Time series analysis using explainable AI

Abstract

In the last couple of years, great leaps have been made in the field of Machine Learning. Despite this, understanding how and why a machine learning model makes a decision is still a challenge faced by non-expert users, for which solutions are being actively developed. Moreover, studies on techniques which may be used to evaluate such solutions quantitatively are very scarce. In this research, Explainable AI techniques are used on LSTM model predictions for forecasting customer depositing data using a time-series dataset which was created for the purpose of this study. An index was also developed in order to quantitatively evaluate the quality and stability of such predictions. In order to achieve this, raw data was extracted from a customer transactional database and analysed in order to create a suitable time-series dataset. This time-series was then applied to an LSTM model as well as a Naive model and an ARIMA model for benchmark purposes. Results showed that the ARIMA model outperformed the LSTM model in most cases. LSTM predictions were generated using a Monte Carlo simulation in order to get measures on prediction confidence. LIME and SHAP explanations were generated for these predictions. The explanations were evaluated both quantitatively, by using the stability index created, as well was qualitatively through an evaluation by human domain experts. The quantitative evaluation performed concluded that SHAP explanations are generally more stable than LIME explanations and that there is no correlation between a prediction’s accuracy and the stability index. Through the qualitative evaluation, it was found that the stability index helped increase the trust of the domain experts in the explanations and predictions.