendation system generally improves the high-quality results of recommendation by mining the historical behavior data of user-item interaction, and most of them use graph convolutional neural networks to enhance the embedding representation of user-item interaction and capture collaborative signals for recommendation. In the recommendation system based on deep learning, the use of neural network learning can effectively represent complex user behavior and item features, and explore user interest to provide accurate recommendation. Graph representation learning can more effectively capture multi-hop connections and adjacent node information between users and items, and learn the representation of users and items through graph neural networks.

[0003] Most recommendation methods often have difficulty in comprehensively learning feature representation in context when improving performance, 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, is particularly critical. Only by better learning the feature representation in the context can the influence of these biases be effectively reduced, thereby improving the performance and accuracy of the recommendation system. SUMMARY

[0004] The present application aims to at least solve the technical problems existing in the prior art, and particularly innovatively provides an intelligent recommendation verification method fusing supergraph and context information.

[0005] In order to achieve the above-mentioned purpose of the present application, the present application provides an intelligent recommendation verification method fusing supergraph and context information, comprising the following steps:

[0006] S1, extracting bias item features, item popularity features and item semantic features from the interaction graph of users and items, combining the bias item features, item popularity features and item semantic features of individuals and neighbor nodes, and then using supergraph convolution to encode the representation of users and items to obtain the encoded embedding, wherein the encoded embedding includes user embedding representation and item embedding representation;

[0007] S2, the encoded embedding is supergraph convolution enhanced by supergraph convolution, the signal propagation is enhanced by supergraph convolution enhancement, the supergraph convolution can effectively select the corresponding row in the superedge and effectively learn the context information in the superedge; reduce the computational complexity. Get the user embedding representation and the embedding representation of the item optimized by supergraph convolution; then calculate the score of the user recommendation interaction item, and arrange the score in descending order to generate a recommendation list of related candidate items.

[0008] Preferably, S1 comprises the following steps:

[0009] S1-1, the user-item interaction graph information is converted into the embedding representation of the user-item interaction, and the historical interaction sequence H of the user u and the item i is derived u,i , determine the user u interacting with the current item i, and then obtain the embedding representation y of the item by uniform distribution random sampling, wherein the embedding value of the item is randomly generated in the uniform distribution range;

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

[0011] S1-3, the embedding representation of the item interacting with the user is integrated, which is according to the item bias feature, item popularity feature and item semantic information in the user interaction process, the user bias in the item embedding representation is separated out, the biased interaction is fitted, and the self-attention mechanism is enhanced to obtain the final user embedding feature representation, and the event sequence representation (x, y) of the user-item interaction feature is obtained.

[0012] Preferably, the calculation formula of the user embedding representation and the item embedding representation is as follows:

[0013]

[0014] Wherein, x is the user embedding representation;

[0015] y is the embedding representation of the item;

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

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

[0020] sign() is a sign function;

[0021] ⊙ is a Hadamard product symbol;

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

[0023] is an adjacency matrix;

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

[0025] Preferably, the hypergraph convolution enhancement includes:

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

[0027]

[0028] Wherein, M l is the output of the lth convolution layer after enhancement;

[0029] is an ELU activation function;

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

[0031] B is a bipartite graph adjacency matrix;

[0032] D represents a diagonal matrix;

[0033] W2 is a parameter weight matrix of convolution propagation;

[0034] S2-2, model the user's context representation using hypergraph convolution, enhance the signal propagation in the convolution process by learning the context representation from the user hyperedge in the hypergraph, adjust the critical matrix and the feature matrix according to the introduced hypergraph information, and select the specific row and column corresponding to the user;

[0035]

[0036] Wherein, ψ u is the hyperedge of user u;

[0037] T is a transpose symbol;

[0038] Norm is an L2 regularization calculation;

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

[0040] is an ELU activation function;

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

[0042] W Q , W K is the corresponding parameter weight matrix;

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

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

[0045]

[0046] wherein W3 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 of the input embedding of the lth layer after being processed by different attention heads;

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

[0051]

[0052]

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

[0054] M i previously processed embedding;

[0055] i∈H u,i is the selection of the embedding belonging to the user-item interaction sequence;

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

[0057] H u,i represents the historical interaction sequence of the user u and the item i;

[0058] y is the embedding representation of the item optimized by hypergraph convolution;

[0059] y i is the embedding representation of the ith item, and y is the item interacting with the user u;

[0060] sign() is a sign function;

[0061] ⊙ is a Hadamard product symbol;

[0062] r u is the embedding vector of the user individual bias of the original user-item interaction graph;

[0063] ||r u || is the length of r u .

[0064] Preferably, the hypergraph convolution enhanced loss function is

[0065]

[0066] wherein, is a user set;

[0067] is the true label vector of the user u interacting with item 1;

[0068] is the true label vector of the user u interacting with item N;

[0069] α u,1 ......α u,N is the predicted score of the user u interacting with item 1 to item N;

[0070] σ is a sigmoid activation function.

[0071] Preferably, it further comprises a contrastive self-supervised optimization: using a contrastive self-supervised learning method, the multi-feature embedding representation of the user-item interaction is enhanced by cross-view contrastive learning, and the personalized features of different nodes are learned to prevent bias; and a hard negative supervision contrastive learning loss optimization strategy is adopted, and samples with different labels but similar samples in the current range are selected as negative samples.

[0072] Preferably, the contrastive self-supervised optimization comprises:

[0073] First, a sample selection strategy based on random probability is designed to calculate the distribution probability (0-1) of positive samples 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 multiple different negative sample selection structures are set in the same input sample space to assist the selection of difficult negative samples in the contrastive learning framework;

[0074]

[0075] wherein, p(z +is the probability distribution of positive samples under different conditions;

[0076] z + , z - are positive samples and negative samples, respectively;

[0077] z is a sample;

[0078] p U (z - ) is the probability distribution of negative samples in the sample set under different conditions;

[0079] represents the probability distribution of selecting positive samples;

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

[0081] and represent the positive sample set and the negative sample set, respectively;

[0082] 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, a difficult negative sample set is constructed, and the negative probability distribution p HU (z - ) of difficult negative sample selection is obtained:

[0083]

[0084] wherein, is the expected value of selecting negative samples within the current limited range, i.e. Ψ HU

[0085] Ψ HU is the specific range of the current conditional sample selection;

[0086] exp() is the exponential operation with the natural base as the base;

[0087] w is the sample label;

[0088] is the sample label set;

[0089] d(z,ξ) is the similarity between the current selected node (sample z) and the anchor point;

[0090] is the threshold value for controlling the sampling hardness;

[0091] Under the sample space combined with the same anchor point, the method for selecting samples similar to the anchor point but of different types is used to construct a difficult negative sample set and obtain the corresponding negative probability distribution:

[0092] ​

[0093] wherein p S (z - ) is the negative probability distribution of the negative sample selection under the condition of

[0094] the label information w z of the sample and the label information w ξ of the anchor point are comprehensively considered; finally, the two difficult negative sample sets are combined to construct a difficult negative sample selection set based on the contrast learning framework; this set selects negative samples within a specified range, which meets two conditions: similar to the anchor point but with different label types, and meets the selection condition within the specified selection range

[0095]

[0096] wherein p HS (z - ) is the negative probability distribution of the difficult negative sample meeting the condition;

[0097] Ψ HS , Ψ S , Ψ HU respectively represent different limited ranges in the sample selection process.

[0098] Preferably, the contrast loss function of the difficult negative sample is:

[0099]

[0100] wherein d(·,·) represents the similarity;

[0101] n is the number of negative samples;

[0102] k is the negative sample index;

[0103] represents the difficult negative sample belonging to the current limited range;

[0104] In addition, in order to effectively balance the user's bias feature of the item, the user's current interaction item feature and the user's future possible interaction item feature, and prevent the problem of node embedding not being smooth. By changing the weighting scheme to balance the gradient, the recommendation of the user's true preference is reduced because of the high popularity of the item, at this time, the loss function is represented as follows:

[0105]

[0106] wherein λ is a hyperparameter for balancing the weight of the item embedding matrix interacted with the user;

[0107] λ1 is used for the item embedding matrix interacted by the user;​

[0108] ⊙ is the Hadamard product symbol;

[0109] η is a weight vector, which is used to weight the loss of different items;

[0110] Θ is a regularization coefficient to control the regularization weight;

[0111] is the previous loss function;

[0112] is the set of users;

[0113] H u,i is the historical interaction sequence of user u and item i.

[0114] Preferably, the score formula for recommending interaction items for a user is as follows:

[0115] β = <x||μ·r u ,y||μ·r i > (14)

[0116] where β represents the score of recommending interaction items for a user;

[0117] <·,·> is the inner product operation;

[0118] || is the concatenation operation;

[0119] <x||μ·r u ,y||μ·r i > is calculated by first concatenating each other and then performing the inner product operation;

[0120] x and y represent the user embedding representation and the item embedding representation, respectively;

[0121] μ represents a hyperparameter that adjusts the influence of popularity bias;

[0122] r u and r i represent the bias vectors of the user and the item, respectively.

[0123] Preferably, it further includes verifying the results of the generated recommendation list of related candidate items by calculating the recall rate, precision rate, cumulative gain, loss cumulative gain, and normalized loss cumulative gain.

[0124] To sum up, by adopting the technical scheme, the self-supervised signal of the hard negative supervision contrast learning is used to optimize the user item embedding representation, and is combined with the multi-label cross entropy loss optimization strategy, so that the learning of the model to the user preference and the distinctiveness in the embedding space are enhanced, and the context feature is enriched; in particular, the beneficial effects are: in order to effectively alleviate the user interest deviation and the popularity deviation problem, high-quality graph collaborative interaction information is obtained through contrast self-supervised learning, and the self-supervised signal is introduced to enhance the model performance, so that the context feature information is better learned. This method not only optimizes the modeling of the user individual deviation, but also effectively alleviates the problem that the popularity bias is intensified due to too much noise in the interaction information. In addition, the self-supervised signal of the hard negative sample supervision contrast learning is used to optimize the embedding representation of the user and the item, and is combined with the multi-label cross entropy loss optimization strategy, so that the perception ability of the model to the user preference is further enhanced, the distinctiveness of the embedding space is improved, and the learning of the context feature is enriched.

[0125] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0126] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the description of the embodiments, taken in conjunction with the following drawings in which:

[0127] Figure 1 is a DDCLRec framework diagram of the present application.

[0128] Figure 2 is a performance comparison diagram of Precision@k, Recall@k and NDCG@k of DDCLRec, EEDN and CaDRec on four data sets of ML-1M, Yelp2018 and the like in a sparse user item interaction graph scenario; Figure 2 (a) is ML-1M, Figure 2 (b) is Yelp2018, Figure 2 (c) is Yelp2018, Figure 2 (c) is Douban-book, Figure 2 (d) is Foursquare.

[0129] Figure 3 is a performance comparison diagram of Precision@k, Recall@k and NDCG@k of DDCLRec, EEDN and CaDRec on four real data sets of ML-1M, Yelp2018 and the like in a 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.

[0130] Figure 4 is a plot of analyzing the change of Precision@k of DDCLRec compared with other models on four real datasets, ML-1M, Yelp2018, etc. Figure 4 (a) is ML-1M, Figure 4 (b) is Yelp2018, Figure 4 (c) is Douban-book, Figure 4 (d) is Foursquare.

[0131] Figure 5 is a plot of analyzing the influence of λ1 and τ and λ2 and τ on Recall@k and NDCG@k of DDCLRec on dataset ML-1M. Figure 5 (a) is λ1 and τ on Recall@10 of ML-1M, Figure 5 (b) is λ1 and τ on Recall@20 of ML-1M, Figure 5 (c) is λ1 and τ on NDCG@10 of ML-1M, Figure 5 (d) is λ1 and τ on NDCG@20 of ML-1M, Figure 5 (e) is λ2 and τ on Recall@10 of ML-1M, Figure 5 (f) is λ2 and τ on Recall@20 of ML-1M, Figure 5 (g) is λ2 and τ on NDCG@10 of ML-1M, Figure 5 (h) is λ2 and τ on NDCG@20 of ML-1M. DETAILED DESCRIPTION

[0132] Embodiments of the present application are described in detail below with reference to the attached drawings, wherein same or similar components have the same or similar reference numbers throughout the several views. The embodiments described below are examples of the present application, and are not intended to limit the present application.

[0133] Compared with traditional recommendation methods, recommendation systems based on graph neural networks can better capture complex user-item interaction relationships. By utilizing the rich information of the interaction graph structure, the complex interaction relationships between users and items are modeled, the embedding encoding representation of users and items is optimized, the data sparsity problem is effectively addressed, and the interpretability of the recommendation system is enhanced. However, excessive use of graph neural networks also brings high computational complexity, data sparsity, and interpretability issues, and there are potential challenges in constructing high-quality interaction graphs.

[0134] The introduction of contrastive learning greatly increases the inclusiveness of recommendation methods, and is more 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 relies on the selection of negative samples, and improper selection strategies can affect the learning of features by the model, and there are high requirements for the selection of negative samples.

[0135] 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 supervised, which is beneficial for various types of downstream tasks.

[0136] 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 more rich relationship information through hyperedges, effectively handle multi-type data relationships, and capture high-order relationships between nodes.

[0137] To better solve the problem of node embedding over-smoothing in recommendation systems, combining the advantages of diffusion models and contrastive learning, selectively retaining more valuable embedding representation information, improving recommendation performance while promoting recommendation diversity, a diffusion model-based contrastive learning debiased recommendation method DDCLRec (Diffusion-based debiased contrastive learning for recommendation) is proposed.

[0138] The user-item interaction process is affected by factors such as popularity, exposure, and the like, and the recommendation effect is affected by the difference in user rating standards. Some users will give high scores, some users will give low scores, and some users 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 certain dimensions, which further widens the semantic distribution of the item and expands the difference. In actual recommendation scenarios, user interest bias will fluctuate due to factors such as historical behavior accumulation effect and short-term interest hotspots, which will further interfere with the true semantics of the item and put higher requirements on the robustness of the recommendation method. However, the user interest bias factor can be used as a learnable disturbance, which can be used as an important basis for changes in item semantic distribution, and corresponding strategies can be made according to the changes, which is more conducive to improving the robustness of the recommendation method. Under the simulation of the interaction of the supergraph convolution enhancement and diffusion model, important features can be more effectively learned, and the generation of user-item embedding representation is reconsidered, the user interest bias in the recommendation process is balanced, the gradient is balanced through different weighting schemes to update the project embedding, and regularization is further used to balance the update of the embedding to prevent popular projects from occupying too large a weight in the embedding space and to mine the true preferences of users.

[0139] The key symbols used in the present application are shown in Table 1.

[0140] Table 1 Mathematical symbols

[0141]

[0142] The task is as follows: input: user-item interaction graph, output: use the denoised user-item interaction embedding representation, generate a recommendation list containing relevant candidate items for the user, and predict the item that the user ultimately wants to interact with according to the user's preferences.

[0143] For the user-item interaction graph, the bias item features of individuals and neighbor nodes, item popularity features and item semantic features are combined to obtain a multi-feature interaction representation of the user to the item. The representation of the user and the item is encoded using supergraph convolution to obtain an enhanced embedding representation of the user-item interaction, and a diffusion model is introduced to integrate various semantic information and bias features to enhance the embedding representation of the user and the item, and to model the user's preferences. Secondly, the process of signal propagation is enhanced using supergraph convolution, and the context information in the superedge is effectively learned. The supergraph convolution is combined with the self-attention mechanism to assist information diffusion and aggregate the semantic information of neighboring nodes. Finally, the multi-feature embedding representation of the user-item interaction is enhanced through cross-view contrast learning by comparing the self-supervised learning method. A hard negative supervision contrast learning loss optimization strategy is used, and the multi-label cross-entropy loss and NCE loss are combined as the objective function to better utilize the context information. The multi-label cross-entropy loss corresponds to NCE loss is a commonly used loss function in contrastive learning, which is included in . Corresponding to the difficult negative sample sampling contrast optimization loss.

[0144] The overall framework of the proposed DDCLRec is shown in Figure 1 : For the user-item interaction graph, the bias item features of individuals and neighbor nodes, item popularity features, and item semantic features are combined to encode the representations of users and items using hypergraph convolution. Considering three different token information personal bias information, semantic information, and popular item information, they are injected into the hypergraph convolution operation as learnable perturbations to decouple the popularity bias and individual bias in user-item interactions and generate unbiased user and item representations. A diffusion model is introduced to incorporate various semantic information and bias features to enhance the embedding representations of users and items. The diffusion process first proceeds forward, gradually adding Gaussian noise to destroy the user's interaction history. In this process, noise is limited to preserve user personalization information. In the reverse process, the original interaction data is recovered from the noise-destroyed interaction history through a parameterized neural network, and the interaction probability is recovered for ranking and recommending uninteracted items.

[0145] Second, hypergraph convolution is used to enhance the process of signal propagation and effectively learn the context information in the hyperedge. The slicing operation is used to effectively select the corresponding row in the hyperedge, reducing the computational complexity. The attention mechanism is injected into the convolution process as a perturbation to select effective neighbor nodes, assist information diffusion, and aggregate the semantic information of adjacent nodes. The multi-head hypergraph convolution network layer learns features in multiple subspaces to generate diverse representations, and finally learns the context information through average pooling and normalization.

[0146] Finally, through the contrastive self-supervised learning method, the multi-feature embedding representation of user-item interaction is enhanced through cross-view contrastive learning to learn individualized features of different nodes to prevent bias. The hard negative supervision contrastive learning loss optimization strategy is adopted to preferentially select samples with different labels but similar samples within the current range as negative samples, and the multi-label cross-entropy loss and NCE loss are combined as the objective function to learn the features of real data, distinguish real data and noise data, and improve the generalization ability and robustness of the model.

[0147] 1. Multi-feature interaction representation learning

[0148] 1.1 Extracting multi-feature embedding representation of user-item interaction

[0149] According to the interaction information of different users on the interaction graph, the obtained information is converted into entity embedding representation and relationship embedding representation, the bias item feature, item popularity feature and item semantic feature of the user nodes and neighbors on the graph are embedded, and a multi-feature interaction embedding representation is constructed. The user-item interaction graph information is converted into an embedding representation of user-item interaction, and the historical interaction sequence H of the user u and the 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 by uniformly distributed random sampling, wherein the embedding value of the item is randomly generated within a uniform distribution range. In the historical behavior sequence of the current user-item interaction, according to the items interacted with the current user, the embedding vectors of all interaction items are integrated to obtain the user embedding representation x. The embedding representation of the item interacted with the current user is separated from the user bias in the item embedding representation according to the item bias feature, item popularity feature and item semantic information in the user interaction process, the biased interaction is fitted, and the self-attention mechanism is enhanced to obtain the final user embedding feature representation, and the event sequence representation (x, y) of the rich user-item interaction feature is obtained.

[0150]

[0151] wherein x is the user embedding representation;

[0152] y is the embedding representation of the item;

[0153] y i is the embedding representation of the i-th item, and is the item interacted with the user u.

[0154] H u,i is the historical interaction sequence of the user u and the item i.

[0155] ||r u || is the module length, This calculation is a regularization calculation;

[0156] sign() is, for example, sign(x), which outputs 1 if x is positive, outputs -1 if x is negative, and outputs 0 if x is 0. The purpose is to extract the sign information of x, and perform element-level multiplication with the following content.

[0157] ⊙ is element-level multiplication, Hadamard product. The corresponding elements in two vectors or matrices are multiplied, and the result is a new vector or matrix, whose each element is the product of the corresponding elements of the original two vectors or matrices.

[0158] Hgc(·) is to multiply the normalized user embedding and the item embedding, combine the information of the two through the encoder, input the updated embedding representation, fuse the user embedding and the adjacency matrix, and generate the final user embedding representation.

[0159] r u The embedding vector of individual user biases in the original user-item interaction graph;

[0160] It is an adjacency matrix;

[0161] Encoder() is a hypergraph convolutional encoding representation with self-attention mechanism injected.

[0162] To better model user preferences, semantic information is introduced to separate item bias. A preference assessment is set for the current user, calculated using the inner product of the user and the item.

[0163] α =<x,y> (3)

[0164] Where x is the user's embedded representation, y is the representation of the item interacting with it, and <> is the inner product operation.

[0165] 1.2 Probability Diffusion Paradigm of Interactions

[0166] For the obtained event sequence representation (x, y) containing rich user-item interaction features, a diffusion model is introduced to enhance user preference modeling and reduce the impact of irrelevant features in the recommendation process. The collaborative signals and feature information of user-item interactions are unified. By disrupting the original user-item interactions, iterative learning is used to restore the initial state through probability diffusion. Iterative denoising training integrates information into the user-item interaction embedding representation, mitigating the negative impact of noise features. A diffusion process is introduced into the user-item interaction graph. First, Gaussian noise is gradually introduced to disrupt the original user-item interaction graph, progressively destroying the interaction between users and items, simulating the negative impact of noise features. Second, in the reverse process, the focus is on learning and denoising the damaged graph connection structure, aiming to progressively refine the damaged interaction information to restore the original interaction between users and items.

[0167] During the forward diffusion process, define the interaction z between the currently selected user and a set of items. u =(x,y) u =[(x,y0),(x,y1),…,(x,y2)] I )], Initialize the diffusion process γ0=z u .

[0168] By progressively introducing Gaussian noise in the S-step, and using s as an index, the entire process is parameterized, gradually introducing noise into the user's interaction with items, thus transforming the original interaction into a noisy state.

[0169]

[0170] Wherein, δ(γ) s |γs-1 ) is a forward process that corrupts the user's interaction history by adding Gaussian noise step by step.

[0171] is a Gaussian distribution;

[0172] s∈{0,…,S} is the current state γ s-1 is transformed into state γ s by adding noise. Here the noise follows a Gaussian distribution with mean and variance μ s ·I.

[0173] Two hyperparameters are introduced and The noise at the current state is controlled by a linear noise scheduler, and the noise is added step by step to imitate the randomness in user-item interactions.

[0174]

[0175] where γ s is the state of the current step in diffusion;

[0176] γ0is the state of the initial step;

[0177] is a scaling factor, representing the intensity of the added noise;

[0178] is the scale of the Gaussian noise added at each step (i.e., s step), which takes a value in the range (0, 1);

[0179] εis a standard Gaussian noise, following a Gaussian distribution with mean 0 and covariance matrix I

[0180] The linear noise scheduler is performed for the current state.

[0181]

[0182] where, is the scale of the Gaussian noise added at each step (i.e., s step), which takes a value in the range (0, 1);

[0183] The hyperparameter c∈[0,1] controls the scale of the noise, and the two hyperparameters η max and η min represent the upper and lower bounds of the added noise.

[0184] In the backward diffusion process, the current noise state is eliminated and the initial state is restored, making the diffusion model effectively capture subtle changes in complex generation processes. From the current state γ sInitially, the denoising transition step gradually restores user item interactions.

[0185]

[0186] where p θ (γ s-1 | γ s ) represents the reverse process in diffusion, predicting the conditional probability distribution of the previous state given the current state. θ (γ s , s) is the mean θ (γ s , s) is the covariance matrix, both of which are predicted by the parameterized neural network from the current state and time step.

[0187] represents the Gaussian distribution, and the content in the brackets specifically demonstrates the gradual addition of noise under different states to finally conform to the Gaussian distribution.

[0188] γ s-1 represents the state of the previous time step;

[0189] χ θ (γ s , s) and θ (γ s , s) are the mean and covariance representing the predicted Gaussian distribution, respectively, generated by two neural networks with learnable parameters.

[0190] 1.3 Diffusion optimization

[0191] To effectively guide the learning of θ in the reverse graph diffusion process, the evidence lower bound (ELBO) of the negative log-likelihood of user interactions is maximized. This is achieved by optimizing the log-likelihood function of the latent variable model, maximizing the log-likelihood of the result to enhance the generative ability of the model and maintain the stationarity and diversity of the model. Define reflects the model training objective at different time steps s in the diffusion model optimization process. According to the changes at different steps, it is optimized at different stages to effectively obtain the necessary information from the noise.

[0192]

[0193] where, is the KL divergence between the denoising distribution predicted by the model and the true posterior distribution;

[0194] is the reconstruction term, representing the ability of the model to reconstruct the original data under a given state, and the calculation is the expected value of the log-likelihood of the initial state under the condition of the current time step;

[0195] is the denoising matching term, which describes the denoising step, and describes how to recover the user's interaction history step by step in the reverse process. At each time step, the KL divergence between the denoising distribution predicted by the model and the true posterior distribution is measured;

[0196] logp(γ0) represents the log-likelihood of the original user interaction data, which measures the recovery of the original user interaction data by the model in the reverse process;

[0197] p θ (γ0|γ1) is the reverse process in diffusion, which predicts the conditional probability distribution of the previous state given the current state.

[0198] D KL (δ(γ s-1 |γ s ,γ0)||p θ (γ s-1 |γ s )) is the difference between the true posterior distribution δ(γ s-1 |γ s ,γ0) and the model-predicted posterior distribution p θ (γ s-1 |γ s );

[0199] δ(γ s-1 |γ s ,γ0) represents the true posterior distribution, which refers to the true distribution of the previous state γ s given the current state γ s-1 and the initial state γ0 in the diffusion process, and the Gaussian noise is added step by step to define it;

[0200] γ0 represents the initial state;

[0201] 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.

[0202] When s = 0,

[0203] s is the time step;

[0204] γ s is the state at the current step.

[0205] The Gaussian distribution in the diffusion process reflects the inverse update of the entire process, ensuring that the inverse sampling at each step can completely recover to the state before the noise is introduced, and the complete user-item interaction information is found from the noise.

[0206] The precision of the reverse update is changed by adjusting the parameters therein, and the current diffusion state and the initial diffusion state are calculated using the mean value, so as to ensure the effectiveness of the reverse recovery and the diversity of the data.

[0207]

[0208]

[0209] where ω (γ s , γ0, s) is the mean value expression of each step in the reverse process, describing how to calculate the next step mean value given the current noise state, the original state and the time s.

[0210] ω θ (γ s , s) is described for the convenience of understanding the transition process of the reverse process, and the final change is to push formula 11 to approach formula 10;

[0211] is the noise proportion parameter of time step s, which controls the amount of noise added at each step;

[0212] is the cumulative noise proportion parameter, which represents the cumulative noise proportion from the initial state to the current time step s;

[0213] s is the time step length;

[0214] ω is the mean value of the state of the last time step predicted under the given current state γ s and the initial state.

[0215] is the prediction of the initial state γ0 based on the existing state γ s . The embedding of the current state and the time step is received as input, and the initial state γ0 is output.

[0216] The reverse distribution probability of the time step s to s-1 is calculated under the current diffusion state. The KL divergence is optimized to measure the approximate distribution of the reverse probability distribution, and the expression of the reverse probability distribution is repaired using the Bayesian rule.

[0217]

[0218] where δ (γ s-1 | γ s , γ0) represents how to estimate the conditional probability distribution of the previous state given the current state and the initial state.

[0219] ∝ represents proportional, that is, the probability density on the right side is proportional to the conditional probability on the left side.

[0220] denotes a Gaussian distribution (normal distribution).

[0221] ω(γ s denotes the expression of the mean value of each step in the reverse process, which describes how to calculate the mean value of the next step given the current noise state, the original state and the time s.

[0222] I denotes the unit matrix.

[0223] variance Randomness is added to the sampling process to ensure the diversity of generated data.

[0224] When the initial state is transformed into a state at a certain time step, the model learns to capture complex dynamic processes. To effectively improve the training efficiency, prevent the introduction of noise instability, and avoid the model falling into a local minimum affecting the final training effect. Simplify the state change of the reverse process, ensure the training efficiency and stability, and instantiate based on the step size s s The initial state γ0is predicted based on the current state and the time step embedding, and the loss function of each time step is defined.

[0225]

[0226] where ‖‖2 is the two-norm;

[0227] is the prediction of the initial state γ0based on the existing state γ s The initial state γ0is predicted based on the current state and the time step embedding, and the loss function of each time step is defined.

[0228] 1.4 De-bias optimization

[0229] In order to better adjust the user and item embedding representation in the training process, the popularity bias is introduced to effectively improve the learning ability of the model. The popularity bias of users and items is added in the training process, but only the representation of users and items is considered in the test stage. The popularity feature information and interaction item bias information of the user context are combined to enrich the user embedding representation, and the rating of the recommended interaction item for the user is set.

[0230] β = <x||μ·r u ,y||μ·r i > (14)

[0231] where β represents the rating of the recommended interaction item for the user;

[0232] <·,·> is the inner product operation;

[0233] || is the concatenation operation;

[0234] x||μ·r u y||μ·r i The inner product operation is performed after the splicing of each;

[0235] x and y represent user embedding representation and item embedding representation, respectively;

[0236] μ represents a hyperparameter that adjusts the influence of popularity bias; changes the level of influence on popularity.

[0237] r u and r i represent the bias vectors of users and items, respectively.

[0238] According to the historical behavior data of user past interactions and the relevance of user-item interactions in the interaction graph, the item embedding

[0239]

[0240] wherein, is the embedding matrix of the past user interaction items;

[0241] is the embedding matrix of the items that the user may interact with in the future.

[0242] The score of the recommended item of the user is obtained by the inner product of the user embedding representation containing only the past user interaction items and the item representation that the user may interact with, and the cross-entropy loss function is calculated.

[0243]

[0244] wherein, W1 T is the transpose of W1, and W1 is the parameter weight matrix for calculating the cross-entropy loss function;

[0245] σ is an activation function.

[0246] Based on the cross-entropy loss function, this paper adjusts the loss gradient by combining parameter weights and regularization. This enables the model to learn more accurate user-item interactions that are unbiased by popular items or user preferences, reducing the risk of biased predictions of user-interacting items due to personal bias. A gradient update parameter weight matrix Z is defined. Through partial derivative operations, specific gradient descent and other optimization algorithms are used to adjust the parameters, minimizing the overall loss function and improving the model's recommendation accuracy. This allows the model to adapt to data features while reducing popularity bias. The parameter matrix adjustment process is divided into three categories: existing user-item interactions in the current interaction scenario, possible user-item interactions in future interaction scenarios, and no interaction scenarios. An embedding matrix of items that may interact with users in the future is introduced to correct for potential biases. The hyperparameter weight matrix Z is used to compare and analyze the impact of bias under different interaction scenarios. Gradient calculations are used to adjust the parameter matrix Z to reduce the impact of bias errors, making the model more robust.

[0247]

[0248] 2. Enhanced Contextual Representation and Contrast

[0249] 2.1 Hypergraph Convolution Enhancement

[0250] The attention mechanism is used as a perturbation injection into the hypergraph convolution operation to learn features in multiple subspaces, effectively select neighbor nodes, and consider context and sequence context during propagation.

[0251] The user embedding is enhanced by using graph convolution. First, feature information is extracted and the user embedding representation is enhanced by propagating through multiple layers of convolution.

[0252]

[0253] in, It is the enhanced output of the l-th convolutional layer;

[0254] It is the ELU activation function;

[0255] D -1 / 2 BD -1 / 2 The calculation is of the symmetric normalized adjacency matrix;

[0256] Given a bipartite graph adjacency matrix B, if the same user and item i... f and items i g Interaction, then b in the matrix f,g The position is 1 if it is not a position, otherwise it is 0.

[0257] This represents a diagonal matrix.

[0258] is the parameter weight matrix of convolution propagation.

[0259] The context representation of the user is modeled using hypergraph convolution, the signal propagation in the convolution process is enhanced by learning the context representation from the user hyperedge in the hypergraph, the critical matrix and the feature matrix are adjusted according to the introduced hypergraph information, and the specific row and column corresponding to the user can be selected. The hypergraph convolution is combined with the attention mechanism to enhance the ability of selecting effective neighbors, improve the information diffusion in the hypergraph neural network, and consider the multi-layer meaning of structure and order to promote different representations related to the context.

[0260]

[0261] wherein ψ u is the hyperedge of the user u;

[0262] T is the transpose symbol;

[0263] Norm is the L2 regularization calculation;

[0264] d is the dimension in the self-attention mechanism, which is the dimension of the query matrix Q and the key matrix K, used for the calculation of the self-attention mechanism.

[0265] is the ELU activation function;

[0266] Q, K, V represent the query, key and value matrices in the attention mechanism, respectively;

[0267] W Q , W K is the corresponding parameter weight matrix;

[0268] Φ is a hyperparameter that controls the degree of attention.

[0269] Then, multiple subspace information is aggregated through the multi-head attention mechanism.

[0270]

[0271] wherein W3 is the weight of each head in the multi-head attention mechanism;

[0272] k is the value of the number of attention heads;

[0273] n is the total number of attention heads;

[0274] represents the result of the input embedding of the lth layer after being processed by different attention heads;

[0275] Then, the user embedding representation optimized by the hypergraph convolution is obtained through average pooling and normalization.

[0276]

[0277] M i Previously processed embeddings;

[0278] i∈H u,i is the selection of the embedding belongs to the user-item interaction sequence;

[0279] || is the absolute value symbol;

[0280] H u,i denotes the historical interaction sequence of user u and item i;

[0281] The hypergraph convolution enhancement process is optimized by a multi-label cross-entropy loss function.

[0282]

[0283] wherein, is the true label vector of the current user-item interaction;

[0284] is the true label vector of the user-item interaction; and each dimension of the vector is equal to 1 when the user has visited the item; otherwise, 0.

[0285] α u,1 , α u,N are the predicted ratings of the model for user u on item 1 to item n, each representing the predicted rating of the user on the item.

[0286] σ is a sigmoid activation function.

[0287] 2.2 Contrastive self-supervised optimization

[0288] In the self-supervised recommendation system, the introduction of the contrastive learning framework combined with the difficult negative sample sampling strategy can significantly improve the learning performance. A self-supervised recommendation method based on difficult negative sample sampling is adopted, which fully utilizes the label information and the selection of difficult negative samples in 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 samples 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 considered as a negative sample. Further, multiple different negative sample selection structures are set within the same input sample space to assist the selection of difficult negative samples in the contrastive learning framework. The similarity between samples is calculated by an auxiliary function, and the final loss function is inferred accordingly to optimize the learning process of the model. This process is similar to unsupervised contrastive learning, which randomly selects negative samples in the entire input space.

[0289]

[0290] wherein p(z + ) is the probability distribution of positive samples under different conditions;

[0291] z + , z - are positive samples, negative samples, respectively;

[0292] z is a sample;

[0293] p U (z - ) is the probability distribution of negative samples in the sample set under different conditions;

[0294] represents the probability distribution of selecting positive samples;

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

[0296] represent the positive sample set and the negative sample set, respectively;

[0297] Under the current negative sample selection structure, the selection of difficult negative samples is performed. After the anchor point is selected, the cosine similarity between the anchor point and the sample is calculated according to the sample space set, a difficult negative sample set is constructed, and the negative probability distribution p HU (z - ) of the selection of difficult negative samples is obtained.

[0298]

[0299] wherein, is the expected value of selecting negative samples within the current limited range, i.e. Ψ HU .

[0300] Ψ HU is the specific range of the current condition sample selection.

[0301] exp() is the exponential operation with the natural base as the base;

[0302] w is the sample label;

[0303] is the sample label set;

[0304] d(z,ξ) is the similarity between the current selected node (sample z) and the anchor point;

[0305] is the threshold value for controlling the sampling hardness.

[0306] 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 selected to construct a difficult negative sample set and obtain the corresponding negative probability distribution.

[0307]

[0308] where p S (z - ) is the negative probability distribution of the negative sample selection under the condition

[0309] The label information w z of the sample and the label information w ξ of the anchor point are comprehensively considered. Finally, the above two difficult negative sample sets are combined to construct a difficult negative sample selection set based on the contrast learning framework. This set selects negative samples within a specified range, satisfying two conditions: similar to the anchor point but having different label types, and satisfying the selection condition within the specified selection range

[0310]

[0311] where p HS (z - ) is the negative probability distribution of the difficult negative sample satisfying the condition

[0312] Ψ HS , Ψ S , and Ψ HU represent different limited ranges in the sample selection process.

[0313] Based on the above difficult negative sample selection preparation, the difficult negative sample contrast loss function is established, allowing the sample selection set satisfying the above two requirements to join 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, and narrowing the similar samples and pushing away the dissimilar samples in the current sample space.

[0314]

[0315] where d(·,·) represents the similarity;

[0316] n is the number of negative samples;

[0317] k is the negative sample index;

[0318] represents the difficult negative sample belonging to the current limited range;

[0319] ​To effectively balance the user's bias towards the item features, the user's current interaction with the item features and the user's future possible interaction with the item features, prevent the problem of non-smooth node embedding. By changing the weighting scheme to balance the gradient, reduce the influence of the high popularity of the item on the user's true preference.

[0320]

[0321] where λ is a hyperparameter that balances the weight of the item embedding matrix interacted with the user;

[0322] λ1is used for the item embedding matrix that the user has interacted with;

[0323] ⊙ is an element-wise multiplication, Hadamard product. The corresponding elements in two vectors or matrices are multiplied, and the result is a new vector or matrix, each element of which is the product of the corresponding elements of the original two vectors or matrices.

[0324] η is a weight vector used to weight the loss of different items;

[0325] Θ is a regularization coefficient to control the regularization weight;

[0326] is the loss function in the front;

[0327] is the user set;

[0328] H u,i is the historical interaction sequence of user u and item i;

[0329] λ2is used for the item embedding matrix that the user may interact with in the future.

[0330]

[0331] 3. Experimental analysis

[0332] Experiments are conducted on different datasets by comparing with various state-of-the-art recommendation methods to evaluate the performance of DDCLRec, and the goal is to answer the following research questions:

[0333] · RQ1: How does DDCLRec perform when competing with different types of recommendation methods?

[0334] · RQ2: How does DDCLRec perform in alleviating the problem of data sparsity?

[0335] · RQ3: How effective is DDCLRec in alleviating the noise problem?

[0336] · RQ4: How robust is the recommendation of DDCLRec?

[0337] • RQ5: What are the contributions of different key modules in DDCLRec to the overall performance?

[0338] • RQ6: How do the hyperparameters set in DDCLRec affect the changes in performance metrics Recall and NDCG?

[0339] 3.1 Datasets

[0340] To evaluate the effectiveness of DDCLRec, experiments were conducted on four public datasets collected from different real-life platforms to reflect real-life scenarios: ML-1M, Yelp2018, Douban-book, and Foursquare, which differ in size and sparsity, and the dataset statistics are shown in Table 3.

[0341] Table 3 Statistics of experimental datasets

[0342]

[0343] 3.2 Evaluation metrics

[0344] For performance evaluation, three representative metrics were used: Precision@k, Recall@k, and NDCG@k to measure the performance of the proposed DDCLRec. The top-k recommendations, with k values set to 5, 10, and 20. The method DDCLRec was implemented using PyTorch, and its performance was compared with various baseline methods using official or third-party code. The calculation formulas for Precision, Recall, and NDCG are as follows:

[0345] 1) Recall (Recall rate) size reflects how much information users are interested in being perceived, R(u) represents the Top-k recommendation list made for the user based on the user's behavior on the training set;

[0346] T(u) represents the set of items actually selected by the user after the system recommends items to the user.

[0347]

[0348] 2) Precision (Precision), 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 actually selected by the user after the system recommends items to the user.

[0349]

[0350] 3) CG (Cumulative Gain), the relevance score of each recommendation result in the recommendation list is accumulated; however, it is possible that in a list, high scores are placed in low positions, and low scores are placed in high positions.

[0351]

[0352] 4) DCG (Discounted Cumulative Gain): Based on CG, DCG introduces the positional factor to "discount" the recommendation effect of lower-ranked recommendations.

[0353]

[0354] 5) NDCG (Normalized Discounted Cumulative Gain): Evaluating a recommender system involves comprehensively assessing the recommendation lists of all users across the entire test set, normalizing the evaluation scores for different users' recommendation lists. NDCG refers to the best recommendation list returned by the recommender system for a given user.

[0355]

[0356] 3.3 Baseline Method

[0357] Performance evaluations of DDCLRec on Recall@5, @10, @20 and NDCG@5, @10, @20 on four datasets including ML-1M and Yelp2018 are compared with those of 16 existing popular recommendation system methods such as LightGCN and AutoCF.

[0358] Graph Convolutional Neural Networks:

[0359] LightGCN: Uses a lightweight convolutional graph encoder for better representation learning and model training.

[0360] • LCFN: A triple training framework based on self-supervised learning that integrates users' social information to enhance the learning of the multi-view encoder and uses self-supervised signals generated by other users to iteratively optimize representation learning in the recommendation system.

[0361] Learning by comparison with diagrams:

[0362] AutoCF: Learnable generative self-supervised learning enables automatic data augmentation, improving the representation capabilities of collaborative filtering models.

[0363] ·HCCF: A cross-view contrastive learning architecture for hypergraph enhancement to jointly capture local and global collaborative relationships.

[0364] • NCL: Incorporates potential neighbors in the structural and semantic spaces into contrast pairs, enhancing the performance of graph collaborative filtering.

[0365] • SGL: generates multiple views and maximizes the similarity between different views of the same node, enhancing representation learning.

[0366] • XSimGCL: uses noise-based embedding enhancement to generate views.

[0367] • LightGCL: a simplified graph contrastive learning paradigm that enhances contrast through singular value decomposition.

[0368] Debiasing deep learning:

[0369] • DICE: causal inference separates user interests and conformity behavior, learning independent representations.

[0370] • InvCF: uncovers decoupled representations that are not affected by changes in popularity distribution, reflecting latent preferences and popularity semantics.

[0371] • STaTRL: Transformer captures long-range dependencies in user check-in sequences.

[0372] Diffusion model:

[0373] • DiffRec: denoising process learns the generative process of user interactions.

[0374] Mitigating over-smoothing:

[0375] • IMP-GCN: interest-aware message passing graph convolutional network.

[0376] • GDE: simple and effective graph denoising encoder.

[0377] • EEDN: enhanced encoder-decoder network that combines hybrid hypergraph convolution to enhance the aggregation of graph convolution steps.

[0378] • CaDRec: introduces hypergraph convolution operations with structure and sequence context to select effective neighbors, mitigating over-smoothing.

[0379] 3.4 Comparative analysis experiment with baseline methods of recommendation system (RQ1)

[0380] In Table 4, the Recall@k and NDCG@k performance evaluation of DDCLRec compared with 16 baseline methods such as LightGCN, AutoCF, etc. on four real datasets ML-1M, Yelp2018, Douban-book, Foursquare, where k is set to 5, 10, 20.

[0381] Table 4 shows the best performance comparison of Recall@k and NDCG@k of DDCLRec and 16 baseline methods on four datasets ML-1M, Yelp2018, etc.

[0382]

[0383]

[0384] The experimental results in Table 4 show that DDCLRec outperforms other baselines in all cases, which verifies the effectiveness of introducing diffusion model enhanced embedding representation and hard-negative supervised contrastive learning optimization strategy. The diversity of the evaluation dataset varies with the interaction graph sparsity, knowledge graph features, and recommendation scenarios. The research results prove the universality and flexibility of DDCLRec. Overall, the progress made by DDCLRec can be attributed to two aspects:

[0385] 1) The hypergraph convolutional network is used to encode the embedding representation of users and items to capture complex multi-edge relationship structures. The diffusion model is introduced to add random noise, preserve and enhance the diversity of the original data, so that the embedding representation has greater differentiation. It further improves the modeling ability of user preferences and promotes the derivation of the unbiased recommendation.

[0386] 2) The embedding representation of users and items is optimized by the self-supervised signal of hard-negative sample contrastive learning, and combined with the multi-label cross-entropy loss optimization rate, the model's perception of user preferences and the distinctiveness in the embedding space are enhanced, and the context features are enriched.

[0387] According to the content of Table 4, it can be seen that DDCLRec achieves better performance than most recommendation methods on four real datasets such as ML-1M, Yelp2018, etc. For example, for the dataset ML-1M, DDCLRec improves LightGCN by about 11.2% in Recall@20 and about 7% in NDCG@20. Compared with AutoCF, it improves by about 27.5% in Recall@20 and about 22.4% in NDCG@20. For the dataset Douban-book, DDCLRec improves LightGCN by about 42.4% in Recall@20 and about 56.3% in NDCG@20. Compared with AutoCF, it improves by about 54.2% in Recall@20 and about 57.9% in NDCG@20. Good performance is also achieved on the other two datasets. It confirms the effectiveness of combining hypergraph convolution coding with diffusion model to solve popularity bias and other problems in recommendation system. Existing recommendation methods mainly use the associated items in the current user historical behavior data for collaborative filtering to calculate recommendations, often ignoring a lot of edge set relevance information. But the advantage of DDCLRec on some datasets is not obvious, indicating that the noise introduced by the diffusion model affects the analysis of user individual bias, leading to distortion of user-item interaction data, and the selection strategy of difficult negative samples may not be perfect enough, leading to user interest bias. DDCLRec still has a lot of room for improvement in these aspects.

[0388] 3.5 Analysis of alleviating data sparsity experiments (RQ2)

[0389] To explore the robustness of the model in the data sparse scene, the user-item interaction graph is randomly deleted with a 10% proportion of interaction edges for sparsity experiments. According to a 10% proportion of edge information reduction of the current user-item interaction graph, in the absence of sufficient interaction information, the performance of DDCLRec and other different recommendation methods in this scenario is evaluated, especially the performance of handling user data sparsity problems in the absence of sufficient interaction information as shown in Figure 2 Precision@k, Recall@k and NDCG@k indicators are used, where k is set to 10, 20.

[0390] In Figure 2In the scenario of 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 problem of data sparsity. The advantage of DDCLRec in alleviating data sparsity is more prominent on the dataset ML-1M, but the advantage on Yelp2018 is weaker. The possible reason is that Yelp2018 contains more complex node and interaction edge information, and the removal of a certain proportion of user-item interaction edge information leads to greater user individual bias, and the bias problem caused by less recommendation basis. This shows that DDCLRec is not very good at dealing with complex network node mining and large amounts of unlabeled data training, and there is still a lot of room for improvement. At the same time, it is also observed that the lower the completeness of the interaction graph, the lower the performance. Slightly reducing the completeness of the interaction graph is also beneficial to alleviate the overfitting problem and improve the recommendation performance, which shows that reasonable selection of interaction information ratio and learning rate of interaction graph in the training process is more conducive to improving the generalization ability of the model.

[0391] 3.6 Noise alleviation analysis experiment (RQ3)

[0392] Noise in recommendation may distort user-item interaction data, leading to recommendation bias, popularity bias and other problems. High noise level may reduce the signal-to-noise ratio in the user-item interaction graph, and reduce the recommendation quality and stability. After encoding the user-item interaction using hypergraph convolution, a 10% random scaling tensor is added to the output of the encoder to introduce noise, which is included in the recommendation basis. The robustness of the recommendation method is verified through performance evaluation as shown in Figure 3 Precision@k, Recall@k and NDCG@k indicators are used, where k is set to 5, 10 and 20.

[0393] In Figure 3In the figure, the horizontal axis is different data sets, and the vertical axis is Precision@5, @10, @20, Recall@5, @10, @20 and NDCG@5, @10, @20. For the existing embedding representation set, a certain proportion of random scaling tensors are added, which will cause the recommendation to deviate when the recommendation basis is included. In the current noise scene, DDCLRec has an advantage on four data sets such as ML-1M, Yelp2018, etc., which proves the effectiveness of DDCLRec in alleviating noise disturbance. It shows the necessity of the combination of hypergraph convolution coding and diffusion model of DDCLRec, which can effectively alleviate the user interest bias problem in the noise disturbance scene. Among them, the advantage on the ML-1M data set is more obvious, while the advantage on the Foursquare data set is weak. It may be because there are more user nodes and fewer item nodes in the Foursquare data set, and the node interaction information is less. In the process of storing hyperedge information in complex hypergraph convolution, adding noise is more likely to cause user interest bias, so that the real preference of the user is covered by the high popularity of the item, and it is difficult to discover the real interest of the user. It shows that in terms of alleviating noise disturbance, DDCLRec still has a lot of room for improvement in improving the coding method and embedding vector calculation method.

[0394] 3.7 Robustness analysis experiment (RQ4)

[0395] In the evaluation of the recommendation system, Precision is used to measure the accuracy of the recommendation result, defined as the proportion of truly relevant items in the recommended items. High Precision indicates that the proportion of relevant items in the recommendation result of the recommendation system is high, indicating that the accuracy of the recommendation system is high, which generally decreases with the increase of the K value of Top-K. 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 perform multiple experiments on four data sets such as ML-1M, Yelp2018, etc. The Precision changes of different models are analyzed as shown in Figure 4

[0396] In Figure 4 ​It is observed that the precision of different models gradually decreases with the increase of K value of Top-K on four real datasets such as ML-1M, Yelp2018, etc. In the initial stage, the precision of CaDRec is higher in some datasets, but as K value increases, the precision of DDCLRec changes more stably compared with EEDN and CaDRec, especially on datasets Yelp2018 and Foursquare. This shows that DDCLRec can maintain high accuracy in longer recommendation lists, and can effectively mine the real interests of users even in longer recommendation lists. It proves that DDCLRec can maintain good recommendation quality and robustness when facing more recommendation options. However, the advantage of DDCLRec on dataset ML-1M is not obvious, and the node and interaction information of ML-1M is less than that of other datasets, which shows that there is still a lot of room for improvement in the strategy of effectively mining user interest preferences in the process of supergraph convolution and multi-layer graph neural network propagation.

[0397] 3.8 Ablation experiments (RQ5)

[0398] DDCLRec outperforms other baseline methods in terms of Recall@k and NDCG@k performance indicators. The superior results of DDCLRec can be attributed to the following factors:

[0399] The supergraph convolution network is used to encode and embed users and items to capture complex multi-relation. The diffusion model is introduced to preserve and enhance the diversity of the original data, distinguish embedded representations, strengthen user preference modeling, and achieve unbiased recommendation. In addition, the optimization strategy of hard negative supervised contrastive learning loss and multi-label cross-entropy loss is combined to optimize embedded representations, enrich context features, and enhance the model's ability to recognize user preferences in the embedding space.

[0400] To study the effectiveness of the key components of DDCLRec, three model variants are used for ablation study, as shown in Table 5. Two DDCLRec method variants are as follows:

[0401] 1) DR: Only use supergraph convolution to encode embedded representations and perform multi-layer graph neural network propagation enhancement. Do not introduce diffusion model and discard hard negative contrastive learning loss optimization strategy.

[0402] 2) DDR: Use supergraph convolution to encode embedded representations, introduce diffusion model, and perform multi-layer graph neural network propagation enhancement. Discard the hard negative contrastive learning loss optimization strategy.

[0403] 3) DCLR: using hypergraph convolutional coding embedding representation, multi-layer graph neural network propagation enhancement, taking hard negative contrastive learning loss optimization strategy. No diffusion model is introduced.

[0404] The comparative analysis experiment of Recall@k and NDCG@k indexes of DDCLRec and three method variants on four real data sets such as ML-1M is shown in Table 5, where the value of k is set to 5, 10, 20.

[0405] Table 5 Evaluation results of Recall@k and NDCG@k indexes of DDCLRec and three method variants on four real data sets

[0406]

[0407] In Table 5, DDCLRec has superior performance in Recall@5, @10, @20 and NDCG@5, @10, @20 indexes on four data sets such as ML-1M, Yelp2018. It shows the necessity of combining hypergraph convolutional coding and diffusion model, verifies the importance of combining hard negative contrastive learning loss optimization strategy and multi-label cross-entropy loss optimization strategy when evaluating user preferences, and proves that DDCLRec is reasonable and effective for solving popularity bias and other problems in recommendation systems.

[0408] 3.9 Hyperparameter analysis experiment (RQ6)

[0409] DDCLRec has three important hyperparameters. Important weight parameters λ1, λ2 and hard negative sampling depth τ in the loss optimization process. The influence of different combinations of λ1 and τ and different combinations of λ2 and τ on the recommendation performance is studied, respectively from (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) range of τ, λ1 and λ2, taking ML-1M data as an example, the influence of hyperparameters on Recall@k and NDCG@k indexes of DDCLRec is shown in Figure 5 , where the value of k is set to 10, 20.

[0410] In Figure 5 (a) and Figure 5 (b), it is found that when τ takes the value of 10 -1 , λ1 takes the value of 0.5, DDCLRec has the best performance, and the performance decreases significantly when the values of the two parameters are changed. In Figure 5 (c) and Figure 5 (d), it is found that when τ takes the value of 10 -2, λ2 takes the value of 0.41, the performance of DDCLRec is optimal. According to the experimental results, it is found that DDCLRec is more sensitive to τ. When λ1 is appropriately increased, the convergence is faster and helps to improve the performance of the model. However, too high λ2 will affect the performance, which shows that the participation and change of τ can greatly improve the recommendation performance, but the overall excellent performance still needs the balance of λ1 and λ2. It proves the necessity of hard negative contrastive learning for the optimization of recommendation method. Therefore, according to the consideration of weight parameters in different ranges, an effective weighting scheme is synthesized, τ, λ1 and λ2 are set to 10 -1 , 0.5, 0.41 respectively, τ, λ1 and λ2 are set to 10 -2 , 0.46, 0.41 respectively, which can make DDCLRec achieve better performance.

[0411] In summary, the application proposes a diffusion model-based contrast learning debiased recommendation method DDCLRec. First, the hypergraph convolution network is combined with the context information to capture the multi-edge complex relationship between users and items, encode the user-item interaction, and effectively pass valuable information to the modeling process of user-item interaction. At the same time, the diffusion model is introduced to retain data diversity, differentiate embedding representation, enhance user preference modeling and derive debiased representation. Secondly, different combinations of regularization and weighting schemes are adopted to enhance the embedding representation of users and items through hard negative supervised contrast learning self-supervised signals, further improve the method's ability in user preference learning, and the performance in user preference perception and embedding space distinguishability, and enrich the method's understanding of context features. In order to verify the effectiveness and feasibility of the method, multiple comparative analysis experiments are carried out on four real datasets such as ML-1M, Yelp2018, and the DDCLRec is compared with 16 existing popular recommendation methods. The experimental results show that: in terms of recommendation system performance evaluation, the proposed DDCLRec is 0.8% to 1.4% better than the latest CaDRec method in Recall@k index and 0.5% to 1.1% better than the latest CaDRec method in NDCG@k index on the ML-1M dataset, which highlights the accuracy and superiority of DDCLRec; in the scene of sparse user-item interaction graph, DDCLRec has certain advantages in Precision@k, Recall@k and NDCG@k on four real datasets such as ML-1M and Yelp2018, which proves that DDCLRec can better recommend items for users in the processing of sparsity problem; in the scene containing noise, DDCLRec is superior to existing popular recommendation methods EEDN and CaDRec in Precision@k, Recall@k and NDCG@k on four real datasets such as ML-1M and Yelp2018; in terms of recommendation stability, DDCLRec performs more stably with the increase of recommended item list, which proves that DDCLRec has more stable and effective method performance in the scene of a large number of item recommendations; in terms of ablation analysis and hyperparameter sensitivity, the influence of different hyperparameters on Recall@k and NDCG@k of DDCLRec is explored, which proves the necessity of introducing diffusion model to enhance hypergraph convolution coding embedding representation of DDCLRec, and the importance of hard negative supervised contrast learning loss optimization strategy. The above experiments fully demonstrate the advantages of DDCLRec in solving the key problems faced by the recommendation system, and highlight the effectiveness of DDCLRec.

[0412] However, there is still much room for improvement in DDCLRec, especially in handling long-tail items and user bias. Distinguishing between user's real preference items and popular items is crucial to alleviate user interest bias. In addition, the application of contrastive learning in recommendation systems can improve the accuracy and efficiency of recommendation systems, which requires the combination of 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 an increase in memory requirements and time in the case of larger datasets. The 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 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 to effectively save time and space resources, which is more conducive to the development of recommendation system research. Therefore, developing lightweight recommendation methods and exploring effective bias-removing recommendation methods to mine users' real preferences are crucial for the high-quality development of recommendation systems.

[0413] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for intelligent recommendation verification by fusing hypergraph and context information, characterized in that, The method comprises the following steps: S1, extracting the bias item features, item popularity features and item semantic features from the interaction graph of user items, combining the bias item features, item popularity features and item semantic features of individuals and neighbor nodes, and using hypergraph convolution to encode the representations of users and items to obtain encoded embeddings, wherein the encoded embeddings comprise user embedding representations and item embedding representations; S2, performing hypergraph convolution enhancement on the encoded embeddings using hypergraph convolution to obtain user embedding representations and item embedding representations optimized by hypergraph convolution; and then calculating the scores of recommended interaction items of the user, arranging the scores in descending order to generate a recommendation list of relevant candidate items; the hypergraph convolution enhancement comprises: S2-1, enhancing the user embedding representations by extracting feature information through multi-layer convolution propagation; ， wherein, is the output of the first convolutional layer; is an ELU activation function; is a symmetric normalized adjacency matrix; is the two-part graph adjacency matrix; denotes a diagonal matrix; is a parameter weight matrix for convolution propagation; S2-2, modeling the context representation of the user using hypergraph convolution, enhancing the signal propagation in the convolution process by learning the context representation from the user hypergraph in the hypergraph, and adjusting the critical matrix and the feature matrix according to the introduced hypergraph information to select specific rows and columns corresponding to the user; ， wherein, is a hyperedge of users ​ is a transpose symbol; is the L2 regularization computation; is a dimension in the self-attention mechanism; is an ELU activation function; , , respectively represent the query, key and value matrices in the attention mechanism; , is the corresponding parameter weight matrix; is a hyperparameter that controls the degree of attention; S2-3, aggregating multiple subspace information through a multi-head attention mechanism: ， wherein, is the weight of each head in the multi-head attention mechanism; is the value of the number of attention heads; is the total number of attention heads; represents the input embedding of the i-th layer after being processed by different attention heads; S2-4, obtaining the user embedding representations and item embedding representations optimized by hypergraph convolution through average pooling and normalization; ， ， wherein, is a user embedding representation optimized through hypergraph convolution; is an embedded object that has already been processed; is embedded in the selected user-item interaction sequence; is the absolute value symbol; representing a user historical interaction sequence with an item ​ is an embedding representation of the item optimized by hypergraph convolution; For the first Embedding representation for the item that the user u interacted with; is a sign function; is the Hadamard product symbol; an embedding vector for a user individual bias for the original user item interaction graph; For the module length. 2.The intelligent recommendation verification method of fusing supergraph and context information according to claim 1, wherein, S1 comprises the following steps: S1-1, transforming the user-item interaction graph information into an embedded representation of user-item interaction, deriving the user and the historical interaction sequence of the item , determining the user interacting with the current item , and then obtaining the embedded representation of the item through uniform distribution random sampling , wherein the embedded value of the item is randomly generated within the uniform distribution range;​ S1-2, in the historical behavior sequence of user-item interaction, according to the current user interaction item, the embedding vectors of all interaction items are integrated to obtain a user embedding representation ; S1-3, the integrated embedded representation of the item interacting with the user, is fitted with biased interaction according to the item bias feature, the item popularity feature and the item semantic information in the user interaction process, so as to obtain the final user embedded feature representation and obtain the event sequence representation of the user item interaction feature . 3.The intelligent recommendation verification method of fusing supergraph and context information according to claim 2, characterized in that, The calculation formula of the user embedding representations and the item embedding representations is as follows: ， ， wherein, is a user-embedded representation; an embedded representation of an item; For the first Embedding representation for the item that the user u interacted with; for a user with an item historical interaction sequence; an embedding vector for a user individual bias for the original user item interaction graph; For the module length; is a sign function; is the Hadamard product symbol; For hypergraph convolution; is an adjacency matrix; is an encoding representation for hypergraph convolution. 4.The intelligent recommendation verification method of fusing supergraph and context information according to claim 1, wherein, The loss function of the hypergraph convolution enhancement is : ， wherein, is a set of users; is a user a true label vector for the interaction with the item 1; is a user to an item a true label vector for the interaction; ... is a prediction score for the user for the item 1 to the item ​ is a sigmoid activation function. 5.The intelligent recommendation verification method of fusing supergraph and context information according to claim 1, characterized in that, The method further comprises contrastive self-supervised optimization: using a contrastive self-supervised learning method to enhance cross-view contrastive learning of the multi-feature embedding representations of user-item interactions, and learning individualized features of different nodes to prevent bias; And using a hard negative supervision contrastive learning loss optimization strategy, selecting samples with different labels but similar types within the current range as negative samples. 6.The intelligent recommendation verification method of fusing supergraph and context information according to claim 5, characterized in that, The contrastive self-supervised optimization comprises: First, a sample selection strategy based on random probability is designed to calculate the distribution probability 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 considered as a negative sample; and in the same input sample space; ， wherein, is the probability distribution of positive samples under different conditions; , positive and negative samples, respectively; for the sample; P (Y = y | X = x) is the probability distribution of the negative samples under different conditions in the sample set; P (Y = 1 | X) represents the probability distribution of selecting a positive sample; a probability distribution representing the selection of negative samples; , positive sample set and negative sample set, respectively; Under the current negative sample selection structure, the selection of difficult negative samples is carried out: after the anchor point is determined, the cosine similarity between the anchor point and the samples is calculated according to the sample space set, a difficult negative sample set is constructed, and a negative probability distribution of difficult negative sample selection is obtained : ， wherein, is the negative sample expectation selected from the current defined range, i.e. the negative sample expectation of the current defined range. is a specific range of the current condition sample selection; is an exponential operation with a natural base; is a sample label; is a set of sample labels; similarity of the current selected node and the anchor node; threshold value for controlling the sampling hardness; In the sample space combined with the same anchor point, a method of selecting samples similar to the anchor point but different in type is used to construct a difficult negative sample set and obtain a corresponding negative probability distribution: ， wherein, is the negative probability distribution for the negative sample selection under conditions. Label information of the sample Label information of the anchor point The two difficult negative sample sets are finally combined to construct a difficult negative sample selection set based on a contrast learning framework; this set selects negative samples within a specified range, satisfying two conditions: similar to the anchor point but having different label types, and satisfying the selection condition within the specified selection range ; ， wherein, difficult negative sample negative probability distribution that satisfies the condition; , , respectively represent different defined ranges in the sample selection process.

7. The method of claim 6, wherein, The contrastive loss function of the difficult negative samples is: ， wherein, represents a similarity degree; is the number of negative samples; is a negative sample index; represents a difficult negative sample belonging to the current defined range; In addition, by changing the weighting scheme to balance the gradient, the influence of high-popularity items on the real preferences of users in the recommendation is reduced, and at this time, the loss function is represented as follows: ， wherein, a hyperparameter balancing the embedding matrix weights of the item interacting with the user; is the Hadamard product symbol; is a weight vector for weighting the loss for different items; regularization weight is controlled by the regularization coefficient; a set of users; for the user with the item historical interaction sequence. 8.The intelligent recommendation verification method of fusing supergraph and context information according to claim 1, wherein, The calculation formula of the scores of recommended interaction items of the user is as follows: ， wherein, represents a score of recommending the interactive item to the user; is an inner product operation; stitching operation; The inner product operation is performed after the stitching operation for each of the calculations. , respectively denote user embedding representations and item embedding representations; a hyperparameter representing the influence of the regulation of the popularity bias; and and respectively represent the bias vectors of users and items. 9.The intelligent recommendation verification method of fusing supergraph and context information according to claim 1, wherein, The method further comprises verifying the generated recommendation list of relevant candidate items by calculating the recall rate, the precision rate, the cumulative gain, the normalized cumulative gain, and the normalized cumulative gain.

Images