Interpretable approach to re-ranking system

Abstract

Recommendation systems plays a crucial role inthedomainofArtificialIntelligence,with the expansion of internet widespread use of smart devices, the requirement to create advanced ranking and re-ranking recommendation systems have become a necessity for businesses to maintain their competitiveness. Interpretability of any machine learning system has now become a trend as well as a legal requirement in many industriesthatuse machine learning models for analytical purposes. Recent advancements in interpretable machine learning systems including ranking systems shows significant potential in this domain. In this research, we developed an intrinsically interpretable re-ranking model based on Tensorflow Ranking Generalized Additive Model (TFR GAM) architecture to score previously recommended offers. This model was compared against an industry leading Neural Collaborative Filtering model which is a black box Model. The evaluation metrics used to compare two models were Normalized Discounted Cumulative Gain (NDCG), Hit Ratio and Precision. These were used at top 2 and top 3 which gave two variations for each metric. Even though performances of these metrics are between 40%-50% for all matrices, all metrics show that both the models performed in similar mannereventhoughtheTFRGAMmodelhadaninterpretablemodelarchitecture.Overall, in this study metrics suggest that for re-ranking tasks we can use the TFR GAM model which provides similar results to that of an industry leading NCF Model. The study also suggest future research directions to improve the model's performance,suchhastryingout different combinations for the number of hidden layers, activation functions, and regularization techniques.the studyconcludesthatthereisgreaterpotentialforfurtherwork in this area and continued research development could lead to significant advancement in thisdomainofinterpretablemodels.