A Evaluation-Based Large Model for Multi-Step Behavior Prediction in Intelligent Consumer Electronics

Behavior prediction is a fundamental task in intelligent consumer electronics. Its goal is to predict potential future behaviors based on known behaviors, playing a crucial role in understanding user intentions and optimizing device performance. However, existing methods lack effective evaluation measures, causing models to overly rely on local features, which restricts them to predicting only the next immediate behavior, thereby limiting their ability to make long-term future inferences. To address these issues, we propose an a evaluation-based large model (ELM) for multi-step behavior prediction. Our approach is to enhance the model’s multi-step reasoning ability by evaluating the results it generates iteratively. Specifically, by generating sub-behaviors through a large model and constructing new sequences of behaviors, we can extend the range of future behaviors. We then designed a chain-level evaluator that iteratively assesses the chain-level representation of each generated behavior sequence using a large model to evaluate the authenticity of the behavior sequence. Additionally, to prevent the large model from getting trapped in a localized understanding of the behavior sequence, we enhanced the model’s chain-level perception ability through two human-like learning strategies. Experimental results on four real-world datasets show that our approach can generate more reliable sub-behavior sequences.