Kaggle Competition | Porto Seguro’s Safe Driver Prediction

From the competition homepage.

In this competition, you’re challenged to build a model that predicts the probability that a driver will initiate an auto insurance claim in the next year. While Porto Seguro has used machine learning for the past 20 years, they’re looking to Kaggle’s machine learning community to explore new, more powerful methods. A more accurate prediction will allow them to further tailor their prices, and hopefully make auto insurance coverage more accessible to more drivers.

Data Description

In the train and test data, features that belong to similar groupings are tagged as such in the feature names (e.g., ind, reg, car, calc). In addition, feature names include the postfix bin to indicate binary features and cat to indicate categorical features. Features without these designations are either continuous or ordinal. Values of -1 indicate that the feature was missing from the observation. The target columns signifies whether or not a claim was filed for that policy holder.

Notebook Content

iPython Notebook here

• About Missing Data
• Drop Redundant Features & Replace Missing Data
• Data Preparation
• Feature Selection (Random Forest Classifier)
• Train A Model (Logistic Regression)
• Predict & Output
• Kaggle Score

Reference and Future Work

For the random forest classifier part in this project , I referenced codes from Prof. Ravi Shroff 's Machine Learning class at CUSP.

I think there are two ways to improve this project:

First of all, when creating dummy variables for categorical features, my code used a lot of computing resources and create sparse matrix due to the 104 unique values feature ps_car_11_cat has. Although the random forest classifier reduced dimensionality later in the project, I am still searching for better ways to convert categorical features at this step.

Secondly, try more prediction algorithms, obviously. I can add more models in the future working on this project.

Analysis Highlights

• Missing Data

• Correlation between features

• Remaining Feature Data Distribution

• Use Random Forest Classifier, the top 20 features contributing to a claim to be filed (target = 1)

	feature_select	feature_importance
1	ps_car_13	0.098262
2	ps_reg_03	0.093183
3	ps_car_14	0.057226
4	ps_ind_03	0.051262
5	ps_ind_15	0.051083
6	ps_reg_02	0.049227
7	ps_ind_01	0.036896
8	ps_car_15	0.036607
9	ps_reg_01	0.036353
10	ps_car_12	0.027303
11	ps_car_11	0.012646
12	ps_car_01_cat_11.0	0.008632
13	ps_car_09_cat_2.0	0.008630
14	ps_ind_04_cat_0.0	0.008376
15	ps_ind_02_cat_1.0	0.008298
16	ps_ind_04_cat_1.0	0.008273
17	ps_ind_02_cat_2.0	0.008004
18	ps_ind_16_bin	0.008000
19	ps_car_09_cat_0.0	0.007793
20	ps_car_01_cat_7.0	0.007719

• AUC score (produce probabilistic predictions) on training dataset

  0.59645323858

• Accuracy score (predict the class)

  0.963615286396

• Train A Model: Logistic Regression to Predict Output sample (first 10 rows)

id	target
0	0.027729
1	0.032712
2	0.022809
3	0.020033
4	0.035388
5	0.030271
6	0.019810
8	0.019596
10	0.066153
11	0.043696

• Submit to Kaggle Competition

  Current normalized Gini score: 0.241