Multistage comparison recommendation method based on hyperbolic space

Abstract

The invention relates to the technical field of recommendation systems, in particular to a hyperbolic space-based multi-level comparison recommendation method, which comprises the following steps of: embedding and mapping a user and an article into a hyperbolic space, capturing a hierarchical structure and a complex relationship, and improving recommendation accuracy; node embedding with differentiated characteristics is generated on the feature space level through noise disturbance, and semantic information of original characterization is reserved at the same time; based on a contrast learning module fusing node-level and affinity-level double view angles, collaborative optimization is carried out on the node-level and affinity-level contrast view angles, so that global semantic association in a graph structure is mined while the model keeps local similarity of nodes, and a hyperbolic representation space with discrimination and generalization capabilities is formed; the weight of different loss functions is controlled to ensure the recommendation effect, the method designs a multi-level comparison perspective, the distance between nodes can be adjusted from different levels, similar nodes are closer, irrelevant nodes are pushed farther, and the problem of node crowding in the hyperbolic space is relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and particularly to a multi-level contrast recommendation method based on hyperbolic space. Background Art

[0002] Recommendation systems are one of the most successful commercial applications of artificial intelligence and play an important role in online shopping, social networks, and video sharing platforms. The core method of these systems is collaborative filtering (CF), which recommends content or products by analyzing user behavior (such as browsing, purchasing, rating) and identifying similarities with the behavior of other users. In recent years, collaborative filtering systems based on graph neural networks (GNNs) have attracted wide attention. Since user-item interactions can be represented as a bipartite graph, GNNs can utilize this bipartite structure to learn node representations, thereby improving recommendation performance.

[0003] In real-world datasets, the long-tail distribution is a common phenomenon, that is, a small number of items are very popular, while the individual frequencies of a large number of items are low, but they generally account for a relatively large proportion. In addition, recent research has shown that the historical records of users usually present a tree-like or ring-like structure, but modeling these structures in Euclidean space may lead to a certain degree of distortion. To specifically address these two challenges, recommendation methods based on hyperbolic geometry have been proposed. The capacity of hyperbolic space grows exponentially and has a natural advantage in dealing with scale-free or power-law distributed data compared with traditional Euclidean geometry. In addition, different from Euclidean geometry, hyperbolic space can more accurately capture the intrinsic structure in the user-item graph, thereby reducing information loss.

[0004] Due to the geometric properties of hyperbolic space and the inherent characteristics of data interaction in recommendation systems, nodes are distributed outward from the origin according to their popularity. Specifically, nodes with higher popularity are closer to the origin, while nodes with lower popularity are distributed farther away. This phenomenon not only reduces the distinguishability of node embeddings but also limits the effective utilization of the capacity characteristics of hyperbolic space that grow exponentially with the radius. Existing research attempts to solve this problem by pushing the overall embedding away from the origin to enhance space utilization or introducing a distance-based node pair optimization mechanism in the loss function to adaptively adjust the distance distribution of nodes in hyperbolic space. Although these methods have improved the model performance to a certain extent, the mainstream hyperbolic embedding models still fail to effectively alleviate the problem of spatial congestion of nodes with different popularity levels. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-level comparative recommendation method based on hyperbolic space, aiming to solve the problem that the existing hyperbolic embedding model fails to effectively alleviate the space congestion of nodes with different popularity levels.

[0006] To achieve the above object, the present invention provides a multi-level comparison recommendation method based on hyperbolic space, comprising the following steps:

[0007] Map the embedding of users and items into hyperbolic space to capture hierarchical structures and complex relationships, improving recommendation accuracy;

[0008] Generate node embeddings with differentiated characteristics in the feature space through noise perturbation while retaining the semantic information of the original representation;

[0009] Based on the contrastive learning module integrating the dual perspectives of node level and affinity level, the contrastive perspectives of node level and affinity level are optimized collaboratively, so that the model can mine the global semantic association in the graph structure while maintaining the local similarity of nodes, thus forming a hyperbolic representation space with both discriminative and generalization capabilities;

[0010] Control the weights of different loss functions to ensure recommendation effect.

[0011] The specific method of mapping the embedding of users and items into a hyperbolic space to capture hierarchical structures and complex relationships and improve recommendation accuracy is as follows:

[0012] A hyperbolic Gaussian sampling method is used to initialize the user and item embeddings in the hyperbolic space. Given a predefined origin, the user or item is embedded in the tangent space at the origin.

[0013] Project the initial embedding to the origin tangent space through geometric mapping to aggregate information;

[0014] After obtaining user and item embeddings through hyperbolic neighbor aggregation, the prediction layer calculates the priority scores using a hyperbolic distance function.

[0015] The method of generating node embeddings with differentiated characteristics at the feature space level through noise perturbation while retaining the semantic information of the original representation includes noise generation, noise processing and view generation operations.

[0016] The node-level contrast loss is composed of user node-level loss and item node-level loss.

[0017] The affinity refers to the similarity strength between users or items, which is quantified by a specific similarity measurement indicator, and the Jaccard similarity coefficient is used as the similarity calculation benchmark.

[0018] The multi-level contrastive recommendation method based on hyperbolic space of the present invention maps the embeddings of users and items into hyperbolic space, captures hierarchical structures and complex relationships, and improves the recommendation accuracy; generates node embeddings with differentiated characteristics at the feature space level through noise perturbation while preserving the semantic information of the original representation; based on a contrastive learning module that fuses the node-level and affinity-level perspectives, collaboratively optimizes the node-level and affinity-level contrastive perspectives, enabling the model to mine global semantic associations in the graph structure while maintaining local node similarity, and forming a hyperbolic representation space with both discriminative and generalization capabilities; controls the weights of different loss functions to ensure the recommendation effect. This method first uses the geometric properties of hyperbolic space to capture the complex hierarchical and non-linear relationships between users and items in the hyperbolic collaborative filtering stage, then uses hyperbolic noise perturbation to generate different contrastive views, then adopts multi-level contrast to form a hyperbolic representation space with both discriminative and generalization capabilities, and finally jointly optimizes the weights of the recommendation task and the contrast task to ensure the recommendation effect. This method introduces multi-level contrastive learning into the hyperbolic collaborative filtering model, combines the advantages of contrastive learning and hyperbolic learning, and significantly improves the effect of representation learning. It not only conducts node-level contrast, but also mines similarity information in user behavior patterns and item attributes by constructing a collaborative affinity graph to further optimize the embedding distribution. By contrastive learning, it ensures that the geometric relationships in the embedding space are consistent with the semantic or structural relationships of the nodes, improving the performance and generalization ability of the model. This method also designs multi-level contrastive perspectives. The model MHGCL can adjust the distances between nodes at different levels, making similar nodes closer while pushing unrelated nodes farther away. This effectively alleviates the problem of node congestion in hyperbolic space and reduces the distortion of node embeddings. It has significant advantages in the recommendation task and overcomes the limitations of existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings without creative efforts based on these drawings.

[0020] Figure 1 is a flowchart of the multi-level contrastive recommendation method based on hyperbolic space provided by the present invention.

[0021] Figure 2 is a diagram of the MHGCL model.

[0022] Figure 3 is an optimization idea diagram of the present invention.

[0023] Figure 4 is a similarity calculation method of the present invention.

[0024] Figure 5 It is a flowchart of the multi-level contrast recommendation method based on hyperbolic space provided by the present invention.

[0025] Figure 6 It is a flowchart of a specific method for mapping the embeddings of users and items into hyperbolic space to capture hierarchical structures and complex relationships and improve the recommendation accuracy. Detailed implementation manners

[0026] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.

[0027] Please refer to Figures 1 to 6 , the present invention provides a multi-level contrast recommendation method based on hyperbolic space, including the following steps:

[0028] S1 Map the embeddings of users and items into hyperbolic space to capture hierarchical structures and complex relationships and improve the recommendation accuracy;

[0029] Specific method:

[0030] S11 Use the hyperbolic Gaussian sampling method to initialize the embeddings of users and items in hyperbolic space. Given a predefined origin, the user or item is embedded in the tangent space at the origin;

[0031] In the embodiments of the present invention, for hyperbolic encoding:

[0032] Use the hyperbolic Gaussian sampling method to initialize the embeddings of users and items in hyperbolic space; given a predefined origin o = [-1, 0,..., 0], the embedding of the user or item in the tangent space at the origin is represented as and

[0033] So the initialized hyperbolic node embeddings and are defined as follows:

[0034]

[0035] where z 0 = (0, x) means setting the first element of the vector x sampled from the multivariate Gaussian distribution to 0 to match the definition of the tangent space at the origin.

[0036] S12 Project the initial embedding onto the origin tangent space through geometric mapping for information aggregation;

[0037] In the embodiments of the present invention, after obtaining the initial embedding representations of users and items, a hyperbolic neighborhood aggregation mechanism is adopted to model the high-order user-item interaction relationship. The embeddings of each user and item are first transformed into Euclidean vector representations through logarithmic mapping, and a weighted linear combination of neighborhood nodes is performed in this space:

[0038]

[0039] where and are the neighbor sets of user u and item i respectively, and are the number of users and the number of items respectively;

[0040] After performing L graph convolutions, a sum pooling function is used as the readout function to fuse the representations of all layers:

[0041]

[0042] After completing the aggregation of neighborhood information in the tangent space, the learned user and item embedding vectors in the tangent space are re-embedded into the hyperbolic space through exponential mapping:

[0043] e u = exp o (z u ), e i = exp o (z i ).

[0044] After obtaining the user and item embeddings through hyperbolic neighbor aggregation in S13, the prediction layer first calculates the priority score using the hyperbolic distance function.

[0045] In the embodiments of the present invention, in order to infer the user's preference for an item, after obtaining the user and item embeddings through hyperbolic neighbor aggregation, the prediction layer uses the hyperbolic distance function to calculate the priority score:

[0046]

[0047] where, The calculation method is as follows, and k generally takes the value of 1:

[0048]

[0049] where,

[0050] The adaptive margin loss function, as a common loss function for training user and item embeddings in the hyperbolic recommendation model, its core optimization strategy includes a triple mechanism:

[0051] First, the embedding representations of positive sample pairs (user-item interaction pairs) are brought closer;

[0052] Then push the negative sample pairs outside the dynamic boundary;

[0053] At the same time, based on the hierarchical characteristics of hyperbolic space, a larger boundary threshold is given to nodes closer to the origin of the space (root node);

[0054] The mathematical form of the loss function can be expressed as:

[0055]

[0056] Where the subscript (u, i) represents a positive pair, and (u, j) represents a negative sample pair;

[0057] m ui is an adaptive boundary, defined as follows:

[0058]

[0059] S2 generates node embeddings with differentiated characteristics at the feature space level through noise perturbation while retaining the semantic information of the original representation;

[0060] In the embodiment of the present invention, noise generation, noise processing and view generation operations are included.

[0061] Noise Generation:

[0062] The package normal distribution is a probability distribution defined in hyperbolic space. Given the hyperbolic space The mean vector in and the covariance matrix This distribution can be obtained by Perform node embedding sampling to obtain

[0063] Noise processing:

[0064] Due to noise Generated in hyperbolic space, and the nodes are embedded in the propagation process Located in the tangent space at the origin, the noise needs to be projected into the tangent space through logarithmic mapping;

[0065] In order to maintain the consistency of the embedding direction while introducing random perturbations, the sign function sign(·) is used to control the noise direction, which is defined as follows:

[0066]

[0067] The operation of injecting noise into node embeddings is defined as follows:

[0068]

[0069] Among them, and respectively represent the perturbation noises added to user u and item i in the l-th convolutional layer;

[0070] View generation:

[0071] Finally, by maintaining the node information propagation and projection process consistent with S12, two contrast views with different semantics, e' and e", are generated.

[0072] S3 is a contrast learning module based on the combined node-level and affinity-level perspectives, which collaboratively optimizes the contrast perspectives at the node level and the affinity level, enabling the model to mine the global semantic associations in the graph structure while maintaining the local similarity of nodes, and forming a hyperbolic representation space with both discriminative and generalization capabilities;

[0073] In the embodiment of the present invention, node-level contrast:

[0074] By maximizing the representational consistency of the same node in different enhanced views, the distribution characteristics of the embedding space are optimized, and a node-level contrast loss is constructed to provide self-supervised signals, thereby maximizing the consistency between node embeddings from different views;

[0075] This mechanism prompts nodes with similar semantics to form a compact clustering structure in the hyperbolic space, while promoting different categories or weakly related nodes to form separate distribution regions in the space;

[0076] User node-level contrast loss:

[0077]

[0078] where e' u , e" u are user embeddings of different views, represents the set of users in the batch training data. τ is the temperature coefficient of the softmax function;

[0079] Item node-level contrast loss:

[0080]

[0081] where e' i , e" i are item embeddings of different views, represents the set of items in the batch training data. τ is the temperature coefficient of the softmax function;

[0082] The final node-level contrast loss is jointly composed of the user node-level loss and the item node-level loss, and is defined as follows:

[0083]

[0084] Affinity-level comparison:

[0085] The affinity-level comparison deeply mines the high-order topological information of graph data, constructs a comparison optimization objective based on the multi-hop association relationship between nodes, and captures the potential collaborative signals in the user-item interaction graph. This mechanism guides nodes with implicit associations to approach each other in the hyperbolic space;

[0086] To deeply mine the collaborative association signals contained in the user-item interaction graph, a collaborative affinity graph is constructed By integrating the dual concepts of collaboration and affinity, this graph emphasizes the core role of user and item similarity measurement in the collaborative filtering method;

[0087] Among them, affinity specifically refers to the similarity strength between users or items, which is usually quantified by specific similarity measurement indicators. The Jaccard similarity coefficient is used as the benchmark for similarity calculation, as Figure 4 shown, its mathematical expression is defined as:

[0088]

[0089] Furthermore, by setting a threshold K to control the number of neighbors of each node, the collaborative affinity graph is as follows:

[0090]

[0091] Next, based on the collaborative affinity graph, an affinity-level loss is constructed;

[0092] User affinity-level comparison loss:

[0093]

[0094] where denotes the neighbors of user u in , and τ is the temperature coefficient;

[0095] Item affinity-level comparison loss:

[0096]

[0097] where denotes the neighbors of item i in , and τ is the temperature coefficient;

[0098] The final affinity-level comparison loss is jointly composed of the user affinity-level loss and the item affinity-level loss, and is defined as follows:

[0099]

[0100] S4 controls the weights of different loss functions to ensure the recommendation effect.

[0101] In the embodiments of the present invention, the adaptive boundary loss in S13 and the node-level and affinity-level contrast losses in S3 are jointly used as the loss function of the model and optimized;

[0102] Among them, the weights of each loss are controlled by corresponding hyperparameters:

[0103]

[0104] Among them, λ1 and λ2 control the weights of the node-level contrast loss and the affinity-level contrast loss;

[0105] The proposed model MHGCL is significantly better than the baseline model under various settings. Compared with the strongest baseline model, MHGCL improves Recall@10 and Recall@20 on the Amazon-CD dataset by 7.55% and 4.65% respectively, and improves NDCG@10 and NDCG@20 by 8.11% and 6.48% respectively. Similar performance improvements are also observed on the other two datasets, further verifying the effectiveness of the method;

[0106]

[0107] Experimental studies verify the effectiveness of the proposed multi-level contrast learning in improving recommendation performance, and this improvement is attributed to the combined advantages of hyperbolic graph representation learning and contrastive graph learning.

[0108] The above-disclosed is only the preferred embodiment of the multi-level contrast recommendation method based on the hyperbolic space of the present invention. Of course, the scope of the rights of the present invention cannot be limited by this. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A multi-level contrastive recommendation method based on hyperbolic space, characterized in that It includes the following steps: Map the embeddings of users and items into the hyperbolic space to capture the hierarchy and complex relationships, and improve the recommendation accuracy; Generate node embeddings with differentiated characteristics at the feature space level through noise perturbation, while preserving the semantic information of the original representation; Based on the contrastive learning module that fuses the node-level and affinity-level perspectives, co-optimize the contrastive perspectives of the node-level and the affinity-level, so that the model can mine the global semantic associations in the graph structure while maintaining the local similarity of nodes, and form a hyperbolic representation space with both discriminative and generalization abilities; Control the weights of different loss functions to ensure the recommendation effect.

2. The multi-level contrastive recommendation method based on the hyperbolic space according to claim 1, wherein; The specific method of mapping the embeddings of users and items into the hyperbolic space to capture the hierarchy and complex relationships, and improve the recommendation accuracy: Adopt the hyperbolic Gaussian sampling method to initialize the embeddings of users and items in the hyperbolic space. Given a predefined origin, the user or item is embedded in the tangent space at the origin; Project the initial embedding to the origin tangent space through geometric mapping for information aggregation; After obtaining the embeddings of users and items through hyperbolic neighbor aggregation, the prediction layer uses the hyperbolic distance function to calculate the priority score.

3. The multi-level contrast recommendation method based on hyperbolic space according to claim 1, characterized in that ； The generation of node embeddings with differentiated characteristics at the feature space level through noise perturbation, while preserving the semantic information of the original representation, includes noise generation, noise processing, and view generation operations.

4. The multi-level contrast recommendation method based on hyperbolic space according to claim 1, wherein ； The node-level contrastive loss consists of the user node-level loss and the item node-level loss.

5. The multi-level contrast recommendation method based on hyperbolic space as described in claim 1, It is characterized in that; The affinity refers to the similarity strength between users or between items, which is quantified by a specific similarity metric, and the Jaccard similarity coefficient is used as the similarity calculation benchmark.

Images