Overall plan68 / 68 checkpoints
CurrentCatBoost identity/spatial-grid within-slot bag scored 0.8304 and was observed at public rank #1
- Features fully examined
- 36 / 36
Physical hierarchy screens retained the granular features - DrivenData submissions
- 21Total
- Best score
- 0.8304CatBoost identity/spatial-grid within-slot bag
- Target score
- 0.8260Stretch
- Used today (UTC)
- 3 / 3
Public leaderboard result
Observed 5 September 2026 · rank #1 · score 0.8304 · submission 321154
Now
Preserve the fixed portfolio and stop adaptive tuning.
Recorded: the fixed slate scored 0.8302, 0.8304 and 0.8303; the CatBoost identity/spatial-grid within-slot bag, submission 321154, was observed at rank #1.
- Preserve all three exact validated CSVs, the manifest, component weights and SHA-256 values.
- Keep the labelled-only reconstruction evidence and historical local evidence beside the public scores.
- Retain the locked submission order, IDs and unchanged score record.
- Record private-leaderboard performance when it becomes available.
Boundary: rank #1 is a time-specific observation; the three recorded scores and validated artefacts are the durable evidence. The image alongside is explicitly the earlier rank-2 milestone.
Next: move attention back to assessed course work rather than tune against public feedback.
Course workflow
Ten course steps. Iteration is expected; checkpoints advance only when evidence exists.
DoneActiveQueued
| Course step | Status | Progress | Evidence or next gate |
|---|---|---|---|
| Step 01Define the goal & scope | Done | 3 / 3 | Problem, users, metric, competition rules and data boundary recorded. |
| Step 02Gather the data | Done | 2 / 2 | Training features, labels and test inputs are present; source structure and column policy are recorded. |
| Step 03Explore the data | Done | 4 / 4 | Maintained overall report and 117 predictor notebooks cover structure, distributions, limitations, relationships and feature families. |
| Step 04Clean & preprocess the data | Done | 3 / 3 | Guarded cleanup and fold-fitted preprocessing are implemented and have run inside a complete pipeline across all five development folds. |
| Step 05Select & engineer features | Done | 36 / 36audit · separate metric | All candidates are audited; the final synthesis retains accepted features plus the supported identity, frequency and spatial-height representations. |
| Step 06Define the ML task | Done | 4 / 4 | Multiclass classification, accuracy, class imbalance, label integrity and exact feature/label ID alignment are recorded. |
| Step 07Partition the data | Done | 3 / 3 | A stratified 20% local test set and five development folds are frozen with seeds and fingerprints; the majority reference is recorded. |
| Step 08Select & train candidate methods | Done | 3 / 3 | Eleven classifier families are reproducibly evaluated; seven bounded MLP configurations and 84 no-refit ANN blends produced no promotion. |
| Step 09Evaluate & interpret results | Done | 3 / 3reference + 20 workflows | The historical seed-20260824 deep archive retained 81.898% development and 81.397% used-local-test evidence; fresh all-label OOF reconstruction admitted two conservative within-slot bags and their fixed combination as exploratory candidates. |
| Step 10Deploy & iterate | Done | 3 / 3 | The fixed three-entry slate scored 83.02%, 83.04% and 83.03%; submission 321154 established the 83.04% public best and was observed at rank #1. |
- Training rows
- 59,400
- Test rows
- 14,850
- Candidate predictors
- 36
- Target classes
- 3
- Metric
- Accuracy
- Locally evaluated
- 21