Explaining Taxi Demand Prediction Models Based on Feature Importance

The prediction of city-wide taxi demand is used to proactively relocate idle taxis. Often neural network-based models are applied to tackle this problem, which is difficult due to its multivariate input and output space. As these models are composed of multiple layers, their predictions become opaque. This opaqueness makes debugging, optimising, and using the models difficult. To address this, we propose the usage of eXplainable AI (XAI) - feature importance methods. In this paper, we build and train four city-wide taxi demand prediction models of commonly used neural network types on the New York City Yellow Taxi Trip data set. To explain their predictions, we select three existing XAI techniques - reduced Layerwise Relevance Propagation, Local Interpretable Model-agnostic Explanation, and Shapely Additive Explanations - and enable their usage on the specified problem. In addition, we propose a suite of five quantitative evaluation metrics suitable for explaining models that tackle regression problems with multivariate input and output space. Lastly, we compare the selected XAI techniques through the proposed evaluation metrics along four real-world scenarios.