Drifting towards new domain of ML

First research work in Recommendation System-> Cd-HRNN: Content Driven HRNN for Session-based Recommendation

While dealing with the User-Item interaction matrix in the context of a recommendation system, there is always a blurry region in capturing the patterns. The doubt comes with the notion of using item Ids (i.e each item is represented with a unique number); and unique numbers in the context of pattern recognition adds stringent restrictions as the algorithm expects the occurrence of exact same item-Id to consider it as a similar pattern. The idea doesn't make much sense in the real world and a supervised learning algorithm struggles when it finds sequels of the same show/ movie as completely different items. This restrains the algorithm in creating and capturing the patterns and ends up with poor performance. To overcome the stated issue, we proposed a mechanism of considering item description along with Item-Ids and applied it on Hierarchical Recurrent Neural Network (HRNN) and witnessed the improvements in the performance. HRNN is the derivative of RNN models that work with sequential data and captures multiple folds of user behavior (a). The user behavior in a particular session (b). The users' holistic behavior on the platform across all the time. HRNN does this with the help of two hierarchies of RNN units (to capture intra-session behavior + inter-session behavior ).

We, the team of recommendation systems in Sony Research India, recently proposed this work in the main conference of IJCNN-2023, Goldcoast Australia. If you find it interesting and meaningful to your domain, please have a look at our paper titled "Cd-HRNN: Content-Driven HRNN to Improve Session-Based Recommendation System".