Data Science Projects in Bangalore

Public COVID-19 patient records from the Karnataka Health Department (2020–2021) containing 45,000 anonymised cases are analysed to identify mortality risk factors. Logistic regression with odds ratios, chi-square tests for categorical variables, and Kaplan-Meier survival curves are used to quantify the association of age, comorbidities, and vaccination status with 30-day mortality. The analysis confirms that unvaccinated patients above 60 with diabetes had a 4.7× higher mortality odds ratio. An interactive Plotly dashboard presents the findings with filterable demographic breakdowns.

One million GPS probe records from a ride-hailing dataset covering Bengaluru's road network are processed to compute link-level travel time indices for 500 road segments across 24 hours and 7 days. Temporal clustering with k-means identifies four distinct congestion regimes: morning peak, evening peak, off-peak, and weekend. A choropleth map built in Folium visualises the worst-performing corridors. The analysis quantifies that the Silk Board junction adds an average of 22 minutes to southbound commutes during the 5–8 PM window.

Transaction data from a UK-based online retailer (541,909 invoices, UCI repository) is used to compute Recency, Frequency, and Monetary value scores for 4,372 customers. K-means clustering with the elbow method identifies five customer segments: champions, loyal customers, at-risk, hibernating, and lost. Segment profiles are visualised with radar charts. A targeted marketing strategy is designed for each segment, estimating a 12% revenue uplift from re-engagement campaigns directed at the at-risk and hibernating groups.

Daily AQI readings for PM2.5, PM10, NO2, and SO2 from four CPCB monitoring stations in Bengaluru (2015–2023) are modelled using SARIMA and Facebook Prophet. Prophet captures weekly and annual seasonality and Diwali festival spikes as special events. The model achieves a mean absolute error of 8.3 AQI units on a 90-day forecast horizon. A Streamlit application displays historical trends, 30-day forecasts, and health advisory thresholds with colour-coded risk levels for each pollutant.

Ball-by-ball data from 950 IPL matches (2008–2023) is aggregated into match-level features: team batting average, bowling economy, head-to-head win rate, venue advantage, and toss decision. A Random Forest classifier predicts the match winner at the start of the second innings with 71.4% accuracy. Feature importance analysis reveals that the target score and venue-specific run-rate are the strongest predictors. A live prediction widget accepts current match state and outputs win probability for both teams with a confidence interval.

The MIMIC-III clinical database is used to predict 30-day hospital readmission for diabetic patients. Structured features (lab values, diagnosis codes, length of stay) are combined with TF-IDF embeddings of discharge summary notes in a late-fusion model. XGBoost on structured features achieves AUC 0.71; adding text features improves AUC to 0.76. SHAP values identify HbA1c levels, number of prior admissions, and discharge disposition as the top predictors. The project includes an IRB-compliant data handling protocol and de-identification pipeline.

100,000 tweets collected via the Twitter Academic API on the topic of electric vehicles in India are preprocessed and analysed using Latent Dirichlet Allocation topic modelling. Ten coherent topics are identified including charging infrastructure, government subsidies, range anxiety, and brand sentiment. Time-series plots show topic volume trends over a six-month period. A sentiment analysis layer using VADER scores each topic's emotional valence. The findings are presented in a Tableau-style Plotly dashboard with interactive topic filters and tweet-level drill-down.

Weekly sales data for 45 Walmart stores across 99 departments (Kaggle competition dataset) is decomposed into trend, seasonal, and residual components using STL decomposition. An LSTM network with two layers and dropout regularisation is trained on rolling 52-week windows to forecast the next 4 weeks of sales. The model achieves a weighted mean absolute error of 2,847 on the competition metric, placing in the top 30% of historical submissions. A Dash dashboard visualises store-level forecasts with confidence bands and holiday effect annotations.

A dataset of 13,000 residential property listings from a Bengaluru real-estate portal is cleaned, geocoded, and analysed to answer three research questions: which localities have the highest price per square foot, how proximity to metro stations affects pricing, and whether BHK configuration or total area is a stronger price predictor. Geospatial visualisations in Folium, correlation heatmaps, and ANOVA tests are used. The analysis finds that properties within 500 metres of a metro station command a 14% premium on average, controlling for locality and size.