Customer Satisfaction Prediction: Practical Applications of Machine Learning and Deep Learning

Build an application for predicting customer satisfaction, using a combination of machine learning and deep learning models, focusing on handling class imbalance issues, and selecting the optimal model through multi-metric evaluation

• Original Author/Maintainer: AnkurBhatt07
• Source Platform: GitHub
• Original Title: Customer-Satisfaction-Score-Prediction-using-ML-DL
• Original Link: https://github.com/AnkurBhatt07/Customer-Satisfaction-Score-Prediction-using-ML-DL
• Publication Date: 2026-06-11

In a highly competitive business environment, the cost of acquiring new customers is 5-25 times that of retaining existing ones. Customer Satisfaction Score (CSAT) is a core metric for measuring customer experience, directly impacting customer retention, word-of-mouth, and final revenue.

Traditional satisfaction surveys rely on post-event questionnaires, which have a lag. Predictive analytics can identify risks before customers express dissatisfaction, giving businesses the opportunity to intervene proactively. This project demonstrates how to build an end-to-end customer satisfaction prediction system using a combination of machine learning and deep learning technologies.

Based on customers' historical behavior data, transaction records, and service interaction information, predict customers' satisfaction scores for services (usually 1-5 points or a binary classification of satisfied/dissatisfied).

Class Imbalance

• Satisfied customers are usually far more than dissatisfied ones
• Extreme scores (1 or 5) may be more common than middle scores
• Standard models tend to predict the majority class

Feature Complexity

• Customer data includes numerical features (consumption amount, usage duration) and categorical features (region, product type)
• Time-series features (trends in purchase frequency changes)
• Text features (customer service chat records, comments)

Data Quality Issues

• Missing values (some customers did not fill in certain information)
• Outliers (large abnormal transactions)
• Data entry errors

Missing Value Handling

• Numerical features: fill with median or mean, or predict and fill based on other features
• Categorical features: fill with mode or create an "Unknown" category
• Features with high missing rate (>50%): consider deletion or special handling

Outlier Detection and Handling

• IQR method: identify data points outside 1.5 times the interquartile range
• Z-score: mark outliers with |z|>3
• Business rules: e.g., a single transaction exceeding 10 times the customer's historical average

Data Type Conversion

• Convert date strings to datetime objects
• Categorical encoding: One-hot or Label encoding
• Text vectorization: TF-IDF or word embedding

Time Feature Extraction

• Customer lifecycle: number of days since first purchase
• Activity: number of days since last purchase (Recency)
• Frequency: number of purchases in the past 30/90/365 days
• Amount: average order value, total consumption amount

RFM Model Features

• Recency: number of days since the customer's last purchase
• Frequency: number of purchases
• Monetary: cumulative consumption amount
• RFM is a classic framework for customer value analysis

Interaction Features

• Create feature combinations, e.g., "consumption amount × purchase frequency"
• Capture non-linear relationships

Feature Scaling

• Standardization (StandardScaler): mean 0, variance 1
• Normalization (MinMaxScaler): scale to [0,1]
• Especially important for neural networks

Oversampling

• Random Oversampling: duplicate minority class samples; simple but prone to overfitting

• SMOTE (Synthetic Minority Over-sampling Technique): generate new samples by interpolating between minority class samples; alleviates overfitting issues

• ADASYN (Adaptive Synthetic Sampling): adaptively generate samples, focusing on hard-to-learn samples

Undersampling

• Random Undersampling: randomly delete majority class samples; may lose important information

• Tomek Links: delete pairs of samples from different classes that are nearest neighbors to each other; clean class boundaries

• Edited Nearest Neighbors: delete majority class samples that are misclassified

Hybrid Strategies

• SMOTE + Tomek Links
• Oversample first then undersample