LEMF: an end-to-end model for intention recognition in multivariate time with missing data

Abstract

The processing and application of time series are widespread, including tasks like weather forecasting, traffic flow prediction and intention recognition. However, in reality, missing data often occurs due to target occlusion or sensor failures. Many deep learning models are designed for uniformly sampled complete data and cannot be directly applied to scenarios with missing values. Traditional data preprocessing methods, such as imputation and interpolation, introduce additional noise. To address these challenges, we propose an end-to-end model with Learnable Embedding and capture Multidimensional Features (LEMF). LEMF can directly handle real-world time series with missing values. We utilize the LE module to extract richer temporal information, compensating for the limitations of missing data. The MF module can extract features related to the relationships between variables. We leverage these hidden representations for intention recognition, which is the time series classification task. We thoroughly evaluate our model on a self-constructed intention dataset. Compared to baseline model, the LEMF model achieved an average of 10% higher accuracy at each missing ratio. Additionally, we validate the model’s generalization capabilities on two real-world datasets. Our model also shows optimal or suboptimal performance.