RAPID: Reinforcement Adaptive Policy-tuned Insight Determination for Dynamic User Feedback Analysis

Financial insights - such as trends, anomalies, forecasts, key performance indicatiors (KPIs) and their contributors - can be valuable in growing a business. Recently, significant amount of research has gone into autonomously extracting relevant insights from financial data. However, not all types of insights are equally important or interesting for a company at a given time. Moreover the user interests in insight types can also evolve over time. To this end, this paper introduces a novel policy gradient-based reinforcement learning method to identify a personalized set of relevant business insights on multiple Key Performance Indicators (KPIs) given user feedback. Additionally, we leverage a Kolmogorov Smirnov Windowing test to capture changes in user preferences dynamically, enabling users to access their current most important features of their business performance from multiple perspectives at any given time. A distinct feature of our model is it’s adaptability to nonstationarity in user preferences, as well as the ability to identify the “most interesting subset” of insights for a particular company, distinguishing our approach from more static, traditional models. Experimental results demonstrate rapid convergence and accuracy of the approach in identifying the most relevant insight types as well as fast adaptation to shifts in user interest, in the presence of sparse data, compared to the current literature. Beyond this specific application, our framework can be effortlessly integrated into any system with a feedback mechanism, highlighting its versatility and adaptability.