Shapelet Extraction Using Visibility Graphs

Time-series shapelets give transparent, example-based explanations but classic discovery is slow and brittle. I present a scalable, interpretable pipeline that discovers shapelets from visibility-graph (NVG/HVG) encodings of sliding windows. Each window becomes a compact 10-D graph descriptor; CLARA-style k-medoids then yields medoid subsequences as prototypes. Series are embedded via minimum z-normalized distances to these prototypes and classified with lightweight models (linear SVMs, tree ensembles). The design supports multivariate data. Across UCR/UEA benchmarks, the method achieves competitive accuracy while remaining auditable: predictions trace back to real subsequences. On GunPoint, the top shapelets recover the known “overshooting” motif, providing qualitative validation. Parallel VG extraction, CLARA subsampling, and batched transforms deliver strong runtime efficiency versus time-contracted baselines.