A lightweight meta-embedding recommendation method and system based on ID

Abstract

The present invention discloses a lightweight meta-embedding recommendation method and system based on ID, which relates to the field of personalized recommendation technology. The present invention constructs a graph based on the direct interaction between users and items. A two-layer virtual node structure is designed. Coarse meta-embedding nodes are superimposed to directly connect all real user and item nodes, thereby capturing a wide range of shared semantics. Subsequently, fine-grained nodes are established and connected to coarse-grained nodes. This hierarchical and selective connection ensures that the model learns a wide range of semantic information at the coarse-grained level while focusing on fine-grained personalized features. In order to effectively combine the two stages of the embedding learning method, the SparsePCA method is used to selectively generate random entries in the coarse-grained meta-embedding code set, and the coarse-grained meta-embedding and multiple fine-grained meta-embeddings are dynamically matched according to semantic relevance through a weight bridge update strategy, ensuring that the present invention is still efficient and effective when representing complex information.

Description

Technical Field

[0001] The present invention relates to the field of personalized recommendation technology, and more particularly to an ID-based lightweight meta-embedding recommendation method and system. Background Art

[0002] Recommender systems (RSs) have revolutionized the way we navigate the vast amount of available information and products by tailoring recommendations to individual users' preferences. Powered by collaborative filtering, recommender systems map users and items into a high-dimensional latent space for personalized recommendations using strategies such as dot products, multi-layer perceptrons, and graph neural networks. As data scale grows, the complexity and resource requirements of these systems increase dramatically. This problem is particularly acute in ID-based recommender systems, where entities are used to uniformly represent all users and items, each requiring a unique embedding vector. Consequently, traditional recommender systems suffer from significant memory overhead and struggle to scale. To address the challenge of optimizing memory usage without compromising recommendation accuracy, researchers have turned to developing lightweight recommender systems using deep neural networks. However, these session-based recommender systems face the daunting task of precisely selecting the optimal embedding size and ensuring accurate representation within a vast search space. Furthermore, under memory constraints, learning from a teacher model requires denser embeddings, resulting in a higher probability of duplicate embeddings. Conversely, selecting a larger number of embeddings to reduce redundancy can result in sparse embeddings, hindering the acquisition of sufficient knowledge from the teacher model.

[0003] To address this issue, parameter sharing techniques, such as hashing-based methods, have been proposed. The basic idea is to represent entities (such as user / item IDs) using combinations of meta-embedding vectors determined by hash functions. These meta-embedding vectors can be shared across entities. However, existing meta-embedding methods often focus on coarse-grained learning and fail to capture richer, finer-grained semantic nuances. This limitation severely impacts the effectiveness of meta-embeddings in representing complex information about users and items, leading to suboptimal recommendation performance.

[0004] Therefore, how to provide a lightweight ID-based meta-embedding recommendation method and system, a new GNN-based model, efficiently learn semantics of different granularities, and enhance the representation ability of coarse-grained meta-embedding through fine-grained meta-embedding is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0005] In view of this, the present invention provides an ID-based lightweight meta-embedding recommendation method and system. Based on a new GNN model, a hierarchical and selective connection graph is constructed to ensure that the model learns a wide range of semantic information at a coarse level, while focusing on fine-grained personalized features at a fine level, so as to promote multi-granularity semantic learning of meta-embedding, and enhance the representation of ID-based recommendations by coarse-grained meta-embedding through fine meta-embedding.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a lightweight meta-embedding recommendation method based on ID, comprising:

[0007] Obtain a benchmark data corpus;

[0008] constructing a user-item interaction graph based on the benchmark data corpus;

[0009] Build a GNN-based meta-embedding multi-granularity semantic model, adopt a two-layer virtual node structure to capture coarse-grained and fine-grained semantics, and use GNN to learn the embedding representation of users and items;

[0010] In the user-item interaction graph, coarse-grained nodes are superimposed to directly connect all user nodes and item nodes;

[0011] On the basis of coarse-grained nodes, fine-grained nodes are introduced and used as auxiliary nodes to connect coarse-grained nodes.

[0012] Initializing the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjusting the sparsity using a soft threshold activation function;

[0013] A weighted bridge update strategy is used to dynamically match coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance.

[0014] Preferably, a two-layer virtual node structure is used to capture coarse-grained semantics and fine-grained semantics, including:

[0015] Capture coarse-grained semantic information through the coarse layer and learn coarse element embedding of coarse-grained semantics;

[0016] Fine-grained semantic information is learned through the fine layer, and sparse fine element embedding of fine-grained semantics is learned based on the coarse element embedding; and the representation of the coarse element embedding is strengthened by the fine element embedding.

[0017] Preferably, a coarse component embedding is obtained based on the coarse element embedding code set and the coarse assignment matrix;

[0018] Add the coarse assignment matrix to the user-item interaction graph to obtain the coarse interaction graph;

[0019] By stacking the coarse component embedding and the coarse meta-embedding code set as the input embedding of the coarse interaction graph, we obtain the graph propagation coarse representation and coarse meta-embedding code set of the complete entity set;

[0020] Extract graph propagation embeddings of entities from the full entity set and compute similarity scores between each user-item pair;

[0021] In each iteration, a fixed assignment matrix is used to compute a Bayesian personalized ranking loss based on the prediction scores of the training batch, thereby facilitating backpropagation to update the set of coarse meta-embeddings.

[0022] Update the coarse assignment matrix according to the updated coarse meta-embedding code set.

[0023] Preferably, on the basis of the coarse-grained nodes, fine-grained nodes are introduced, and the fine-grained nodes are used as auxiliary nodes to connect the coarse-grained nodes, including:

[0024] Build detailed interaction graphs;

[0025] By improving the method of initialization process of coarse element embedding codebook;

[0026] Select r coarse element embeddings from the coarse element embedding code set to form a new coarse element embedding matrix;

[0027] The fine-grained representation is initialized based on the selected coarse-grained element embedding, thereby enhancing the semantic relationship between entities from the very beginning.

[0028] Preferably, initializing the meta-embedding multi-granularity semantic model based on SparsePCA comprises:

[0029]

[0030] in, represents the reconstruction error, ||·||1 represents the sparsity penalty term, and α is a hyperparameter that controls sparsity; represents the new coarse element embedding matrix; the goal of SparsePCA is to find a d r dimensional sparse representation And the corresponding sparse coefficient matrix W r ;

[0031] Use zero-fill SparsePCA to reduce the sparse vector to the specified d-dimensional vector.

[0032] Preferably, a soft threshold activation function is used to dynamically adjust the sparsity level, including:

[0033] The soft threshold activation function is expressed as:

[0034]

[0035] Where sign(x) is a sign function that returns -1 if x < 0, 1 if x > 0, and 0 if x = 0. |·| represents the absolute value, and λ is the threshold parameter. For fine-grained representation.

[0036] Preferably, the coarse components of the fine-grained representation are embedded Expressed as:

[0037]

[0038] in, is the refined composition embedding, w cr is the weight parameter, S c represents the rough assignment matrix; S r It is the indirect assignment matrix between coarse element embedding and fine element embedding;

[0039] The input embedding of the refined GNN will be and Stacking them up, we get:

[0040]

[0041] in, Represents the input embedding of GNN, according to the fine interaction graph A r The adjacency matrix of , the propagation operation of the l+1 layer is:

[0042]

[0043] in, It's A r The diagonal matrix of

[0044] The final GNN embedding is obtained by averaging the embeddings of each layer:

[0045]

[0046] Through H r Split and obtain the complete entity set;

[0047] Extract graph propagation embeddings of entities from the complete entity set and compute the similarity score between each user-item pair, fixing the assignment matrix S in each iteration. c , S r and

[0048] A Bayesian personalized ranking loss is computed based on the prediction scores of the training batches, which facilitates backpropagation to update the set of refined meta-embeddings.

[0049] Preferably, a weighted bridge update strategy is used to dynamically match coarse-grained meta-embedding and fine-grained meta-embedding according to semantic relevance, and a bridge update strategy of allocating matrices from coarse to fine is adopted, including:

[0050] The indirect assignment matrix S between coarse element embedding and fine element embedding r Expressed as:

[0051]

[0052] in, yes The pseudo-rebellion, yes The pseudo-inverse of , both of which are calculated using the Moore-Penrose inverse algorithm; represents a coarse meta-embedding code set; Represents a set of refined meta-embedded codes;

[0053] The indirect assignment matrix S between coarse element embedding and fine element embedding r Perform sparse processing so that each coarse element embedding uses exactly t at any time r Element embedding:

[0054]

[0055] in, Returns the top-(t r ) item;

[0056] The indirect assignment matrix S between coarse element embedding and fine element embedding r Perform initialization and sparse processing;

[0057] The indirect assignment matrix S between coarse element embedding and fine element embedding r The formula for initialization is expressed as:

[0058]

[0059] in, yes The pseudo-rebellion, yes of pseudo-rebellion; For fine-grained representation; Embed code set for crude element.

[0060] Preferably, a lightweight meta-embedding recommendation system based on ID includes:

[0061] A data acquisition module, used to obtain a benchmark data corpus;

[0062] an interaction graph construction module, configured to construct a user-item interaction graph based on the benchmark data corpus;

[0063] The model building module is used to build a GNN-based meta-embedding multi-granularity semantic model. It adopts a two-layer virtual node structure to capture coarse-grained and fine-grained semantics and uses GNN to learn the embedding representation of users and items.

[0064] The coarse-grained node overlay module is used to overlay coarse-grained nodes in the user-item interaction graph and directly connect all user nodes and item nodes;

[0065] A fine-grained node connection module is used to introduce fine-grained nodes based on coarse-grained nodes and use the fine-grained nodes as auxiliary nodes to connect the coarse-grained nodes;

[0066] An initialization and sparsification processing module, configured to initialize the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjust the sparsity level using a soft threshold activation function;

[0067] A weight update module is used to dynamically match coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance based on a weight bridge update strategy.

[0068] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a lightweight meta-embedding recommendation method and system based on ID, which constructs a graph based on the direct interaction between users and items. Subsequently, a two-layer virtual node structure is designed. First, coarse meta-embedding nodes are superimposed to directly connect all real user and item nodes to capture a wide range of shared semantics. Subsequently, fine-grained nodes are established, which are specifically linked to these coarse-grained nodes, so that more detailed semantics can be mapped. This hierarchical and selective connection ensures that the model learns a wide range of semantic information at the coarse-grained level while focusing on fine-grained personalized features. This coarse-to-fine learning strategy helps to fully understand the semantics of users / items from both global and local perspectives.

[0069] To effectively combine the two stages of the embedding learning method, we use an optimized coarse meta-embedding as a springboard for developing sparse fine-grained embeddings. SparsePCA is used to selectively generate random entries in the coarse meta-embedding code set. This targeted initialization not only preserves sparsity but also ensures the association between users and items. Furthermore, a soft threshold is designed during training to dynamically adjust the sparsity level. Unlike the rigid approach of SparsePCA, the soft thresholding method can adapt to the data and enhance robustness. This combination establishes a fine-tuned dynamic embedding process that remains sensitive to fine-grained semantics. Finally, a novel weight bridge update strategy is used to dynamically match the coarse meta-embedding with multiple fine-grained meta-embeddings based on semantic relevance. This innovation ensures that the present invention remains efficient and effective in representing complex information.

[0070] The present invention has the following beneficial effects:

[0071] (1) This paper proposes a novel GNN-based meta-embedding multi-granularity semantic model, whose hierarchical and selective connections ensure that the model can learn both coarse-grained broad semantic information and fine-grained personalized features.

[0072] (2) The present invention adopts an initialization method by optimizing coarse element embedding and SparsePCA, and selectively adds it to the embedding code library, which can maintain sparsity while maintaining the association relationship between users and items.

[0073] (3) The present invention develops a soft thresholding technique to dynamically adjust the sparsity during training, while providing a more adaptable and robust model to cope with various complex data.

[0074] (4) This paper develops an innovative weight bridge update strategy to dynamically adjust the coarse-grained meta-embedding and multiple fine-grained meta-embeddings according to semantic relevance to ensure the effectiveness of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0076] Figure 1 A flowchart of a lightweight meta-embedding recommendation method based on ID is provided in an embodiment of the present invention.

[0077] Figure 2A schematic diagram of a fine element embedding initialization strategy provided by an embodiment of the present invention.

[0078] Figure 3 A schematic diagram of a weighted bridge update strategy provided by an embodiment of the present invention.

[0079] FIG4( a ) is a schematic diagram of the NDCG@5 effect of the model component in Gowallo provided by an embodiment of the present invention.

[0080] FIG4( b ) is a schematic diagram of the Recall@5 effect of the model component in Gowallo provided by an embodiment of the present invention.

[0081] Figure 4(c) is a schematic diagram of the NDCG@5 effect of the model component in Yelp2020 provided by an embodiment of the present invention.

[0082] Figure 4(d) is a schematic diagram of the Recall@5 effect of the model component in Yelp2020 provided by an embodiment of the present invention.

[0083] Figure 5 The figure is a schematic diagram of the performance effect of an embodiment of the present invention under different hyperparameter settings.

[0084] FIG6( a ) shows the performance of coarse element embedding and fine element embedding in NDCG@5 provided by an embodiment of the present invention.

[0085] FIG6( b ) shows the performance of the coarse element embedding and the fine element embedding in NDCG@20 provided by an embodiment of the present invention.

[0086] FIG6( c ) shows the performance of coarse element embedding and fine element embedding in Recall@5 provided by an embodiment of the present invention.

[0087] FIG6( d ) shows the performance of the coarse element embedding and the fine element embedding in Recall@20 provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0089] State-of-the-art recommender systems have shifted their attention towards efficient recommendations, such as on-device recommendations under memory constraints. To this end, existing methods either focus on lightweight embeddings for users and items, or involve on-device systems that use compact embeddings to enhance reusability and reduce space complexity. However, these methods focus only on the coarse-grained embeddings while ignoring the fine-grained semantic nuances, which adversely affects the ability of meta-embeddings to capture the complex relationships between users and items, leading to suboptimal recommendation results. This paper aims to investigate how meta-embeddings can efficiently learn semantics at different granularities, and how fine-grained meta-embeddings can enhance the representational power of coarse-grained meta-embeddings.

[0090] To address the above problems, an embodiment of the present invention develops a novel recommendation system based on graph neural networks (GNNs), in which each user and item is treated as a node, directly connected to a coarse-grained virtual node, and indirectly connected to a fine-grained virtual node to ensure semantic learning at different granularities, while revealing: 1) Compared with coarse-grained semantics, fine-grained semantics can be adaptively better captured by sparse meta-embeddings 2) How to strike a balance between maintaining the uniqueness of embeddings and memory constraints. In addition, a SparsePCA-based initialization method is proposed, and a soft threshold activation function is used to achieve the sparsity of meta-embeddings. A weight bridging update strategy is proposed, which focuses on matching each coarse-grained meta-embedding with multiple fine-grained meta-embeddings based on the semantics of the user / item. A large number of experiments have verified the superiority of the method described in the embodiment of the present invention over the existing baseline methods.

[0091] The embodiment of the present invention discloses a lightweight meta-embedding recommendation method based on ID, such as Figure 1 Shown, including:

[0092] Obtain a benchmark data corpus;

[0093] constructing a user-item interaction graph based on the benchmark data corpus;

[0094] Build a GNN-based meta-embedding multi-granularity semantic model, adopt a two-layer virtual node structure to capture coarse-grained and fine-grained semantics, and use GNN to learn the embedding representation of users and items;

[0095] In the user-item interaction graph, coarse-grained nodes are superimposed to directly connect all user nodes and item nodes;

[0096] On the basis of coarse-grained nodes, fine-grained nodes are introduced and used as auxiliary nodes to connect coarse-grained nodes.

[0097] Initializing the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjusting the sparsity using a soft threshold activation function;

[0098] A weighted bridge update strategy is used to dynamically match coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance.

[0099] The principles of the embodiments of the present invention are as follows:

[0100] In ID-based recommendation, the user set and item sets The total number of entities is The interaction matrix between users and items is Represents, where 0 indicates no interaction and 1 indicates that there is interaction between the user and the item.

[0101] To improve the memory efficiency of the embedding layer, a common approach is to utilize a compact set of meta-embedding codes, i.e. Where d represents the dimension of the embedding. In the case where the number of meta-embeddings m is significantly smaller than N, this codebook can replace the wider full embedding matrix To generate embeddings for each entity, a common approach is to use an assignment matrix Each row of this matrix corresponds to a user or item entity, where the non-zero elements specify which meta-embeddings are selected from the m available options and their respective weights in forming the entity embedding. Therefore, the constituent embeddings of all entities (denoted as ) is calculated as follows:

[0102]

[0103] Building on this, recent developments have shown that graph neural networks (GNNs) can effectively propagate collaborative signals between interconnected entities. Leveraging this, GNN-based recommendation systems construct user-item interaction graphs to capture the semantic associations between entities, which are reflected in their interactions.

[0104] GNN-based recommendation systems usually use an adjacency matrix A to represent these connections in the user-item interaction graph, facilitating the propagation of collaborative signals between connected entities and enabling embedded learning. The structure of the adjacency matrix A is as follows:

[0105]

[0106] Where R represents the user-item interaction matrix. For each user and items For example, their embedding values e u and e i All from , and serves as the input to the user-item interaction graph A. To improve adaptability, a combined embedding approach considers the semantic relationships between entities and provides a more refined, weighted meta-embedding combination for each entity. Meta-embeddings are represented as additional virtual nodes connected to a set of nodes in the user-item interaction graph A. The assignment matrix S represents the connection between entities and the set of meta-embedding codes. Combining S with A yields an expanded interaction graph A':

[0107]

[0108] in, The input embedding of GNN (0th layer embedding) is embedded by stacking and Meta Embed Code Set E meta Generated:

[0109]

[0110] in The propagation operation of the l+1 layer is defined as:

[0111]

[0112] Where A' is the adjacency matrix of the extended interaction graph, is the diagonal matrix of A'. A symmetric normalized adjacency matrix is created. The final GNN embedding value is obtained by averaging the embedding values of all layers:

[0113]

[0114] By splitting H as follows, the complete entity set can be obtained and meta embed code sets The graph propagation representation of is:

[0115] H full ←H[:N,:],H neta ←H[N:,:] (7)

[0116] Meta-embedding H meta Including collaborative information from semantically similar user / item neighbors can assign more effective meta-embeddings to each user / item. In the recommendation task, full You can get the user's image embedding h u and the graph embedding h of the item i , and calculate their recommended affinity prediction values:

[0117]

[0118] This prediction feeds the Bayesian Personalized Ranking loss (BPR) that is used to optimize the relevant meta-embedding for each entity:

[0119]

[0120] in, Refers to the entire training set or training batch; (u,i + ,i - ) is the interaction item i observed by the sampled user u + and unobserved item i - The triples composed of ||Θ|| 2 Denotes the L2 regularization term applied to the trainable parameters; λ determines the regularization weight in the loss function. With the assignment matrix S fixed in each iteration, the extended interaction graph A' uses the backpropagation of Formula 9 to update the meta-embedding code set E meta The value in .

[0121] With Meta Embed Code Library E meta To learn E, the weights in S must be continuously updated to achieve a more accurate mapping between entities and their meta-embedding assignments. For example, the embeddings of two closely related entities should consist of a set of similar meta-embeddings, which reflects the core principle of collaborative filtering. However, jointly updating S and E meta This makes it difficult to find a stable solution. To address this challenge, after propagating through L layers in the extended interaction graph A', the mapping of each entity to its corresponding meta-embedding is retained, which is the same as the mapping in the GNN input layer. Therefore,

[0122] H full =SH meta (10)

[0123] in, and is the graph propagation embedding shown in Formula 7. On this basis, H is calculated by Moore-Penrose inverse meta The pseudo-inverse (i.e. ), and then calculate the assignment matrix S:

[0124]

[0125] in, is the identity matrix, is the inverse matrix of the square matrix Σ containing the diagonal singular values, is the identity matrix The conjugate transpose of . Updating S via a gradient-free learning strategy can improve computational efficiency. In addition, this approach preserves the semantic associations between entities in the meta-embedding assignment. Specifically, H fullSimilar rows in will produce similar results after the multiplication process in Equation 11.

[0126] The core of the embodiments of the present invention lies in three aspects: 1) learning coarse element embedding of coarse-grained semantics; 2) learning sparse fine element embedding of fine-grained semantics based on coarse element embedding; 3) how fine element embedding can enhance the representation of coarse element embedding.

[0127] Specifically, the coarse graph propagates meta-embedding:

[0128] Construct a coarse graph propagation meta-embedding, including:

[0129] First, use the crude meta embed code set and the rough assignment matrix S C Compute coarse component embedding as follows:

[0130]

[0131] Then, S c Add to the user-item interaction graph A to get the coarse interaction graph

[0132]

[0133] By stacking and As A c The input embedding is used to obtain the complete entity set Graph propagation coarse representation and coarse meta-embedding code set Then, from the complete entity set Extract the graph propagation embedding of the entity and calculate the similarity score between each user-item pair as defined in Formula 8. A fixed assignment matrix S is used in each iteration. C , enter the prediction score of the training batch into formula 9 to calculate the BPR loss This facilitates backpropagation to update the coarse element embedding code set In addition, the embodiment of the present invention updates S by the above-mentioned non-gradient learning strategy c .

[0134] Specifically, fine-grained graph propagation meta-nesting:

[0135] Building upon the coarse element embeddings, it is crucial to capture finer representations to improve them. To learn more refined semantic information, embodiments of the present invention connect these fine element embeddings as auxiliary nodes to the coarse element embeddings, rather than directly connecting the coarse element embeddings to a set of entity nodes for representation learning. This avoids learning redundant information at the coarse stage.

[0136] Building a detailed interaction diagram as follows:

[0137]

[0138] in, is the indirect assignment matrix between two meta-embeddings. Before discussing how to improve the coarse meta-embedding code set, we need to learn fine-grained representations.

[0139] like Figure 2 The fine element embedding initialization strategy diagram shown in FIG, which retains sparsity and maintains the association relationship between users and items. The embodiment of the present invention proposes an improved coarse element embedding codebook Method to initialize the process. Select r coarse elements to embed and form a new matrix Instead of random assignment. Through this selection process, the coarse element embeddings can be initialized according to these selected Thus, the semantic relationship between entities is enhanced from the beginning. In addition, when refining the coarse element embedding, the goal of the embodiment of the present invention is to avoid learning the same dense vector as the coarse element embedding, because this will lead to a dense matrix On the contrary, in order to This approach serves two purposes. First, it ensures that the coarse-grained stage focuses on capturing coarse-grained semantic information, while the fine-grained stage processes fine-grained details as auxiliary information. Second, it meets the goal of compressing the total number of meta-embeddings in the two stages. In order to achieve more flexible sparsity Initialization, the embodiment of the present invention uses SparsePCA (Principal Component Analysis) for initialization. SparsePCA inherits the advantages of PCA, which can effectively capture the main variation patterns and reduce computational complexity and feature redundancy. In addition, it also introduces the function of controlling the sparsity of the extracted components. This control of sparsity is crucial and can adjust the sparsity of the meta-embedding matrix. Using SparsePCA, the initialization process is customized to better induce the sparsity of the meta-embedding matrix, thereby improving the overall efficiency and effectiveness of the learning process.

[0140]

[0141] in, represents the reconstruction error (Frobenius norm), ||·||1 represents the sparsity penalty term, and α is a hyperparameter that controls sparsity. The goal of SparsePCA is to find a d r dimensional sparse representation and the corresponding sparse coefficient matrix Then, the sparse vector obtained by SparsePCA dimensionality reduction is filled with zeros Restore it to the specified d-dimensional vector However, although SparsePCA can effectively generate sparse representations, it suffers from the problem of indistinguishability and cannot maintain the sparsity of embeddings during training. To solve this problem, the embodiment of the present invention uses a soft threshold method, which is defined as follows:

[0142]

[0143] Where sign(x) is the sign function that returns -1 if x<0, 1 if x>0, and 0 if x=0, |·| represents the absolute value, and λ is the threshold parameter. The integration of soft thresholding techniques is to dynamically adjust the sparsity during training, thereby providing a more adaptable and robust model that can cope with complex data changes by smoothly setting small coefficients to zero while retaining important coefficients. This approach ensures that the generated embeddings remain sparse and computationally efficient, which helps in better convergence during training. With the help of The coarse component in formula 1 is embedded It can be rewritten as follows:

[0144]

[0145] in, is the refined composition embedding, w cr is the weight parameter. The input embedding of the refined GNN is and Stacking up:

[0146]

[0147] in, According to the refined interaction diagram A r The adjacency matrix of the l+1 layer is

[0148]

[0149] in, It's A r The final GNN embedding is obtained by averaging the embeddings of each layer:

[0150]

[0151] Through H r Perform the following split to obtain the complete entity set Coarse Meta Embed Code Set and a granular set of meta embed codes Graph propagation representation of :

[0152]

[0153] Then, from the complete entity set Extract the graph propagation embedding of the entity and calculate the similarity score between each user-item pair as defined in Formula 8. In each iteration, the assignment matrix S is fixed. c , S r and Enter the prediction scores of the training batch into Formula 9 to calculate the BPR loss Facilitates back-propagation to update refined meta-embedding code sets

[0154] Specifically, learn the assignment weights from coarse to fine:

[0155] In the rough stage, S can be directly calculated using formula 11 c The fine interaction graph is different and is constructed hierarchically. Therefore, it is necessary to calculate the weight matrix S between the two types of meta-embedding tables. r . In order for each coarse meta-embedding to effectively select the corresponding fine meta-embedding as semantic refinement. However, directly calculating the similarity score between the coarse meta-embedding and the fine meta-embedding will ignore the semantics of the entity. As mentioned above, Formula 10 shows that after L layers of propagation, the mapping relationship between each entity and its corresponding meta-embedding is still the same as the mapping relationship in the GNN input layer. In this spirit, entities can act as bridges to connect each coarse meta-embedding with multiple fine meta-embeddings based on semantics. Therefore, the embodiment of the present invention proposes the following bridging update strategy from coarse to fine allocation matrix (such as Figure 3 As shown in

[15] , the strategy dynamically aligns the coarse meta-embedding with multiple fine meta-embeddings according to the semantic relevance of users / items).

[0156]

[0157] in, yes The pseudo-rebellion, yes The pseudo-inverse of , both are calculated using the Moore-Penrose inverse algorithm. Since Formula 22 is a dense matrix, in order to keep S r The sparsity of S is processed to preserve the r The three largest non-zero weights in each row ensure that each coarse element embedding uses exactly t at any time. r Element embedding:

[0158]

[0159] in, Returns the top-(t r) item. The embodiment of the present invention only retains S for those fine element embeddings that are considered to be crucial for the coarse element embedding. r Finally, follow the above process to calculate the weight of S r Initialize and reformulate Equation 22 as

[0160]

[0161] in, yes The pseudo-rebellion, yes Then, it is sparsified by formula 23. The algorithm process of the coarse-grained training stage is shown in Table 1:

[0162] Table 1

[0163]

[0164]

[0165] The algorithm process of the fine-grained training stage is shown in Table 2:

[0166] Table 2

[0167]

[0168] The pseudocode provided in Tables 1 and 2 summarizes the two phases of the algorithm: the coarse phase (Algorithm 1) and the fine phase (Algorithm 2). In the coarse phase, the coarse element embedding matrix is initialized and the assignment matrix S c The update frequency of this stage (using f c denoted) is used to control the frequency of allocation updates. In addition, the adjacency matrix A is calculated according to a specific equation r In Bayesian Personalized Ranking During the iterative optimization process of the loss, the allocation matrix is updated regularly and A is regenerated c , to reflect any changes in the entity-meta embedding association. When transitioning to the refinement phase, the refined meta embedding matrix is initialized and the coarse-grained to fine-grained assignment matrix S r Here, the frozen coarse element embedding matrix Similar to the coarse-grained stage, the update frequency f of this stage is determined r After initialization, calculate the adjacency matrix A r , to capture the subtle relationships between entities and meta-embeddings. During the loss process, the distribution matrix is updated from coarse to fine, and according to f r Regenerate A at a specified time interval r, to ensure consistent improvement and adaptation to changing data dynamics.

[0169] In a specific embodiment of the present invention, a series of experiments will be conducted to verify its effectiveness. The experiments are organized around the following research questions (RQs):

[0170] Q1: How accurate is the recommendation of the embodiment of the present invention compared with other baselines?

[0171] Q2: What is the function of the main ingredients?

[0172] Q3: How sensitive are embodiments of the present invention to hyperparameters?

[0173] Q4: How does fine-grained meta-embedding enhance lightweight implementations?

[0174] 1.1 Experimental Setup

[0175] 1.1.1 Experimental Set

[0176] During the evaluation process, this embodiment of the present invention used three different benchmark data corpora: Gowalla, Yelp2020, and Amazonbooks. Their notable features are open accessibility and diversity in terms of scale, departments, and user-project interaction density. This embodiment of the present invention split the dataset into training / testing / validation. Table 3 lists the details of the datasets used in the experiment:

[0177] (1) Gowalla: This dataset originates from the social networking platform Gowalla and captures user-generated location check-ins.

[0178] Table 3

[0179]

[0180] (2) Yelp2020: This dataset comes from Yelp’s 2020 challenge database and includes entities such as local service providers, restaurants, and bars.

[0181] (3) Amazon Books: This subset is extracted from the Amazon review ecosystem, which is a common benchmark in product recommendation research, with a particular focus on books.

[0182] 1.1.2 Benchmark Algorithm

[0183] To demonstrate the superiority of the present invention, we compared it with typical recommendation models, including PEP, QR, and LEGCF, which allow for the generation of embedding layers to meet different memory targets. AutoEmb, ESAPN, OptEmbed, CIESS, DHE, and NimbleTT ignore whether the final embedding structure meets the memory budget, as these methods do not formally consider memory targets as optimization objectives. The present invention also provides results for the uncompressed version of LightGCN with relatively large dimensions (i.e., UD-dim128 and UD-dim64) to compare the performance tradeoffs of various lightweight embedding methods. Because the recommenders on these devices are designed for session-based recommendations, their ID-based recommendation results are not available. Tables 4-1, 4-2, and 4-3 show the performance comparisons of the present invention's methods and the baselines. "#Param" indicates the total parameter size of the embedding layer for each method. "UD-dim128" and "UD-dim64" represent the complete embedding table settings with uniform dimensions d = 128 and d = 64, respectively. In each column, the best results achieved by the lightweight embedding method are indicated in bold.

[0184] Table 4-1

[0185]

[0186] Table 4-2

[0187]

[0188] Table 4-3

[0189]

[0190]

[0191] 1.1.3 Evaluation Metrics

[0192] The following metrics are used to evaluate the compared methods, with N set to {5, 10, 20} to observe the performance of the recommendation system.

[0193] (1) NDCG@N (Normalized Discounted Cumulative Gain): This metric evaluates ranking quality by measuring the hierarchical relevance of recommended items. Unlike precision-based metrics, NDCG@N takes into account the decay of relevance, as users are less likely to consider items ranked lower in their feeds. It assigns the largest gain to top-ranked relevant recommendations, with the gain value decreasing logarithmically as the ranking decreases.

[0194] (2) Recall@N: This reflects the proportion of the user’s preferred items that appear in the first N segments of the recommendation list. Recall@N is very instructive for judging the coverage breadth of the recommendation system. An ideal system can discover all the items of interest in the user’s preferences.

[0195] 1.1.4 Implementation Details

[0196] In this embodiment of the present invention, the embedding size d of the meta-embedding is set to 128. To create the training set, 5 negative items are randomly sampled for each user-positive item interaction. The meta-embeddings and recommendation weights are initialized using Xavier initialization and trained using the ADAM optimizer. The learning rate is selected from {10-2, 10-3, 10-4}, and the L2 penalty factor λ is selected from {0, 5×10-4, 10-3, 5×10-3, 10-2}. A 3-layer GNN is used for the Gowalla dataset, and a 4-layer GNN is used for the Yelp2020 and Amazon-book datasets. The codebook is divided into two parts: 300 code buckets for coarse meta-embedding and 100 code buckets for fine meta-embedding. The overall performance evaluated using this configuration is discussed in RQ1. For the baseline with varying sizes, the total number of parameters used in the 500-bucket configuration is first calculated. Then, using this memory budget, an appropriate sparsity target or number of hash buckets is set for methods such as PEP and QR to ensure that their embedding layers do not exceed the specified memory limit. For baselines that lack control over the final parameter size, a suitable search space is established to avoid excessive memory usage while providing flexibility for optimal setting selection. All experiments are conducted using the PyTorch framework and run on an NVIDIA A6000. Figure 4(a)-Figure 4(d) The performance of the embodiment of the present invention on different model components.

[0197] 1.2 Overall Performance (Q1)

[0198] In the embodiment of the present invention, the size m of the coarse element embedding bucket c =300, the size of the fine element embedding bucket m r = 100. Tables 4-1, 4-2, and 4-3 summarize the overall performance of the embodiments of the present invention compared to the baseline methods. The embodiments of the present invention achieved higher NDCG@{10, 20} and Recall@{10, 20} scores on all three datasets, while also using fewer parameters. This demonstrates the effectiveness of the embodiments of the present invention in improving coarse element embedding and enhancing combined full embedding.

[0199] In contrast, variable-size methods such as PEP, QR, and LEGCF ignore fine-grained semantics, resulting in suboptimal combined full embedding of user / item representations. Among single-size methods, NimbleTT and CIESS have similar parameter sizes to variable-size methods, but their performance cannot compare to the embodiments of the present invention. Compared with the embodiments of the present invention, techniques such as OptEmbed, ESAPN, and AutoEmb generate embedding layers with much larger parameter sizes. DHE performs the worst on all datasets, indicating that small-dimensional hash coding cannot improve the uniqueness of embeddings.

[0200] Compared to the uniform dimension (UD) setting, the embodiments of the present invention achieve an excellent balance between performance and parameter efficiency on the Gowalla, Yelp2020, and Amazon-book datasets. The embodiments of the present invention maintain highly competitive performance while significantly reducing the parameter size. This efficiency proves that the embodiments of the present invention can effectively refine rough meta-embeddings into high-quality combined full embeddings, making them suitable for scenarios with strict memory constraints. The significant reduction in parameter size without a proportional drop in performance highlights the robustness and practicality of the embodiments of the present invention in various applications.

[0201] 1.3 Model Component Analysis (Q2)

[0202] To verify the performance gain of each key component, the present invention conducts ablation studies on several innovative components, including the fine element embedding initialization strategy, the soft threshold sparse strategy, and the weight bridging update strategy. Figure 4(a)-Figure 4(d) shown.

[0203] (1) Fine element embedding initialization: The embodiment of the present invention improves fine element embedding in two aspects. Initialization strategy. First, instead of randomly initializing the embedding value, a part of the coarse element embedding is selected to form (Case A). Second, the present invention uses SparsePCA to preserve the main mode and control the sparsity (Case B). Experiments of the present invention show that the lack of these two components (Case A and B) will lead to a significant performance degradation. This shows that initializing with semantic and sparse attributes It is crucial to train an accurate meta-embedding code set to produce high-quality entity embeddings. This fact further reveals the core contribution of the embodiments of the present invention - improving the coarse meta-embedding with fine semantic information.

[0204] (2) Soft threshold sparse strategy: In order to evaluate the effect of soft threshold, the embodiment of the present invention removes the soft threshold and directly uses To improve (Case C). The performance degradation without soft thresholding demonstrates that the integration of soft thresholding allows for dynamic adjustment of sparsity by smoothly zeroing out small coefficients while preserving important coefficients. This embodiment of the present invention ensures that the generated embeddings maintain sparsity and computational efficiency, thereby promoting better convergence during training.

[0205] (3) Weighted bridge update strategy: The embodiment of the present invention uses formula 22 to calculate S r , the embodiment of the present invention abandons it and matches the two meta-embeddings without considering the complete embedding of the user / item to verify the importance of the bridging strategy of the embodiment of the present invention (case D). The strategy proposed by the embodiment of the present invention aims to well match each coarse meta-embedding with multiple fine meta-embeddings based on the semantics of the user / item.

[0206] 1.4 Hyperparameter Analysis (Q3)

[0207] In order to study the sensitivity of the embodiment of the present invention to hyperparameters, the number of new features required in SparePCA is d r , the threshold parameter λ in soft thresholding, the weight w of fine element embedding cr and the number of fine element embeddings t assigned to each coarse element embedding r Hyperparameter analysis was performed.

[0208] Figure 5 This is the performance of the embodiment of the present invention under different hyperparameter settings.

[0209] (1) The number of new features d r : and The embedding size d is set to 128. r The values are set to {40; 60; 80; 100; 120}. The performance results of the two datasets are shown in Figure 5 As shown in the first column of the figure group in . The study found that increasing the number of non-zero values does not guarantee performance improvement. This is mainly because the fine stage focuses on learning fine-grained semantic information. Dense and fine meta-embeddings may capture redundant semantics similar to those in the coarse stage, thereby ignoring finer details. On the contrary, too sparse embeddings (such as d r =40) will affect the ability to effectively express semantic nuances. By trying different degrees of initialization sparsity, we strive to strike a balance between capturing detailed semantics and avoiding redundancy.

[0210] (2) Threshold λ of soft threshold processing: In order to study the influence of the soft threshold parameter λ on the embedding sparsity and representation quality, the embodiment of the present invention conducted experiments with λ values set to {0.5; 1; 3; 5; 10}. These values determine the different proportions of non-zero values in the fine element embedding. Figure 5 As shown in the second column of figures in , the embodiments of the present invention achieve the best performance on both datasets when the refined meta-embeddings remain sparse. Specifically, for Gowalla and Yelp2020, the performance is best when λ = 3, with about half of the values being non-zero. This further verifies that maintaining an optimal sparsity is crucial for capturing fine-grained semantic information without generating redundancy. The results of the embodiments of the present invention also show that the design of the refinement stage can work under tight memory budgets.

[0211] (3) Weight w of fine meta-embedding cr :like Figure 5 The third column of graphs in ,shows the balance between coarse element embedding and fine element embedding.,The w cr The set of values includes {0.1, 0.3, 0.5, 0.7, 0.9}. It can be seen that for the Gowalla dataset, a weight of 0.5 performs best, while for Yelp2020, the optimal weight is 0.3. Lighter weights result in less competitive performance. This phenomenon demonstrates that the refinement stage is crucial for enhancing coarse meta-embeddings, and that sparse fine meta-embeddings can perform well even under memory constraints.

[0212] (4) The number of fine element embeddings t for each coarse element embedding r :like Figure 5 The fourth column of the graphs in t shows the performance of the number of fine element embeddings assigned to each coarse element embedding. r The value of comes from the set {2, 3, 4, 5, 6}. For the Gowalla dataset, the embodiment of the present invention is in t r The best performance was achieved under the larger setting of t = 5, while for Yelp2020, the embodiment of the present invention achieved the best performance under the larger setting of t r The best performance was achieved with a larger setting of ∑ = 3. This phenomenon occurs because larger datasets employ deeper GNNs for training, which gives fine meta-embeddings greater representational power. Consequently, fewer fine meta-embeddings are sufficient to effectively refine the coarse embeddings. Consequently, smaller datasets may require more fine meta-embeddings to achieve optimal performance.

[0213] Figure 6(a)-Figure 6(d) The figure shows the performance of the embodiment of the present invention under different ratios of coarse element embedding and fine element embedding.

[0214] 1.5 Space Complexity (Q4)

[0215] The memory consumption of the framework of the embodiments of the present invention is evaluated by adjusting the ratio between the coarse element embedding and the fine element embedding. In LEGCF, the space complexity of its lightweight embedding is O(tN + md), where t is the number of meta-embeddings for each entity and m is the number of meta-embedding codebooks. Compared with the space complexity of the complete embedding table (where t << m), the embodiments of the present invention have a space complexity of O(tN + md) << O(Nd). In the embodiments of the present invention, the space complexity of the rough stage is O(t c N + m c d), which is similar to LEGCF. However, as discussed in Section 1.4, in the fine stage, setting half of the fine element embedding values to zero can achieve the best performance. Therefore, the space complexity of this stage can be calculated as Considering m c + m r < m, m r << m c , and in the settings of the embodiments of the present invention t r << d, the overall complexity is In the case of memory constraints, the smaller m c , the lower the space complexity. The fine stage helps to balance embedding uniqueness and embedding fidelity. Figure 6(a)-Figure 6(d) This is proven.

[0216] The embodiments of the present invention propose a novel GNN-based model for multi-granularity semantic learning of lightweight meta-embeddings for ID-based recommendation systems. This hierarchical and selective connection ensures that the model of the embodiments of the present invention learns broad semantic information at the rough level, while focusing on fine-grained personalized features at the fine level. The embodiments of the present invention adopt a two-layer virtual node structure to capture coarse-grained and fine-grained semantics; adopt an innovative initialization method that relies on SparsePCA to maintain sparsity and correlation relationships; adopt a dynamic soft threshold during training to adjust the sparsity level; and adopt a weight bridging update strategy to adjust the coarse-grained and fine-grained meta-embeddings according to semantic relevance. The experimental results on the benchmarks verify the advantages of the model of the embodiments of the present invention compared with the state-of-the-art models.

[0217] The embodiments of the present invention also disclose an ID-based lightweight meta-embedding recommendation system, including:

[0218] A data acquisition module for acquiring a benchmark data corpus;

[0219] An interaction graph construction module for constructing a user-item interaction graph based on the benchmark data corpus;

[0220] The model building module is used to build a GNN-based meta-embedding multi-granularity semantic model. It adopts a two-layer virtual node structure to capture coarse-grained and fine-grained semantics and uses GNN to learn the embedding representation of users and items.

[0221] The coarse-grained node overlay module is used to overlay coarse-grained nodes in the user-item interaction graph and directly connect all user nodes and item nodes;

[0222] A fine-grained node connection module is used to introduce fine-grained nodes based on coarse-grained nodes and use the fine-grained nodes as auxiliary nodes to connect the coarse-grained nodes;

[0223] An initialization and sparsification processing module, configured to initialize the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjust the sparsity level using a soft threshold activation function;

[0224] A weight update module is used to dynamically match coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance based on a weight bridge update strategy.

[0225] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0226] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightweight meta-embedding recommendation method based on ID, characterized by: include: Obtain a benchmark data corpus; constructing a user-item interaction graph based on the benchmark data corpus; Build a GNN-based meta-embedding multi-granularity semantic model, adopt a two-layer virtual node structure to capture coarse-grained and fine-grained semantics, and use GNN to learn the embedding representation of users and items; In the user-item interaction graph, coarse-grained nodes are superimposed to directly connect all user nodes and item nodes; On the basis of coarse-grained nodes, fine-grained nodes are introduced and used as auxiliary nodes to connect coarse-grained nodes. Initializing the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjusting the sparsity using a soft threshold activation function; Dynamically matching coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance based on a weighted bridge update strategy; A two-layer virtual node structure is used to capture coarse-grained semantics and fine-grained semantics, including: Capture coarse-grained semantic information through the coarse layer and learn coarse element embedding of coarse-grained semantics; Learn fine-grained semantic information through fine layers, learn sparse fine element embeddings of fine-grained semantics based on coarse element embeddings; and strengthen the representation of coarse element embeddings through fine element embeddings; Get the coarse component embedding based on the coarse element embedding code set and the coarse assignment matrix; Add the coarse assignment matrix to the user-item interaction graph to obtain the coarse interaction graph; By stacking the coarse component embedding and the coarse meta-embedding code set as the input embedding of the coarse interaction graph, we obtain the graph propagation coarse representation and coarse meta-embedding code set of the complete entity set; Extract graph propagation embeddings of entities from the full entity set and compute similarity scores between each user-item pair; In each iteration, a fixed assignment matrix is used to compute a Bayesian personalized ranking loss based on the prediction scores of the training batch, thereby facilitating backpropagation to update the set of coarse meta-embeddings. updating the coarse assignment matrix according to the updated coarse meta-embedding code set; On the basis of coarse-grained nodes, fine-grained nodes are introduced and used as auxiliary nodes to connect coarse-grained nodes, including: Build detailed interaction graphs; By improving the method of initialization process of coarse element embedding codebook; Select r coarse element embeddings from the coarse element embedding code set to form a new coarse element embedding matrix; Initialize the fine-grained representation based on the selected coarse element embedding.

2. The ID-based lightweight meta-embedding recommendation method according to claim 1, characterized in that: Initializing the meta-embedding multi-granularity semantic model based on SparsePCA, including: in, represents the reconstruction error, ||·||1 represents the sparsity penalty term, and α is a hyperparameter that controls sparsity; represents the new coarse element embedding matrix; the goal of SparsePCA is to find a d r dimensional sparse representation And the corresponding sparse coefficient matrix W r ; Use zero-fill SparsePCA to reduce the sparse vector to the specified d-dimensional vector.

3. The ID-based lightweight meta-embedding recommendation method according to claim 2, characterized in that: The soft threshold activation function is used to dynamically adjust the sparsity level, including: The soft threshold activation function is expressed as: Where sign(x) is a sign function that returns -1 if x < 0, 1 if x > 0, and 0 if x = 0. |·| represents the absolute value, and λ is the threshold parameter. For fine-grained representation.

4. The ID-based lightweight meta-embedding recommendation method according to claim 3, characterized in that: Embedded coarse components based on fine-grained representation Expressed as: in, is the refined composition embedding, w cr is the weight parameter, S c represents the rough assignment matrix; S r It is the indirect assignment matrix between coarse element embedding and fine element embedding; The input embedding of the refined GNN will be and Stacking them up, we get: in, Represents the input embedding of GNN, according to the fine interaction graph A r The adjacency matrix of , the propagation operation of the l+1 layer is: in, It's A r The diagonal matrix of The final GNN embedding is obtained by averaging the embeddings of each layer: Among them, L represents the number of propagation layers; by r Split and obtain the complete entity set; Extract graph propagation embeddings of entities from the complete entity set and compute the similarity score between each user-item pair, fixing the assignment matrix S in each iteration. c , S r and A Bayesian personalized ranking loss is computed based on the prediction scores of the training batches, which facilitates backpropagation to update the set of refined meta-embeddings.

5. The ID-based lightweight meta-embedding recommendation method according to claim 1, characterized in that: Based on the weighted bridge update strategy, coarse-grained meta-embedding and fine-grained meta-embedding are dynamically matched according to semantic relevance, and a bridge update strategy of coarse-to-fine distribution matrix is adopted, including: The indirect assignment matrix S between coarse element embedding and fine element embedding r Expressed as: in, yes The pseudo-rebellion, yes The pseudo-inverse of , both of which are calculated using the Moore-Penrose inverse algorithm; represents a coarse meta-embedding code set; Represents a set of refined meta-embedded codes; Represents a complete entity set; The indirect assignment matrix S between coarse element embedding and fine element embedding r Perform sparse processing so that each coarse element embedding uses exactly t at any time r Element embedding: in, Returns the top-(t r ) item; The indirect assignment matrix S between coarse element embedding and fine element embedding r Perform initialization and sparse processing; The indirect assignment matrix S between coarse element embedding and fine element embedding r The formula for initialization is expressed as: in, yes The pseudo-rebellion, yes of pseudo-rebellion; For fine-grained representation; Embed code set for crude element.

6. A lightweight meta-embedding recommendation system based on ID, characterized by: include: A data acquisition module, used to obtain a benchmark data corpus; an interaction graph construction module, configured to construct a user-item interaction graph based on the benchmark data corpus; The model building module is used to build a GNN-based meta-embedding multi-granularity semantic model. It adopts a two-layer virtual node structure to capture coarse-grained and fine-grained semantics and uses GNN to learn the embedding representation of users and items. The coarse-grained node overlay module is used to overlay coarse-grained nodes in the user-item interaction graph and directly connect all user nodes and item nodes; A fine-grained node connection module is used to introduce fine-grained nodes based on coarse-grained nodes and use the fine-grained nodes as auxiliary nodes to connect the coarse-grained nodes; An initialization and sparsification processing module, configured to initialize the meta-embedding multi-granularity semantic model based on SparsePCA and dynamically adjust the sparsity level using a soft threshold activation function; A weight update module, which dynamically matches coarse-grained meta-embeddings and fine-grained meta-embeddings according to semantic relevance based on a weight bridge update strategy; A two-layer virtual node structure is used to capture coarse-grained semantics and fine-grained semantics, including: Capture coarse-grained semantic information through the coarse layer and learn coarse element embedding of coarse-grained semantics; Learn fine-grained semantic information through fine layers, learn sparse fine element embeddings of fine-grained semantics based on coarse element embeddings; and strengthen the representation of coarse element embeddings through fine element embeddings; Get the coarse component embedding based on the coarse element embedding code set and the coarse assignment matrix; Add the coarse assignment matrix to the user-item interaction graph to obtain the coarse interaction graph; By stacking the coarse component embedding and the coarse meta-embedding code set as the input embedding of the coarse interaction graph, we obtain the graph propagation coarse representation and coarse meta-embedding code set of the complete entity set; Extract graph propagation embeddings of entities from the full entity set and compute similarity scores between each user-item pair; In each iteration, a fixed assignment matrix is used to compute a Bayesian personalized ranking loss based on the prediction scores of the training batch, thereby facilitating backpropagation to update the set of coarse meta-embeddings. updating the coarse assignment matrix according to the updated coarse meta-embedding code set; On the basis of coarse-grained nodes, fine-grained nodes are introduced and used as auxiliary nodes to connect coarse-grained nodes, including: Build detailed interaction graphs; By improving the method of initialization process of coarse element embedding codebook; Select r coarse element embeddings from the coarse element embedding code set to form a new coarse element embedding matrix; Initialize the fine-grained representation based on the selected coarse element embedding.

Images