seq2vec: Analyzing sequential data using multi-rank embedding vectors

Abstract

The fields of machine learning and deep learning witnessed significant advances in the past few decades. However, progress in the development of methods to analyze sequential data (e.g. sensor data and event data) has been relatively stagnant. There are six major challenges encountered when conducting sequential-data analysis: high dimensionality, time variance, categorical variables, interpretability, data integration, and privacy. We propose a new multi-rank embedding (MRE) method for sequential data to address these problems. The first rank (row) of the MRE contains the compressed temporal information of the original data, and each row thereafter represents an embedding vector of a specific block (time span) of the original data. Our experiment results indicate that analysis of seq2vec representations can deliver similar performances to those obtained using original raw data for the purposes of clustering household electricity usage patterns, classifying human activities, and forecasting crop yield data at a greatly reduced storage cost. Furthermore, the embedded data does not contain sensitive personal information and can be shared without critical privacy concerns.