| Summary: | Recommender systems filter through the vast pool of information and provide personalized recommendations. However, with the dynamic nature of user preference, it is essential to design recommender systems that can adapt to the continuous changes in user preferences and the evolving environment.
In this thesis, we conduct a review of existing studies in recommendation models. More importantly, we perform a systematic analysis of training and evaluation protocols in recommender system research. Our analysis reveal that current setups often overlook the global timeline, leading to data leakage issues. To assess the impact of data leakage, we conduct carefully designed experiments where we gradually introduce increasing leaked data in training. The results show that data leakage results in unpredictable and inconsistent recommendation accuracy, which poses challenges in estimating a recommendation model's actual performance. Hence, we underscore the importance of following the global timeline in both training and evaluation stages of recommendation models.
Furthermore, we demonstrate that the ignorance of the global timeline and data leakage hinders recommender systems from accurately modeling the temporal context. We specifically investigate this point using "popularity" of items. It is showed that failure to capture the accurate temporal context results in less accurate recommendations. While the importance of adhering to the global timeline is emphasized, the methods for incorporating temporal context into recommender systems remain unclear. To address this gap, we conduct experiments to explore the relationship between user interactions and temporal context. The experimental results provide insights into which interactions and how many interactions should be included in the training set to improve recommendation performance within specific temporal contexts. Through extensive analysis, we find that recent interactions, which are more relevant to the target time context, should be prioritized. These findings offer guidance on effectively integrating temporal context into recommender systems to enhance recommendation accuracy in specific time contexts.
From the recent interactions, we learn a user's latest preference. However, relying solely on recent interactions for training may lead to overfitting, causing the recommendation model to overlook long-term preferences that are essential for accurate recommendations. To address this, we propose the use of incremental learning techniques to retain both the short-term and long-term preferences of users. Specifically, we introduce an incremental learning framework that retrains the GCN-based model with the disentanglement of the two types of preferences. This approach ensures effective recommendations.
In summary, despite existing research efforts in the field of recommender systems, we contend that there is a lack of study from the perspective of the global timeline. In this thesis, we emphasize the importance of following the global timeline to avoid data leakage issues and effectively model temporal dynamics over time. Furthermore, we propose a retraining framework that not only rigorously considers the global timeline during training and learns user preferences from the most relevant interactions but also retains the long-term characteristics of users to enhance recommendation performance. Finally, we discuss potential areas for future research in the concluding chapter.
|
|---|