Adaptive Federated Stream Itemset Mining under Drift and Unequal Client Data Quality

Abstract

Distributed transaction streams can change over time while client data quality and arrival volumes differ. We present AdaFed-TPM, a federated stream-itemset design with windowed counting, local change alarms, digest suppression, and consensus-based weighting. In the evaluated realization, clients mine unordered itemsets in transaction windows; temporal order and episode constraints are not modelled. A reported 10,000-transaction UCI Adult-derived stream contains six planted three-itemset groups, common drift times at transactions 3333 and 6666, and six clients with item-corruption probabilities between 2 and 20 percent. Across ten seeds, mean active-pattern recall is 0.988 versus 0.869 for periodic retraining (reported paired Wilcoxon W=0,p=0.0020), with 30.0 percent fewer uplink bytes than the fixed-window federation. The fixed-window recall of 0.983 is not significantly different; no equivalence test was performed. False re-discovery is higher, at 0.115 versus 0.033 for periodic retraining, and first-drift adaptation is slower than fixed-window refresh. Nine of ten ablation seeds have identical checkpoint outputs, leaving no reliable component-specific benefit established for quality weighting or event synchronization. Geometric decay, asynchronous site-specific drift, a stability floor, and privacy are untested design extensions.