Neural processes for recommendation system : algorithms and applications

Abstract

With the rapid expansion of the Internet, users are overwhelmed with diverse content, including articles, videos, and products. This abundance of choices makes it challenging for individuals to identify their desired items. Consequently, mining user preferences through their behavior data is crucial for enhancing content delivery services and helping users find their favorite items more efficiently. Recommendation systems play a key role in matching content with potential users. However, these systems face challenges such as the cold-start problem and the need to adapt to evolving user preferences. This thesis addresses these challenges by advancing research on the cold-start problem and developing methods to capture diverse evolving user preferences.

In Chapters 3 and 4, we explore the meta-learning methods to alleviate the cold-start problem and capture the diverse preferences via hierarchical intention learning. Specifically, we propose two novel neural process models, MINSP and DISMP. MINSP considers the recommendation process for each user as a stochastic process, which defines distributions over functions and is capable of rapid adaptation to different users. To capture the user’s intention on different levels, an iterative additive algorithm is proposed to minimize the approximation error by backfitting the residuals of previous approximations. This approach achieves superior performance compared to the state-of-the-art baselines. Then, DISMP considers recommendations for users and items as stochastic processes so that it can adapt to new users and items efficiently. Meanwhile, it includes a two-level intention filtering for the hierarchical intention representations of cold-start users and items. Experiment results demonstrate the effectiveness of this method on cold-start recommendation.

User behaviors can be represented in a series of session sets over time, where each session set consists of the items interacted within a short timeframe (known as a session). In Chapter 5, we explore building a neural process model that can accurately describe evolving user preferences in sequential recommendation. In particular, we propose a session-aware pre-training of neural process transformer (SessNPT). SessNPT first treats each short sequence of a session as a stochastic process. A pre-training objective of session set completion task is derived to learn complex understanding of inter-session and intra-session relations. Then, a sequential adaptation is performed on the pre-trained SessNPT for sequential recommendation. Extensive experiments are conducted to demonstrate the superior performance compared to the state-of-the-art baselines.

Overall, this thesis contributes to the family of neural process models by providing novel solutions to the cold-start and sequential recommendation. Additionally, it offers new insights for future research in this field.