Research & Papers

Researchers predict bus destinations from tap-in data alone

New model infers alightings from 838K tap-ins, beating baselines by 20%

Deep Dive

Researchers at [anonymized institution] developed a Hierarchical Bayesian Latent-Destination (HBLD) model to infer bus trip destinations from tap-in-only smart-card data, addressing the common issue of missing alighting records in transit systems.

The HBLD model combines network topology, time-of-day patterns, hourly weather data, and passenger travel history to predict trip destinations as probability distributions over feasible downstream stops. Trained on 838,305 bus tap-ins from Changzhou in May 2025, the model uses stochastic variational inference to fit parameters and outperformed the strongest baseline by 20% in destination prediction accuracy. A Bayesian personalization layer leverages prior card trips for frequent riders, reverting to a shared trip-level distribution when history is unavailable, improving predictions even for occasional riders.

Key Points
  • HBLD infers destinations from 838,305 tap-ins using a hierarchical Bayesian approach with passenger history and network topology
  • Model outperformed baselines by 20% and handles uncertainty in trip-chain evidence
  • Enables OD matrix generation for transit planning, scheduling, and resource allocation

Why It Matters

This could revolutionize transit planning by unlocking destination insights from existing tap-in data alone.

📬 Get the top 10 AI stories daily