Attending to Customer Attention: A Novel Deep Learning Method for Leveraging Multimodal Online Reviews to Enhance Sales Prediction

Abstract

Review helpfulness has been measured commonly relying on quantitative indicators at the review level. Helpful reviews qualified by such simple indicators, however, may not necessarily yield accurate sales predictions, owing to the ever-evolving review information quality, customer demand, and product attributes. Positing that reviews with higher customer attention should be more influential to customers’ purchase intention and product sales, we propose to leverage customer attention to better realize the potential of multimodal reviews for sales prediction. We conceptualize customer attention at the holistic review set, review subset, individual review, and review element levels, respectively, and induce four indicators of customer attention, that is, timeliness, semantic diversity, voting-awareness, and varying multimodal interaction. We then propose a novel deep learning method, which incorporates these customer attention indicators using neural network attention mechanisms specifically designed for multimodal-review-based sales prediction. Empirical evaluation based on a large data set in a case study predicting hotel sales (specifically, monthly occupancy rate) shows that, in terms of both prediction performance and representation learning performance, our proposed method outperformed benchmarked state-of-the-art deep learning methods. As multimodal reviews become increasingly prevalent, this method serves as a tool for adequately leveraging such multimodal data to support business decision making.