HT-ConvFormer: A Hierarchical Framework for Multi-scale Temporal Modeling in Table Tennis Analytics

Predicting player attributes from wearable sensor data is a challenging task, particularly when faced with limited datasets and complex, hierarchical time-series signals. This paper addresses the problem of identifying table tennis player characteristics (gender, dominant hand, experience, and skill level) from raw inertial measurement unit (IMU) data recorded during strokes. The high dimensionality and multi-scale nature of the data comprises fine-grained intra-stroke dynamics and coarse-grained inter-stroke patterns, which poses significant risks of overfitting and instability. To overcome these challenges, we propose the Hierarchical Temporal ConvFormer (HT-ConvFormer), a two-stage framework that models these temporal levels separately. The first stage, CNNFormer, extracts features from individual strokes, while the second stage, Refiner, models dependencies across a sequence of strokes. Our methodology is further enhanced by a progressive task simplification strategy to stabilize learning. To maximize performance, we introduce an iterative self-distillation scheme where the knowledge from the macro-level Refiner is progressively transferred to the micro-level CNNFormer. Combined with a balanced data processing pipeline and ensemble learning, our approach robustly learns discriminative representations from limited, imbalanced sensor data.