CLT: Cooperative Lottery Ticket Hypothesis in Live Streaming Sales Prediction

Abstract

Sales volume prediction, vital for strategic business planning, has always been a daunting task. With today’s data landscape brimming with diverse content from text to audio, and the rising influence of live streaming on sales data, the challenge becomes even more significant. The dense nature of live streaming data, combined with the computational challenges of deep learning, often results in slow and inefficient models. However, recent strides in the Lottery Ticket Hypothesis, especially in weight and dataset pruning across image and audio, present a promising avenue. This study pioneers the integration of this hypothesis into live-stream sales forecasting, introducing a novel approach that jointly prunes datasets and network weights. Concretely, we design a visual lottery ticket component (VLT) and audio lottery ticket component (ALT) for visual and audio network/data jointly pruning. Through rigorous experimentation and detailed ablation studies, we affirmed the potency of our algorithm, underscoring the practicality of pruning live-stream data. This paves the way for a robust and efficient framework tailored for future live-stream sales prediction.