Motif Discovery and Classification Using Multi-Feature Chainlets
Makaleler > Tam Makale · 2026
Yazarlar: Vasfi Tataroglu, Gurhan Gunduz, Hasan BULUT
IEEE Access, Uluslararası, Wos, 2026, cilt 14, ss 45338-45353
DOI: 10.1109/access.2026.3674875
Özet
Bitcoin’s 24/7 trading generates recurring intraday price patterns, and public blockchains enable time-aligned analysis of transaction activity. We define bullish and bearish price motifs within a fixed 24-hour windowand test whether the on-chain transaction-graph signals observed before a motif begins can distinguish these patterns. Transaction activity is summarized using chainlet representations on a fixed input–output grid, enriched with economic content through fee- and amount-weighted chainlets. To prevent label leakage, all features are extracted only from the left (pre-motif) window. We also introduce FOA, a ratio-based measure that combines transaction fees and transferred amounts (fee/amount) to capture the fee pressure relative to the moved value. Experiments on 10-minute blocks from 2017 to 2024 were conducted to evaluate scaling, dimensionality reduction (including principal component analysis), and 11 classical machine learning classifiers. Seasonal analyses show that motif frequencies cluster in specific periods. Overall, the pre-motif on-chain structure and economic signals are informative for motif discrimination, with the highest accuracy achieved by Fee + PCA + ExtraTrees.