Hi,
I'm Faizan Bhutto
Data Scientist

×

Gym Customer Churn Analysis and Clustering Project

Problem Statement: The gym was facing high customer churn rates, which was impacting revenue and overall growth. The goal of this project was to analyze customer data to understand the factors contributing to churn, predict which customers were at risk of leaving, and segment the customer base to design targeted retention strategies.

Approach: This project involved several key steps:

• Exploratory Data Analysis (EDA): We conducted a thorough analysis of the dataset, exploring the distributions of features such as customer age, contract length, lifetime membership, and class attendance. Visualizations were created to understand the relationships between features and their impact on churn, providing insights into key factors influencing customer behavior.
• Feature Engineering: We engineered features from raw data to enhance model performance, including metrics like engagement score, average monthly spending, and customer lifetime value.
• Churn Prediction Modeling: We built multiple machine learning models, including Logistic Regression, Random Forest, Gradient Boosting, and Decision Trees, to predict customer churn. Each model was evaluated using key metrics such as Accuracy, Precision, Recall, F1 Score, and ROC AUC to determine the best-performing model for churn prediction.
• Customer Segmentation Using Clustering: We applied K-means clustering to segment customers into five distinct groups based on their characteristics. Clustering allowed us to identify high-risk customers, moderately engaged customers, and loyal customers. The clusters were visualized using PCA to reduce dimensionality and better understand customer groupings.
• Business Recommendations: Based on the model predictions and cluster analysis, we provided targeted business recommendations aimed at reducing churn and improving customer engagement. These included personalized offers for high-risk customers, loyalty rewards for loyal customers, and new fitness challenges to keep customers engaged.

Outcome: The project provided several actionable insights:

• Model Performance: The Gradient Boosting model outperformed others with an accuracy of 92.5%, precision of 87.77%, recall of 81.68%, F1 score of 84.61%, and ROC AUC of 97.25%. This model was recommended for ongoing churn prediction efforts to help identify at-risk customers before they decided to leave.
• Customer Segmentation: The K-means clustering revealed five distinct customer groups:
  • Cluster 0: Represented the highest churn risk, with a churn rate of 93.22%. These customers were largely disengaged, indicating a need for personalized interventions to prevent further churn.
  • Cluster 1: Had a moderate churn rate of 43.33%. Customers in this group showed partial engagement, and targeted engagement initiatives, such as fitness challenges or loyalty programs, could help retain them.
  • Clusters 2 and 4: Represented the most loyal customer segments, with churn rates of 1.37% and 0.32%, respectively. These customers had high engagement levels and should be rewarded with loyalty programs, premium services, or VIP experiences to maintain their loyalty.
  • Cluster 3: Showed a low to moderate churn risk of 26.68%, with potential for increased engagement through targeted offers.
• Business Impact: Our analysis and recommendations provided the gym with a structured approach to reducing customer churn. By focusing on high-risk clusters and rewarding loyalty, the gym could implement effective retention strategies, ultimately leading to increased customer lifetime value and better overall customer satisfaction.

Technologies Used: The project utilized various technologies and tools, including:

• Python: Used for data analysis, modeling, and implementation of clustering techniques.
• Pandas: For data manipulation, cleaning, and feature engineering.
• Matplotlib & Seaborn: For visualizing data distributions, correlations, and model performance.
• Scikit-learn: For building machine learning models, including Logistic Regression, Random Forest, Gradient Boosting, and Decision Trees, as well as for clustering.
• SciPy: For additional analysis, including hierarchical clustering and dendrogram visualization.
• Jupyter Notebook: Provided an interactive environment to write and execute code, visualize data, and document findings.

×

Telecom Churn Analysis and Modelling Project

Problem Statement: The telecom company was experiencing high customer churn rates, leading to significant revenue loss. The goal of this project was to predict customer churn and provide insights to improve customer retention by focusing on high-value customers and identifying churn risks.

Approach: The project followed a structured approach to analyze churn:

• Exploratory Data Analysis (EDA): A detailed analysis of customer demographics, subscription details, service usage, and churn reasons was performed. Key insights were derived to understand how features like tenure, monthly charges, and contract types influenced customer churn.
• Feature Engineering: Important features such as customer tenure, contract type, and service usage were engineered to improve model performance. Categorical variables were converted to numerical format using encoding techniques.
• Churn Prediction Modeling: Multiple machine learning models, including Logistic Regression, Random Forest, SVM, KNN, Gradient Boosting, XGBoost, Naive Bayes, Decision Tree, and a Deep Learning Neural Network, were built to predict customer churn. Each model was evaluated based on metrics like Accuracy, Precision, Recall, and F1 Score to determine the best model for predicting churn.
• Feature Importance Analysis: Feature importance analysis was conducted to identify the key drivers of churn, helping the business understand which factors had the most impact on customer retention.
• Business Recommendations: Based on the model predictions and feature importance analysis, specific business recommendations were provided to enhance customer retention and minimize churn.

Outcome: The project provided several actionable insights:

• Model Performance: The Random Forest model provided the best accuracy of 86.4%, with a precision of 91% and a recall of 89%. XGBoost and Deep Learning models also performed well, providing reliable predictions for customers likely to churn.
• Key Insights:
  • Churned customers were more likely to be on month-to-month contracts and had lower tenure.
  • Offer E had a higher churn rate, suggesting that it may not align with customer expectations.
  • Customers with additional internet services (e.g., online security) were less likely to churn.
  • Fiber optic internet users had a higher churn likelihood, indicating potential quality issues.
• Business Impact: By focusing on high-value customers and providing targeted loyalty programs, the telecom company could improve customer satisfaction and retention. Recommendations such as promoting long-term subscriptions and enhancing service quality for fiber optic internet users were also provided.

Technologies Used: The project utilized various technologies and tools, including:

• Python: Used for data analysis, modeling, and implementation of machine learning algorithms.
• Pandas: For data manipulation and cleaning.
• Matplotlib & Seaborn: For visualizing data distributions, correlations, and model performance.
• Scikit-learn: For building machine learning models such as Logistic Regression, Random Forest, and SVM.
• XGBoost: For building a powerful gradient boosting model to predict churn.
• TensorFlow/Keras: Used to develop a Deep Learning Neural Network for churn prediction.
• Jupyter Notebook: Provided an interactive environment to write and execute code, visualize data, and document findings.

×

Automated Customer Complaints Classification

Problem Statement: Automate the classification of customer complaints to improve response times and enhance service quality using NLP and Machine Learning techniques.

Approach: This project leveraged advanced NLP and Machine Learning methods to classify customer complaints:

• Exploratory Data Analysis (EDA): Analyzed complaint data to uncover patterns and trends.
• Topic Modeling: Identified underlying themes in the complaints using Non-Negative Matrix Factorization (NMF).
• Supervised Classification: Built classification models such as Logistic Regression, Random Forest, and XGBoost to categorize complaints into predefined categories.

Outcome: The Logistic Regression model achieved the highest accuracy of 92%, providing reliable classifications for complaints.

Key Insights:

• Identified common themes like credit card issues, payment disputes, and loan problems.
• Pinpointed regional variations in complaints, highlighting areas for service improvement.

Technologies Used:

• Python for data manipulation and modeling
• Spacy and NLTK for NLP tasks
• Scikit-learn and XGBoost for machine learning models

×

Multi-Source Fitness Log Analysis

Data Access: The dataset is sourced from the PMData: A Sports Logging Dataset on Kaggle, containing multi-source fitness data from 16 participants.

Objective: Analyze and visualize sports logging data to understand physical activity, calorie expenditure, sleep patterns, and health metrics, providing actionable insights for improvement.

Dataset Description:

• User Data: Demographic details like age, height, and gender.
• Calories Data: Logs of calories burned over time.
• Exercise Data: Details on exercises, duration, steps, heart rate, and calories burned.
• Sleep Data: Information on sleep duration and stages.
• Data Sources: Fitbit, PMSys App, and Google Forms.

Exploratory Data Analysis (EDA):

• Analyzed activity patterns, calorie expenditure, and sleep quality.
• Visualized relationships using heatmaps, bar charts, and histograms.

Machine Learning Models:

• Clustering: K-means clustering to group activity and sleep patterns.
• Predictive Modeling: Regression models to predict sleep score based on activity metrics.
• Models Used: Random Forest, Gradient Boosting, and Linear Regression.

Key Insights:

• High-calorie activities include treadmill running and outdoor biking.
• Sleep duration averaged 7.5 hours, with light sleep making up 54%.
• Gender differences in activity intensity and sleep quality were observed.

Recommendations:

• Optimize exercise routines to maximize calorie burn.
• Encourage better sleep hygiene for improved recovery.
• Diversify activities to balance fitness data.

Technologies Used:

• Python for analysis and modeling
• Pandas for data manipulation
• Seaborn & Matplotlib for visualization
• Scikit-learn for clustering and regression models

×

Bearing Failure Detection

Overview: This project explores bearing failure detection by processing vibration data and implementing anomaly detection models such as Isolation Forest, One-Class SVM, Local Outlier Factor, and Autoencoders.

Objectives:

• Process and analyze vibration signal data from multiple test sets.
• Extract statistical features for effective anomaly detection.
• Implement various models to detect bearing failures and evaluate their performance.

Dataset Description:

• Test Set 1: Inner race defect in Bearing 3 and roller element defect in Bearing 4. Duration: Oct 22 - Nov 25, 2003 (2,156 files, 8 channels).
• Test Set 2: Outer race failure in Bearing 1. Duration: Feb 12 - Feb 19, 2004 (984 files, 4 channels).
• Test Set 3: Outer race failure in Bearing 3. Duration: Mar 4 - Apr 4, 2004 (4,448 files, 4 channels).

Models Used:

• Isolation Forest: Detects anomalies by isolating data points.
• One-Class SVM: Uses support vector machines for anomaly detection.
• Local Outlier Factor: Identifies anomalies based on local density comparison.
• Autoencoders: Reconstruct input data and detect anomalies based on reconstruction errors.
• Mahalanobis Distance: A statistical baseline for outlier detection.

Key Insights:

• Test Set 1: Anomalies detected in Bearing 3 (inner race) and Bearing 4 (roller element).
• Test Set 2: Outer race failure in Bearing 1 indicated by a rise in vibration signals.
• Test Set 3: Outer race failure in Bearing 3 detected by a sharp vibration increase.

Business Recommendations:

• Implement predictive maintenance using Isolation Forest and Autoencoders.
• Monitor vibration signals continuously to prevent failures.
• Use statistical feature extraction to enhance detection accuracy.

Results:

Model	Test Set 1 Anomalies	Test Set 2 Anomalies	Test Set 3 Anomalies
Isolation Forest	Optimal Detection	Optimal Detection	Optimal Detection
One-Class SVM	High False Positives	Adjusted Parameters	Adjusted Parameters
Local Outlier Factor	Tuned Parameters	Tuned Parameters	Tuned Parameters
Autoencoders	Best Reconstruction	Best Reconstruction	Best Reconstruction
Mahalanobis Distance	Statistical Baseline	Statistical Baseline	Statistical Baseline

Technologies Used:

• Python: Pandas, NumPy, TensorFlow, Scikit-learn
• Matplotlib and Seaborn for data visualization
• Google Colab for computation

×

End2End Price Prediction System

Overview: This project involves building a complete machine learning pipeline for price prediction using ZenML and MLflow for experiment tracking and model deployment.

Key Features:

• Set up an end-to-end machine learning pipeline from scratch.
• Integrates ZenML for pipeline management and orchestration.
• Uses MLflow for experiment tracking and model deployment.

Prerequisites:

• Python 3.8 or higher
• Virtualenv (optional but recommended)
• Git (to clone the project)
• ZenML for pipeline management

Setup Instructions:

1. Clone the repository.
2. Create a virtual environment and activate it.
3. Install ZenML and set up required integrations.
4. Run the pipeline to train, evaluate, and deploy the price prediction model.

Technologies Used:

• ZenML: For managing machine learning pipelines.
• MLflow: For tracking experiments and model deployment.
• Python: Core programming language for implementation.

×

Machine Learning Ongoing Projects

Overview: A series of ongoing machine learning projects aimed at solving real-world problems in diverse domains.

Key Areas:

• Time series forecasting for demand prediction.
• Image recognition using convolutional neural networks.
• Text sentiment analysis with deep learning approaches.

Technologies Used:

• Python
• TensorFlow and Keras for building neural networks
• Scikit-learn for traditional machine learning algorithms

Goals: To deliver innovative solutions and enhance the understanding of complex machine learning workflows.