Churn Prediction

1. Connection Establishment

We kick off by establishing a connection to the Snowflake database using Snowpark Python.

2. Data Selection

We curate customer demographics, shows, and subscription data from Snowflake for a comprehensive analysis.

3. Data Preprocessing

Before diving into machine learning, we perform crucial data preprocessing steps:

• Calculated Columns: Create a churn column using open and close date columns.
• Feature Engineering: Transform existing columns like view duration to derive meaningful insights.
• Data Cleaning: Tidy up the dataset by handling missing values, ensuring consistency, and imputing null values strategically.
• Encoding: Prepare the data for model training by encoding categorical columns.

4. Discover who's sticking around and who's saying goodbye

5. Feature Selection and Model Training

Prior to training machine learning models, we conduct feature selection using chi-square test. The selection is based on significant columns identified through chi-square testing. Finally, we train a Random Forest model, evaluating performance metrics like MSE and R2.

6. Exploring Features 

Dive into the impact each feature adds to our churn prediction game:

Heatmap of Confusion Matrix: It's a glimpse into the accuracy and missteps of our churn predictions.

The Churn Risk Spectrum: We're not just predicting churn; we're classifying risk levels! Low, medium, or high.

For detailed instructions, please refer to CF_Churn_prediction.ipynb on GitHub.

Time Series Forecasting

1. Connection Establishment

Similar to churn prediction, we establish a seamless connection to the Snowflake database using Snowpark Python.

2. Data Selection

For time series forecasting, we focus on historical customer and subscription data, preparing it for in-depth analysis.

3. Data Transformation and Aggregation

Transforming Snowflake data into a Pandas DataFrame, we aggregate based on 'view date' and derive additional insights such as the days a customer spent before leaving the platform.

4. Model Selection and Tuning

Choosing the XGBoost Regressor for its prowess in handling intricate data relationships, we fine-tune parameters like the learning rate, maximum depth, and number of estimators.

5. Streamlined Training and Forecasting

To streamline the entire process, we implement functions for creating datasets at the customer level, training models, and forecasting for various time periods.

6. Forecast for Next Month

A sneak peek into the future of Spotflix:

Now our XGBoost model is ready to start forecasting customer viewing patterns for the next 6 months or 1 year. This will give us crucial insights to make better decisions. Witness how we predict engagement levels and pinpoint those high-risk viewers:

For detailed instructions, please refer to CF_time_series_forecasting.ipynb on GitHub.

Guide to use .tml files

For step-by-step instructions on how to import .tml files into your ThoughtSpot cluster, please refer to How to use TML files.