Description: The main purpose of this project is to classify the category of crime based on Location and Time at which crime was commited .
The steps followed are as follows:
- Step-1: Download dataset from https://www.kaggle.com/c/sf-crime/data, store it in dataframe.
- Step-2: Create a new dataframe containing Latitude, Longtitude, Category, Hour, Week, Month and Year of crime.
- Step-3: Create a new dataframe X by dropping Category. Create a variable Y containing the category of crime.
- Step-4: Split the data(X_train, y_train and X_test, y_test data(80:20)).
- Step-5: Check Different Machine Learning Algorithm to get best accuracy.
- Step-6: Check Different feature Engineering techniques on a particular algorithm(I have tried on KNN) to get better accuracy.
- Convert to 3 Dimensional point(x, y, z) from Latitude and Longtitude and use as input.
- Use only Latitude and Longtitude as input data.
- Use Hour, Week, Month, Year, Latitude, Longtitude as input data.
- Use the columns of dataframe that has positive correlation with category of crime.
- Try on a smaller part of dataset maybe the model is overfitting.
- Step-7: Use One Hot Encoder Technique to submit predicted value into required submission format
- Out of the feature engineering methods suggested the one which works best on KNN is simply by taking latitude and longtitude as input data.
- Since input data contains negative values we cannot use OVA Naive Bayes.
- OVA SVM with linear and rbf kernel and CrammerSVM
- The Algorithm gives the best accuracy is KNN for the given dataset.
| Algorithm | Parameters | Accuracy |
|---|---|---|
| KNN | K= 50 | 27.65% |
| OVA Naive Bayes | - | - |
| OVA Logistic Regression | - | 19.96% |
The reason for such less accuracy is because the dataset contains around 800k rows and ML algorithm work well only till 100k. A strong proof for this can be seen with the shoot of accuracy to 83.97% by reducing the dataset to 20% after randomly shuffing it. So we need to use Neural Network approach.