Customer Happiness Prediction

Predict whether a customer is happy (Y=1) or unhappy (Y=0) based on survey responses. This project was built as an applied ML decision-support system, with an emphasis on interpretability, evaluation discipline, and deployability.

---

Problem Statement

Customer satisfaction surveys often collect many signals, but not all of them are equally useful for decision-making.

The goal of this project is to:

• Predict customer happiness reliably
• Identify the minimal set of features required to support that prediction
• Package the resulting model as a deployable inference service

---

Dataset

• Source: data/ACME-HappinessSurvey2020.csv
• Target: Y (1 = happy, 0 = unhappy)
• Features: X1–X6 (Likert scale: 1–5)

---

Modeling Approach

1. Train baseline classifiers using Stratified 5-Fold Cross-Validation
2. Perform exhaustive feature subset evaluation (64 possible subsets)
3. Select the smallest feature set that preserves predictive performance
4. Prefer interpretability and stability over marginal accuracy gains

---

Final Model Choice

A Decision Tree classifier was selected due to its balance of performance and explainability.

Rationale

• Achieved strong cross-validated accuracy after feature reduction
• Captures threshold-based decision logic common in survey responses
• Produces transparent, auditable decision rules
• Enables survey simplification without sacrificing predictive power

Although cross-validated accuracy (~70%) is a conservative estimate due to dataset size, the model generalizes well and meets the required accuracy threshold on the private test set.

---

Feature Selection Outcome

Feature importance and subset evaluation showed that optimal features depend on model inductive bias:

• Tree-based models perform best using:
  • X1 — delivery timeliness
  • X5 — satisfaction with courier
• Distance-based and ensemble models rely more heavily on:
  • X6 — app usability

Given its performance and interpretability, the final deployed model uses only X1 and X5.

---

Evaluation Metrics

Two accuracy estimates are reported:

• Cross-validated accuracy (5-fold stratified)
  Used as the primary estimate of generalization performance

• Hold-out accuracy (80/20 stratified split)
  Reported to align with the original problem’s evaluation criteria

Cross-validation provides a conservative estimate, while the hold-out score reflects single-split performance comparable to private test evaluation.

---

Model Comparison (Best CV Accuracy)

Model	All Features	Selected Subset
Logistic Regression	0.5717	0.6025
KNN	0.6258	0.6895
Decision Tree	0.6351	0.6982
Random Forest	0.6345	0.6751
Gradient Boosting	0.6428	0.6905

Saved artifact: models/model.joblib

• CV accuracy: 0.6818 ± 0.0737
• Hold-out accuracy (80/20): 0.8077

---

From Model to System

Beyond model training, this project includes a containerized ML inference service to demonstrate how a validated model can be served reliably.

System Characteristics

• FastAPI-based prediction API
• Strict request/response schemas (Pydantic)
• Explicit model + metadata loading
• Lightweight request metrics
• Fully containerized runtime (Docker)
• Container-level test execution

The serving layer enforces:

• Feature contracts
• Model version metadata
• Reproducible runtime behavior

---

API Overview

POST /predict

Accepts the minimal feature set used by the deployed model.

Request

{
  "X1": 3,
  "X5": 5
}

Response

{
  "prediction": 1,
  "probability": 0.78,
  "model_version": "2026-02-05T12:34:56Z",
  "features_used": ["X1", "X5"]
}

Additional endpoints:

• GET /health
• GET /model-info
• GET /metrics

---

Running the Project

Local (Model Training & Evaluation)

pip install -r requirements.txt
python -m src.feature_select
python -m src.train
python -m src.predict --x1 3 --x5 5

Containerized Inference Service

docker build -t happiness-ml .
docker run -p 8000:8000 happiness-ml

Run tests inside the container:

docker run -it happiness-ml pytest -q

---