Stochastic Event Prediction via Temporal Motif Transitions

Researchers İbrahim Bahadır Altun and Ahmet Erdem Sarıyüce have introduced STEP (STochastic Event Predictor), a novel AI framework that fundamentally changes how we predict events in dynamic networks. Unlike traditional methods that treat temporal link prediction as simple binary classification, STEP reformulates it as a sequential forecasting problem in continuous time. The core innovation lies in modeling event dynamics through discrete temporal motif transitions governed by Poisson processes, maintaining evolving sets of open motif instances that update as new interactions arrive.

At each prediction step, STEP decides whether to initiate a new temporal motif or extend an existing one, selecting the most probable event through Bayesian scoring of temporal likelihoods and structural priors. The framework also generates compact, temporal motif-based feature vectors that can enhance existing temporal graph neural network outputs without requiring architectural changes. In experiments across five real-world datasets, STEP demonstrated remarkable performance gains—achieving up to 21% better average precision in classification tasks and reaching 0.99 precision in next-k sequential forecasting, all while maintaining consistently lower runtime than competing motif-aware methods.

The practical implications are significant for domains where understanding interaction sequences matters. In social networks, STEP could predict friendship formations or information cascades; in finance, it could forecast transaction patterns; in biology, it might model protein interactions. The framework's ability to handle both evolving topology and temporal ordering while producing interpretable motif-based features represents a substantial advance over current state-of-the-art approaches that often discard the sequential nature of real-world interactions.