M.Tech projects in machine learning, with base papers

We matched each paper to its Crossref DOI record on 12 November 2026; titles, journals and years appear exactly as IEEE registered them, and each card says only what the paper’s abstract says. For base papers in more domains, see the 28 IEEE base papers by domain; for M.Tech topics beyond machine learning, see our M.Tech project ideas.

What makes a good M.Tech machine learning project?

Three things: a base paper you can rebuild, one change you can measure, and results someone else could repeat.

• A base paper you can rebuild. Recent, open access, on data you can download, and clear about its split, settings and metric. Code from the authors helps a lot. Our base paper selection guide walks through every check.
• One measured extension. A single change, such as a new model, a second dataset, a robustness test, calibrated uncertainty or a lighter model, tested on the same split as the original. One change measured well is worth more than five changes nobody can separate.
• Reproducibility. Fixed random seeds, the exact dataset version, library versions written down and every run logged. Your methodology chapter is written from that log.

This is where M.Tech differs from B.Tech. A B.Tech project can end with a working demo of a known method. An M.Tech thesis has to show, in a results table, that your change did something the base paper did not.

Which machine learning projects suit M.Tech in 2026-27?

Thirteen ideas in four groups. Each card states the problem and what you build and measure. Open “The base paper and the M.Tech extension” to see the paper, linked by its DOI, then the data, the tools and one change to measure, with the table it produces.

Tabular data and time series

Most of these run on a laptop CPU. When the data is ordered in time, split it by time and never at random, or the model is tested on a past it has already seen the future of.

• Topic idea · M.Tech · Tabular and time seriesGradient boosting or deep learning on tabular data: XGBoost, TabNet and TabRProblem: Most records in Indian banks, hospitals and government offices are tables, and it is still unsettled when a deep model is worth its extra cost over gradient-boosted trees.Output: XGBoost, TabNet and the retrieval-based TabR trained on the same benchmark datasets and compared on accuracy and AUC for classification and on MAE, RMSE and R² for regression.
  The base paper and the M.Tech extension

  Base paper: Tabular Data Classification and Regression : XGBoost or Deep Learning with Retrieval-Augmented Generation, J. Pasaribu et al., IEEE Access, 2024
  Data, tools and M.Tech extension: Data and tools: the benchmark datasets listed in the paper’s experiments section; Python with XGBoost, the pytorch-tabnet package and the public TabR code from its original authors. M.Tech extension: give all three models the same tuning budget with Optuna and record training time, to test whether TabR’s lead in classification survives fair tuning; the thesis table shows every metric and the training time for each model, under the paper’s settings and under equal tuning.

• Topic idea · M.Tech · Tabular and time seriesProbabilistic household load forecasting with a quantile LSTMProblem: Indian distribution companies are fitting smart meters in homes, and a single-number load forecast hides how uncertain each household’s demand really is.Output: Households grouped by load pattern, then a quantile LSTM with a log-cosh quantile loss that gives a day-ahead load interval at 15-minute and 1-hour steps, scored on prediction interval coverage against quantile regression forest, QRNN and quantile gradient boosting.
  The base paper and the M.Tech extension

  Base paper: Enhancing Short-Term Electric Load Forecasting for Households Using Quantile LSTM and Clustering-Based Probabilistic Approach, Z. Masood et al., IEEE Access, 2024
  Data, tools and M.Tech extension: Data and tools: the abstract does not name the household dataset, so check the paper, or use a public multi-household smart-meter set such as the London smart meter data on Kaggle and say so; Python with PyTorch, scikit-learn and tslearn for the clustering. M.Tech extension: add a conformal calibration step on top of the quantile LSTM so the intervals reach their stated coverage; the thesis table gives coverage (PICP) and average interval width for every model, before and after calibration.

• Topic idea · M.Tech · Tabular and time seriesAircraft engine remaining useful life with a lighter Transformer fusion modelProblem: Maintenance planners want to know how many cycles an engine has left, and fixed service intervals ignore how each engine is actually wearing.Output: A Transformer with CNN layers in its encoder and LSTM layers in its decoder that predicts the remaining cycles of each engine from its sensor history, scored on RMSE and the C-MAPSS score against plain LSTM and CNN models.
  The base paper and the M.Tech extension

  Base paper: Novel Transformer-based Fusion Models for Aero-engine Remaining Useful Life Estimation, Q. Hu et al., IEEE Access, 2023
  Data, tools and M.Tech extension: Data and tools: NASA’s C-MAPSS turbofan simulation data, in four subsets from FD001 to FD004; Python with PyTorch. M.Tech extension: prune the fusion model, or replace its full attention with a lighter one, to halve its parameters and measure what accuracy it gives up; the thesis table lists RMSE, score, parameter count and inference time on all four subsets for an LSTM baseline, the paper’s model and your lighter version.

Deep learning for health data

Keep each patient’s records in one split only. A model that has seen other heartbeats or EEG windows from the same patient in training will look better than it is.

• Topic idea · M.Tech · Health dataHeartbeat arrhythmia classification that holds up in noisy ECGProblem: Rural clinics in India can record an ECG but often have no cardiologist on hand, so a screening model must run on cheap hardware and still catch rare abnormal beats.Output: A two-layer 1D CNN with squeeze-and-excitation attention that labels each heartbeat from per-beat features, with SMOTE on the training data only, reported as per-class precision and recall, the parameter count and the INT8 model size.
  The base paper and the M.Tech extension

  Base paper: A Lightweight Channel-Attention CNN for Robust Beat-Level Arrhythmia Detection, M. S. Tahsin et al., IEEE Access, 2026
  Data, tools and M.Tech extension: Data and tools: the four PhysioNet databases the paper tests on, which are MIT-BIH Arrhythmia, MIT-BIH Supraventricular Arrhythmia, INCART-2 and the Sudden Cardiac Death Holter database; Python with the wfdb package, and TensorFlow Lite or PyTorch for the quantisation. M.Tech extension: add electrode-motion and muscle noise from PhysioNet’s MIT-BIH Noise Stress Test Database at fixed signal-to-noise ratios, then train with noise augmentation; the thesis table gives macro F1 for the paper’s model and your noise-trained model on clean records and at each noise level.

• Topic idea · M.Tech · Health dataEpileptic seizure detection from raw EEG with fewer false alarmsProblem: Reading a long EEG recording for seizures by eye takes a neurologist hours, and a detector that raises frequent false alarms adds to the work instead of cutting it.Output: A CNN that takes raw EEG straight in and looks at it at several scales through inception and residual blocks, with no feature extraction step, tested per subject and across subjects on the Bonn and CHB-MIT datasets, with F1-score as the main measure.
  The base paper and the M.Tech extension

  Base paper: Less Parameterization Inception-Based End to End CNN Model for EEG Seizure Detection, K.-K. Shyu et al., IEEE Access, 2023
  Data, tools and M.Tech extension: Data and tools: the Bonn EEG recordings, which are public, and CHB-MIT, the children’s scalp EEG database on PhysioNet; MNE reads its EDF files into Python, and PyTorch trains the model. M.Tech extension: add post-processing that raises an alarm only after several consecutive seizure windows, and report the measures a clinician asks for; the thesis table shows sensitivity, false alarms per hour and detection delay for each CHB-MIT patient, with and without the post-processing.

• Topic idea · M.Tech · Health dataExplainable diabetes risk prediction: comparing LIME and SHAPProblem: Diabetes is common in India, and a doctor will not act on a risk score unless the model can show which of the patient’s answers raised it.Output: Logistic regression and random forest models that predict diabetes from health survey answers, with LIME and SHAP explanations for single patients and for the whole model, and a comparison of where the two explainers agree and differ.
  The base paper and the M.Tech extension

  Base paper: A Comparative Analysis of LIME and SHAP Interpreters With Explainable ML-Based Diabetes Predictions, S. Ahmed et al., IEEE Access, 2025
  Data, tools and M.Tech extension: Data and tools: the 253,680 survey responses the abstract describes, with the source named in the paper; Python with scikit-learn and the lime and shap packages. M.Tech extension: measure how stable the explanations are by rerunning LIME and SHAP with different seeds and background samples, and add a gradient-boosted model; the thesis table gives accuracy and recall for each model beside how often its top five features stay the same across runs and across the two explainers.

NLP and large language models

The classical NLP ideas run on a CPU; the LLM ideas need a GPU. State the model size and the GPU you used in the thesis, because examiners ask.

• Topic idea · M.Tech · NLP and LLMsSentiment analysis of Hindi movie reviews: a stacked ensemble against a transformerProblem: Indian film fans write many reviews in Hindi, yet most sentiment tools are built for English and Hindi has few labelled datasets.Output: Tuned classical classifiers on TF-ISF n-gram features, combined in a stacked ensemble that labels a Hindi review positive or negative, compared with each classifier alone on three review datasets.
  The base paper and the M.Tech extension

  Base paper: Toward Machine Learning Based Binary Sentiment Classification of Movie Reviews for Resource Restraint Language (RRL)—Hindi, A. Sharma and U. Ghose, IEEE Access, 2023
  Data, tools and M.Tech extension: Data and tools: the authors’ Hindi Language Movie Review set of 10,000 reviews (ask whether they share it) and the public IIT Patna movie and product review sets; Python with scikit-learn and the Indic NLP Library. M.Tech extension: fine-tune a multilingual transformer such as MuRIL or IndicBERT on the same splits; the thesis table compares accuracy, precision, recall, F1 and training time for the stacked ensemble and the transformer on each dataset.

• Topic idea · M.Tech · NLP and LLMsDetecting AI-written text with LoRA-tuned open LLMs, in English and HindiProblem: Colleges in India now receive assignments written with chatbots, and zero-shot detectors give unstable results.Output: Open LLMs fine-tuned with LoRA, QLoRA and PLoRA to label a text as human- or machine-written, compared with the same models used zero-shot on accuracy, memory use and training time.
  The base paper and the M.Tech extension

  Base paper: Enhancing Text-Based Deepfake Detection: A Comprehensive Evaluation of LoRA Fine-Tuned Large Language Models, A. Mohammed and R. Kora, IEEE Access, 2026
  Data, tools and M.Tech extension: Data and tools: the four multilingual benchmarks named in the paper; Python with Hugging Face Transformers and PEFT, using small open models with QLoRA on a free Colab or Kaggle GPU. M.Tech extension: build a Hindi and Hinglish test set from human-written text and from text you generate with two LLMs, and measure how a detector tuned on the benchmarks holds up; the thesis table gives accuracy and F1 for zero-shot, LoRA and QLoRA on the original test sets and on your Hindi and Hinglish sets.

• Topic idea · M.Tech · NLP and LLMsRAG, fine-tuning or both: how much the retriever decides the answerProblem: Fine-tuning an LLM on a university’s rule book costs GPU time, retrieval adds a step to every answer, and few projects measure which one actually answers better.Output: Retrieval-augmented generation, fine-tuning and a hybrid of the two, built on one open model and compared on SQuAD, MS MARCO and SQL CREATE TABLE questions for answer quality and compute cost.
  The base paper and the M.Tech extension

  Base paper: Unveiling the Power of Large Language Models: A Comparative Study of Retrieval-Augmented Generation, Fine-Tuning, and Their Synergistic Fusion for Enhanced Performance, G. Budakoglu and H. Emekci, IEEE Access, 2025
  Data, tools and M.Tech extension: Data and tools: the public SQuAD and MS MARCO question sets, plus the SQL CREATE TABLE statements the authors also tested on; Python with Hugging Face Transformers, PEFT for the fine-tuning and FAISS for the vector index. M.Tech extension: swap the retriever between BM25, dense embeddings and a hybrid of both, and trace how retrieval quality changes the final answer; the thesis table gives the retriever’s recall at 5 beside exact match, F1 and time per question for each RAG variant, fine-tuning and the hybrid.

• Topic idea · M.Tech · NLP and LLMsModerating code-mixed Dravidian comments: does transliteration help?Problem: Comments under Indian videos and posts often mix Tamil, Malayalam or Kannada with English in Roman script, and moderation tools built for English miss the abuse in them.Output: Embeddings from seven multilingual models, including MuRIL, LaBSE and IndicBERT, fed to classical classifiers with higher weights for the rare classes, compared on six code-mixed datasets.
  The base paper and the M.Tech extension

  Base paper: Detection of Hate Speech and Offensive Language CodeMix Text in Dravidian Languages Using Cost-Sensitive Learning Approach, K. Sreelakshmi et al., IEEE Access, 2024
  Data, tools and M.Tech extension: Data and tools: the six public sets the paper uses, from DravidianLangTech 2021, HASOC 2020 and HASOC 2021, plus the authors’ new Malayalam-English test set; Python with Hugging Face Transformers and scikit-learn. M.Tech extension: transliterate the Roman-script comments into Tamil, Malayalam or Kannada script with an open tool such as IndicXlit before embedding, and measure the effect; the thesis table gives macro F1 for each dataset on the original text and on the transliterated text.

Trustworthy and efficient ML

These ideas ask how far a model can be trusted and how cheaply it runs. That makes the extension easy to measure: coverage, robustness, trainable parameters, memory or time per prediction.

• Topic idea · M.Tech · Trust and efficiencyHER2 breast cancer scoring that says when it is unsure, with conformal predictionProblem: HER2 status decides whether a breast cancer patient is offered targeted therapy, and a classifier that gives a confident wrong answer on an equivocal slide does more harm than one that asks for a confirmatory test.Output: Colour, local binary pattern and Haralick features with tree-based classifiers, wrapped in conformal prediction so that uncertain images get a set of labels and are referred for testing, measured by accuracy on the decided cases and the share referred.
  The base paper and the M.Tech extension

  Base paper: Conformal Prediction for Uncertainty Quantification and Reliable HER2 Status Classification in Breast Cancer IHC Images, S. Pintawong et al., IEEE Access, 2025
  Data, tools and M.Tech extension: Data and tools: the paper’s 2,570 IHC images from 94 cases come from one pathology institute in Bangkok, so check whether they are shared, and the authors’ code is on GitHub; Python with scikit-learn, XGBoost and the MAPIE conformal prediction library. M.Tech extension: replace the handcrafted features with features from a pre-trained CNN and keep the conformal step the same; the thesis table gives accuracy on the decided cases and the referral rate at two significance levels for both feature sets.

• Topic idea · M.Tech · Trust and efficiencyHow easily medical image classifiers are fooled, and what one defence costsProblem: Hospitals are starting to use deep learning to read scans, and a small change to an image that no doctor would notice can flip the model’s answer.Output: FGSM and PGD attacks applied to whole images and to patches of different sizes, run against four deep learning models on three medical image datasets, with accuracy recorded at each attack strength.
  The base paper and the M.Tech extension

  Base paper: The Impact of Simultaneous Adversarial Attacks on Robustness of Medical Image Analysis, S. Pal et al., IEEE Access, 2024
  Data, tools and M.Tech extension: Data and tools: the three medical datasets and four models named in the paper’s experiments section; Python with PyTorch and an attack library such as torchattacks or the Adversarial Robustness Toolbox. M.Tech extension: add adversarial training with PGD examples to one model and measure what it costs in clean accuracy and training time; the thesis table gives clean accuracy and accuracy under FGSM and PGD at each attack strength, before and after the defence.

• Topic idea · M.Tech · Trust and efficiencyLung nodule classification with LoRA: finding the smallest adapter that worksProblem: Fully fine-tuning a large pre-trained vision model on lung CT nodules needs more GPU memory and labelled scans than most M.Tech projects have.Output: A large vision model pre-trained with self-supervised learning, adapted with LoRA to label nodules benign or malignant, and compared with full fine-tuning on ROC AUC, trainable parameters and training time.
  The base paper and the M.Tech extension

  Base paper: Low-Rank Adaptation of Pre-Trained Large Vision Models for Improved Lung Nodule Malignancy Classification, B. P. Veasey and A. A. Amini, IEEE Open Journal of Engineering in Medicine and Biology, 2025
  Data, tools and M.Tech extension: Data and tools: LIDC, which is public on The Cancer Imaging Archive, and NLSTx, whose annotations the authors released on GitHub; Python with PyTorch and the PEFT library. M.Tech extension: vary the LoRA rank and the layers that get adapters, and find the smallest setting that keeps the result; the thesis table gives ROC AUC, trainable parameters, GPU memory and training time for full fine-tuning and for LoRA at three ranks.

Prefer to start from a dataset rather than a paper? See our projects with Kaggle datasets, then look for a recent IEEE paper on the same data.

How does an M.Tech machine learning project run, stage by stage?

Most universities split the M.Tech project into two stages across the second year. The names differ, such as Phase I and Phase II or Dissertation Part I and Part II, but the work falls the same way.

Stage I: literature and reproduction

1. Choose the base paper and confirm that its data is available and the compute fits your time.
2. Read around it. Collect the papers it builds on and the ones that cite it, and note what each one changed. This becomes your literature survey.
3. Reproduce the main result. Match the dataset, split, preprocessing and metric, and put your number beside the paper’s. Our guide on how to implement a base paper covers each step and what to do when the numbers differ.
4. Propose the extension. Name the one change, the metric and the table you will fill in Stage II.
5. Present Stage I. The review panel wants the paper, your reproduced result and the plan. The base paper presentation guide shows the slides to make.

Stage II: extension, results, thesis and paper

1. Build the extension and run it on the same split as your reproduction.
2. Fill the comparison table: the paper’s number, your reproduction and your extension on the same metric, with more than one random seed where training is unstable.
3. Write the thesis. The results chapter is built around that table. Our M.Tech thesis format guide gives the chapter order and the formatting.
4. Write a paper if your university or guide asks for one, usually in the IEEE two-column layout. Our IEEE paper format guide lists the rules.
5. Prepare for the viva. Be ready to explain every number in your table, and why your change helped or did not.

Which datasets and how much compute do you need?

Every idea here uses public data or data the authors describe in their paper. Where a card says to check the paper, do that before you commit, because the project cannot start without its data.

• Where public datasets live. PhysioNet for ECG and EEG, The Cancer Imaging Archive for CT scans, Kaggle and the UCI Machine Learning Repository for tabular data, and Hugging Face Datasets for text. Read each dataset’s licence and cite it.
• Free GPUs. Google Colab and Kaggle notebooks give free GPU time, with limits on session length and weekly use that change from time to time. Save checkpoints to Drive or a Kaggle dataset, so a dropped session does not cost you a day’s training.
• LLM projects. Use small open models, load them in 4-bit with QLoRA, and state in the thesis which model size and GPU you used.
• When you need more. 3D CT work, large ensembles and long fine-tuning runs are easier on a college GPU server or a paid cloud GPU rented for a few days.

Compute is our estimate for an M.Tech-sized run, not a figure from the papers. Check each dataset’s licence and access terms before you commit.

How can we help with your M.Tech machine learning project?

We help M.Tech students in CSE, IT and AI from Stage I to the final submission. We check the base paper and its data first, reproduce the main result, then design one extension, run it on the same split and write it up. You get the code and results, the thesis in your university’s format, an IEEE-format paper if your guide wants one, review presentations and viva preparation, delivered stage by stage to match your reviews.

Start with our M.Tech projects for CSE and AI page. It sets out the scope, the two stages and what you receive. The consultation is free, and you get a written plan and a fixed quote.

More M.Tech domains, each built on verified IEEE base papers: M.Tech projects in image processing, M.Tech projects in embedded systems and IoT, M.Tech projects in cyber security. For the full service, see our M.Tech projects for CSE and AI.

Frequently asked questions