Personalized recommendation systems are crucial for online platforms like e-commerce and video/audio hosting services, improving user experience and boosting the revenue of the organizations. While traditional systems rely on explicit user-item ratings, recent studies favour implicit signals such as item clicks and dwell time on the website. This shift has led to two main approaches: session-aware and session-based recommendations.
Session-aware recommendations use historical interaction sequences to predict future interactions but struggle when user profiles are absent. On the contrary, session-based recommendations overcome this limitation by focusing solely on the current session's interactions.

Recent deep learning models in session-based recommendations differentiate between long-term and short-term user interests. Recognizing that users' interests are dynamic and evolve over time, even within a session, deep learning models differentiate between long-term and short-term interests. The long-term module models the entire session, capturing overarching user behaviour. In contrast, the short-term module focuses on a fixed set of recent items, reflecting immediate interests. In our first study, we attempt to enhance the modeling of short-term interests by dynamically selecting items related to the most recent item, and also incorporating item metadata like category and price, leading to an improved accuracy in real-world datasets.
Session-based recommendation algorithms traditionally focus on predicting user interests with high accuracy. While accuracy is important, it is also imperative to incorporate novelty and diversity of recommendations to improve user satisfaction and business outcomes. Recently, there is a growing emphasis on addressing the bias towards popular items ('short-head') in recommendations, which often overlooks less popular but relevant 'long-tail' items. To tackle this, in our second study, we introduce a novel neural architecture that enhances the performance of long-tail recommendations without compromising accuracy. This method encodes item popularity into embeddings and leverages collaborative information from sessions with similar diversity. Extensive testing on three real world datasets, namely, Yoochoose, Diginetica, and Cosmetics datasets demonstrated its effectiveness in increasing recommendation diversity while maintaining accuracy.

Our third study extends the research scope to session-aware recommendations, which leverage historical session data for a deeper understanding of users' overall purchasing behaviour. Recent studies in this context excel at identifying long-term trends but are less suited to predicting the next purchased item. Therefore, in our third study, we employ a two-tower deep learning architecture where one tower analyses past purchases for long-term preferences, while the other focuses on recent activities for immediate interests. This method effectively balances between consistent purchasing trends and short-term user behaviour. Our experiments with Walmart data demonstrated that this model significantly outperforms existing benchmarks, showcasing its ability to capture and predict user purchase interests effectively.

In summary, our research advances session-based recommendation systems by leveraging deep learning to mitigate biases and enhance the recommendation accuracy and diversity. These innovations cater to evolving user interests, setting new standards in e-commerce recommendation systems.