Feature Extraction of CNN-GRU based Multivariate Time Series Data for Regional Clustering

Time-series clustering is usually done through statistical analysis, which does not fully reflect the data’s characteristics. With multivariate data it is harder still, because each variable contributes differently.

• Proposes a network that extracts per-variable features with CNN and spending trend over time with GRU
• Preprocessing removes noise with a moving average and normalizes per-industry spending to 0–1 before generating sub-sequences
• Clusters regions with similar per-industry spending trends using two years of real card data

What separates this from existing clustering is treating the trend over time as a sequence, not just the amount.