Interpretable Time Series Autoregression for Periodicity Quantification

Abstract: Time series autoregression is a classical statistical model for capturing auto-correlations and identifying temporal patterns such as periodicity and seasonality. In this work, we propose a novel sparse autoregression framework from an interpretable machine learning perspective and the model interpretability for periodicity quantification is reinforced by $\ell_0$-norm induced sparsity constraints. On the time-varying time series data, we reformulate the sparse autoregression and convert the involved optimization problem into a mixed-integer optimization (MIO). To accelerate it, we develop a subspace pursuit based decision variable pruning (DVP) strategy to reduce the search space. On the multidimensional time series that involves complicated spatial and temporal dimensions, we propose a spatially- and time-varying sparse autoregression model and resolve the corresponding MIO problem by developing a two-stage optimization scheme. In particular, the proposed scheme makes the model scalable to large problems even with millions of decision variables. Empirically, we conduct extensive experiments to evaluate the proposed models on real-world time series data. First, we demonstrate that the MIO solver can be drastically accelerated through the DVP strategy, while maintaining the same solution quality as a full MIO solver. Applying the time-varying sparse autoregression model to ridesharing trip data, we uncover both daily and weekly periodicities and reveal long-term changes in regularity of human mobility. Second, we demonstrate the spatial patterns of yearly seasonality in climate variable time series such as temperature and precipitation across the past four decades, and our model allows to discover dynamic climate patterns and identify climate phenomena such as El Nino in sea surface temperature.