50 machine learning dissertation topics organised by method, from tree ensembles and imbalanced classification to CNNs, forecasting, clustering, interpretable AI, transfer learning and edge or federated systems.
By Siddharth Gupta · Updated November 2026
A strong machine learning dissertation makes an evidence-based comparison, not just a complex model. A simple baseline, a leakage-safe validation strategy, appropriate performance metrics and reproducible data preparation often matter more than adding architecture complexity.
This method-led collection provides a question, proposed study design, suggested model comparison, starting dataset and related university thesis for each topic. The theses illustrate related research, not necessarily the same algorithm or dataset. Access ranges from open downloads to registration and approval.
| Your data or research problem | Starting method family |
|---|---|
| Tabular, structured rows and columns | Tree ensembles for tabular data |
| A rare event or anomaly class | Imbalanced data and anomaly detection |
| Images or visual classification | Deep learning for images |
| Clinical datasets and patient prediction | Machine learning for health prediction |
| Signals, sequences or forecasts | Sequence models and forecasting |
| Unlabelled observations | Clustering and representation learning |
| High-dimensional features | Feature selection and dimensionality reduction |
| Need explanations, uncertainty or group fairness | Interpretability, uncertainty and fairness |
| Limited labelled images | Transfer learning and generative models |
| Deployment efficiency, multiple sites or sequential decisions | Efficient, federated and reinforcement learning |
Open means publicly accessible. Free login requires registration; Apply requires approval. Model training needs vary with dataset size and architecture.
Most real data is a table. Random forests and gradient boosting are the strongest place to start.
Design, Development and Benchmarking of Machine Learning Algorithms in Biomedical ApplicationsQi Sun · University of Kentucky · 2022 · Related research example
Feasibility note: The public Home Credit competition does not include a reliable application date for straightforward out-of-time validation. This study compares reproducible cross-validation strategies instead.
Temporal Machine Learning for Predicting Accidents and Violations in the Mining IndustryNathan T. Kelley · University of Kentucky · 2025 · Related research example
Application Of Machine Learning in Transportation EngineeringAli Gorji Sefidmazgi · University of Kentucky · 2024 · Related research example
Nondestructive Multivariate Classification of Codling Moth Infested Apples Using Machine Learning and Sensor FusionNader Ekramirad · University of Kentucky · 2022 · Related research example
Machine Learning and Bioinformatic Insights into Key Enzymes for a Bio-Based Circular EconomyJapheth E. Gado · University of Kentucky · 2021 · Related research example
When the class you care about is rare, accuracy lies. Use precision, recall and the PR curve instead.
Fault Identification on Electrical Transmission Lines Using Artificial Neural NetworksChristopher W. Asbery · University of Kentucky · 2020 · Related research example
Feasibility note: Use a patient-level split: heartbeat-level random splitting can leak characteristics of the same patient into train and test sets.
Mispronunciation Detection and Diagnosis in Mandarin Accented English SpeechSubash Khanal · University of Kentucky · 2020 · Related research example
Prediction of Blast-Induced Ground Vibrations: A Comparison Between Empirical and Artificial-Neural-Network ApproachesLuis F. Velasquez · University of Kentucky · 2023 · Related research example
Rule Mining and Sequential Pattern Based Predictive Modeling with EMR DataOrhan Abar · University of Kentucky · 2019 · Related research example
Harmful algal bloom prediction modeling in the face of a changing climateJordan E. Spears · University of Kentucky · 2025 · Related research example
CNN topics suit students with Python skills. Small benchmark sets train on a laptop.
Image-Based Roadway Assessment Using Convolutional Neural NetworksWeilian Song · University of Kentucky · 2019 · Related research example
Mammogram and Tomosynthesis Classification Using Convolutional Neural NetworksXiaofei Zhang · University of Kentucky · 2017 · Related research example
An Optical Character Recognition Engine for Graphical Processing UnitsJeremy Reed · University of Kentucky · 2016 · Related research example
Image Geo-localization with Cross-AttentionConnor Greenwell · University of Kentucky · 2022 · Related research example
Leveraging Overhead Imagery for Localization, Mapping, and UnderstandingScott Workman · University of Kentucky · 2018 · Related research example
Health ML topics need careful validation. Use patient-level splits and report calibration.
Multi-stream Longitudinal Data Analysis using Deep LearningSajjad Fouladvand · University of Kentucky · 2021 · Related research example
Feasibility note: Free-text discharge notes are in MIMIC-IV-Note, a separate credentialed resource requiring its own data-use agreement.
Multi-Modal Medical Imaging Analysis with Modern Neural NetworksGongbo Liang · University of Kentucky · 2020 · Related research example
Novel Machine Learning and Wearable Sensor Based Solutions for Smart Healthcare MonitoringRajdeep Kumar Nath · University of Kentucky · 2021 · Related research example
Novel Applications of Machine Learning in BioinformaticsYi Zhang · University of Kentucky · 2019 · Related research example
Feasibility note: Diagnosis-code prediction from discharge notes additionally needs MIMIC-IV-Note, with separate access approval. Verify code-label alignment by admission.
Deep Neural Networks for Multi-Label Text Classification: Application to Coding Electronic Medical RecordsAnthony Rios · University of Kentucky · 2018 · Related research example
Sequence topics use signals over time. Compare deep models with simple baselines before you claim a win.
Precision Meteorological Prediction Employing A Data-Driven, Adaptive, Real-Time (DART) ApproachSujit Sinha · University of Kentucky · 2023 · Related research example
Forecasting the Workload with a Hybrid Model to Reduce the Inefficiency CostXinwei Pan · University of Kentucky · 2017 · Related research example
A recurrent neural network architecture for biomedical event trigger classificationJeevith Bopaiah · University of Kentucky · 2018 · Related research example
Feasibility note: This UCI dataset includes fixed-length pre-windowed sensor sequences. Testing arbitrary new window lengths may require original unsegmented recordings instead.
Accelerometer-Based Vigilance State Classification in Dairy CowsEvan King · University of Kentucky · 2021 · Related research example
Feasibility note: Compare the provided engineered feature tables with the linked inertial-signal files; keep all windows from each subject in the same validation split.
Machine Learning Approach for Vigilance State Classification in MiceAnik Muhury · University of Kentucky · 2021 · Related research example
Unsupervised topics find structure without labels. Always test whether clusters are stable.
Difs and Bayescluster: Novel Methods for Single_cell RNA Sequencing AnalysisKun Liu · University of Kentucky · 2024 · Related research example
Clustering States to Improve the Strategic Highway Safety PlanVictoria Cambron · University of Kentucky · 2023 · Related research example
Neural Representations of Concepts and Texts for Biomedical Information RetrievalJiho Noh · University of Kentucky · 2021 · Related research example
Biomedical Word Sense Disambiguation with Neural Word and Concept EmbeddingsAKM Sabbir · University of Kentucky · 2016 · Related research example
Estimating Free-Flow Speed with LiDAR and Overhead ImageryArmin Hadzic · University of Kentucky · 2020 · Related research example
With many features and few rows, selection matters more than the model. These topics test that.
Scalable Feature Selection and Extraction with Applications in Kinase PolypharmacologyDerek Jones · University of Kentucky · 2018 · Related research example
Relation Prediction over Biomedical Knowledge Bases for Drug RepositioningMehmet Bakal · University of Kentucky · 2019 · Related research example
End-To-End Prediction of Weld Penetration in Real Time Based on Deep LearningWenhua Jiao · University of Kentucky · 2020 · Related research example
Image Processing and Machine Learning using LiDAR and 360 Camera in Underground MinesAna Fabiola Patricia Tejada Peralta · University of Kentucky · 2025 · Related research example
Machine Learning for Advancing Automation and Quality Control in Robotic WeldingJoseph Kershaw · University of Kentucky · 2023 · Related research example
Examiners now ask why a model predicts what it does. These topics open the black box.
Accuracy and Uncertainty in Traffic and Transit Ridership ForecastsJawad Mahmud Hoque · University of Kentucky · 2022 · Related research example
Using Natural Language Processing and Machine Learning Techniques to Characterize Configuration Bug Reports: A StudyWei Wen · University of Kentucky · 2017 · Related research example
Deep Neural Architectures for End-to-End Relation ExtractionTung Tran · University of Kentucky · 2020 · Related research example
On Generative Models and Joint Architectures for Document-level Relation ExtractionAviv Brokman · University of Kentucky · 2024 · Related research example
Feasibility note: The UCI Adult dataset does not identify US states. Age-group transfer testing uses variables actually present in the dataset.
Context Classification of Roadways Using Nationally Available GIS DataLaura Wright · University of Kentucky · 2021 · Related research example
Transfer learning lets you train good models on little data. Generative models create new data.
Deep Learning Models for CT Image StandardizationMd Selim · University of Kentucky · 2023 · Related research example
Feasibility note: GANs and diffusion models can be computationally intensive. Use small models and reduced training budgets for a feasible student project.
Advancing Generative Methods for Multimodal Data AnalysisRabeya Tus Sadia · University of Kentucky · 2026 · Related research example
Weakly Supervised Learning for Multi-Image SynthesisMuhammad Usman Rafique · University of Kentucky · 2021 · Related research example
A Machine Learning Approach to Artificial Floorplan GenerationGenghis Goodman · University of Kentucky · 2019 · Related research example
Transfer Learning-Enhanced Transformer for Virtual Sensing Applications in Resistance Spot WeldingEthan York · University of Kentucky · 2024 · Related research example
These topics study how ML runs in the real world, on devices, across sites and through trial and error.
Feasibility note: Simulated client partitions on Fashion-MNIST support an experimental federated benchmark; they do not represent real hospital or phone networks.
Federated Data Engineering and Learning for Edge Intelligence SystemsAfsaneh Mahanipour · University of Kentucky · 2026 · Related research example
Advancing Edge Intelligence: Strategies for Improving Machine Learning on Embedded DevicesPrakash Dhungana · University of Kentucky · 2026 · Related research example
Feasibility note: M5 contains unit sales and prices, not observed replenishment decisions, stockouts or holding costs. Inventory-control claims must rely on explicitly stated simulation assumptions.
A Comparative Analysis of Reinforcement Learning Applied to Task-Space Reaching With a Robotic Manipulator With and Without Gravity CompensationJonathan Fugal · University of Kentucky · 2020 · Related research example
Post-Capture Synthesis of Images Using Manipulable Integration FunctionsPaul Eberhart · University of Kentucky · 2024 · Related research example
Feasibility note: UCI sensor recordings support model-size and inference-speed benchmarking, but true phone latency requires testing on an actual device.
3D Body Tracking using Deep LearningQingguo Xu · University of Kentucky · 2017 · Related research example
Open datasets are publicly accessible, Free login requires an account and Apply requires approval. Some Kaggle competitions have licensing conditions; check them before research use.
Thousands of ready-to-use machine learning datasets with standard tasks and benchmark results. Use it to compare algorithms across many datasets at once.
70,000 small greyscale images of clothing in 10 classes. Use it for image classification, autoencoder and generative model topics.
60,000 colour images in 10 classes, the standard small image benchmark. Use it for CNN, transfer learning and model compression topics.
48 half-hour ECG recordings with every heartbeat labelled by cardiologists. Use it for sequence models and rare-class detection topics.
569 tumour samples with 30 cell features and a benign or malignant label. Use it for small-sample classification and feature selection topics.
Loan applications with hundreds of features and a default label. Use it for gradient boosting, feature engineering and fairness topics.
48,000 census records with income above or below 50,000 dollars. Use it for tree ensemble, interpretability and fairness topics.
284,000 card transactions where only 0.17 percent are fraud. Use it for imbalanced data and anomaly detection topics.
US health survey with lab tests, body measures and diagnoses for thousands of adults. Use it for health prediction and clustering topics.
Hospital records with vitals, labs, notes and diagnosis codes for over 300,000 patients. Use it for clinical prediction and medical coding topics. Free-text notes require the separately approved MIMIC-IV-Note dataset.
Two years of transactions from a UK online retailer. Use it for customer clustering and embedding topics.
Daily unit sales for 3,049 Walmart products over five years. Use it for machine learning forecasting topics.
Phone accelerometer and gyroscope signals from 30 people doing six activities. Use it for sensor sequence and on-device model topics.
A printable list is being prepared. Meanwhile, all 50 questions, research designs, methods, datasets and related theses are available on this page.
PDF download coming soonFor broad project ideas, see the general data science collection. This page focuses specifically on machine learning methodology and benchmark design.
Bring your shortlist, data and planned evaluation strategy to a free 15-minute consultation about baselines, model choice, leakage and methodology.
Book a free 15-minute callAsk about my machine learning analysis on WhatsApp
15 minutes · No cost · No pressure
A good topic answers a clear research question with accessible data, a relevant baseline and fair evaluation. Imbalanced classification, interpretability and transfer learning are useful areas to explore.
Not always. Many tabular benchmarks and small image classifiers run on a laptop or hosted notebook, but diffusion models, transformers, large hyperparameter searches and clinical datasets can require more computing or memory.
Choose a simple task-appropriate baseline, such as logistic regression or a decision tree for classification, and then compare a few stronger candidates such as random forest, gradient boosting or a neural network when justified.
Use train/validation/test splits or cross-validation appropriate for the data. For imbalanced classification consider PR-AUC, precision, recall, F1 and calibration. Forecasting and clustering need different metrics.
Python with scikit-learn, XGBoost and PyTorch is widely used. R with tidymodels is also suitable for statistical learning and reproducible model evaluation.
The printable Topic and Method list is being prepared. Every topic, method and dataset link is currently available on this page.