Cross-domain novelty seeking trait mining for sequential recommendation

Fuzhen Zhuang, Yingmin Zhou, Fuzheng Zhang, Xiang Ao, Xing Xie, Qing He

Introduction

Sequential recommendation systems aim to predict a user's next interaction based on their historical behavior. However, these systems frequently encounter challenges related to data sparsity and "cold starts," particularly when a user has limited history in a specific target domain. While cross-domain recommendation techniques attempt to mitigate this by leveraging data from a source domain, they often rely on the assumption of consistent rating patterns across areas.

This paper proposes that such approaches overlook two fundamental aspects of human behavior: the sequential nature of choices and the underlying psychological drivers. Specifically, the authors focus on the novelty-seeking trait—an innate psychological desire for unfamiliar or complex experiences. By modeling this trait as a stable characteristic that transcends specific domains, the proposed Cross-Domain Novelty-Seeking Trait mining (CDNST) framework can more effectively transfer knowledge and improve recommendation accuracy.

The CDNST Framework

The CDNST model is a probabilistic graphical framework designed to capture and share a user's novelty-seeking distribution across domains.

Shared Trait Distribution

At the heart of the model is a shared multinomial distribution ($\theta$) representing the user's novelty-seeking level. Unlike models that treat domains in isolation, CDNST learns this distribution simultaneously from both source and target domains. This assumes that a user's tendency to seek novelty in one area (e.g., niche indie music) is a strong predictor of similar behavior in another (e.g., experimental films).

The CDNST graphical model illustrates how the novelty-seeking distribution ($\theta$) is shared to facilitate knowledge transfer between source ($s$) and target ($t$) domains.

Dynamic Choice Novelty (DCN)

To operationalize "novelty" within a sequence, the authors introduce the Dynamic Choice Novelty (DCN) matrix. This metric quantifies the novelty of an item at a specific point in time by considering its global popularity and the frequency of transitions from the user's previous choice. By incorporating context information such as tags or directors, the DCN provides a granular view of how "new" an item is relative to the user's current trajectory.

The DCN component calculates novelty by analyzing sequential transitions and context information across user choice histories.

Experimental Evaluation

The researchers evaluated CDNST using real-world data from Douban, covering movies, music, and books. The dataset included thousands of users active across multiple domains, allowing for the creation of 14 distinct transfer recommendation scenarios (e.g., Music $\rightarrow$ Movie).

Key Findings

  • Performance Gains: CDNST consistently outperformed state-of-the-art baselines, including standard Markov Chains (MC) and the single-domain Novel Seeking Model (NSM), across metrics such as nDCG, MRR, and Precision@k.
  • Effective Knowledge Transfer: The study found that simply pooling data from two domains (the "U" variants of baselines) often degraded performance. This underscores the importance of the structured transfer mechanism provided by the shared trait distribution in CDNST.
  • Directed Sequential Behavior: Transfer effectiveness was found to be asymmetrical and temporally sensitive. For instance, transferring from Music $\rightarrow$ Movie was often more effective than the reverse, likely because users frequently consume a soundtrack before watching the associated film.
  • Mitigating Negative Transfer: The authors introduced a relatedness measure ($Sim$) using external web data and word2vec to predict transfer effectiveness, helping the model avoid "negative transfer" from unrelated auxiliary data.

Conclusion

The CDNST framework demonstrates that modeling psychological traits like novelty-seeking provides a powerful bridge for cross-domain sequential recommendation. By moving beyond raw rating correlations to capture the "psychological DNA" of users, the model effectively addresses data sparsity. Future research will explore additional psychological traits, such as harm avoidance or reward dependence, and seek to develop more flexible models for handling complex temporal dependencies.

Summary updated

Abstract

Transfer learning has attracted a large amount of interest and research in last decades, and some efforts have been made to build more precise recommendation systems. Most previous transfer recommendation systems assume that the target domain shares the same/similar rating patterns with the auxiliary source domain, which is used to improve the recommendation performance. However, to the best of our knowledge, almost these works do not consider the characteristics of sequential data. In this paper, we study the new cross-domain recommendation scenario for mining novelty-seeking trait. Recent studies in psychology suggest that novelty-seeking trait is highly related to consumer behavior, which has a profound business impact on online recommendation. Previous work performing on only one single target domain may not fully characterize users' novelty-seeking trait well due to the data scarcity and sparsity, leading to the poor recommendation performance. Along this line, we proposed a new cross-domain novelty-seeking trait mining model (CDNST for short) to improve the sequential recommendation performance by transferring the knowledge from auxiliary source domain. We conduct systematic experiments on three domain data sets crawled from Douban (www.douban.com) to demonstrate the effectiveness of the proposed model. Moreover, we analyze how the temporal property of sequential data affects the performance of CDNST, and conduct simulation experiments to validate our analysis.
BibTeX
@misc{zhuang2018cross,
  author = {Fuzhen Zhuang and Yingmin Zhou and Fuzheng Zhang and Xiang Ao and Xing Xie and Qing He},
  title = {{Cross-domain novelty seeking trait mining for sequential recommendation}},
  year = {2018},
  eprint = {1803.01542},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.01542},
}
Figure