Design and performance evaluation of multi-task transfer learning model for time series data

This paper proposes a dynamic relationship aware temporal multi task transfer learning framework (Dyna MTL) to address the issue of reduced generalization performance caused by domain shift and task conflicts in temporal data prediction. This framework achieves adaptive decoupling of temporal trends, seasons, and residual components through discrete wavelet transform, and introduces a temporal aware MMD variant (T-MMD) for cross domain feature alignment, effectively alleviating domain shift. Furthermore, the model uses graph neural network (GNN) to construct a dynamic task relationship graph, and uses attention mechanism to adjust the task weight in real time to improve the efficiency of multi-task collaboration. Experiments on 13 tasks in three categories: industrial fault prediction, medical risk early warning and financial fraud detection show that Dyna-MTL achieves an average AUC of 87.8%, which is significantly better than the baseline methods such as LSTM, DANN and MTL-MMOE, and has lower computational cost and stronger cross-domain adaptability.