Business Problem
When an incident is opened, everyone asks: “How long until it is fixed?” Today the answer is a guess. Support, status pages and managers need a realistic ETA, and on-call teams need an early warning when an incident will be long.
Goal
Build a machine learning system that predicts the time to resolve (in minutes) of a new incident, using only information known at the moment the incident is opened.
Dataset
Here is the dataset that you can use to train the ML Regression Model. Conduct a experiment using ML regression models and build a Jupyter Notebook showing the ML Training and Evaluation and report the best model and params.
Regression Problem
The model must predict a continuous number (ETA in minutes), not a class.
Technologies and Tools
- Data analysis: Pandas, NumPy, Matplotlib
- Models: LinearRegression, DecisionTreeRegressor, RandomForestRegressor (optional: Ridge, GradientBoostingRegressor)
- Saving: joblib
- Interface: Jupyter Notebook
Tasks
- Explore. Shape, info, describe, missing values, target distribution (mean vs median). List the data problems you find.
- Clean. Duplicates, inconsistent text, date formats, impossible target values, missing values (fill using the training set only).
- Allowed features. Decide which columns are known when the incident is opened. Reject the others with evidence.
- Features. Extract useful features from the date and time, encode text columns, and add at least one idea of your own.
- Baseline and split. Mean and median baselines; 80/20 split (random_state=42).
- Train and compare. Linear Regression, Decision Tree, Random Forest. Tune hyper-parameters on the training set. Try predicting log of the target. Check overfitting.
- Evaluate. R2, MAE, median absolute error. Plot predicted vs actual and the residuals. Find where the model is weakest.
- Inference. Save the best model and test the inference loading the saved model artifact.
Notes
Submit: Once the ML system is ready and tested, share the GitHub link (README, notebook, code, predictions.csv) and comment below which model is the best, with its R2, MAE and median absolute error.
