The aim of the project is to build a Machine Learning Model to predict whether an owner will initiate an auto insurance claim in the next year.

The Machine Learning model mainly consist of two phases:

1. EDA (Exploratory Data Analysis):

Analyze the datasets to summarize their main characteristics (with visual methods). A statistical model can be used, primarily EDA can be used to see what the data can tell us beyond the formal modeling or hypothesis testing task.

Following tasks can be performed as a part of EDA:

o Scaling/Normalization

o Fill the missing values

o Feature selection & engineering

2. Machine Learning Modeling:

After EDA, the modeling comes into the process. The process of training an ML model involves providing an ML algorithm (that is, the learning algorithm) with training data. The term “ML model” refers to the model artifact that is created by the training process.

The training data must contain the correct answer (target or target attribute). The learning algorithm finds patterns in the training data that maps the input data attributes to the target (the answer that you want to predict), and it outputs an ML model that captures these patterns. You will use the ML model to get predictions on new data for which you will not know the target.

Following tasks can be performed as a part of Modeling:

• Start with the basic model but eventually move towards ensemble

• Use Deep Learning with sklearn MLPClassifier and check if the Neural Network Model is better than traditional models

• Arrival at a model with best f1-score

Following are the deliverables (.ipynb files), which needed to be developed with respect to Exploratory Data Analysis:

1. Write at least 3 important inferences from the data above

2. Is the data balanced? Meaning are targets 0 and 1 in the right proportion?

3. How many categorical features are there?

4. How many binary features are there?

5. Write inferences from data on interval variables.

6. Write inferences from data on ordinal variables.

7. Write inferences from data on binary variables

8. Check if the target data is proportionate or not. Hint: Below than 30% for binary data is sign of imbalance

9. What should be the preferred way in this case to balance the data?

10. How many training records are there after achieving a balance of 12%?

11. Which are the top two features in terms of missing values?

12. In total, how many features have missing values?

13. What steps should be taken to handle the missing data?

14. Which interval variables have strong correlation?

15. What's the level of correlation among ordinal features?

16. Implement Hot Encoding for categorical features

17. In nominal and interval features, which features are suitable for StandardScaler?

18. Summarize the learnings of ED

Following are the deliverables (.ipynb files), which needed to be developed with respect to Modeling :

1. The Simple Logistic Regression Model seems to have high accuracy. Is that what we need at all? What is the problem with this model?

2. Why do you think f1-score is 0.0?

3. What is the precision and recall score for the model?

4. What is the most important inference you can draw from the result?

5. What is the accuracy score and f1-score for the improved Logistic Regression model?

6. Why do you think f1-score has improved?

7. For model LinearSVC play with parameters – dual, max_iter and see if there is any improvement

8. SVC with Imbalance Check & Feature Optimization & only 100K Records → is there improvement in scores?

9. XGBoost is one the better classifiers -- but still f1-score is very low. What could be the reason?

10. What is the increase in number of features after one-hot encoding of the data?

11. Is there any improvement in scores after encoding?

12. If not missing a positive sample is the priority which model is best so far?

13. If not marking negative sample as positive is top priority, which model is best so far?

14. Do you think using AdaBoost can give any significant improvement over XGBoost?

15. MLPClassifier is the neural network we are trying. But how to choose the right no. of layers and size?

16. At what layer size we get the best f1-score?