University of Turku Data Analytics Master’s Thesis Guide: DTEK1002, DTEK1104, 30 ECTS and UTUGradu

Quick answer: what is the University of Turku Data Analytics thesis route?

The current Data Analytics specialisation track in the University of Turku Master’s Degree Programme in Information and Communication Technology is a 120 ECTS, two-year Master of Science (Technology) programme in the Faculty of Technology. The controlling curriculum object is ICTDATA2427 / programme 99359. The exact thesis is DTEK1002 Master’s Thesis in Technology, 30 ECTS, graded 0-5, with TTDK1308, 0 ECTS as the maturity examination. The current track-specific thesis seminar is DTEK1104 Master’s Thesis in Technology Seminar, Data Analytics, 5 ECTS, Pass/Fail, with a live Fall 2026-Spring 2027 Moodle instance.

1. Start from ICTDATA2427 / programme 99359

Use the current 2024-2027 Peppi object for course-level planning. The public programme page gives the correct high-level degree structure, while Peppi resolves the exact modules and course objects. Do not copy the Cyber Security or Cryptography packaging simply because all belong to the ICT parent programme.

2. The degree is 120 ECTS

The public programme page defines the degree as 120 ECTS over two academic years. Peppi can display a 115-125 ECTS root selection range because some minor and other-study blocks have minimum and maximum ranges. That technical range does not replace the formal 120 ECTS degree requirement.

3. Public and Peppi advanced-study totals align cleanly here

The public Data Analytics page gives 50 ECTS advanced-level studies in the major subject plus a 30 ECTS thesis. Peppi models an 80 ECTS Advanced Studies module as 20 ECTS Data Analytics Core, 10 ECTS Data Analytics Advanced, 20 ECTS Common Studies in ICT and 30 ECTS thesis. Unlike Cyber Security, there is no 40-versus-42 ECTS major-content discrepancy to resolve.

4. Know the 50 ECTS pre-thesis advanced structure

The 20 ECTS Core Module contains Data Analysis and Knowledge Discovery, Machine Learning and Pattern Recognition, Statistical Data Analysis and Evaluation of Machine Learning Methods. The 10 ECTS Advanced Module contains Introduction to Deep Learning and Algorithm Design. The 20 ECTS Common Studies block includes Knowledge and Innovation Management, language study and a 10 ECTS choice between Capstone and Lean Digital Business Design.

5. DTEK1104 is separate from the 30 ECTS thesis

Peppi places DTEK1104 in Other Studies, not inside the compulsory 30 ECTS thesis category and not inside the 20 ECTS Common Studies category. DTEK1002 nevertheless requires presentation in a related Master’s thesis seminar. Treat DTEK1104 as the current Data Analytics thesis-seminar process, but confirm its exact placement in your HOPS rather than counting it inside DTEK1002.

6. DTEK1002 is the exact thesis course

DTEK1002 Master’s Thesis in Technology is 30 ECTS, Advanced Studies, available in Finnish or English and graded 0-5. It requires scientific work, research-method mastery, knowledge of the field and scientific writing. The public Data Analytics page adds an engineering framing: identify a relevant problem, propose/design a solution, test and analyse it, then evaluate its usefulness.

7. Build the plan with supervisors before modelling

DTEK1002 requires a thesis plan made with assigned supervisor or supervisors. A company can commission the work, a company representative can co-supervise, or the thesis can sit inside a University research group. For data analytics, the plan should specify the research question, unit of analysis, dataset/provenance, train-validation-test logic, statistical or predictive estimand, primary metrics, privacy constraints and reproducibility strategy before extensive model tuning begins.

8. Know the examiner and approval route

At least two examiners evaluate DTEK1002 using University grading guidance. Final acceptance and grading are decided by the head of the department based on those evaluations. Seminar completion, maturity-test completion and HOPS placement are related administrative obligations but should not be confused with examiner grading of the 30 ECTS thesis itself.

9. TTDK1308 is the 0 ECTS maturity examination

The thesis category also contains TTDK1308 Degree Qualifying Examination for Master’s Degree, 0 ECTS. Normally the thesis abstract or another suitable section serves as the maturity test. A conditional written-exam route can apply depending on prior degree and Finnish/Swedish educational-language background, so verify your individual route.

10. DTEK1104 is a real current 2026-2027 seminar

DTEK1104 is 5 ECTS, Advanced Studies and Pass/Fail. Its Peppi object describes thesis-structure analysis, field-specific research methods, a previous-thesis analysis presentation, the student’s own thesis presentation, peer participation and a research-methods exercise. The live Moodle instance is explicitly Fall 2026-Spring 2027 and names Csaba Raduly-Baka as teacher.

11. Use the seminar to challenge the evaluation design

The seminar should help you test whether the method can actually answer the question. For a model-based thesis, present not just architecture diagrams but the data split, target population, baseline, tuning procedure and final evaluation. For statistical work, present assumptions and the inferential target. For algorithmic work, separate correctness, efficiency and empirical performance.

12. Write the prediction or inference question before choosing a model

A common weak workflow is to try many models first and formulate the research question afterward. Reverse that order. State whether you are estimating an association, testing a hypothesis, predicting new observations, comparing algorithms, estimating uncertainty or designing a system. The evaluation design should follow from that question.

13. Data leakage is a research-validity problem

Preprocessing can leak information even when the final classifier never directly sees the test labels. Scaling, imputation, feature selection, dimensionality reduction or representation learning fitted on the full dataset can transfer test-set information into training. Fit data-dependent preprocessing inside the training process and apply the learned transformation to validation/test data.

14. Match the split to the deployment question

TKO_7092 explicitly teaches that different generalisation questions require different resampling designs. New measurements from already-seen subjects are not the same as entirely new subjects. Nearby geospatial observations are not the same as new distant locations. Drug-target or other pairwise problems can have several distinct cold-start cases. Define the intended future case before choosing folds.

15. Separate tuning from final performance estimation

If hyperparameters are selected on the same evidence later reported as the final test performance, the estimate can be optimistically biased. Use a separate validation set or a nested resampling design when model selection and performance estimation must both be done from limited data. Record the selection rule before looking at the final evidence where practical.

16. Keep the final test set genuinely final

Repeatedly checking a held-out test set and then changing features, thresholds or architecture turns it into another development set. Preserve one final evaluation surface that is not used to make modelling decisions. If this is impossible because the dataset is very small, state the limitation and use a resampling design appropriate to the question.

17. Overfitting and underfitting are different failure modes

TKO_3120 explicitly teaches both. Overfitting means the model captures training-specific patterns that do not generalise; underfitting means the model is too limited or poorly specified to capture the relevant structure. Training loss alone cannot diagnose generalisation. Compare training and validation behaviour and use the final evaluation only after the modelling choices are fixed.

18. Always include a meaningful baseline

A sophisticated model is not automatically a useful contribution. Compare against a simple reference appropriate to the task: majority or prevalence baseline, linear/logistic model, simple tree, established algorithm or current operational method. Use the same split and metric so any claimed gain is attributable to the method rather than a different evaluation protocol.

19. Choose metrics from the decision problem

For imbalanced classification, accuracy can hide severe failure on the minority class. Define the positive class and explain why precision, recall, specificity, F-score, ROC-AUC, PR-AUC, calibration or another metric fits the research use. False positives and false negatives can carry very different costs, so a single headline score can be insufficient.

20. Calibration and discrimination are not the same

A model can rank cases well while its predicted probabilities are systematically too high or too low. If the thesis interprets probabilities as risks or decision inputs, check calibration rather than only ranking metrics. Report uncertainty around performance estimates where feasible, especially when the sample is small or the split is unstable.

21. Predictive performance is not statistical significance

TKO_7093 and TKO_7092 support different parts of this distinction. A statistically significant association does not guarantee useful out-of-sample prediction, and strong prediction does not by itself identify a causal mechanism. Write conclusions using the evidence type actually produced by the analysis.

22. Distinguish exploratory and confirmatory analysis

Exploratory plots, correlations and subgroup checks are valuable for understanding data, but repeated inspection can influence what is later tested. If the same dataset is used to generate and confirm a hypothesis, state that limitation or use a new confirmatory sample where feasible. Do not present a post-hoc discovery as if it was fully prespecified.

23. Statistical assumptions belong in the method section

State why the chosen test or model is suitable for the data structure. Check relevant distributional, independence or variance assumptions rather than treating a software default as the method. When observations are repeated within subjects, locations, organisations or time periods, account for the dependence relevant to the claim.

24. Multiple comparisons and model search need disclosure

Trying many variables, transformations, subgroups or models raises the chance of a favourable result appearing by selection. Record the search process and use appropriate correction, validation or independent confirmation where the inferential claim requires it. At minimum, separate exploratory search from the final evidence used to support the conclusion.

25. Deep-learning reproducibility requires more than a model name

For TKO_7094-style work, preserve architecture, optimiser, learning rate, preprocessing, regularisation, stopping rule, batch settings and random-seed policy. Record framework/library versions and compute environment. One favourable run should not be treated as deterministic if stochastic training materially affects the result.

26. Transfer learning needs a traceable starting point

Identify the pretrained model or immutable source, dataset/domain where known, and what was frozen or fine-tuned. If pretrained representations may overlap with the evaluation domain, discuss possible contamination or prior exposure where it matters. Separate gains from pretraining from gains attributable to your own method.

27. Algorithm correctness and efficiency are separate claims

TKO_3108 explicitly teaches proof of correctness and efficiency properties. A fast implementation can still be wrong, while a correct algorithm can be impractical. If the thesis contribution is algorithmic, state which properties are proved theoretically and which are measured empirically. Runtime benchmarks do not substitute for a correctness argument.

28. Benchmark algorithms under controlled conditions

Record input size/distribution, hardware, software/compiler/library versions and repeated-run aggregation when timing noise matters. Use comparable implementations and the same inputs where you claim one method is faster. Avoid generalising from one machine or one input family to universal complexity conclusions.

29. Probabilistic analysis should preserve uncertainty

TKO_7106 teaches complete reproducible probabilistic workflows. Preserve model specification, priors or assumptions, computational settings and posterior summaries. Do not reduce the entire result to one point estimate if uncertainty is central to the research question. Explain how convergence or computational diagnostics were assessed where relevant.

30. Text data needs provenance and quality control

TKO_8964 and TKO_7095 emphasise corpus sourcing, annotation and quality. Record where documents came from, inclusion/exclusion rules, language/domain, annotation instructions and quality checks. Duplicates, near duplicates, repeated authors or documents derived from the same source can create leakage across train/test splits.

31. NLP claims must be bounded to language and domain

A model evaluated on one language, genre or corpus is not automatically validated for another. Identify the exact pretrained model, fine-tuning data, task definition and split. If the thesis claims cross-domain or cross-lingual performance, evaluate that change directly rather than inferring it from in-domain results.

32. Computer vision needs subject/site/frame-aware splitting

Video frames or repeated images from the same subject can be nearly identical. Random image-level splits can therefore produce optimistic results when the intended claim concerns new subjects, sites or environments. Design the split around the independent unit that matches deployment. Record acquisition conditions where they can change performance.

33. Sensor fusion adds calibration assumptions

For camera/radar/LiDAR or other multi-sensor work, state the sensor modalities, calibration/registration assumptions, synchronization and fusion level. A performance gain can depend on these engineering conditions. Do not attribute it only to the fusion algorithm if acquisition or alignment changed between comparisons.

34. Clustering and embeddings are not self-validating

Unsupervised clusters do not come with ground-truth meaning automatically. Explain the validation rationale and whether stability, external labels, domain interpretation or another criterion is used. A visually separated 2D embedding is useful exploratory evidence but is not by itself proof that discrete populations exist in the original high-dimensional space.

35. Missing-data handling can leak too

Describe how missingness was assessed and how imputation was performed. If imputation parameters are estimated from the full dataset before splitting, test information can enter the training pipeline. Put data-dependent imputation inside the same resampling boundary as the model when generalisation performance is the target.

36. Data provenance should be end-to-end

Keep a trace from raw or received data to the final model-ready dataset: acquisition/source, snapshot or version, exclusions, cleaning, transformations, label construction and split membership. This matters especially for company datasets that can be updated continuously. A final CSV without provenance is difficult to audit or reproduce.

37. Personal data changes the data-engineering plan

Research datasets can contain personal data even after obvious names are removed. Text, images, location traces, IDs or behavioural records may remain re-identifiable. Define controller/access arrangements, minimisation, retention and publication boundaries. Pseudonymisation is useful protection but should not be described as anonymisation when re-identification remains possible.

38. Ethics and research permits depend on the actual study

A purely technical benchmark does not automatically need human-subject ethics review. Surveys, behavioural experiments, sensitive personal-data work or interventions can require additional review depending on the design. A host-organisation research permit can also be necessary, but it does not replace consent, ethics or legal requirements where those separately apply.

39. Company data needs a publication plan early

The programme explicitly supports company-commissioned theses. Before modelling, agree what raw data, derived features, performance figures, error examples and model artifacts can be published. Keep confidential raw material separate from the public manuscript. Synthetic or de-identified examples can preserve methodological transparency without releasing protected records.

40. Use AI under the DTEK1002 disclosure rule

DTEK1002 permits generative AI but requires its use to be clearly documented so the student’s own contribution can be identified and graded. Check AI-generated code, references, statistical explanations and prose against primary sources and actual execution. Do not upload confidential partner data or restricted personal data to uncontrolled AI services.

41. Turnitin checks originality, not analytical validity

Turnitin is mandatory for degree theses and the master’s thesis check is approved through UTUGradu. A low similarity score does not prove leakage control, statistical assumptions, model generalisation or reproducibility. Treat originality, research integrity and analytical validity as different quality gates.

42. UTUGradu is the final institutional route

UTUGradu manages originality checking, examination, approval, electronic publication and archiving for higher-degree theses. Before upload, confirm that the examiner-ready manuscript matches the final code/data outputs, confidentiality decisions are settled, DTEK1104/HOPS obligations are in order, and the current maturity-test route is correct for your background.

43. Freeze a claim-to-evidence table before submission

List each headline claim and the evidence supporting it: statistical test, held-out prediction estimate, nested CV, algorithm proof, runtime benchmark, posterior analysis, NLP evaluation, computer-vision experiment or literature synthesis. Put the strongest defensible verb beside each claim. This catches category errors such as using significance to claim prediction or one benchmark to claim universal superiority.

44. Reproduce one headline result from scratch

From the documented data snapshot and environment, regenerate one central table, metric, figure or algorithm benchmark without copying intermediate notebook state. Verify split definitions, preprocessing, seed policy, model parameters, units and aggregation. If the regenerated value differs materially, investigate the pipeline instead of manually forcing the manuscript number to match.

Before calling the reproduction successful, compare the regenerated result with the exact claim written in the manuscript, not only with an older notebook cell or exported screenshot. Confirm the dataset snapshot, row exclusions, target definition, fold membership, preprocessing objects, feature list, hyperparameters, model checkpoint and package environment. If the analysis uses stochastic optimisation, repeated cross-validation or posterior simulation, define in advance what numerical variation is acceptable and what would count as a material discrepancy.

Also verify that every table or figure uses the same evaluation population described in the methods. It is easy for a final chart to be regenerated from a slightly different filtered dataset than the headline metric. Keep machine-readable metadata with important outputs where practical: code revision, data version, run identifier, date, environment and key parameters. This makes it possible to trace a manuscript number back to the computation that produced it.

For company or sensitive research data, reproducibility does not require public release of the raw material. Preserve protected internal inputs under the agreed access controls and create a synthetic, de-identified or schema-compatible demonstration dataset where useful. The public thesis should explain the transformation pipeline and evaluation logic clearly enough to audit the method while respecting contractual, privacy and security restrictions.

Finally, review DTEK1104 and HOPS status together with the manuscript. Because the 5 ECTS seminar is a separate course object in Other Studies, a completed 30 ECTS DTEK1002 thesis does not by itself prove that the seminar/course-placement obligations are complete. Verify registrations, seminar deliverables, maturity route and degree-credit placement before the final UTUGradu submission.

If you release code or a public replication package, test it from a fresh environment and from the documented public inputs rather than from private local paths. Remove cached outputs that could make a broken pipeline appear successful. Check that README commands, environment files, data-download instructions and expected outputs refer to the final frozen thesis version. Where a dependency cannot be redistributed, state the exact acquisition or access requirement instead of silently assuming the examiner has your local setup.

45. Final Data Analytics checklist

Confirm ICTDATA2427 / programme 99359, 120 ECTS, the aligned 50 ECTS advanced-major structure, DTEK1002 30 ECTS, DTEK1104 5 ECTS Pass/Fail, TTDK1308 0 ECTS, at least two examiners and department-head final acceptance/grading. Then verify HOPS placement of the separate seminar, data provenance, leakage-safe evaluation, baseline/metrics, privacy/ethics, company confidentiality, AI disclosure, Turnitin and UTUGradu.