רוברט מושקוביץ'

Frequent Time Intervals-Related Pattern (TIRP) mining from symbolic time-interval series has attracted increasing research interest in recent decades, leading to significant algorithmic advances. By using temporal abstraction to transform heterogeneous multivariate temporal data, whether sampled regularly or irregularly, into symbolic time-interval series, TIRP discovery can be applied to diverse types of temporal variables. This makes TIRP mining broadly applicable to real-world data, for tasks such as temporal knowledge discovery, classification, and continuous event prediction, while yielding inherently explainable patterns. However, existing approaches typically index first 2-sized TIRPs and rely on exhaustive candidate generation to extend them, incorporating redundancy and inefficiency in the mining process. We introduce BITIRP, a novel and efficient algorithm for complete frequent TIRP mining. BITIRP incrementally grows patterns from 1-sized TIRPs by using the newly introduced concepts of a TIRP's prefix and suffix. In each iteration, the suffixes of the current frequent TIRPs are matched with corresponding indexed prefixes, and the resulting Join operation creates new TIRPs in a fully data-driven manner. Efficient pruning strategies and dedicated data structures further reduce redundant computation and avoid repetitive support counting throughout the mining process. Extensive experiments on twelve benchmark datasets show that BITIRP achieves substantial runtime improvements while maintaining comparable memory consumption relative to state-of-the-art TIRP mining algorithms.