Warming Up Cold-Start CTR Prediction by Learning Item-Specific Feature Interactions

Yaqing Wang,Hongming Piao,Daxiang Dong,Quanming Yao,Jingbo Zhou

2024-07-14

Abstract:In recommendation systems, new items are continuously introduced, initially lacking interaction records but gradually accumulating them over time. Accurately predicting the click-through rate (CTR) for these items is crucial for enhancing both revenue and user experience. While existing methods focus on enhancing item ID embeddings for new items within general CTR models, they tend to adopt a global feature interaction approach, often overshadowing new items with sparse data by those with abundant interactions. Addressing this, our work introduces EmerG, a novel approach that warms up cold-start CTR prediction by learning item-specific feature interaction patterns. EmerG utilizes hypernetworks to generate an item-specific feature graph based on item characteristics, which is then processed by a Graph Neural Network (GNN). This GNN is specially tailored to provably capture feature interactions at any order through a customized message passing mechanism. We further design a meta learning strategy that optimizes parameters of hypernetworks and GNN across various item CTR prediction tasks, while only adjusting a minimal set of item-specific parameters within each task. This strategy effectively reduces the risk of overfitting when dealing with limited data. Extensive experiments on benchmark datasets validate that EmerG consistently performs the best given no, a few and sufficient instances of new items.

Information Retrieval

What problem does this paper attempt to address?

The paper aims to address the cold start problem in recommendation systems, particularly how to accurately predict the click-through rate (CTR) in the absence of user interaction records for new items. Specifically, the paper proposes a new method called EmerG, which improves the accuracy of CTR prediction during cold start by learning the feature interaction patterns specific to items. Existing methods typically adopt a global feature interaction approach, which can cause new items with sparse data to be overshadowed by older items with rich data, thereby affecting prediction performance. EmerG utilizes a hypernetwork to generate item-specific feature maps based on item characteristics and captures arbitrary-order feature interactions through a graph neural network (GNN) with a customized message-passing mechanism. Additionally, the method designs a meta-learning strategy to optimize the parameters of the hypernetwork and GNN, while adjusting only a small number of item-specific parameters in each task to reduce the risk of overfitting. Experimental results show that EmerG performs excellently on benchmark datasets for new items at different stages (no interaction records, few interaction records, and sufficient interaction records).