Enhancing the prediction accuracy and robustness of collaborative filtering models by leveraging user rating credibility and temporal data

<p dir="ltr">A recommender system is a type of information filtering system that offers tailored suggestions for items that best suit a specific user. Recommender systems are indispensable for businesses, aiding in targeting products and services effectively. They guide individuals through a multitude of choices, facilitating decisions on purchases, music selection, and content consumption. Their utility spans across diverse domains, from playlist creation to product recommendations and content suggestions online.</p><p dir="ltr">Recommender systems commonly employ either collaborative filtering, content-based filtering, or a combination of both. Collaborative filtering analyzes a user’s past actions, such as purchases or ratings, along with similar decisions by other users to predict items of interest. Content-based filtering, on the other hand, uses specific attributes of items to recommend others with similar properties.</p><p dir="ltr">This research focuses on collaborative filtering, a popular technique for recommender systems. It’s favored for its reliance on user-item interaction data alone, avoiding the need for item attribute data, which can be challenging to gather, organize, and keep up-to-date. Essential to the success of collaborative filtering is the ability to deliver accurate recommendations. This domain, however, faces two notable challenges: the limited amount of research there is into user rating behaviour, which is key to uncovering the patterns that impact accuracy; and the tendency to overlook the temporal dynamics of user preferences.</p><p dir="ltr">In this research, our first focus was on understanding user rating behaviour to establish user credibility. We integrated credibility scores into a memory-based collaborative filtering model, the result was more precise and robust recommendations. User credibility was determined using past ratings, alongside demographic data and ontological semantics that measured the similarities between users and items. This approach addresses issues like data sparsity and cold starts. Plus, experiments conducted on real-world datasets, such as MovieLens and Yahoo! Movies, demonstrate a significant enhancement in recommendation quality compared to the state-of-art memory-based methods.</p><p dir="ltr">In the second part of this study, we turned to the dynamic nature of user preferences. User tastes evolve over time, and failing to account for this can lead to less effective recommendations and user dissatisfaction. To tackle this issue, we introduced a novel temporal collaborative filtering model based on a graph convolutional networks (GCN). This model incorporates both baseline signals, i.e., long term temporal signals, and transient temporal signals via relative and absolute time distance functions. Thus, the model is designed to capture changes in user preferences. By considering these temporal dynamics, this model provides more accurate and personalised recommendations. Experiments conducted on the MovieLens datasets showcase the superiority of this approach over static models, reducing the root mean square error (RMSE) by 4.7% and 1.7%, while also enhancing the F1_score@10 by 4.21% and 4.12% on the ML100k and ML1M datasets, respectively. In addition, the model achieved 6.6x to 15.8x speedup over pure static GCN collaborative filtering model.</p><p dir="ltr">In conclusion, this study makes substantial contributions to the field of collaborative filtering by investigating user rating behaviour and accounting for the temporal evolution of user preferences. These insights do not only enhance the accuracy and relevance of recommendations but also pave the way for future research in optimising recommendation systems. The findings underscore the importance of integrating user credibility and temporal factors into collaborative filtering models, offering a more robust and nuanced understanding of user behaviour in the context of recommendation systems.</p>