We propose a novel time encoding approach that learns effective reference points and incorporates temporal and intervariable dependencies to enhance classification performance on irregular multivariate time series.