MSc · BSc · Data science · Machine learning
50 data science dissertation questions across classification, explainability, fairness, NLP, recommenders, vision, forecasting and healthcare AI. Each has a method, dataset and university thesis reference.
50 Research topics
10 Sub-fields
13 Datasets
5 Universities cited
2020–26 Thesis years
By Siddharth Gupta · Statssy
Most data science students reach for the newest model first. Then they cannot show whether it truly outperformed a simpler baseline.
From 12 years of working with data, I have found that a focused research question, well-chosen evaluation metrics and reproducible comparisons matter more than model complexity alone.
Use the questions below to shortlist a topic, check whether the data are suitable and plan an honest test before training a model.
How to use this list Select two or three questions in the field you want to study. Open the dataset and inspect its size, labels, licence and access requirements. Read the related thesis, focusing on its baselines, data splits and evaluation metrics. Define your validation protocol before model training; avoid data leakage and use time-ordered splits for forecasting. Bring your shortlist to your supervisor or book a free 15-minute call . Jump to a sub-field 01. Classification and predictive modelling 02. Imbalanced data and fraud detection 03. Explainable AI and model interpretability 04. Fairness and bias in machine learning 05. Natural language processing 06. Recommender systems 07. Computer vision and medical imaging 08. Time series forecasting 09. Clustering and customer analytics 10. Machine learning in healthcare Data science research area 01 / 10
Classification and predictive modelling Classification topics compare models on a clear task. The best dissertations test whether a complex model is worth its extra cost.
TOPIC 01 Does gradient boosting beat logistic regression for predicting income? Study design Model comparison Method XGBoost compared with logistic regression (AUC, F1) Dataset Adult Income (UCI) Open TOPIC 02 Which machine learning model predicts loan default best? Study design Model comparison Method Logistic regression, random forest, XGBoost and a neural network Dataset Default of Credit Card Clients (UCI) Open TOPIC 03 Does feature engineering matter more than model choice for churn prediction? Study design Experimental comparison Method Ablation study with cross-validation Dataset Telco Customer Churn (Kaggle) Free login TOPIC 04 Can a neural network predict bike demand better than a random forest? Study design Model comparison Method Multilayer perceptron compared with random forest (RMSE) Dataset Bike Sharing (UCI) Open TOPIC 05 Does hyperparameter tuning give real gains on tabular data? Study design Experimental comparison Method Grid search compared with Bayesian optimisation Dataset Adult Income (UCI) Open Data science research area 02 / 10
Imbalanced data and fraud detection Fraud and default are rare events. These topics test how to find the few cases that matter without drowning in false alarms.
TOPIC 06 Does SMOTE improve fraud detection more than class weights? Study design Experimental comparison Method SMOTE compared with cost-sensitive learning (precision-recall AUC) Dataset Credit Card Fraud Detection (Kaggle) Free login TOPIC 07 Can anomaly detection find fraud without using labels? Study design Unsupervised learning Method Isolation forest compared with an autoencoder Dataset Credit Card Fraud Detection (Kaggle) Free login TOPIC 08 Which metric should guide model choice on imbalanced credit data? Study design Methods comparison Method ROC AUC compared with precision-recall AUC and F2 score Dataset Default of Credit Card Clients (UCI) Open TOPIC 09 Does cost-sensitive learning cut money lost to fraud? Study design Experimental comparison Method Cost-weighted XGBoost with expected loss analysis Dataset Credit Card Fraud Detection (Kaggle) Free login TOPIC 10 Does undersampling or oversampling work better for churn prediction? Study design Experimental comparison Method Resampling strategies with stratified cross-validation Dataset Telco Customer Churn (Kaggle) Free login Data science research area 03 / 10
Explainable AI and model interpretability Explainable AI topics ask why a model made its call. They are popular with examiners because they link technical work to real decisions.
TOPIC 11 Do SHAP values agree with logistic regression on what drives default? Study design Explainability study Method SHAP values compared with regression coefficients Dataset Default of Credit Card Clients (UCI) Open TOPIC 12 Can LIME explain single churn predictions in a way managers trust? Study design Explainability study Method LIME with a user evaluation Dataset Telco Customer Churn (Kaggle) Free login TOPIC 13 Do pneumonia models look at the lungs or at the image edges? Study design Explainability study Method Grad-CAM heatmaps on a CNN Dataset Chest X-Ray Pneumonia (Kaggle) Free login TOPIC 14 Which words drive a transformer’s sentiment predictions? Study design Explainability study Method Attention maps compared with integrated gradients Dataset IMDB Reviews (Hugging Face) Open TOPIC 15 How much accuracy do you lose with an interpretable model? Study design Model comparison Method Explainable boosting machine compared with XGBoost Dataset Adult Income (UCI) Open Data science research area 04 / 10
Fairness and bias in machine learning Fairness topics audit whether models treat groups differently. The COMPAS and Adult datasets are the standard test beds.
TOPIC 16 Is the COMPAS risk score fair across race groups? Study design Bias audit Method Equalised odds and calibration by group Dataset ProPublica COMPAS Data Open TOPIC 17 Does removing sex and race from the features remove bias in income prediction? Study design Bias audit Method Fairness metrics with and without protected attributes Dataset Adult Income (UCI) Open TOPIC 18 Which bias mitigation method gives the best fairness and accuracy trade-off? Study design Experimental comparison Method Reweighing, adversarial debiasing and threshold adjustment Dataset Adult Income (UCI) Open TOPIC 19 Does a fair model stay fair when the data shifts over time? Study design Experimental comparison Method Fairness metrics across time splits Dataset ProPublica COMPAS Data Open TOPIC 20 Do recommender systems favour popular movies over niche ones? Study design Bias audit Method Popularity bias metrics on collaborative filtering Dataset MovieLens (GroupLens) Open Data science research area 05 / 10
Natural language processing NLP topics turn text into insight. Review and sentiment datasets are large, labelled and free.
TOPIC 21 Does a fine-tuned BERT beat TF-IDF with logistic regression on sentiment? Study design Model comparison Method Fine-tuned BERT compared with TF-IDF baseline Dataset IMDB Reviews (Hugging Face) Open TOPIC 22 Can review text predict the star rating a customer gives? Study design Prediction model Method Text regression with transformer embeddings Dataset Amazon Reviews 2023 Open TOPIC 23 Do word embeddings group similar products together? Study design Representation learning Method Word2Vec compared with sentence-BERT similarity Dataset Amazon Reviews 2023 Open TOPIC 24 Can a model detect abusive or toxic reviews? Study design Classification Method Transformer classifier with error analysis Dataset Amazon Reviews 2023 Open TOPIC 25 Can large language models classify sentiment with no training examples? Study design Model comparison Method Zero-shot LLM compared with a fine-tuned model Dataset IMDB Reviews (Hugging Face) Open View the printable 50-topic guide
Data science research area 06 / 10
Recommender systems Recommender topics predict what a user wants next. MovieLens is the benchmark every examiner will recognise.
TOPIC 26 Does matrix factorisation beat neighbourhood methods on MovieLens? Study design Model comparison Method SVD compared with item-based collaborative filtering (RMSE, precision at k) Dataset MovieLens (GroupLens) Open TOPIC 27 Do graph neural networks improve movie recommendations? Study design Model comparison Method LightGCN compared with matrix factorisation Dataset MovieLens (GroupLens) Open TOPIC 28 Can review text solve the cold-start problem for new products? Study design Hybrid model Method Content-based model with text embeddings Dataset Amazon Reviews 2023 Open TOPIC 29 Can association rules power a simple product recommender? Study design Transaction data Method Apriori rules with hit-rate evaluation Dataset Online Retail II (UCI) Open TOPIC 30 Does adding diversity to recommendations cost accuracy? Study design Experimental comparison Method Re-ranking with diversity and accuracy metrics Dataset MovieLens (GroupLens) Open Data science research area 07 / 10
Computer vision and medical imaging Vision topics teach models to read images. Chest X-rays are a free, well-labelled starting point.
TOPIC 31 Does transfer learning beat a CNN trained from scratch on chest X-rays? Study design Model comparison Method ResNet transfer learning compared with a custom CNN Dataset Chest X-Ray Pneumonia (Kaggle) Free login TOPIC 32 Does data augmentation help when medical images are scarce? Study design Experimental comparison Method Training with and without augmentation on data subsets Dataset Chest X-Ray Pneumonia (Kaggle) Free login TOPIC 33 Does self-supervised pretraining help when labels are few? Study design Experimental comparison Method Contrastive pretraining compared with supervised training Dataset Chest X-Ray Pneumonia (Kaggle) Free login TOPIC 34 Does a model trained on one image source work on another? Study design Domain shift study Method Cross-dataset evaluation with domain adaptation Dataset Chest X-Ray Pneumonia (Kaggle) Free login TOPIC 35 Can federated learning match central training on medical images? Study design Experimental comparison Method Federated averaging compared with centralised training Dataset Chest X-Ray Pneumonia (Kaggle) Free login Data science research area 08 / 10
Time series forecasting Forecasting topics compare classic and deep learning models on real demand data. Taxi and bike data are hourly and free.
TOPIC 36 Can an LSTM forecast hourly taxi demand better than SARIMA? Study design Model comparison Method LSTM compared with SARIMA (MAPE, RMSE) Dataset NYC TLC Trip Record Data Open TOPIC 37 Does adding weather data improve bike demand forecasts? Study design Experimental comparison Method Gradient boosting with and without weather features Dataset Bike Sharing (UCI) Open TOPIC 38 Can Prophet forecast weekly retail sales well? Study design Model comparison Method Prophet compared with ETS and ARIMA Dataset Online Retail II (UCI) Open TOPIC 39 Do taxi trip patterns show anomalies around holidays and events? Study design Anomaly detection Method Seasonal decomposition with residual-based detection Dataset NYC TLC Trip Record Data Open TOPIC 40 Can a convolutional model forecast taxi fares from trip sequences? Study design Model comparison Method Temporal CNN compared with gradient boosting Dataset NYC TLC Trip Record Data Open Data science research area 09 / 10
Clustering and customer analytics Clustering topics find groups without labels. The challenge is proving the groups are real and useful.
TOPIC 41 Does k-means or DBSCAN give more useful customer segments? Study design Unsupervised learning Method K-means compared with DBSCAN (silhouette score) Dataset Online Retail II (UCI) Open TOPIC 42 Do customer segments stay stable from one year to the next? Study design Unsupervised learning Method Gaussian mixture model with stability analysis Dataset Online Retail II (UCI) Open TOPIC 43 Do customers who churn fall into distinct groups? Study design Unsupervised learning Method Hierarchical clustering with profile analysis Dataset Telco Customer Churn (Kaggle) Free login TOPIC 44 Can topic modelling group reviews into clear complaint themes? Study design Text mining Method LDA compared with BERTopic Dataset Amazon Reviews 2023 Open TOPIC 45 Do movie fans cluster into distinct taste groups? Study design Unsupervised learning Method Spectral clustering on rating embeddings Dataset MovieLens (GroupLens) Open Data science research area 10 / 10
Machine learning in healthcare Healthcare ML topics predict outcomes for real patients. MIMIC-IV is free but needs CITI training, so start the paperwork early.
TOPIC 46 Can machine learning predict ICU death better than standard severity scores? Study design Prediction model Method XGBoost compared with SOFA score (AUC) Dataset MIMIC-IV (PhysioNet) Apply TOPIC 47 Which method handles missing lab values best in clinical prediction? Study design Methods comparison Method Multiple imputation compared with model-native handling Dataset MIMIC-IV (PhysioNet) Apply TOPIC 48 Can clinical notes predict hospital readmission? Study design NLP prediction Method Clinical BERT on discharge notes Dataset MIMIC-IV (PhysioNet) Apply TOPIC 49 Does a clinical model perform equally well across patient groups? Study design Bias audit Method Subgroup AUC and calibration analysis Dataset MIMIC-IV (PhysioNet) Apply TOPIC 50 Can an early warning model flag sepsis hours before diagnosis? Study design Prediction model Method Recurrent neural network on vital sign sequences Dataset MIMIC-IV (PhysioNet) Apply View the printable 50-topic guide
13 data science and machine learning datasets All sources have public information pages, but download requirements differ. Some datasets need a free login and MIMIC-IV requires formal access approval. Review the variables and labels before selecting a model.
Highly imbalanced credit-card transactions for fraud detection and rare-event classification.
Credit histories and default labels for risk prediction and interpretability.
Customer demographics, billing, contracts and churn labels.
Transaction data for segmentation, recommendation and forecasting.
Movie ratings and user-item interactions for recommender experiments.
Labelled movie-review sentiment for NLP evaluation.
Product reviews, metadata and ratings; toxicity annotations are not included by default.
Chest radiographs labelled for pneumonia; verify patient-level independence.
Detailed clinical records requiring credentialed access and training. Clinical notes have separate access.
Taxi trips, dates, locations and fares for forecasting and anomaly research.
Hourly and daily rentals with weather and calendar features.
Income classification and fairness analyses with US census attributes.
Criminal risk scores and outcomes for fairness audits.
Printable data science dissertation guide
Get all 50 topics in one PDF The printable guide will include the 50 research questions, study designs, machine learning methods, dataset links and related university thesis references. The download link will be added after the PDF is generated and uploaded.
PDF download coming soon How to turn one of these into your own topic Form a comparison question. Specify the task, models and practical reason for comparing them.Establish a baseline. Start with a simple, appropriate model or forecasting benchmark.Select an evaluation metric. Use precision-recall AUC for rare events, RMSE for continuous forecasts, and appropriate calibration and fairness metrics for audits.Test fairly. Hold out unseen data, prevent data leakage, tune within the training split and report uncertainty.Contribute an angle. Explainability, drift, cost or fairness can strengthen a standard benchmark.See dissertation topics using regression analysis and business and management research topics for related applied research ideas.
Need a second pair of eyes?
Choose a data science dissertation you can evaluate. Bring your topic shortlist, dataset and baseline ideas to a free 15-minute call. We can review model choice, metrics and evaluation plans.
Book a free 15-minute call 