TITARL Tutorial

Abstraction: Data mining algorithm extracting temporal association rules from time sequences

Key points:

• TITARL extracts temporal association rules from symbolic time sequences and scalar time series; rules have confidence (probability of prediction), temporal inaccuracy (interval constraints), and support (fraction of events explained)
• Algorithm proceeds in four stages: create initial unit rules for all event-type pairs, divide rules to improve temporal accuracy, refine rules with Gaussian blur/threshold, and add conditions via entropy gain
• To separate interlaced rules in dense datasets, TITARL uses a co-occurrence matrix (2D histogram of predictions) rather than density distributions; splits rules via graph coloring (DSAT algorithm) or hierarchical matrix clustering
• Rules support negation, scalar time-series conditions, non-connected temporal constraints, and probability distributions on the head temporal constraint
• Minimum support threshold (typically 5%) prunes rules too sparse to be useful; body conditions selected by maximizing entropy gain
• Applications include medical monitoring (predicting patient instability from vital signs), complex system understanding, temporal planning, and system comparison

Connections: Titarl · Temporal Data Mining · Association Rules · Time Series

Source: http://mathieu.guillame-bert.com/titarl_tutorial