1.1 Applied Etiq + ChatGPT assisted preprocessing to automatically detect feature issues and generate correction rules.\n1.2 Removed the previous deduplication step used after merging tables — aggregation was needed, and deduplication would incorrectly discard valuable records.\n1.3 Manually corrected time formats based on distribution inspection and actual value patterns.\n1.4 Used Sweetviz for EDA to visualize distributions and detect anomalies.\n1.5 For Light, Weather, and other categorical attributes, performed custom normalization and format correction based on observed values.
\n\"other\" class to improve model robustness.2.1 The main framework used was AutoGluon, which provided strong baseline performance with minimal manual tuning.
\nThe public leaderboard and local cross-validation were heavily misaligned, causing significant difficulty during model development. After post-competition analysis, I found that local CV was more reliable.\nPost-competition, I combined OOF files from my notebook and open-source notebook, and then applied GPU-accelerated Stacking, which led to a substantial improvement in performance.
", "comments": [], "comment_count": 0, "messages": [ { "id": 3344363, "postDate": "2025-11-22T13:58:45.690Z", "content": "1.1 Applied Etiq + ChatGPT assisted preprocessing to automatically detect feature issues and generate correction rules.\n1.2 Removed the previous deduplication step used after merging tables — aggregation was needed, and deduplication would incorrectly discard valuable records.\n1.3 Manually corrected time formats based on distribution inspection and actual value patterns.\n1.4 Used Sweetviz for EDA to visualize distributions and detect anomalies.\n1.5 For Light, Weather, and other categorical attributes, performed custom normalization and format correction based on observed values.
\n\"other\" class to improve model robustness.2.1 The main framework used was AutoGluon, which provided strong baseline performance with minimal manual tuning.
\nThe public leaderboard and local cross-validation were heavily misaligned, causing significant difficulty during model development. After post-competition analysis, I found that local CV was more reliable.\nPost-competition, I combined OOF files from my notebook and open-source notebook, and then applied GPU-accelerated Stacking, which led to a substantial improvement in performance.
", "rawMarkdown": "\n\n#### **1. Data Preprocessing**\n\n1.1 **Applied Etiq + ChatGPT assisted preprocessing** to automatically detect feature issues and generate correction rules.\n1.2 Removed the previous deduplication step used after merging tables — **aggregation was needed**, and deduplication would incorrectly discard valuable records.\n1.3 **Manually corrected time formats** based on distribution inspection and actual value patterns.\n1.4 Used **Sweetviz for EDA** to visualize distributions and detect anomalies.\n1.5 For **Light, Weather, and other categorical attributes**, performed **custom normalization** and format correction based on observed values.\n1.6 Unified **rare categories** and categories **absent in the test set** into an `\"other\"` class to improve model robustness.\n\n---\n\n#### **2. Modeling**\n\n2.1 The main framework used was **AutoGluon**, which provided strong baseline performance with minimal manual tuning.\n2.2 Experimented with **automated feature processing and Stacking using OOF** (out-of-fold) predictions.\n\n---\n\n### **Notes & Observations**\n\nThe **public leaderboard and local cross-validation were heavily misaligned**, causing significant difficulty during model development. After post-competition analysis, I found that **local CV was more reliable**.\n\nPost-competition, I combined **OOF files from my notebook and open-source notebook**, and then applied **GPU-accelerated Stacking**, which led to a substantial improvement in performance.\n ", "votes": 1 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T16:27:27Z", "competition_id": 117343, "competition_slug": "etiq-roadsense", "team_id": 14590933, "solution_url": "/competitions/etiq-roadsense/writeups/solution-overview-44th-place" }