shaktipanda1235 / kpmg_virtual_internship

Sprocket Central is medium size bike company which requires analytical insights regarding marketing strategy and which customers to target from both current and future customers. A final visualisation input needed to be given to get a sign-off to work further.

ScreenRecorderProject1.mp4

Inspiration

After working with more than 70 toy datasets, this internship provided a platform to get a detailed overview of how industry works. As most of dataset I worked on used to be cleaned ones, this internship helped in getting data in it's raw form. Starting from basic collected data to model building and collecting insigths from it was a full fledged work I was interested in.

Data Quality Assessment

• As data was in its raw form, it was necessary to clean it and observe any disperancy in it before proceeding to deriving insigts, as insights from a vague value may in a misleading conclusions.
• Dataset was evaluated based on Data Quality Framework Table and a draft was generated which was sent to client for double check. Also strategies were given to mitigate such disperancy.

Feature Engineering

• A dataset was downloaded from Australian Bureau of Statistics where each pincode was segregated to different states. Most important finding of such feature engineered column was segregating Australian Capital Territory from 'New South Wales.

Insights on current customers

• Pareto Principle was checked for current customers and it failed.

• A RFM analysis was carried out on all the customers and following conclusions were derived:

The company was loosing 39 important customers

Data Preparation before modelling

• Dataset had no target variable, so a column was created from RFM analysis using monetary habits, i.e They were divided into two category based on spending.
• As there were 12.6% missing value in job_industry_category which was found to be an important indicator in predicting importance of customer, instead of dropping the column a target based encoding is done which took care of the missing values in data. i.e. WOE(Weight of Evidence) Encoding

Model Building

• Best model was found to be a Stacking Classifier which had two KNNClassifier as base models and GausianNB as final_estimator.
• A model was created using full datasets as previously train-test split was compromising patterns present in train dataset.

Prediction

• The new customer dataset was converted to similar structure to that of train dataset and prediction was carried out.
• Deciding factor job_industry_category of 'Manufacturing' has an important potential to tap on.

Visualisation

• A dashboard was created to make client understand about what was happening and what is going to happen among their current customers and new customers.