Research on Key Technology for Modeling Cross-Domain Sequential Behaviors in Recommender Systems

In the era of artificial intelligence, recommender systems have become core modules and standard configurations of many online service platforms (e.g., e-commerce, digital media, education, healthcare, travel, social entertainment, and more) for providing users with personalized services and suggestions. In recent years, modeling users’ sequential behavior data and non-behavioral information has attracted extensive attention from academia and industry, because they can be exploited to more effectively learn users’ real interests and recent preferences for items, thereby enabling the system to provide more accurate personalized recommendation services.

However, most of the research on users’ sequential behavior modeling only considers the case of a single domain, and there are few modeling methods and recommendation algorithms designed specifically for cross-domain sequential behavior data. In a real recommender system (such as an online shopping platform), user behaviors not only have a sequential nature but also have cross-domain characteristics (such as browsing and purchasing behaviors of books and movie tickets in one day).

Therefore, effectively modeling cross-domain sequential behaviors (CDSB) can be crucial to deeply explore and learn user behavior data, further improving the performance of personalized services of recommender systems, which is in demand in many practical application scenarios. Unlike general single-domain sequential behavior modeling, modeling cross-domain sequential behaviors faces more prominent challenges, including (i) the diversity of users’ preferences across domains, (ii) the dependency of users’ behaviors across domains, (iii) the transferability of preference knowledge across domains, (iv) the cross-domain fairness issues, (v) the cross-domain biased dependency in user behaviors, and (vi) the transfer of sensitive information in cross-domain knowledge.

This project will center on exploring innovative Transfer Learning (TL) techniques for modeling CDSB. Our objective is to develop an Unbiased Learning (UL) framework for TL and recommendation algorithms, enabling learning user preferences from CDSB in an accurate and fairness-aware manner. It is important to highlight that the introduction of TL and UL paradigms and related technologies will offer fresh perspectives and novel technical approaches to address the above-mentioned challenges, constituting the primary contribution of this project.