Enhancing Sequential Recommendation with Graph Contrastive Learning

Abstract

Learning useful interactions between input features is crucial for tabular data modeling. Recent efforts start to explicitly model the feature interactions with graph, where each feature is treated as an individual node. However, the existing graph construction methods either heuristically formulate a fixed feature-interaction graph based on specific domain knowledge, or simply apply attention function to compute the pairwise feature similarities for each sample. While the fixed graph may be sub-optimal to downstream tasks, the sample-wise graph construction is time-consuming during model training and inference. To tackle these issues, we propose a framework named Table2Graph to transform the feature interaction modeling to learning a unified graph. Represented as a probability adjacency matrix, the unified graph learns to model the key feature interactions shared by the diverse samples in the tabular data. To well optimize the unified graph, we employ the reinforcement learning policy to capture the key feature interactions stably. A sparsity constraint is also proposed to regularize the learned graph from being overly-sparse/smooth. The experimental results in a variety of real-world applications demonstrate the effectiveness and efficiency of our Table2Graph, in terms of the prediction accuracy and feature interaction detection.

Aspect-guided Syntax Graph Learning for Explainable Recommendation

“…For instance, [39] introduces a graph-based contrastive learning framework that exploits review information to enhance the user-item interaction graph for improving recommendation performance. Moreover, [40] also shows that GNN can help improve the performance of sequential recommendation models. For example, [41] uses GNN to model users' session-based behavior sequences.…”

Temporal Graph Contrastive Learning for Sequential Recommendation

“…For example, SGL (Wu et al 2021) improved recommendation models by random nodes and edges dropout to create contrastive views. GCL4SR (Zhang et al 2022) first performed graph contrastive learning on SR by random neighbor sampling on item transition graphs to get stable item representations in different subgraphs. LightGCL (Cai et al 2023) applied singular value decomposition as graph augmentation to enhance global collaborative relation learning process.…”

“…where P u ∈ R 1×n is the interest matrix of user u, W T ∈ R 3n×n is a trainable weight matrix. AttNet(•) denotes the attention network employed in (Zhang et al 2022). Then the following formulation calculates the possibility that the user u would interact with the expected item v at the (|S u |+1)-th step according to S u :…”

Neighborhood-based Hard Negative Mining for Sequential Recommendation

“…In contrast to the commonly used user-item graph, to make use of global information in behavior sequences, we propose to mine hard negatives on a WITG G(I, E) [27] where E is the edge set. A WITG contains global item transition patterns extracted from all user behavior sequences in the training set D. G is constructed by traversing every sequence in D. For a sequence s ∈ D, if there exists no edge between the items 𝑖 𝑚 and 𝑖 (𝑚+𝑘 ) in G, we connect them and set the edge weight 𝑤 (𝑖 𝑚 , 𝑖 (𝑚+𝑘 ) ) = 1/𝑘, where 𝑘 represents the importance of a target item 𝑖 𝑚 to its 𝑘-hop neighbor 𝑖 (𝑚+𝑘 ) in s. Otherwise, if there is already an edge between them, we update the edge weight as 𝑤 (𝑖 𝑚 , 𝑖 (𝑚+𝑘 ) ) ← 𝑤 (𝑖 𝑚 , 𝑖 (𝑚+𝑘 ) ) + 1/𝑘.…”