MAML-TSC: Model-Agnostic Meta-Learning for Time Series Classification

User portrait technology, involves analyzing historical usage records of household electronic devices, is evolving quickly. Traditional methods using LSTM or CNN for time series classification face challenges in achieving high accuracy with limited data. To address this, our study introduces MAML-TSC, which combines a dual-channel Transformer with model-agnostic meta-learning (MAML). This model features a dual-channel Transformer that leverages a multi-head self-attention mechanism to extract information from both time and channel axes effectively. By incorporating MAML, the model harnesses knowledge transfer to optimize initialization parameters for diverse training tasks, enhancing performance with fewer data samples. Experimental outcomes show MAML-TSC achieving 90% accuracy in the meta-learning phase and maintaining over 80% after 10 iterations, outperforming the DeepConvLSTM baseline by 12.5% and surpassing most current deep-learning approaches.