A contrast collaborative filtering recommendation method based on adaptive decoupling

Abstract

The application discloses a contrast collaborative filtering recommendation method based on adaptive decoupling, and belongs to the technical field of recommendation. The method enriches the preference representation of users and items by modeling global latent intention information, and uses adaptive weights to complete the task of generally minimizing node similarity, thereby relieving the problem of over-smoothing. The method mainly consists of four parts: the first part performs graph information propagation based on a user-item interaction matrix, generates local embeddings of users and items, and uses cross-layer contrast learning to capture user and item representations that are more uniformly distributed; the second part uses a K-means clustering algorithm to construct decoupled intention embeddings of users and items, and obtains the final node representation by fusing information of two levels; the third part uses adaptive weights to control the difference between nodes based on the neighbor harmony score of the node, thereby relieving the problem of over-smoothing; and the fourth part completes recommendation according to item features, decoupled information of users and interest preferences. The application models user preferences in a more fine-grained manner by decoupling intention features, controls the difference between nodes by using adaptive weights, thereby relieving the problem of over-smoothing, and effectively improves the recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a comparative collaborative filtering recommendation method based on adaptive decoupling. Background Art

[0002] In an era of information explosion, the rapid development of the internet has brought many conveniences to people, but it has also exacerbated the problem of information overload. Recommendation systems learn from users' interests through historical interactions and provide personalized recommendations, helping users filter out potentially interesting items from a vast amount of information. Collaborative filtering is a classic recommendation algorithm that primarily predicts user preferences for various items based on historical user behavior (including browsing, clicking, and purchasing).

[0003] Existing collaborative filtering recommendation algorithms focus on how to effectively model user interest preferences. With the booming development of graph neural networks, many people have tried to introduce graph neural networks into collaborative filtering, which can model interaction data as a graph (such as a user-item bipartite graph) and then iteratively propagate neighborhood information to learn effective node representations. This is called graph collaborative filtering. This technology significantly improves the model's ability to capture high-order collaborative signals. Although these methods have achieved relatively good results, they still face some problems. Most previous studies have ignored the fact that user interaction behavior may be affected by many potential intent factors. Recently, some methods have tried to use decoupled intent factors to enrich the embedding representations of users and items. Although these methods have achieved good results, they still have many limitations, namely, they do not take into account the fact that unreasonable integration of additional modeling factors may aggravate the over-smoothing problem.

[0004] Therefore, how to better model and decouple intention factors from a global perspective and reduce the harm caused by unreasonable integration is an urgent problem that needs to be solved. Summary of the Invention

[0005] To address these issues, this paper refines local and global supervisory signals across views and adaptively controls the differences between nodes to mitigate the oversmoothing problem. A contrastive collaborative filtering recommendation method based on adaptive decoupling is designed and invented to predict user preference for candidate items.

[0006] The technical solution adopted in the present invention is:

[0007] A comparative collaborative filtering recommendation method based on adaptive decoupling, the method comprising the following steps:

[0008] Step 1: Based on the user-item interaction matrix in the dataset Learning local embeddings for users and items via graph convolutional networks embedding propagation and neighbor aggregation u and ei ,In order to promote a more uniform distribution of user and item representations, ,cross-layer contrastive learning is used to enhance the supervision signal, and ,the embeddings of the first and last layers are compared based on the ,mutual information maximization principle;

[0009] Step 2: Perform K-means clustering on the local embeddings of users and items to obtain the intent prototype representation with global context, i.e., the cluster center point c u , and then use the local embedding e u With each intent prototype c u The correlation score between them is used to calculate the user's global level decoupled intention embedding h u Finally, the user local embedding e is fused through the combination weight α u and global decoupling embedding h u To obtain the final user representation s u ;

[0010] Step 3: Based on the structural consistency between the final embedded information and its complete neighbor information, calculate the embedded neighbor harmony score a , so as to identify special nodes. The specific calculation formula is as follows:

[0011]

[0012] Among them, e a ,e a' Represents the initialized user and item embeddings, sig(·) represents the sigmoid activation function, and we summarize the k-order neighbor information of the node by summing up to explore the neighbor harmony score of the central node a. The higher the neighbor harmony score, the stronger the structural consistency between the node and its neighboring nodes. a Get the adaptive weight ω of user u u , thereby adaptively controlling the differences between nodes;

[0013] Step 4: Complete the recommendation based on the item features, user’s decoupling information and interest preferences. For each candidate item, the final recommendation probability is based on its feature s i and the current user's embeds u , process it through the softmax function to obtain the final recommendation probability

[0014] The beneficial effects of the comparative collaborative filtering recommendation method based on adaptive decoupling provided by the present invention are:

[0015] (1) This invention decouples user-item interactions, more accurately capturing the user's true behavioral intentions during interaction, thereby achieving better recommendation results.

[0016] (2) The present invention adaptively controls the differences between nodes by comparing the degree of harmony between a node and its complete neighbor information, thereby reducing the impact of oversmoothing on model representation learning;

[0017] (3) The present invention effectively extracts local and global supervisory signals from two perspectives. When the recommendation scenario data is relatively sparse, the problem of data sparsity is solved by supplementing additional modeling factors, thereby improving the recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention, and those skilled in the art can also obtain other drawings based on these drawings.

[0019] Figure 1 A schematic diagram of the overall structure of a model of a comparative collaborative filtering recommendation method based on adaptive decoupling provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0021] Given a user-item interaction matrix If user u has purchased item i before, then otherwise According to the interaction matrix To predict the possibility of candidate user u purchasing item i Collaborative filtering algorithms assume that if two users behave similarly, they are more likely to have similar interests. Figure 1 The following is a detailed introduction to the system process:

[0022] (1) Local collaborative relationship encoding

[0023] First, the user-item interaction matrix It is constructed as a user-item bipartite graph. Following the LightGCN method, it discards the nonlinear activation and feature transformation parts in the propagation process and obtains the user and item feature embedding vectors of each layer. and

[0024]

[0025] in and Represents the neighbor set of target user u and item i, which can offset the impact of node degree imbalance. In order to improve the feature embedding vector, we use multi-layer embedding propagation based on graph neural network and add the node embeddings of all layers to obtain the final user and item local embedding e u and e i :

[0026]

[0027] To promote a more even distribution of user and item representations, we use cross-layer contrastive learning to enhance the supervisory signal. We compare the embeddings of the first and last layers, representing the results before and after aggregating neighborhood information, respectively. Finally, we construct a contrastive loss function based on the principle of maximizing mutual information:

[0028]

[0029] Where s(·) represents the cosine similarity function, which is used to measure the similarity between two embeddings; τ1 is the temperature in softmax, which is used to control the smoothness of the softmax curve;

[0030] (2) Global Decoupling of Multi-Intent Encoding

[0031] The interaction between users and items is often influenced by latent intent factors. We assume that users with similar intents tend to exhibit similar behaviors. Therefore, we use the K-means clustering algorithm to learn the prototype representation of the intent of users and items, and regard the center of the cluster as the intent prototype with global context:

[0032]

[0033] Among them C k represents the kth cluster obtained by clustering, Represents the kth intention prototype from the user's perspective. Considering that a single user's behavior may be affected by multiple intentions, the local embedding e u With each intent prototype c u The final intent representation is calculated using the correlation score between them:

[0034]

[0035] Each user u has two embeddings from different perspectives, namely, e with local dependencies u and h with global dependencies u , the two are fused together by combining the weight α to obtain the final representation of the node:

[0036] s u =(1-α)e u +αhu ;

[0037] (3) Adaptive weight difference module

[0038] To prevent over-smoothing of the supervisory signal due to unreasonable decoupled representation combinations, the model proposes a method to dynamically adjust the learning of embedded representations. That is, by comparing the consistency between the embedded information and its neighboring information, it adaptively determines which nodes are more likely to cause over-smoothing. Our goal is to minimize the following loss function:

[0039]

[0040] The loss consists of two parts: user-node similarity and item-node similarity, where g(·) represents the similarity function between each node and all other nodes. The specific calculation formula is as follows:

[0041]

[0042] Among them, the hyperparameter τ2 is used to adjust the neighborhood harmony score, and ω u Controls the node similarity weight of user u, which is obtained based on the harmony score between the node and its neighbor information:

[0043]

[0044] Among them, 0<ε<1 is a hyperparameter, score a is the harmony score between the original embedding of node a and its neighboring node information, specifically:

[0045]

[0046] Among them, e a ,e a' represents the initialized user and neighbor embeddings, sig(·) represents the sigmoid activation function. We summarize the k-order neighbor information of the node by summing up, thereby exploring the neighbor harmony score of the central node a. The higher the neighbor harmony score, the stronger the structural consistency between the node and its neighboring nodes.

[0047] (4) Model training

[0048] During the training phase, the interaction probability between users and items is calculated based on the final embeddings of users and items, and the softmax function is applied to obtain the final recommendation probability of each candidate item:

[0049]

[0050] The greater the interaction preference score is, the greater the possibility that the user u interacts with the item i is, and the recommendation task is optimized by using the pair-wise marginal loss, specifically:

[0051]

[0052] wherein, all the items that the user u has interacted with before, and i' represents a negative sample that has not been interacted with.

[0053] The above is only a preferred embodiment of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and modifications can be made, which should be included in the protection scope of the present application.

Claims

1. A comparative collaborative filtering recommendation method based on adaptive decoupling, the method comprising the following steps: Step 1: Based on the user-item interaction matrix in the dataset, we learn local embeddings of users and items through graph convolutional network embedding propagation and neighbor aggregation, and use cross-layer contrastive learning to capture more evenly distributed representations. Step 2: Perform K-means clustering on the local embeddings of users and items to obtain the intent prototype representation with global context, i.e., the cluster center point c u , and then use the local embedding e u With each intent prototype c u The correlation score between them is used to calculate the user's global level decoupled intention embedding h u Finally, the user local embedding e is fused through the combination weight α u and global decoupling embedding h u To obtain the final user representation s u ; Step 3: Based on the structural consistency between the final embedded information and its complete neighbor information, calculate the embedded neighbor harmony score a , so as to identify special nodes. The specific calculation formula is as follows: Among them, e a ,e a' Represents the initialized user and item embeddings, sig(·) represents the sigmoid activation function, and we summarize the k-order neighbor information of the node by summing up to explore the neighbor harmony score of the central node a. The higher the neighbor harmony score, the stronger the structural consistency between the node and its neighboring nodes. a Get the adaptive weight ω of user u u , thereby adaptively controlling the differences between nodes; Step 4: Complete the recommendation based on the item features, the user's decoupled information, and interest preferences. For each candidate item, the final recommendation probability is based on its features and the current user preferences, and is processed by the softmax function to obtain the final recommendation probability.

2. The method for recommending collaborative filtering based on adaptive decoupling according to claim 1, characterized in that: In step 1, based on the user-item interaction matrix We follow the LightGCN approach, discarding nonlinear activation and feature transformation during propagation to obtain user and item feature embedding vectors at each layer: in and Represents the neighbor set of target user u and item i, which can offset the impact of node degree imbalance. In order to improve the feature embedding vector, we use multi-layer embedding propagation and add the node embeddings of all layers to obtain the user and item local embedding e u and e i ; To promote a more even distribution of user and item representations, we use cross-layer contrastive learning to enhance the supervisory signal. We compare the embeddings of the first and last layers and construct a contrastive loss function based on the principle of maximizing mutual information: Where s(·) represents the cosine similarity function, which is used to measure the similarity between two embeddings, and τ1 is the temperature in the softmax, which is used to control the smoothness of the softmax curve.

3. The method of claim 1, wherein: In step 2, the interaction between users and items is often affected by potential intention factors. Therefore, the K-means clustering algorithm is used to learn the intention prototype representation of users and items, and the center of the cluster is regarded as the intention prototype with global context: Among them C k represents the kth cluster obtained by clustering, To represent the k-th intention prototype from the user’s perspective, we use the local embedding e u With each intent prototype c u The final intent representation is calculated using the correlation score between them: Finally, the local embedding and the global decoupled embedding are fused by combining the weight α to obtain the final representation of the node: s u =(1-a)e u +ah u。 4. The method of claim 1, wherein: In step 3, the clustering algorithm will cause similar users to have basically the same decoupled information. At the same time, unreasonable integration of additional modeling information may aggravate the oversmoothing phenomenon. Therefore, the harmony score between the embedded information and the neighbor information is used to adaptively control the differences between nodes. The loss function is specifically: where g(·) represents the similarity function between each node and all other nodes, ω u represents the adaptive weight of user u, which is obtained based on the harmony score between the node and the neighbor information: Among them, 0<ε<1 is a hyperparameter, score a is the harmony score between the original embedding of node a and its neighboring node information, specifically: Among them, e a ,e a' denotes the initialized user and item embeddings, sig(·) denotes the sigmoid activation function. We aggregate the k-order neighbor information of a node by summing up, thereby exploring the neighbor harmony score of the central node a. A higher neighbor harmony score indicates a stronger structural consistency between the node and its neighboring nodes.

5. The method of claim 1, wherein: In step 4, the interaction probability between the user and the item is calculated based on the final embedding of the user and the item, and the softmax function is applied to obtain the final recommendation probability of each candidate item: We use pairwise margin loss to optimize the recommendation task, specifically: in, represents all items that user u has interacted with before, and i' represents negative samples that have not interacted with.

Images