Abstract
Collaborative Filtering (CF) approaches have been widely used in various applications of recommender systems. These methods are based on estimating the similarity between users/items by analyzing the ratings provided by users. The existing methods are often domain-specific and have not considered the time of the ratings being made in the calculation of the similarity. However, users' preferences vary over time, and so their similarity. In this paper, a novel method is proposed by re-ranking the users/items neighborhood set considering their future similarity trend. The trend of similarity is predicted, and depending on increased/decreased trend, we update the final nearest neighbor sets that are used in CF formulation. This method can be applied on a broad range of CF methods that are based on similarities between users and/or items. We apply the proposed approach on a set of CF algorithms over two benchmark datasets and show that the proposed approach significantly improves the performance of the original CF recommenders. As the proposed method only re-ranks the neighborhood set, it can be applied to any existing non-temporal similarity-based CF recommenders to improve their performance.
Introduction
Collaborative Filtering (CF) remains a cornerstone of modern recommender systems, powering the "suggested for you" features across major digital platforms. By identifying a "neighborhood" of users with similar historical preferences, these systems can predict an active user's interest in unseen items. However, a significant limitation of traditional CF is its treatment of user similarity as a static, aggregate value.
Standard CF models typically calculate similarity based on the entire history of user ratings as a single block. This approach ignores the temporal dynamics of human behavior; interests evolve, and a user who shared your taste in cinema five years ago may have diverged significantly today. While some "time-aware" models utilize time-decay functions to prioritize recent data, these can inadvertently discard valuable long-term preference patterns. This paper addresses this gap by proposing a method that treats similarity not as a fixed point, but as a trajectory.
Methodology
The proposed approach acts as a "wrapper" for existing similarity-based CF algorithms. It introduces a four-step process to predict future similarity trends and refine neighborhood selection.
1. Sequential Representation
Rather than treating ratings as an unordered set, the system represents each user's history as a time-ordered sequence of (item, rating) pairs. This preserves the chronological context necessary for trend analysis.
2. Similarity Time-Series Generation
Instead of a single similarity score, the algorithm generates a time-series of similarity values. This is achieved by using an expanding time window. For two users $u$ and $v$, the similarity is calculated at multiple points in time, reflecting how their shared interests have developed over their interaction history.
3. Trend Prediction
The resulting similarity time-series is analyzed using simple Linear Regression. The primary metric of interest is the slope ($\beta_1$) of the regression line. A positive slope indicates that the two users are becoming increasingly similar over time, while a negative slope suggests their preferences are drifting apart. Linear regression was chosen for its robustness in the face of sparse data, where more complex deep learning models might overfit.
4. Neighborhood Re-ranking
The core innovation lies in how the neighborhood is re-ranked based on the predicted trend. Neighbors are categorized into four priority classes:
- UP (Upward & Positive): High current similarity with an increasing trend.
- UD (Upward & Negative): Lower current similarity but trending upward.
- DP (Downward & Positive): High current similarity but trending downward.
- DN (Downward & Negative): Low current similarity and trending downward.
By prioritizing "UP" and "UD" users, the system focuses on neighbors whose preferences are most likely to align with the user's future interests.
Experimental Evaluation
The method was evaluated using two benchmark datasets: MovieLens (100k), a dense dataset of movie ratings, and a subset of Goodreads, representing a sparse real-world scenario with only 0.18% density.
The researchers tested the "trend-corrected" version of 18 different CF algorithms, including popular methods like Pearson Correlation (PCC) and Adjusted Cosine (ACOS). The results demonstrated near-universal improvements in recommendation accuracy.
Key Results
- Precision and F1-Score: On the MovieLens dataset, Item-based CF (ICF) achieved a 62.5% improvement in precision. PCC and ACOS saw gains of 39.6% and 37.2%, respectively. The F1-score improved across all 18 tested algorithms on both datasets.
- Computational Efficiency: The proposed method introduces manageable overhead. The runtime is approximately 3x that of standard static CF, making it viable for production environments.
Conclusion and Future Work
This research demonstrates that predicting the trajectory of user similarity significantly enhances the performance of collaborative filtering. By prioritizing neighbors who are "becoming" similar rather than those who "were" similar, the system better captures the evolving nature of human preferences.
Future work will explore the sensitivity of the model to initial window sizes and investigate non-linear trend prediction models to capture cyclical or bursty interest patterns. Additionally, addressing the "cold start" problem for new users with minimal history remains a key area for further development.
Summary updated