Machine Learning Projects in Bangalore

This project builds a fraud detection system trained on the publicly available Kaggle credit card transaction dataset containing 284,807 transactions with a severe class imbalance of 0.17% fraud cases. SMOTE oversampling is applied to balance the training distribution. An ensemble of Random Forest, XGBoost, and LightGBM classifiers is trained and combined via soft-voting. The final model achieves an AUC-ROC of 0.987 and a recall of 0.94 on the minority class, making it suitable for real-time transaction screening. A Flask REST API exposes the inference endpoint.

A multi-output regression model predicts semester GPA and subject-wise scores from demographic, attendance, and prior academic features collected from 1,200 student records. The pipeline includes missing-value imputation, ordinal encoding, and min-max scaling. Ridge regression, Support Vector Regression, and a shallow neural network are compared. SHAP values are computed to explain which features — study hours, parental education, and mid-term scores — contribute most to the prediction, providing actionable insights for academic counsellors.

Historical crop yield data from the Karnataka State Department of Agriculture (2000–2022) is combined with IMD rainfall and temperature records to train a Random Forest regressor predicting district-level yield for paddy, ragi, and maize. Feature engineering extracts growing-season averages, drought indices, and soil-type encodings. The model achieves an RMSE of 3.2 quintals per hectare on the held-out test set. A Streamlit dashboard visualises district-wise predictions on a choropleth map, enabling farmers and policymakers to plan procurement.

Clinical features from the Cleveland Heart Disease dataset — including age, cholesterol, resting ECG, and maximum heart rate — are used to train a binary classifier that stratifies patients into low, medium, and high cardiac risk categories. Logistic regression and a radial-basis-function SVM are evaluated under 10-fold cross-validation. Calibration curves confirm that predicted probabilities align with observed event rates. The project includes a web form where clinicians can enter patient vitals and receive an instant risk score with a SHAP waterfall explanation.

A text classification pipeline processes the SpamAssassin public corpus of 6,000 labelled emails. Raw email bodies are cleaned, tokenised, and vectorised using TF-IDF with bigram features. Multinomial Naive Bayes, Logistic Regression, and a Linear SVM are trained and compared. The SVM achieves 98.4% accuracy and 97.1% F1 on the test set. The project includes a browser-based interface where users paste email text and receive an instant spam probability score along with the top-10 most influential tokens.

A regression model predicts residential property prices in Bengaluru using a dataset of 13,000 listings scraped from a real-estate portal. Features include locality, BHK count, total square footage, number of bathrooms, and proximity to metro stations. Gradient Boosting Regressor and XGBoost are tuned via Optuna. The best model achieves a mean absolute percentage error of 8.3% on the test set. A React frontend allows users to input property details and receive a price estimate with a confidence interval derived from quantile regression.

A churn prediction model is trained on the IBM Telco Customer Churn dataset containing 7,043 customer records with 20 features including contract type, monthly charges, and tenure. Class imbalance is handled with class-weight adjustment. XGBoost with early stopping achieves an AUC of 0.862 and a precision-recall AUC of 0.71. SHAP summary plots identify month-to-month contracts and high monthly charges as the dominant churn drivers. A CRM-style dashboard highlights at-risk customers ranked by churn probability.

This project compares a lexicon-based approach (VADER) with a supervised Logistic Regression classifier on 50,000 Amazon product reviews from the Electronics category. Reviews are preprocessed with contraction expansion, stopword removal, and lemmatisation. The supervised model trained on TF-IDF features outperforms VADER by 6.2 percentage points in F1 on the three-class (positive, neutral, negative) task. A live demo accepts a product URL, scrapes the first 100 reviews, and displays a sentiment distribution chart with representative review excerpts.

Time-series sensor readings from a simulated CNC machine — vibration, temperature, and spindle speed sampled at 10 Hz — are used to train an Isolation Forest anomaly detector. The dataset contains 500,000 readings with 2.1% labelled fault events. Rolling statistical features (mean, standard deviation, kurtosis over 60-second windows) are extracted before training. The model achieves a precision of 0.88 and recall of 0.91 on fault detection. A real-time dashboard plots the sensor stream and highlights anomalous windows in red with an alert log.

A collaborative filtering recommendation engine is built on the MovieLens 25M dataset. Matrix factorisation using Singular Value Decomposition and the Alternating Least Squares algorithm are implemented and compared. The ALS model achieves an RMSE of 0.82 on the held-out test set. A content-based fallback using TF-IDF on movie genres and tags handles cold-start users. The final system is deployed as a FastAPI service with a React frontend where users rate five movies and receive ten personalised recommendations with predicted ratings.