ColdU: User Cold-start Recommendation with User-specific Modulation

Crafting personalized recommendations for users with minimal interaction histories, a prevalent challenge in user cold-start recommendation within recommendation systems (RSs), is characterized by its pervasive nature. This issue is particularly pronounced in modern over-parameterized RSs built on deep networks, heightening the risk of overfitting for cold-start users. The significance of addressing the user cold-start problem extends to user satisfaction, platform growth, and ongoing algorithmic evolution. Recent approaches have modeled this challenge as a few-shot learning task, intending to rapidly generalize to personalized recommendations with limited training samples. However, existing methods are hampered by a high risk of overfitting and the substantial computational cost associated with learning large deep models. In response, this paper introduces ColdU, an innovative approach that leverages the capabilities of a multi-layer perceptron (MLP) to effectively approximate complex functions. To achieve parameter efficiency in modulating sample embeddings, the same MLP is employed for each element of the embeddings, with distinct MLPs used for different layers of the predictor. This design maintains the flexibility of MLPs while reducing the size of learnable parameters, facilitating easy personalization of recommendation models for cold-start users. Extensive experiments conducted on benchmark datasets consistently validate ColdU as a state-of-the-art solution, underscoring its efficacy in providing personalized recommendations for users with limited interaction histories.