on 24-Jun-2022 (Fri)

As machine learning models become ubiquitous in our everyday lives, demand for explaining their predictions is growing [5, 16, 14]. In the context of behaviour prediction, we want to understand how previous consumer actions influence model predictions: How does order probability change when products are put into the cart? Does it decrease significantly if a consumer does not return to a webshop for two days? Answers to these questions are consumer-specific; they depend on the complete consumer history

In principle, one could evaluate the logistic regression model at every single time-step in the consumer history to determine the influence of individual events. However, this would involve the inefficient process of re-calculating features for every time-step. Calculations at timesteps t and t − 1 would be highly redundant: features at t represent the complete history until t and not only what happened in between t − 1 and t. Generally speaking, explaining the predictions of vector-based methods is more difficult than often as- sumed. This holds even for linear models like logistic regression. Features are often preprocessed, for example to binarize counts (Sec. 2). Furthermore, they are typically strongly correlated, making it troublesome to interpret individual coefficients [6]. Table 3 shows exemplary features weights in a logistic regression model used to predict order probabilities. If hundreds of features are utilized and are correlated and preprocessed, explaining the impact of consumer actions becomes a complex and confusing task

We have proposed an approach to apply RNNs to predict future consumer behavior in e-commerce. Consumer behavior is inherently sequential which makes RNNs a perfect fit. We are employing RNNs in production now which offers significant advantages over existing methods: reduced feature engineering; improved empirical performance; and better prediction explanations. In the future, predictions on the level of products and individual tastes will be in our focus, enabling sophisticated recommendation products. This will require richer input descriptions at individual time-steps. Likewise, more sophisticated RNN architectures will be promising for future research