finished on 20th, May 2015
This is the code for Kaggle's 101 game Titanic Survival Prediction. I used logistic regression model with some simple feature engineering. The final score got 0.81340 which ranked 169/3084 at my final submission.
Missing age seldom appears in a survivor's data, so adding a binary feature "AgeIsNa" is necessary. The missing ages are then filled with random values between age_min and age_max.
I also tried to fill missing ages using a random forest but in vain.
Name title may reveal a person's information. This may help decide whether he/she has the chance to get into lifeboat. Using a categorical variable "Title" to record this.
The price's distribution is very skewed. Using Log(fare) can alleviate the spike near low ticket fare, making it fit well to a linear model.
When "Parch" (parent and child number) + "SibSp" (sibling and spouse number) equals 1,2 or 3, a person tend more likely to survive. So a variable "FamilySize" helps.
Make the continuous variables into discrete ones can help logistic models predict well. This information is recorded in variable AgeCat.
The magnitude in a categorical variable like "Title" is meaningless, so dumminize it.
Logistic Regression: 0.81340
{'penalty': 'L1', 'classifier__C': 10, 'classifier__tol': 0.001}
Support Vector Classifier: 0.79349
{'kernel': 'linear', 'C': 1, 'tol': 0.0001}
Random Forest Classifier: 0.78469
{'n_estimators': 200, 'n_jobs': '-1'}