Incorporating Curiosity into Personalized Ranking for Collaborative Filtering (Student Abstract)

Qiqi Ding, Yi Cai, Ke Xu, Huakui Zhang

Introduction

Social-based recommender systems typically rely on the assumption that users share preferences with their friends. While this approach is effective, it often leads to "information cocoons," where users are trapped in echo chambers of repetitive recommendations. Furthermore, traditional models often overlook the psychological drive of curiosity. In social contexts, human interest is frequently piqued not just by agreement, but by disagreement. When friends hold conflicting opinions about an item, it creates an "uncertainty stimulus" that can motivate a user to explore that item. To address this, we propose CBPR (Curiosity Enhanced Bayesian Personalized Ranking), a model that quantifies curiosity to provide more accurate and diverse recommendations.

Methodology

The CBPR model integrates curiosity into the Bayesian Personalized Ranking framework by quantifying social disagreement and its impact on user interest.

Quantifying Curiosity

Following Berlyne’s psychological theory (1960), curiosity is modeled as a response to uncertainty. The model uses information entropy to measure the diversity of ratings among a user's social circle. High entropy in ratings (e.g., a mix of very high and very low scores) indicates a high curiosity score for that item.

Social Influence and Decay

The model extends beyond immediate social links to include "indirect friends" (friends of friends). To account for the diminishing influence of distant connections, a decay factor ($e^{-\mu d}$) is applied based on the distance $d$ in the social graph.

The Wundt Curve and Ranking

To ensure recommendations remain relevant, the model incorporates a curiosity threshold ($\gamma$) based on the Wundt Curve principle, which suggests that while some stimulus is engaging, too much can be overwhelming or irrelevant. The ranking logic is modified to prioritize curiosity-inducing items over standard negative items, alongside the traditional preference for positive items.

Experimental Setup

The CBPR model was evaluated using two real-world social datasets:

  • FilmTrust: A movie-centric social network.
  • CiaoDVDs: A product review dataset featuring "trust" relationships between users.

These datasets provide the necessary combination of user-item interactions and social network topology to test curiosity-driven recommendations.

Results and Discussion

The performance of CBPR was compared against several baseline methods, including standard BPR and social-aware models like SBPR and UC.

Performance comparison of CBPR against baseline methods on FilmTrust and CiaoDVDs datasets.

As shown in the results, CBPR consistently outperforms the baselines in both Precision (P@10) and Recall (R@10) across both datasets. This improvement demonstrates that incorporating social disagreement as a signal for curiosity significantly enhances recommendation accuracy. By moving beyond simple similarity, the model successfully identifies items that users are likely to engage with due to piqued interest.

Conclusion and Future Work

This research highlights the value of psychological factors like curiosity in social recommendation. By treating social conflict as a valuable signal rather than noise, CBPR provides more accurate and diverse suggestions. Future work could explore dynamic curiosity modeling to account for temporal changes in user interest and optimize the computational efficiency of entropy calculations at scale.

Summary updated

Abstract

Curiosity affects users' selections of items, and it motivates them to explore the items regardless of their interests. This phenomenon is particularly common in social networks. However, the existing social-based recommendation methods neglect such feature in social network, and it may cause the accuracy decease in recommendation. What's more, only focusing on simulating the users' preferences can lead to information cocoons. In order to tackle the problem, we propose a novel Curiosity Enhanced Bayesian Personalized Ranking (CBPR) model. Our model makes full use of the theories of psychology to model the users' curiosity aroused when facing different opinions. The experimental results on two public datasets demonstrate the advantages of our CBPR model over the existing models.
BibTeX
@inproceedings{ding2021incorporating,
  author = {Qiqi Ding and Yi Cai and Ke Xu and Huakui Zhang},
  title = {{Incorporating Curiosity into Personalized Ranking for Collaborative Filtering (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17885},
}
Figure