ademyavuzz/algerian-forest-fire-ml

Paired classification and regression study on Algerian forest fire data. 0.967 accuracy with 0.998 ROC AUC on fire detection, 0.984 R2 on the Fire Weather Index. # Algerian Forest Fire ML Two paired machine learning studies on the Algerian Forest Fires dataset, sharing the same cleaning and feature selection pipeline but answering different questions: | Notebook | Task | Target | Best result | | :-- | :-- | :-- | :-- | | `AlgerianForestFireClassification.ipynb` | Binary classification | `Classes`, fire or no fire | Accuracy 0.967, ROC AUC 0.998 | | `AlgerianForestFireRegression.ipynb` | Regression | `FWI`, Fire Weather Index | R2 score 0.984 | ## Dataset 244 daily observations from two regions of Algeria, collected between June and September 2012. Each record combines meteorological readings with the components of the Canadian Fire Weather Index system. | Group | Columns | | :-- | :-- | | Date | `day`, `month`, `year` | | Weather | `Temperature`, `RH` relative humidity, `Ws` wind speed, `Rain` | | FWI system | `FFMC`, `DMC`, `DC`, `ISI`, `BUI`, `FWI` | | Label | `Classes`, fire or not fire | ## Preprocessing Both notebooks apply the same preparation: 1. Drop four malformed rows that carry the region header inside the data body rather than valid observations. 2. Cast the FWI component columns to float, since they load as object type because of trailing whitespace in the source file. 3. Map the `Classes` label to a binary target with `not fire` as 0 and `fire` as 1. 4. Split into training and test sets at 75/25 with a fixed random state. 5. Remove multicollinear features. A helper walks the correlation matrix of the training split and drops any column whose absolute correlation with an earlier column exceeds the threshold. The threshold is computed on the training split only, then the same columns are dropped from the test split so no test information leaks into feature selection. The FWI system components are correlated by construction, which is why this step matters here rather than being a routine precaution. ## Classification study Logistic regression, evaluated at three levels of tuning. | Stage | Accuracy | N …