A Data-Driven Model Selection Approach toSpatio-Temporal Prediction

Spatio-temporal  Predictive  Queries  encompass  a  spatio-temporalconstraint, defining a region, a target variable, and an evaluation metric.  Theoutput of such queries presents the future values for the target variable computedby predictive models at each point of the spatio-temporal region. Unfortunately,especially for large spatio-temporal domains with millions of points, trainingtemporal models at each spatial domain point is prohibitive.  In this work, wepropose a data-driven approach for selecting pre-trained temporal models tobe  applied  at  each  query  point.   The  chosen  approach  applies  a  model  to  apoint according to the training and input time series similarity.  The approachavoids training a different model for each domain point, saving model trainingtime.  Moreover, it provides a technique to decide on the best-trained model tobe applied to a point for prediction.  In order to assess the applicability of theproposed strategy, we evaluate a case study for temperature forecasting usinghistorical data and auto-regressive models.  Computational experiments showthat the proposed approach, compared to the baseline, achieves equivalent pre-dictive performance using a composition of pre-trained models at a fraction ofthe total computational cost.