SS
Back to Projects

Gridlock

Developed a traffic-demand prediction solution for the GridLock hackathon, focusing on distribution shift, model specialization, validation design, and RoadType-aware ensemble routing.

PythonXGBoostCatBoostLightGBM

SYSTEM ARCHITECTURE

SYSTEM ARCHITECTUREGrid-based application with persistent dataUser InputGrid interactionGrid LogicState + processingSQLitePersistent stateSQLData operationsResultUpdated grid stateGrid state is processed by the application logic with SQLite-backed persistence and SQL data operations.

Grid state is processed by the application logic with SQLite-backed persistence and SQL data operations.

Problem

Traffic demand varied strongly across time, geography, and road types, while the leaderboard distribution differed from the dominant patterns in the training data.

Architecture

Feature Set → CatBoost + XGBoost + LightGBM → RoadType Gate → domain-specific prediction blend.

Challenges

  • Detected a mismatch between conventional validation and leaderboard performance
  • Analyzed model errors across hour and RoadType
  • Tested and rejected several approaches that improved local validation but hurt target performance
  • Designed a RoadType-aware routing strategy for the final ensemble

Benchmarks

The repository reports a final Gate B leaderboard score of 91.946, with a post-mortem Gate B test MAE of 0.02996.

Lessons Learned

  • Validation design can matter more than model complexity
  • Different gradient-boosting models can specialize in different regimes
  • Domain information can be used as an ensemble routing signal
  • Failed experiments can reveal the real source of distribution shift