LIA - Suggested Project Ideas

Ideas 30 min read Planning

---

How to Choose Your Project

Selecting the right project is critical. Choose one that matches your skill level and interests. Use the decision flowchart below to help you decide.

Decision Flowchart

Selection Criteria

Factor	Questions to Ask Yourself
Interest	Am I genuinely curious about this problem?
Data Availability	Is the dataset easy to access and well-documented?
Complexity	Can I realistically complete this in 15 hours?
Explainability	Will LIME/SHAP produce meaningful results for this model?
Demonstrability	Can I build a convincing live demo?

Choose Wisely

You cannot change your project after Week 13. Make sure you have access to the dataset and understand the problem before committing.

---

Project Comparison Table

#	Project	Type	Difficulty	Dataset Size	Best For
1	House Price Predictor	Regression	⭐ Beginner	~1,500 rows	Comfort with numbers
2	Sentiment Analyzer	Classification	⭐⭐ Intermediate	~50,000 rows	Interest in NLP
3	Spam Classifier	Classification	⭐ Beginner	~5,500 rows	First ML project
4	Customer Churn Predictor	Classification	⭐⭐ Intermediate	~7,000 rows	Business interest
5	Image Classifier	Classification	⭐⭐⭐ Advanced	~70,000 images	Computer vision interest
6	Medical Risk Assessment	Classification	⭐⭐ Intermediate	~300 rows	Healthcare interest
7	Movie Recommender	Classification	⭐⭐⭐ Advanced	~100,000 ratings	Recommender systems
8	Credit Scoring	Classification	⭐⭐ Intermediate	~30,000 rows	Finance interest
9	Insurance Cost Predictor	Regression	⭐ Beginner	~1,300 rows	Quick start, clean data
10	Employee Attrition Predictor	Classification	⭐⭐ Intermediate	~1,400 rows	HR / people analytics

---

1. House Price Predictor

Predict housing prices based on property features.

Aspect	Details
Type	Regression
Difficulty	⭐ Beginner
Dataset	Boston Housing or Ames Housing
Samples	~1,460 (Ames) / ~506 (Boston)
Features	Square footage, bedrooms, neighborhood, year built, etc.
Target	Sale price (continuous)

Suggested Models

• Linear Regression (baseline)
• Random Forest Regressor
• Gradient Boosting (XGBoost or LightGBM)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ sqft, bedrooms, bathrooms, year_built, neighborhood }	{ predicted_price, confidence_interval }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, r2_score, mae, features }

Evaluation Metrics

• MAE (Mean Absolute Error)
• RMSE (Root Mean Square Error)
• R² Score
• MAPE (Mean Absolute Percentage Error)

Explainability Focus

SHAP is ideal here — waterfall plots clearly show how each feature pushes the predicted price up or down (e.g., larger square footage increases price, older year_built decreases it).

Sample predict request

{
  "square_feet": 1500,
  "bedrooms": 3,
  "bathrooms": 2,
  "year_built": 2005,
  "neighborhood": "CollgCr",
  "garage_cars": 2,
  "overall_quality": 7
}

Response:

{
  "predicted_price": 215400.50,
  "model_version": "1.0.0",
  "confidence": "Prediction based on Random Forest with R²=0.89"
}

---

2. Sentiment Analyzer

Classify text reviews as positive, negative, or neutral.

Aspect	Details
Type	Text Classification (NLP)
Difficulty	⭐⭐ Intermediate
Dataset	IMDB Reviews or Amazon Reviews
Samples	~50,000
Features	Raw text (requires TF-IDF or CountVectorizer)
Target	Sentiment (positive / negative)

Suggested Models

• Logistic Regression with TF-IDF (baseline)
• Naive Bayes (MultinomialNB)
• Random Forest with TF-IDF features

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ text: "This movie was fantastic!" }	{ sentiment, confidence, keywords }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, accuracy, vocabulary_size }

Evaluation Metrics

• Accuracy
• Precision / Recall per class
• F1-Score
• Confusion Matrix

Explainability Focus

LIME excels for text — it highlights which words contribute most to the prediction (e.g., "fantastic" → positive, "terrible" → negative).

Text Preprocessing

Don't forget: lowercase, remove punctuation, remove stopwords, and apply TF-IDF before training. Document your preprocessing pipeline carefully — it must be reproduced at prediction time.

Sample predict request

{
  "text": "The product arrived quickly and works perfectly. Very satisfied with my purchase!"
}

Response:

{
  "sentiment": "positive",
  "confidence": 0.94,
  "model_version": "1.0.0"
}

---

3. Spam Classifier

Classify SMS messages or emails as spam or not spam (ham).

Aspect	Details
Type	Binary Classification (NLP)
Difficulty	⭐ Beginner
Dataset	SMS Spam Collection
Samples	~5,574
Features	Raw SMS text
Target	Spam / Ham

Suggested Models

• Naive Bayes with CountVectorizer (baseline — classic spam filter)
• Logistic Regression with TF-IDF
• Support Vector Machine (SVM)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ message: "You won a free iPhone!" }	{ label, spam_probability }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, accuracy, f1_score }

Evaluation Metrics

• Accuracy
• Precision (critical — avoid false positives marking real messages as spam)
• Recall
• F1-Score

Explainability Focus

LIME highlights the "spammy" words that trigger the classifier (e.g., "free", "winner", "click", "prize").

Why This is Great for Beginners

The SMS Spam dataset is small, clean, and well-documented. The problem is intuitive (everyone knows what spam looks like), and the results are easy to demo and explain.

Sample predict request

{
  "message": "CONGRATULATIONS! You've been selected to receive a $1000 Walmart gift card. Click here to claim now!"
}

Response:

{
  "label": "spam",
  "spam_probability": 0.97,
  "model_version": "1.0.0"
}

---

4. Customer Churn Predictor

Predict whether a customer will leave (churn) a service.

Aspect	Details
Type	Binary Classification
Difficulty	⭐⭐ Intermediate
Dataset	Telco Customer Churn
Samples	~7,043
Features	Tenure, monthly charges, contract type, internet service, payment method, etc.
Target	Churn (Yes / No)

Suggested Models

• Logistic Regression (baseline)
• Random Forest Classifier
• Gradient Boosting (XGBoost)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ tenure, monthly_charges, contract, internet_service, ... }	{ churn_prediction, churn_probability, risk_level }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, auc_roc, features }

Evaluation Metrics

• Accuracy
• AUC-ROC (primary metric for imbalanced classes)
• Precision / Recall
• F1-Score

Explainability Focus

SHAP dependence plots reveal which factors drive churn most (e.g., month-to-month contracts, high monthly charges). Business-relevant insights make this great for your presentation.

Sample predict request

{
  "tenure": 3,
  "monthly_charges": 85.50,
  "total_charges": 256.50,
  "contract": "Month-to-month",
  "internet_service": "Fiber optic",
  "payment_method": "Electronic check",
  "senior_citizen": 0
}

Response:

{
  "churn_prediction": "Yes",
  "churn_probability": 0.82,
  "risk_level": "High",
  "model_version": "1.0.0"
}

---

5. Image Classifier (Simple)

Classify images into categories using pixel features or pre-extracted features.

Aspect	Details
Type	Multi-class Classification
Difficulty	⭐⭐⭐ Advanced
Dataset	Fashion MNIST or CIFAR-10
Samples	~70,000 (Fashion MNIST)
Features	28×28 pixel values (flattened to 784 features)
Target	Category (T-shirt, Trouser, Pullover, etc.)

Suggested Models

• Logistic Regression on flattened pixels (baseline)
• Random Forest on flattened pixels
• SVM with dimensionality reduction (PCA)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ pixels: [0.0, 0.0, ..., 0.5, ...] } (784 values)	{ category, confidence, top_3_predictions }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, accuracy, classes }

Evaluation Metrics

• Accuracy
• Per-class Precision and Recall
• Confusion Matrix (10×10)
• Macro F1-Score

Explainability Focus

LIME for images highlights which pixel regions contributed most to the classification. SHAP can show feature importance across pixel positions.

Advanced Challenge

This project requires handling high-dimensional input (784 features). Consider PCA for dimensionality reduction. The API payload will be larger — plan for validation of array length and value ranges.

Sample predict request

{
  "pixels": [0.0, 0.0, 0.0, 0.0, 0.12, 0.45, 0.78, "... (784 values total)"]
}

Response:

{
  "category": "Ankle boot",
  "confidence": 0.91,
  "top_3_predictions": [
    {"category": "Ankle boot", "confidence": 0.91},
    {"category": "Sneaker", "confidence": 0.06},
    {"category": "Sandal", "confidence": 0.02}
  ],
  "model_version": "1.0.0"
}

---

6. Medical Risk Assessment

Predict the risk of a medical condition based on patient features.

Aspect	Details
Type	Binary Classification
Difficulty	⭐⭐ Intermediate
Dataset	Heart Disease UCI or Diabetes
Samples	~303 (Heart) / ~768 (Diabetes)
Features	Age, blood pressure, cholesterol, glucose, BMI, etc.
Target	Disease presence (Yes / No)

Suggested Models

• Logistic Regression (baseline — interpretable in healthcare)
• Random Forest Classifier
• Support Vector Machine

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ age, blood_pressure, cholesterol, glucose, bmi, ... }	{ risk_level, probability, contributing_factors }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, recall, precision, features }

Evaluation Metrics

• Recall (primary — minimizing false negatives is critical in healthcare)
• Precision
• F1-Score
• AUC-ROC

Explainability Focus

Explainability is essential in healthcare. SHAP force plots and LIME explanations show exactly which patient factors contribute to the risk assessment. This makes your analysis highly relevant for your presentation.

Ethical Considerations

In your report, discuss the ethical implications of deploying an AI system for medical risk assessment. Address topics like: model bias, the importance of recall over precision, and why the model should assist (not replace) medical professionals.

Sample predict request

{
  "age": 55,
  "sex": 1,
  "blood_pressure": 140,
  "cholesterol": 250,
  "fasting_blood_sugar": 1,
  "max_heart_rate": 150,
  "exercise_angina": 0
}

Response:

{
  "risk_level": "High",
  "probability": 0.78,
  "contributing_factors": ["high cholesterol", "elevated blood pressure", "age"],
  "model_version": "1.0.0"
}

---

7. Movie Recommender

Predict whether a user will like a movie based on their preferences.

Aspect	Details
Type	Classification / Recommendation
Difficulty	⭐⭐⭐ Advanced
Dataset	MovieLens 100K
Samples	~100,000 ratings
Features	User demographics, movie genres, historical ratings
Target	Rating prediction or like/dislike binary

Suggested Models

• Logistic Regression on engineered features (baseline)
• Random Forest on user-movie feature matrix
• K-Nearest Neighbors (content-based)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ user_age, genres: ["Action", "Sci-Fi"], avg_rating }	{ will_like, confidence, similar_movies }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, accuracy, genres_supported }

Evaluation Metrics

• Accuracy
• Precision@K
• F1-Score
• Mean Absolute Error (for rating prediction)

Explainability Focus

SHAP can reveal which genres, user demographics, or viewing patterns most influence recommendations. Feature interaction plots show how genre preference interacts with age or viewing history.

Advanced Project

This project requires significant feature engineering — you'll need to create a user-movie feature matrix from raw rating data. Budget extra time for data preprocessing.

Sample predict request

{
  "user_age": 28,
  "preferred_genres": ["Action", "Sci-Fi", "Thriller"],
  "average_rating": 3.8,
  "movie_genre": "Action",
  "movie_year": 2020
}

Response:

{
  "will_like": true,
  "confidence": 0.85,
  "predicted_rating": 4.2,
  "model_version": "1.0.0"
}

---

8. Credit Scoring

Predict whether a loan applicant is likely to default.

Aspect	Details
Type	Binary Classification
Difficulty	⭐⭐ Intermediate
Dataset	German Credit Data or Give Me Some Credit
Samples	~1,000 (German) / ~150,000 (Give Me Some Credit)
Features	Income, age, number of dependents, credit history, loan amount, etc.
Target	Default (Yes / No)

Suggested Models

• Logistic Regression (baseline — standard in finance)
• Random Forest
• Gradient Boosting (XGBoost)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ income, age, debt_ratio, credit_lines, loan_amount }	{ default_risk, probability, risk_category }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, auc_roc, features }

Evaluation Metrics

• AUC-ROC (primary metric in credit scoring)
• Precision / Recall
• F1-Score
• KS Statistic (bonus — industry standard)

Explainability Focus

Explainability is legally required in many jurisdictions for credit decisions. SHAP force plots can explain individual loan decisions, making this project highly relevant for regulatory compliance discussions.

Regulatory Context

In the EU (GDPR) and Canada, individuals have the "right to explanation" for automated decisions. Credit scoring models must be explainable. Discuss this in your report.

Sample predict request

{
  "annual_income": 52000,
  "age": 35,
  "debt_ratio": 0.35,
  "number_of_open_credit_lines": 5,
  "number_of_dependents": 2,
  "months_employed": 48
}

Response:

{
  "default_risk": "Low",
  "default_probability": 0.12,
  "risk_category": "A",
  "model_version": "1.0.0"
}

---

9. Insurance Cost Predictor

Predict individual medical insurance costs based on personal attributes.

Aspect	Details
Type	Regression
Difficulty	⭐ Beginner
Dataset	Medical Cost Personal Dataset
Samples	~1,338
Features	Age, sex, BMI, children, smoker, region
Target	Insurance charges (continuous)

Suggested Models

• Linear Regression (baseline)
• Random Forest Regressor
• Gradient Boosting Regressor

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ age, sex, bmi, children, smoker, region }	{ predicted_cost, model_version }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, r2_score, mae }

Evaluation Metrics

• MAE
• RMSE
• R² Score
• MAPE

Explainability Focus

SHAP clearly shows that smoker status has a massive impact on predicted cost. This makes for an excellent, visually impactful demo. Waterfall plots for individual predictions are very intuitive.

Best Beginner Choice for Regression

This dataset is small, clean, requires minimal preprocessing, and produces very interpretable SHAP results. If you're new to ML and prefer regression, start here.

Sample predict request

{
  "age": 32,
  "sex": "female",
  "bmi": 27.5,
  "children": 1,
  "smoker": "no",
  "region": "northwest"
}

Response:

{
  "predicted_cost": 5765.30,
  "model_version": "1.0.0"
}

---

10. Employee Attrition Predictor

Predict whether an employee is likely to leave the company.

Aspect	Details
Type	Binary Classification
Difficulty	⭐⭐ Intermediate
Dataset	IBM HR Analytics Employee Attrition
Samples	~1,470
Features	Age, job satisfaction, monthly income, overtime, years at company, etc.
Target	Attrition (Yes / No)

Suggested Models

• Logistic Regression (baseline)
• Random Forest Classifier
• Gradient Boosting (XGBoost)

API Endpoints

Endpoint	Method	Input	Output
/predict	POST	{ age, job_satisfaction, monthly_income, overtime, years_at_company }	{ attrition_risk, probability, key_factors }
/health	GET	—	{ status, timestamp }
/model-info	GET	—	{ model, f1_score, features }

Evaluation Metrics

• Accuracy
• F1-Score (important due to class imbalance)
• AUC-ROC
• Recall

Explainability Focus

SHAP reveals which workplace factors drive attrition most (overtime, satisfaction, compensation). HR-relevant insights make the presentation compelling.

Sample predict request

{
  "age": 30,
  "job_satisfaction": 2,
  "monthly_income": 3500,
  "overtime": "Yes",
  "years_at_company": 2,
  "distance_from_home": 25,
  "work_life_balance": 2
}

Response:

{
  "attrition_risk": "High",
  "attrition_probability": 0.73,
  "key_factors": ["low job satisfaction", "overtime", "short tenure"],
  "model_version": "1.0.0"
}

---

Project Difficulty Summary

View Projects by Difficulty Level

Instructor Recommendation

If this is your first end-to-end ML project, choose Spam Classifier, Insurance Cost Predictor, or House Price Predictor. These datasets are clean, small, and well-documented — letting you focus on the deployment pipeline rather than data wrangling. If you want a challenge, Customer Churn or Credit Scoring offer great business-relevant problems with interesting explainability results.

---

Can I Choose My Own Project?

Yes! You may propose your own project, provided it meets these minimum requirements:

Requirement	Minimum
Dataset size	≥ 500 samples
Number of features	≥ 5
Problem type	Classification or Regression
Data accessibility	Publicly available or approved by instructor
Explainability	LIME and/or SHAP must be applicable

Submit your custom project proposal to your instructor by the end of Week 12 for approval.