Time2Vec for Time Series features encoding
Learn a valuable representation of time for your Machine Learning Model
Time is golden information in every Machine Learning problem which engages Time Series. As Data Scientists, we must do our best to extract time patterns and make our data speak themself. During preprocessing, common procedures are for example standardization (stationarity check, autocorrelation removal…), creating and encoding categorical time features (day, week, month, season…), manual feature engineering (Fourier transformations…). Not always our efforts are repaid because, during modelization, our selected model can fail to treat time itself properly as a feature.
In this post, I try to reproduce the approach proposed in the paper ‘Time2Vec: Learning a Vector Representation of Time’, which final scope is to develop a general-purpose model-agnostic representation for time that can be potentially used in any architecture (I adapt this solution developing a Neural Network in Keras). The authors don’t want to propose a new model for time series analysis, but instead, their goal is to provide a representation of time, in form of vector embedding, in order to automatize the feature engineering process and model time in a better way.
