Multi-Patching: Life-Log Classification with the Reconstructed Representation of Multivariate Time Series

Understanding human beings requires analyzing human behavior with their life-log data. The life-log data is typically represented as multivariate time series data. This data is characterized by its extensive volume and the unforeseen emergence of missing values. These characteristics make the analysis challenging. This paper tackles these difficulties by generating reconstructed representations of the life-log data to utilize an LSTM autoencoder. Our method captures and compresses short-term patterns occurring within a single cycle. These reconstructed life-log data are fed into multivariate time series classification (MTSC) backbone models. This method not only improves performance but also efficiently manages memory usage. Experimental results showed that memory usage was reduced by 88.01 % while performance increased by 1.73 %. Achieving a weighted sum of F1-score of 5.91 on the test dataset confirmed the model's effectiveness.