Cold start recommendation method and system based on adaptive aggregated neighbor graph

Abstract

The application relates to the technical field of recommendation systems, in particular to a cold start recommendation method and system based on an adaptive aggregated neighbor graph, which first adopts an SE (self-ensemble) network to learn the relative importance of second-order feature interaction, then obtains high-order attribute feature representation of a user / item through an MLP (multi-layer perceptron) network to obtain an attribute feature vector of the user / item; an extended variational autoencoder is used to reconstruct missing user preferences / item characteristics from the attribute features of the user / item; an attribute graph attention network is used to aggregate neighbor information of the user / item and interaction information between neighbors, so that a final vector representation of the user / item is obtained; and finally, a non-linear interaction between user and item features is introduced through an MLP and an inner product to obtain a score of the user to the item. The application aims to fully mine the complex relationship between users / items by using a graph structure, so as to realize more accurate recommendation in a cold start environment.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and in particular to a cold-start recommendation method and system based on adaptive aggregated neighbor graph. Background Technology

[0002] Recommender systems have been widely applied in online business communities such as e-commerce platforms and news platforms, aiming to alleviate information overload for users and thus improve platform productivity. Their main goal is to analyze user preferences through explicit or implicit user feedback data and predict potential user needs to provide personalized recommendation services. Although much research has been dedicated to developing reliable and efficient recommendation algorithms, these existing works still face the challenge of the cold start problem.

[0003] Typically, the historical interactions between users and items are crucial for inferring a user's interest in a particular item. However, it's quite common to encounter new users or items with no historical records. In such cases, traditional recommendation methods based on historical interaction records are no longer effective.

[0004] In the strictly cold-start recommendation problem, where no historical information is available, many studies have focused on using ancillary information to mitigate the lack of user preferences. These studies treat the cold-start problem as an issue of incomplete user preference data and attempt to reconstruct user preferences using this ancillary information. Therefore, how to comprehensively characterize user preferences and item properties is considered an urgent problem to be solved.

[0005] The current challenge lies in how to effectively extract the complex relationships and information from sparse features to accurately characterize user preferences.

[0006] The first problem is how to overcome the limitations of simple feature interaction methods. Feature interactions are not just simple linear addition relationships, but involve more complex nonlinear relationships. Simple second-order feature interactions usually only consider pairwise combinations of features, which may ignore higher-order interaction information, leading to information loss and insufficient model expressive power.

[0007] The second issue is how to reflect the importance of different feature interactions. Feature interactions learn all cross-features, including useless cross-combinations. Useless cross-features introduce noise and degrade model performance. Therefore, it is necessary to select the most important cross-features from among many features. For example, if the application scenario is predicting the probability that a male user will buy a keyboard, then the cross-feature "gender = male" and "purchase history includes mouse" is likely to be more important than the cross-feature "gender = male" and "age = 30". The model should devote more "attention" to the former feature.

[0008] The third issue is how to fully mine and aggregate neighbor information. Currently widely used Graph Attention Networks (GATs) progressively aggregate neighbor node features to generate node embeddings, typically assuming that adjacent nodes are independent and ignoring the mutual influence between potential neighbors. This leads to insufficient capture of the global graph structure during information propagation, limiting the model's ability to learn node representations. Summary of the Invention

[0009] The purpose of this invention is to solve the problem of how to fully explore user characteristics and their complex relationships in a strict cold start environment. It proposes a cold start recommendation method and system based on adaptive aggregated neighbor graph, abbreviated as A-GAR. Its main purpose is to fully explore the complex relationships between users / items by utilizing graph structure, so as to achieve more accurate recommendations in a cold start environment.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A cold-start recommendation method based on adaptive aggregated neighbor graph, comprising:

[0012] First, SENet is used to learn the relative importance of second-order feature interactions. Then, an MLP network is used to obtain the high-order attribute feature representations of users / items, resulting in the attribute feature vectors of users / items.

[0013] Then, the missing user preferences / item characteristics are reconstructed from the user / item attribute features using an extended variational autoencoder;

[0014] Then, an attribute graph attention network is used to aggregate the neighbor information of users / items and the interaction information between neighbors to obtain the final vector representation of users / items.

[0015] Finally, by introducing a non-linear interaction between user and item features through MLP and inner product, the user's rating of the item is obtained.

[0016] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the process of obtaining the attribute feature vector of user / item is as follows:

[0017] First, the original attributes are learned using an embedding layer and converted into embedded representations; then, second-order interaction vectors of the attributes are extracted through feature learning.

[0018] Next, the attention mechanism of SENet is introduced to perform weighted fusion of the second-order interaction vectors;

[0019] Finally, by combining the linear combination vector with the second-order interaction fusion vector through the multilayer perceptron model, the attribute feature vectors of users / items are obtained.

[0020] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the attention mechanism of SENet is further introduced to perform weighted fusion of second-order interaction vectors, including:

[0021] First, a compression operation is performed by averaging the K-dimensional second-order interaction vector using global average pooling to obtain a numerical value:

[0022]

[0023] Where h = [1,1,…,1], It is the second-order interaction vector between the i-th attribute and the j-th attribute. Represents the second-order interaction vector for user u. The compressed value, z u This represents the set of compressed values ​​of the second-order interaction vectors of user u.

[0024] Next, a two-layer MLP network is used to learn the importance of the second-order interaction vector of user u:

[0025] τ u =Sigmod(W t LeakyReLU(W z z u ))

[0026] Where Sigmod and LeakyReLU are activation functions, W t W z Let τ be the weight vector. u Let it be represented as the set of second-order interaction vector weights for user u;

[0027] Finally, by multiplying the corresponding weights by the second-order interaction vector of user u, the weighted operation on important interaction features is completed, resulting in the weighted second-order interaction fusion vector υ of user u. u .

[0028] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the interaction operation of linear combination vector and second-order interactive fusion vector is further represented as follows:

[0029]

[0030] χ u =LeakyReLU(W v υ u +W k κ u +b)

[0031] Among them, W v W kLet b be the weight matrix, b be the bias vector, LeakyReLU be the activation function, and κ be the weight matrix. u This represents a linear combination vector of users u. This represents the attribute characteristics of user u. This represents the value of the i-th attribute of user u.

[0032] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the reconstruction of missing user preference / item characteristics from user / item attribute features using an extended variational autoencoder includes:

[0033] The variational autoencoder was extended by introducing the constraint of reconstructing the feature approximating the preference vector, thereby deriving an effective representation of user preference / item characteristic features; finally, attribute features and preference / item characteristics were combined to form fused features.

[0034] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, an attribute graph attention network is further used to aggregate the neighbor information of users / items and the interaction information between neighbors to obtain the final vector representation of users / items, including:

[0035] First, construct a user / item graph using user / item attribute information;

[0036] Then, a bilinear aggregator is used to model the interaction relationships between neighbors in the graph, and a graph attention network is used to aggregate neighbor information.

[0037] Finally, we obtain the final vector representation of the user / item.

[0038] According to the cold start recommendation method based on adaptive aggregated neighbor graph of the present invention, the graph attention network aggregated neighbor information further includes: using the graph attention model GAT to capture the complex relationships between nodes in the user / item graph structure, using the attention mechanism to dynamically learn the importance between nodes, and weighting and aggregating neighbor nodes according to the importance of their neighbor nodes to generate a new vector representation;

[0039] The interaction relationships between neighbors in the graph are modeled using a bilinear aggregator, as expressed in the following formula:

[0040]

[0041] Here, ⊙ represents element-wise multiplication. Indicates user u i The neighbor interaction feature vector, user u i neighbor node u m and u n Belongs to set Indicates user u i Neighbor node u m u n Attribute feature vectors; Represents the target node u i The corresponding number of associations, d i =|N(u i | represents user u i degree of nodes, W m W n This is the weight matrix.

[0042] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the final vector representation of user / item is further as follows:

[0043]

[0044] in, User u i The final vector representation, where β is a hyperparameter, User u i The neighbor aggregation feature vector, User u i eigenvectors.

[0045] According to the cold-start recommendation method based on adaptive aggregated neighbor graph of the present invention, the predicted rating of user u for item v is further modeled as follows:

[0046]

[0047] Where, p u It is the final feature representation of user u, q v b is the final characteristic representation of item v. u ,b v μ represents user bias, item bias, and global bias, respectively.

[0048] Furthermore, this invention also proposes a cold-start recommendation system based on an adaptive aggregated neighbor graph, used to implement the cold-start recommendation method based on the adaptive aggregated neighbor graph described above, comprising:

[0049] The attribute fusion module is used to learn the relative importance of second-order feature interactions using SENet, and then obtains the high-order attribute feature representation of users / items through an MLP network to obtain the attribute feature vector of users / items.

[0050] The preference generation module is used to reconstruct missing user preference / item characteristics from user / item attribute features using an extended variational autoencoder;

[0051] The neighbor aggregation module is used to aggregate the neighbor information of users / items and the interaction information between neighbors using an attribute graph attention network to obtain the final vector representation of users / items.

[0052] The rating prediction module is used to introduce a non-linear interaction between user and item features through MLP and inner product to obtain the user's rating for the item.

[0053] Compared with the prior art, the present invention has the following advantages:

[0054] This invention proposes a cold-start recommendation method based on adaptive aggregated neighbor graphs, aiming to solve the problem of fully aggregating various information from sparse attributes to comprehensively and accurately characterize user preferences and improve model expressiveness and prediction accuracy. First, a channel attention mechanism (Squeeze-and-Excitation Network, SENet) and an MLP network are introduced to obtain high-order attribute feature representations of users / items. The channel attention mechanism flexibly captures the importance of different feature interactions, assigning different weights to each interaction, improving the ability to mine latent feature relationships and enhancing the generalization ability of the linear regression model. Then, an extended variational autoencoder is used to reconstruct missing user preferences / item characteristics from user / item attribute features. Second, an attribute graph attention network is used to aggregate neighbor information and interaction information between neighbors. This invention extends Graph Attention Networks (GAT) by introducing a bilinear aggregator (BA) to model the interaction relationships between neighbors, fully considering the interaction between neighbor nodes and integrating it into the embedding representation of the target node, effectively improving the quality of the target node's embedding representation. This invention uses high-quality node embedding representations to generate accurate personalized recommendation results for cold-start users. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the cold start recommendation method based on adaptive aggregated neighbor graph according to an embodiment of the present invention.

[0057] Figure 2 This is an overall framework diagram of the cold start recommendation method based on adaptive aggregated neighbor graph according to an embodiment of the present invention;

[0058] Figure 3 This is an example diagram of user graph construction according to an embodiment of the present invention;

[0059] Figure 4 This is a graph showing the ablation experiment analysis results of an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] This solution focuses on strict cold-start recommendation, where new users and items have no historical data. It addresses the cold-start recommendation problem by utilizing their attribute information. Assume a recommendation system has two types of entities: users and items. The user set is represented as... Item sets are represented as in and These represent the number of users and the number of items, respectively. Taking user u as an example during cold start, for any user... The attribute set is represented as Where k is the number of user attributes, Let be the value of the i-th attribute of user u.

[0062] This solution addresses the cold start scenario, specifically the recommendation problem for newly registered users or newly introduced items. It aims to solve how to effectively recommend items to new users and how to effectively recommend new items to users who might potentially like them. In this scenario, the test data includes new users and new items, while the training data includes attribute information and historical interaction information of existing users and items.

[0063] like Figure 1 and Figure 2 As shown, this embodiment proposes a cold-start recommendation method based on an adaptive aggregated neighbor graph, which includes the following steps:

[0064] Step S101: SENet (Squeeze-and-Excitation Network) is used to learn the relative importance of second-order feature interactions. Then, MLP (Multilayer Perceptron) network is used to obtain the high-order attribute feature representation of users / items, and the attribute feature vector of users / items is obtained.

[0065] Specifically, the method first learns the original attributes using an embedding layer, converting them into embedded representations; then, it extracts the second-order interaction vectors of the attributes through feature learning. Next, it introduces the attention mechanism of SENet to perform weighted fusion of the second-order interaction vectors. Finally, it combines the linear combination vector with the fused second-order interaction vector using a multilayer perceptron (MLP) model to obtain a more representative and comprehensive user / item attribute feature vector.

[0066] Attribute fusion aims to learn the relationships between sparse attributes to fully acquire useful information and learn from cross-combinations of two or more attributes. First, an embedding layer maps the sparse encoding of each user attribute to a low-dimensional, dense real-valued vector, where the i-th attribute of user u is... The embedding vector is represented as Where K is the dimension of the attribute embedding vector. The attribute embedding set of user u is represented as: Then, by introducing the second-order interaction between pairwise attributes of user u through the Hadamard product, the operation is represented as:

[0067]

[0068]

[0069] Here, ⊙ represents element-wise multiplication. The user has k attributes, and pairwise interactions yield the desired result. Two second-order interaction vectors. It is the second-order interaction vector between the i-th attribute and the j-th attribute, e u This represents the set of second-order interaction vectors of attributes.

[0070] Considering the varying degrees of influence of different feature combinations on the prediction target, this solution introduces an attention mechanism. This mechanism filters out the most influential feature combinations from a complex feature set, excluding those that are irrelevant or less relevant. Unlike the attention mechanisms used in traditional natural language processing models, this solution borrows from the image domain's Squeeze-and-Excitation Networks (SENet) mechanism. This algorithm performs exceptionally well in image classification tasks, has fewer parameters and lower computational cost, and can automatically learn the importance of attributes to enhance useful attributes that significantly impact the results while suppressing those with less influence.

[0071] The SENet module dynamically determines which attributes are important through interactions. If the weight value is large, the second-order interaction of that attribute is considered to have high importance. The module first performs a compression operation by averaging the K-dimensional second-order interaction vectors through global average pooling to obtain a value that can represent the summary information of the feature.

[0072]

[0073]

[0074] Where h = [1,1,…,1], Represents the second-order interaction vector for user u. The compressed value, z u This represents the set of compressed values ​​of the second-order interaction vectors of user u.

[0075] Next, a two-layer MLP network is used to learn the importance of the second-order interaction vector of user u.

[0076] τ u =Sigmod(W t LeakyReLU(W z z u (5)

[0077] Where Sigmod and LeakyReLU are activation functions, W t W z Let τ be the weight vector. u Let the weights be represented as the second-order interaction vector weights of user u. By multiplying the corresponding weights by the second-order interaction vectors of user u, the weighting operation on important features is completed, and the formal representation is as follows:

[0078]

[0079] Here, ⊙ represents element-wise multiplication. It is the second-order interaction vector between the i-th and j-th attributes of user u. The weight, υ u Let represent the weighted second-order interaction fusion vector of user u.

[0080] Finally, a fully connected layer is introduced into the weighted attribute second-order interactions and linear combinations to achieve the learning of higher-order feature interactions. The fusion vector interaction operation of the linear combination vector and the second-order interaction is represented as follows:

[0081]

[0082] χ u =LeakyReLU(W v υ u +W k κ u +b) (8)

[0083] Among them, W v Wk Let b be the weight matrix, b be the bias vector, LeakyReLU be the activation function, and κ be the weight matrix. u This represents a linear combination vector of users u. This represents the attribute characteristics of user u. This represents the value of the i-th attribute of user u.

[0084] Step S102: Reconstruct the missing user preferences / item characteristics from user / item attribute features using an extended variational autoencoder.

[0085] Specifically, the variational autoencoder was extended by introducing the constraint that the reconstructed features approximate the preference vector, thereby deriving an effective representation of user preferences / item characteristics. Finally, attribute features and preference / item characteristics were combined to form fused features.

[0086] The preference generation layer aims to explore how to reconstruct missing preference embeddings using attribute features. Specific types of users may show interest in similar items, suggesting that attribute embeddings and preference embeddings are close to each other in the latent space and have similar distribution characteristics, and vice versa. For example, animation is often a primary form of entertainment for teenagers and children; users with similar age attributes may have similar interests. This solution addresses the problem of missing preference information by reconstructing preferences from attribute distributions using a variational autoencoder (VAE) structure. This structure consists of three parts: inference, generation, and approximation. The first two parts are standard VAEs, and the third part is an extension, χ². u Let be the attribute feature vector of user u. In the generation part, the reconstructed preference vector χ... u ' is determined by latent variable s u Generated through an MLP network containing parameters.

[0087]

[0088] In the reasoning section, variational inference approximates the latent variable s by introducing a θ-parameterized inference network. u The true and difficult post-hoc proof.

[0089]

[0090] The goal of variational inference is to optimize the free variational parameters to make the KL divergence... Minimum. Utilize reparameterization techniques to sample ε ~ N(0,I) and reset parameters. In this case, the gradient towards φ can be obtained through the sampled s. u Backpropagation.

[0091] In the approximation, the embedding of constraint reconstruction Approximate preference vector. During the training phase, user behavior embeddings can be obtained by learning from the historical ratings of existing users. Therefore, the reconstructed preference vector is required. With the learned user behavior vector and attribute features They are all similar. Therefore, the reconstruction loss function in the variational autoencoder is defined as follows:

[0092]

[0093] The first two terms are identical to those in a standard variational autoencoder (VAE), while the last term is an extension of the approximation. Finally, the user attribute embeddings and preference embeddings are combined to form the user feature vector. The fused features of user u are formally represented as follows: Where [;] denotes the vector concatenation operation.

[0094] Step S103: Use an attribute graph attention network to aggregate the neighbor information of users / items and the interaction information between neighbors to obtain the final vector representation of users / items.

[0095] Specifically, a user / item graph is constructed using user / item attribute information; then, a bilinear aggregator is used to model the interaction relationships between neighbors in the graph, and a graph attention network is combined to aggregate neighbor information; finally, the final vector representation of users / items is obtained.

[0096] S3.1 User graph construction.

[0097] A user graph network was constructed based on user attribute information (such as age, occupation, etc.), such as Figure 3 As shown, the aim is to aggregate information from neighboring users to enhance the embedded representation of the target user. For each attribute category, users with the same attribute are retrieved to form a user graph network. The user graph construction process is as follows:

[0098] ① Construct an index dictionary where the key is an attribute and the value is a set of users with that attribute, for example, "occupation:1→User1,User2"; "age:50→User2,User3".

[0099] ② For a new user, retrieve their neighbors based on each of their attributes. For example, if the new user User5 has the attributes "occupation:1,age:50", retrieve their neighbors based on these two attributes, and the retrieved neighbors are User1, User2, and User3.

[0100] ③ Calculate the similarity score for each neighbor and select the top-N neighbors. We define the score as the number of attributes from which neighbors can be retrieved. For example, User2 has a score of 2 because it can be retrieved from 2 attributes. This step preserves useful neighbors for subsequent processing.

[0101] ④ The top-N neighbors corresponding to the target node are obtained through the above statistics, and finally the user graph G of the target node is obtained.

[0102] S3.2 Neighbor aggregation.

[0103] In graph structures, not only are there relationships between target nodes and their neighbors, but the interactions between neighboring nodes can also transmit strong signals. To enhance the aggregation capability of node information, this solution extends the Graph Attention Network (GAT) by introducing a bilinear aggregator (BA) to capture the interaction relationships between neighboring nodes.

[0104] Define User Graph in It is a collection of entities including users, items, user attributes, and item attributes. A is a neighbor matrix, when the element Represents user node u i With user node u j There is an edge between user u i The attribute feature embedding is represented as Its neighbor set is defined as The expansion of the neighbor set containing itself is defined as follows: User u i The node degree is d i =|N(u i )|, its corresponding

[0105] First, the graph attention model GAT is used to capture the complex relationships between nodes in the user / item graph structure. The attention mechanism is used to dynamically learn the importance between nodes, and neighbor nodes are weighted and aggregated according to their importance to generate a new vector representation.

[0106]

[0107]

[0108] in, Indicates user u i The neighbor aggregation feature vector, w ij Represents neighbor node u j For user u iThe importance of W represents a trainable parameter vector used to learn the relative importance of the target node and its neighbors. g W e This is the weight matrix.

[0109] Then, in order to target user u i The interactions between neighboring nodes are modeled, and a bilinear aggregator (BA) is introduced to handle the interactions between the target user u. i The interaction relationships between neighboring nodes can be formally represented as follows:

[0110]

[0111] Here, ⊙ represents element-wise multiplication. Indicates user u i The neighbor interaction feature vector, user u i neighbor node u m and u n Belongs to set Indicates user u i Neighbor node u m u n The attribute feature vector, to avoid noise introduced by autocorrelation, is added to u. m n Conditions and restrictions. Represents the target node u i The corresponding number of associations is used to regularize the obtained representation to eliminate the bias of node degree, W. m W n This is the weight matrix.

[0112] In the user graph structure, the Graph Attention Network (GAT) is extended and combined with a bilinear aggregator (BA). This extension fully considers the interactions between neighboring nodes and incorporates these interactions into the feature representation of the target user, as formalized below:

[0113]

[0114] This uses a simple linear combination mechanism. User u i The final vector representation shows that β is a hyperparameter used to balance the GAT network and the bilinear aggregator. User u i eigenvectors.

[0115] ​Step S104: Introduce non-linear interaction between user and item features through MLP and inner product to obtain user ratings for items.

[0116] Using the attribute graph attention network model described above, the final feature representation p of user u is obtained. u and the final characteristic representation q of item v v The predicted rating of user u for item v is modeled as follows:

[0117]

[0118] The MLP function uses a hidden layer to implement a multi-layer perception to capture complex non-linear interactions between users and objects. The second term is the inner product interaction function, which obtains the linear interaction relationship between users and objects. u ,b v μ represents user bias, item bias, and global bias, respectively.

[0119] The prediction layer includes prediction loss and simulation loss. The overall loss function for training is defined as:

[0120] L = L pred +L recon +λ (17)

[0121] Among them, L pred The representative is the rating prediction loss for a specific task; L recon λ represents the reconstruction function loss. λ is a constant weighting factor. For rating prediction loss, we use squared loss as the objective function:

[0122]

[0123] The instance set D is divided into the training set D. train and test set D test These are used for model training and testing, respectively. uv The training set D represents train The actual rating value in the data. It is the predicted score.

[0124] Corresponding to the above method, this embodiment also proposes a cold-start recommendation system based on an adaptive aggregated neighbor graph, comprising:

[0125] The attribute fusion module is used to learn the relative importance of second-order feature interactions using SENet, and then obtains the high-order attribute feature representations of users / items through an MLP network to obtain the attribute feature vectors of users / items.

[0126] The preference generation module is used to reconstruct missing user preference / item characteristics from user / item attribute features using an extended variational autoencoder.

[0127] The neighbor aggregation module is used to aggregate the neighbor information of users / items and the interaction information between neighbors using an attribute graph attention network to obtain the final vector representation of users / items.

[0128] The rating prediction module is used to introduce a non-linear interaction between user and item features through MLP and inner product to obtain the user's rating for the item.

[0129] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0130] (I) Dataset and Experiment Setup

[0131] The proposed A-GAR model was evaluated using three real-world datasets. MovieLens is a widely used movie rating dataset. We used the ML-100K and ML-1M versions. The ML-1M dataset contains 1 million ratings from 6,000 users on 4,000 movies, covering ratings, user information, and movie information. Movie attributes include category, star, director, author, and country, while user attributes include gender, age, occupation, and zip code. Since the original dataset did not provide sufficient movie attribute information, we expanded the movie attributes by obtaining movie titles and release years from the IMDb2 database, thus including star, director, author, and country information.

[0132] Yelp used data from the 2017 version, where local businesses such as restaurants and hotels were considered items in this study. Given the large size and sparsity of the original Yelp data, we preprocessed the data using existing methods, reducing dataset complexity by removing nodes with fewer than 20 ratings. These nodes, similar to newly released movies, had never been observed before the testing phase. We defined item attributes as category, state, and city. Considering the lack of user attribute information in the Yelp data, each row of the social matrix was encoded as a user attribute, where each dimension of the social vector represents a different attribute value. The statistical results of the dataset are shown in Table 1.

[0133] Table 1 Statistical results of the dataset

[0134]

[0135] (II) Evaluation Indicators

[0136] Root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics, both of which are widely used in rating and forecasting tasks. The definitions of RMSE and MAE are as follows:

[0137]

[0138] (III) Settings

[0139] The model's performance was evaluated on cold-start users and cold-start items. On three datasets, 20% of users or items were randomly selected as the test set, while the remaining users or items and their historical interactions were used as the training set. The parameters from the current state-of-the-art FINI paper were fine-tuned. The experimental setup was set with a batch size of 128, LeakyReLU of 0.01, and an initial learning rate of 0.0005. The Adam optimizer was used with an adaptive learning rate, where the embedding dimension K = 40 and the threshold p for the candidate attribute set in graph construction was 5. For the baseline model, if the authors of the paper reported the hyperparameters, the same hyperparameter settings were followed; otherwise, fine-tuning was performed to ensure consistent model performance.

[0140] (iv) Comparison Methods

[0141] Several state-of-the-art cold-start recommendation methods were selected as baselines to validate the performance of our model. All baseline models used the same attribute information as our model.

[0142] NFM is a method that combines a decomposer and a neural network for rating prediction under sparse conditions. This method maps user and item attributes to dense vectors and then uses a multi-layered neural network to model feature interactions.

[0143] DANSER is a graph attention network (GAT) based method used to integrate multiple aspects of social relationships. This method models users' social relationships as user-user graphs and constructs item-item graphs using similarity coefficients between items clicked by the same user.

[0144] DiffNet fuses user and item features with node embeddings. It incorporates a layer-by-layer diffusion process similar to GCN to simulate dynamic diffusion on the user-user social graph for recommendation purposes.

[0145] GCMC utilizes user and item attribute vectors as auxiliary information, combining these attribute vectors into node embedding representations. This method applies the Graph Convolutional Network (GCN) framework to the matrix completion task of user-item graphs.

[0146] The STAR-GCN (Stacked Graph Convolutional Network) method implements a hierarchical approach on user-item interaction graphs using stacked GCN layers. By integrating attribute embeddings and free embeddings to represent individual nodes, this technique enhances information propagation and feature acquisition, thereby improving the overall effectiveness of learning and disseminating relevant information.

[0147] DropoutNet integrates content and preference information into the DNN architecture and introduces a dropout mechanism to address the challenges posed by the strict cold start scenario.

[0148] MetaEmb is a meta-learning method designed to solve the strict cold start problem. It trains an embedding generator through gradient-based meta-learning to generate embeddings for new IDs.

[0149] The AGNN method represents nodes and edges in graph-structured data as a graph and uses a neural network model to learn and infer the nodes in the graph. By introducing an attention mechanism, AGNN can pay more attention to important nodes and edges during the learning process, thereby improving the model's expressive power and performance.

[0150] The FINI method considers feature importance and neighbor node interaction graph neural networks, and adopts a global-local context attention mechanism to dynamically learn the feature weights of each node. This mechanism can simultaneously emphasize global public information and local personalized information. This framework aims to improve the quality of attribute embedding and approximate preference embedding.

[0151] (V) Experimental Results

[0152] Table 2. Performance Comparison Results of Different Models

[0153]

[0154] This section summarizes the recommendation performance of the model in this case study on the MovieLens and Yelp datasets. For ease of comparison, the table shows the performance improvement of A-GAR's RMSE and MAE relative to each benchmark method, quantified by Improvement. The best results for each dataset are underlined. Analysis of the experimental results leads to the following conclusions:

[0155] (1) Our proposed A-GAR model outperforms the best baseline FINI in cold-start scenarios. The A-GAR method shows the best performance on various datasets, outperforming the best baseline FINI model. This result is mainly due to the fact that A-GAR improves upon FINI, combining the advantages of the original method and adopting a channel attention mechanism to obtain the importance of each feature channel through adaptive learning. Compared with other benchmark methods, this method can improve the accuracy of the model to a certain extent.

[0156] (2) Our proposed model outperforms all baseline methods on all three datasets. On the MovieLens dataset, our model improves the RMSE and MAE metrics for item cold start by 1.38% and 2.79% respectively compared to the best baseline FINI. On the Yelp dataset, it also improves the RMSE and MAE metrics for user cold start by 1.23% and 2.72% respectively. These results clearly demonstrate that the introduction of attribute fusion importance enables the model to better distinguish the influence of different features, thereby improving the model's expressive power. Furthermore, by using a bilinear aggregator to model the interaction relationships between neighbors, we extend the Graph Attention Network (GAT) to improve the model's prediction accuracy.

[0157] (vi) Ablation test

[0158] The impact of several key modules of A-GAR on recommendation performance was analyzed through ablation experiments to verify the rationality of the network model design. Four A-GAR variants were designed for comparison to demonstrate the necessity of all modules in A-GAR: (1) the complete A-GAR model; (2) A-GAR with attribute fusion but without attention mechanism (named A-GAR~AT1); (3) A-GAR with neighbor aggregation but without attention mechanism (named A-GAR~AT2); and (4) A-GAR without neighbor aggregation (named A-GAR~GAT).

[0159] Table 3 Comparison of A-GAR and its variant models

[0160]

[0161] Figure 4 The model is shown on the MovieLens dataset in both the ML-100K and Yelp versions. The following conclusions can be drawn from the bar chart:

[0162] ①Attention Mechanism Module

[0163] The proposed method, A-GAR, employs an attention mechanism in two distinct modules: attribute fusion and neighbor aggregation. To analyze the impact of this attention mechanism on recommendation performance, this section compares A-GAR with the following three method variants:

[0164] A-GAR~AT1: This method is a variant of A-GAR that removes the second-order interactive SEnet attention mechanism. Specifically, by setting the weight τ in Equation (6) to 1, this variant treats any combination of attributes equally during the attribute fusion process.

[0165] A-GAR~AT2: This method is a variant of A-GAR that removes the attention mechanism of neighbor aggregation. Specifically, by replacing the graph attention neural network GAT in formula (15) with a graph neural network GNN, this variant aggregates information from its neighbor nodes equally for the user and updates its own information accordingly.

[0166] The recommendation performance of the proposed A-GAR method and the two variants mentioned above on two datasets is compared as follows: Figure 4 As shown in (a) and (b), the experimental results show that the performance of A-GAR on both datasets decreases to varying degrees after removing different attention mechanisms. These results demonstrate that introducing attention mechanisms into attribute fusion and neighbor aggregation has a positive impact on cold-start recommendation.

[0167] ② Neighbor aggregation module

[0168] In this paper, the A-GAR model considers the potential relationships between the target node and its neighbors during the neighbor aggregation process, and employs a bilinear aggregator to effectively capture the interactions between neighbors. Next, we compare A-GAR with models that do not include neighbor node information:

[0169] A-GAR~BA: This method is a variant of A-GAR that removes the interaction between neighboring nodes. This variant only aggregates neighboring node information and updates the feature representation of the nodes, while ignoring the association between neighboring nodes. Specifically, it removes the bilinear aggregator (BA) in formula (14).

[0170] The proposed method A-GAR and its variants are compared in terms of recommendation performance on two datasets. Figure 4 As shown in (c) and (d), the experimental results show that when the interaction relationship between neighboring nodes is removed, the prediction performance of A-GAR on the three datasets will decrease to varying degrees. The results demonstrate the effectiveness and importance of considering the relationship between neighboring nodes in improving the expressive power of node embedding.

[0171] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0172] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0173] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0174] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0175] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cold-start recommendation method based on adaptive aggregated neighbor graph, characterized in that, Include: First, SENet is used to learn the relative importance of second-order feature interactions. Then, an MLP network is used to obtain the high-order attribute feature representations of users / items, resulting in the attribute feature vectors of users / items. Then, the missing user preferences / item characteristics are reconstructed from the user / item attribute features using an extended variational autoencoder; Then, an attribute graph attention network is used to aggregate the neighbor information and interaction information between neighbors of users / items to obtain the final vector representation of users / items. Specifically, this involves: first, constructing a user / item graph using the attribute information of users / items; then, modeling the interaction relationships between neighbors in the graph using a bilinear aggregator, and aggregating neighbor information using a graph attention network; finally, obtaining the final vector representation of users / items. The process of aggregating neighbor information using the graph attention network is as follows: the graph attention model GAT is used to capture the complex relationships between nodes in the user / item graph structure, and the attention mechanism is used to dynamically learn the importance between nodes. Neighbor nodes are weighted and aggregated according to the importance of their neighbor nodes to generate a new vector representation. The interaction relationships between neighbors in the graph are modeled using a bilinear aggregator, as expressed in the following formula: ,in, This indicates element-wise multiplication. Indicates user Neighbor interaction feature vectors, users neighboring nodes and Belongs to set , , Indicates user Neighbor nodes , Attribute feature vectors; Represents the target node The corresponding number of associations, Indicates user degree of nodes, ; , This is the weight matrix; Finally, by introducing a non-linear interaction between user and item features through MLP and inner product, the user's rating of the item is obtained.

2. The cold-start recommendation method based on adaptive aggregated neighbor graph according to claim 1, characterized in that, The process of obtaining the attribute feature vector of a user / item is as follows: First, the original attributes are learned using an embedding layer and converted into embedded representations; then, second-order interaction vectors of the attributes are extracted through feature learning. Next, the attention mechanism of SENet is introduced to perform weighted fusion of the second-order interaction vectors; Finally, by combining the linear combination vector with the second-order interaction fusion vector through the multilayer perceptron model, the attribute feature vectors of users / items are obtained.

3. The cold-start recommendation method based on adaptive aggregated neighbor graph according to claim 2, characterized in that, Introducing SENet's attention mechanism to perform weighted fusion of second-order interaction vectors includes: First, a compression operation is performed by averaging the K-dimensional second-order interaction vector using global average pooling to obtain a numerical value: , ,in, , , It is the second-order interaction vector between the i-th attribute and the j-th attribute. Represents the second-order interaction vector for user u. The compressed value, This represents the set of compressed values ​​of the second-order interaction vectors of user u. ; Next, a two-layer MLP network is used to learn the importance of the second-order interaction vector of user u: Where Sigmod and LeakyReLU are activation functions. , For the weight vector, Let it be represented as the set of second-order interaction vector weights for user u; Finally, by multiplying the corresponding weights by the second-order interaction vector of user u, the weighted operation on important interaction features is completed, resulting in the weighted second-order interaction fusion vector of user u. .

4. The cold-start recommendation method based on adaptive aggregated neighbor graph according to claim 3, characterized in that, The interaction operation between linear combination vectors and second-order interactive vectors is represented as follows: , ,in, , Let b be the weight matrix, b be the bias vector, and LeakyReLU be the activation function. This represents a linear combination vector of users u. This represents the attribute characteristics of user u. This represents the value of the i-th attribute of user u.

5. The cold start recommendation method based on adaptive aggregated neighbor graph according to claim 1, characterized in that, Reconstructing missing user preference / item characteristics from user / item attribute features using an extended variational autoencoder includes: The variational autoencoder was extended by introducing the constraint of reconstructing the feature approximating the preference vector, thereby deriving an effective representation of user preference / item characteristic features; finally, attribute features and preference / item characteristics were combined to form fused features.

6. The cold-start recommendation method based on adaptive aggregated neighbor graph according to claim 1, characterized in that, The final vector representation of the user / item is as follows: ,in, User The final vector representation, It's a hyperparameter. User The neighbor aggregation feature vector, User eigenvectors.

7. The cold start recommendation method based on adaptive aggregated neighbor graph according to claim 1, characterized in that, user For items The predictive rating model is as follows: ,in, User The final characteristic representation, It is an item The final characteristic representation, , , These represent user bias, item bias, and global bias, respectively.

8. A cold-start recommendation system based on adaptive aggregated neighbor graph, characterized in that, A cold-start recommendation method based on adaptive aggregated neighbor graph as described in any one of claims 1-7, comprising: The attribute fusion module is used to learn the relative importance of second-order feature interactions using SENet, and then obtains the high-order attribute feature representation of users / items through an MLP network to obtain the attribute feature vector of users / items. The preference generation module is used to reconstruct missing user preference / item characteristics from user / item attribute features using an extended variational autoencoder; The neighbor aggregation module is used to aggregate the neighbor information of users / items and the interaction information between neighbors using an attribute graph attention network to obtain the final vector representation of users / items. The rating prediction module is used to introduce a non-linear interaction between user and item features through MLP and inner product to obtain the user's rating for the item.

Images