Abstract
Existing recommendation algorithms often rely heavily on raw score information in user rating matrices. However, a user's rating of an item does not always reflect their true interest due to external biases. This paper proposes a collaborative filtering recommendation algorithm (FixCF) based on user rating modification. By analyzing item quality, user memory functions, and social trust networks, FixCF corrects original rating information to isolate genuine user preferences. Experimental results across datasets like Movielens, Epinions, and Flixster demonstrate that FixCF significantly improves recommendation accuracy and catalog coverage compared to classical and matrix-decomposition-based algorithms.
Introduction: The Problem of Noisy Ratings
Most recommendation engines utilize collaborative filtering, assuming that past shared preferences predict future behavior. However, raw star ratings are inherently "noisy." A high rating might be driven by the objective high quality of an item rather than a specific personal interest, or it might be influenced by the anchor effect—a psychological bias where a user's current rating is swayed by their recent rating history. Furthermore, social pressure from trusted friends can distort a user's individual preference. The FixCF algorithm addresses these challenges by preprocessing raw ratings to eliminate these "unfavorable factors."
The FixCF Methodology
The core innovation of FixCF is a preprocessing step that modifies raw ratings ($r_{ui}$) based on three distinct dimensions: item quality, user memory, and social trust.
- Item Quality (Bayesian Estimation): Instead of simple averages, which are vulnerable to outliers in items with few ratings, the researchers used a Bayesian formula (similar to the IMDB Top 250) to provide a stable quality score ($Q$).
- User Memory Function: To account for the anchor effect, a memory factor ($M$) is calculated by analyzing rating sequences. This determines if a user is in a "high-rating" or "low-rating" streak, adjusting the score to find the baseline interest.
- User Trust Network: The model builds a directed graph of social trust. For users without direct connections, a "weight spread" model calculates indirect trust, selecting the path with the highest value ($T$).
Key Insights: Trust Decay and True Interest
A central hypothesis of this research is that a rating is a composite of interest, objective quality, and situational bias. By mathematically subtracting quality and bias, a purer signal of user interest remains.
One critical finding involves the behavior of social influence within the trust network. The researchers confirmed that as the distance between users in a social network increases, their similarity—and thus the reliability of their recommendations—decreases sharply.
Additionally, the study found a "sweet spot" for the modification offset ($\alpha$) at 0.3. This suggests that while correcting for bias is essential, over-correcting can lead to a loss of the original user signal, potentially harming accuracy.
Experimental Results
FixCF was benchmarked against classical algorithms (UserCF, ItemCF) and advanced models (SVD++, PMF, TrustSVD) using five representative subsets designed to test edge cases like cold-start users and biased items.
- Accuracy Gains: FixCF achieved the highest accuracy in most scenarios. On the Epinions dataset, it showed an average RMSE improvement of 11.8% over baselines. On Flixster, accuracy improvements ranged from 2.3% to 15%.
- Coverage: The algorithm maintained high coverage, proving effective at recommending "long-tail" or niche items that standard score-based systems often overlook.
- Quality Validation: The Bayesian method achieved a 0.7 hit rate with the IMDB Top 250 for top-tier movies, significantly outperforming traditional link-analysis methods like HITS (0.2 hit rate).
Conclusion and Future Work
By treating raw ratings as a starting point rather than an absolute truth, FixCF provides a more nuanced understanding of user interest. While highly effective, future work will look into "interest drift"—how user preferences evolve over long periods—and the potential for hybridizing this rating modification approach with content-based recommendation systems to further refine results.
Summary updated