Machine Learning Project Review - Travel insurance prediction

[okothchristopher]

1. You need to create 3 repositories i.e

• Machine Learning Projects (Learning),
• Machine Learning projects(Capstones)
• Machine Learning projects (Personal Practice) Within each repository, you will need a folder for each of the projects, have a standard way of naming them i.e
  1. Week 1 Boston Regression
  2. Week 2 Credit Risk Classification etc. Within each week, have a readme of the project, detail what you did, what was the outcome and what were the key concepts/lessons learnt.

[okothchristopher]

2. The Readme here is incomplete. Your readme needs to have the following sections
   • Title and Description (What the project is about and what it seeks to solve, i.e the objectives)
   • Data Source
   • Insights about the data (Key ones only, accompanied with some charts)
   • The product, what was the ML development steps and the outcomes to the same (output can be a graph eg a confusion matrix plot, with explanations below it)
   • Acknowledgments

[okothchristopher]

3. Notebook Structure is not okay. You need to have a structured way of modelling, ie,
   • Give background (Done)
   • Data Context (When was the data collected, from where have you obtained it)
   • Problem Statement (What are you trying to predict)

[okothchristopher]

4. EDA Steps
   • When doing EDA, I would advice against splitting the data first before exploring it, as some features would need to be explored in the context of the independent variable y
   • What is the full meaning of this "OOPs for plots"
   • For each plot you have to note what you have observed
   • This step 'Summary statistics for each numerical feature' can easily be done by the .info() method
   • Are your features linearly separable, that should be intention of this plot : sns.pairplot(Clients,vars=['Age', 'AnnualIncome','FamilyMembers'], hue='TravelInsurance')
   • This step (Converting the categorial features into binary) can easily be done by an encoder - use that.

[okothchristopher]

5. Feature Engineering
   • There are no derived features in your dataset, which would help with improving the prediction.

[okothchristopher]

6. Data pre-processing
   • You need to have a section title for this.
   • There is no need for this "To check if the spread of splits"

[okothchristopher]

7. Modelling
   • You data is so small relatively speaking, thus instead of going with hold out sets, I would advice that you do cross val scores which give a realistic value for the model performance

[okothchristopher]

8. Model evaluation
   • There is no tuning of parameters that you have done here, so the section " Defining models while tuning their hyperparameters" is misleading
   • This model evaluation metric selected is not the best, you need to use a more balance metric eg F1_score
   • Your model performance on train data is wanting, remember, I mentioned you need to aim for 90% and above, given this work has been done in many platforms.

[okothchristopher]

9. You need to post the score you got on Kaggle.