Getting Cold Start Users Connected in a Recommender System' s Trust Network

Patricia Victor, Martine De Cock, Chris Cornelis, Ankur Teredesai2008 · DOI

TL;DR

This paper tackles the "cold start" problem in recommender systems by leveraging trust networks. They propose and evaluate metrics to quantify the benefit of connecting new users to "key figures" (mavens and frequent raters) within these networks, showing it significantly improves recommendation coverage compared to random connections.

What Problem Does This Paper Address?

Recommender systems (RSs) are great at suggesting items users might like, but they struggle when new users join. This is the "cold start" problem: with no past ratings or interactions, the RS has no data to personalize recommendations. While some RSs ask users to rate more items, this paper explores an alternative: using a "trust network." Imagine a social network where users indicate how much they trust each other. This paper investigates how connecting a new user to trusted individuals within this network can help them quickly access a wider range of recommendations, effectively bypassing the cold start issue.

Methods

The core of the paper lies in defining and measuring the "coverage impact" of connecting a new user to someone in their trust network. They introduce two key metrics:

  • Fragmentation (F(b,a)): This measures how much the network becomes fragmented (i.e., how many items become inaccessible) if a specific user 'b' is removed from user 'a's trust network. A higher fragmentation score for 'b' means 'b' is more crucial for 'a's access to recommendations.
  • Fragmentation Utility (FU(b,a)): This refines the fragmentation score by considering the "room for original contribution" – how unique a user's rated items are compared to others in the network. It normalizes the fragmentation score, allowing for better comparison across different network structures.

These metrics are used to quantify the benefit of connecting to different types of users, specifically "mavens" ( prolific reviewers) and "frequent raters" (users who have rated many items), versus connecting to a random user.

Datasets

The authors evaluate their approach using a large dataset from Epinions.com. This dataset contains user reviews and trust relationships, which are essential for building and analyzing trust networks. The paper does not introduce a new benchmark dataset but rather utilizes an existing, well-established one for their experiments.

Key Insights and Ideas

  • Trust Networks as a Cold Start Solution: The paper strongly supports the idea that trust networks are a viable and effective way to alleviate the cold start problem by propagating trust and recommendations.
  • The Power of Key Figures: Connecting new users to "key figures" like mavens and frequent raters provides a significantly larger boost in recommendation coverage than connecting to a random user.
  • Mavens vs. Frequent Raters: Both mavens and frequent raters offer substantial benefits. Frequent raters provide a lot of ratings, while mavens are more visible and easier for new users to evaluate, making them a good starting point.
  • Activity Matters: The more active a key figure is (more reviews, more ratings), the greater their positive impact on recommendation coverage.

Key Results

The experiments on the Epinions dataset yielded several significant findings:

  • Connecting to a key figure (maven or frequent rater) is, on average, 0.41 higher in utility than connecting to a randomly chosen user.
  • This advantage holds across different levels of user activity (CS1 to CS4, representing users with 1 to 4 initial ratings).
  • The benefit of connecting to a key figure increases with their activity level. For instance, users connected to highly active frequent raters (F-50000+) showed a larger advantage than those connected to less active ones (F-2500).
  • While the paper focuses on coverage, preliminary results suggest that the presence or absence of a key figure in a trust network does not significantly affect recommendation accuracy.

Limitations and Future Work

  • Accuracy vs. Coverage: The paper acknowledges that while coverage is improved, the impact on recommendation accuracy needs further investigation. Preliminary results are promising but not discussed in detail due to space constraints.
  • Direct Trust Links: The current analysis primarily considers direct trust links. The authors suggest future work could explore the impact of "connectors" – users who bridge larger groups through indirect trust propagation.
  • Temporal Information: The evaluation of recommendation appropriateness (based on helpfulness scores) would ideally require temporal information, which was not fully utilized in this study.
  • Graded Trust Scores: Epinions does not support graded trust scores, only binary trust or distrust. This limits the granularity of the trust network analysis.

Summary updated

BibTeX
@misc{victor2008getting,
  author = {Patricia Victor and Martine De Cock and Chris Cornelis and Ankur Teredesai},
  title = {{Getting Cold Start Users Connected in a Recommender System' s Trust Network}},
  year = {2008},
  doi = {10.1142/9789812799470\_0144},
}
Figure