Predicting sales in live streaming: An interpretable framework

Live streaming commerce has achieved great success in recent years. The multimedia factors play a predominant role in this promotion event. Focused on anchor-centric features extracted from multimedia data, our model stands out for its integration of explainable factors, employing a deep learning architecture with LSTM layers as encoder. Our key innovation lies in the incorporation of these anchor-centric features, which significantly enhance predictive accuracy. Specifically, our deep learning model showcases a remarkable 17.97% reduction in Mean Squared Error (MSE) compared to models lacking these features. In addition, the evaluation of the feature importance is conducted by adopting the Shapely value, which can help explain the contributions of the anchor-centric features to the sales during the live streaming commerce. Through interpretability summaries, we identify patterns highlighting the positive impact of various factors such as tone, attribute emphasis, and dynamic body language on sales. Besides, the other textual cues and postural cues also contribute to the promotion, although their contribution is relatively minor. This framework not only serves as a robust predictive tool for anchor performance but also offers trainable cues for the practitioners. By elucidating the significance of verbal strategies and behavioral cues, our model empowers stakeholders to optimize their strategies for more effective live streaming commerce engagements.