Classification Project on Telco Customers

Demo on industrial application of classification ML models.

Description:

Telco is a telecommunications company offering various products. This project uses information about this companies customers, such as various demographic data and information about the services provided to them to help predict churn.

Goals:

The goal of this project is to determine the drivers of churn, or how whether a customer stays or goes, from Telco customer data. After exploring the company data to find drivers of churn, a ML model is made to make reccomendations to the shareholders of Telco, created for generating actionable intelligence on customer churn.

Beginning Assumptions:

Going into the project with churn and its contributing factors as the primary target, I will assess demographic data as well as details of the customers contracts.

Methodology:

• Data acquired from MySQL 'telco_churn' database.
• Dataframe was 7043 rows, 25 columns before cleaning (row:customers, columns: variables).
• SQL query customers data, joining relevant tables on shared IDs, creating function to convert to dataframe.
• Clean data by dropping irrelevant columns.
• Encode useful categorical variables, removing original non-encoded.
• Split data (train/val/test) at ≈ (60/20/20).
• Split stratification left ambiguous per use.
• Model potential relationships with the use of visuals and graphs.
• Use various statistical tests and models to analyze combinations of variables against the target churn.
• Hypotheses to test:
  • There is no significance between contract type and churn
  • There is no significance between single households (M&F) with no dependents and churn.
  • Tenure is unrelated to churn.
  • Monthly bill price is unrelated to churn.
  • Product subscriptions are unrelated to churn.
• After iterating models and selecting those with the most optimal evaluation metrics, evaluate with the best on test data.
• Synthesize exploration into actionable intelligence for shareholders.

Data Dictionary:

Variable Name	Description
churn_encoded	Binary encoding of churn status (0: Not Churn, 1: Churn)
monthly_charges	The monthly charges for the service (float)
tenure	Number of months the customer has been with the company (int)
gender_encoded	Binary encoding of customer gender (0: Female, 1: Male)
dependents_encoded	Binary encoding of customer having dependents (0: No, 1: Yes)
internet_service_type_Fiber optic	Binary encoding if customer has fiber optic internet service
...	Various other encoded values are present in the DataFrame

Reproduction of Results:

• clone github repo
• download .csv of telco_data (for no mysql) included in repo
• run final_report notebook within an environment.

Analysis Results:

Drivers:

• tenure
• contract type
• monthly bill higher than average (whether theyre paying for fiber optic included)

Models:

• RF best prediction tool currently, KNN close second.
• Log. Reg. could outperform with more manipulation.
• RF performs best on train(80%) and validate(82%) sets.
• KNN is a close second. (78%)/(80%). Adjusting hyperparameters might yield a more accurate model than RF.
• Logisitic regression is powerful and would no doubt outpace the others with more tuning of the features.
• RF model performs at 83% Accuracy on test set.

Reccomendations:

• Upgrade fiber optic service quality or lower price.