Tampere University Data Science Master's Thesis Guide: 30 ECTS, CSEE-DSY and Trepo

What this guide covers

Tampere University’s international Data Science option belongs to the Master’s Programme in Computing Sciences and Electrical Engineering (CSEE) and awards Master of Science, not Master of Science (Technology). The programme is 120 ECTS and designed for two years of full-time study. This degree title is the first thesis-rule boundary to get right: even though the current thesis is 30 ECTS, it is not the Technology diplomityö. The relevant university-wide process is Tampere’s non-technical master’s thesis framework.

The current 2026–2027 CSEE curriculum object, CSEE-DSY, gives the decisive structure: 50 ECTS advanced studies in Data Science + 30 ECTS master’s thesis on a topic inside the specialisation + 40 ECTS other studies = 120 ECTS. The other studies can include joint, free-choice and possible complementary studies. For this applicant option, this is the structure to anchor your thesis planning to.

Do not confuse CSEE-DSY with the separate TIM-DSY structure

Tampere currently exposes another Data Science curriculum object, TIM-DSY, under the Master’s Programme in Computer Sciences. TIM-DSY describes 5 ECTS joint studies, 80 ECTS advanced studies including a 30 ECTS pro gradu thesis, and 35 ECTS electives. That object is current, but it is not interchangeable with the CSEE-DSY object tied to the international CSEE study option processed in this guide.

This distinction matters because two pages can both say “Data Science” and both be current while belonging to different programme structures. When checking credits or course requirements, follow the curriculum object attached to your own Sisu study plan and admission route. Never combine the 5+80+35 structure from TIM-DSY with the 50+30+40 CSEE structure unless Tampere has explicitly migrated your study right.

The thesis is 30 ECTS and must belong to Data Science

CSEE-DSY explicitly places a 30 ECTS master’s thesis inside the Data Science specialisation. The topic therefore needs an academically defensible connection to Data Science rather than merely using a dataset or a fashionable AI tool. A thesis can focus on statistical modelling, machine learning, network or complex-systems analysis, data mining, computational methods, forecasting, causal questions, data engineering for research, evaluation of algorithms, or another data-driven problem that fits the specialisation.

A project is not made into a Data Science thesis by adding a predictive model at the end. The research question should determine what data are needed, which assumptions matter, how the method will be evaluated, and what conclusions can legitimately be drawn.

ITC.CEE.800 Tools for theses is a current 5 ECTS requirement to understand

The current 2026–2027 ITC.CEE.800 Tools for theses course is 5 ECTS and pass/fail. Its four parts are basics of research methods, research ethics, information searching skills and career planning. All parts of the completion option are compulsory, and Tampere recommends completing the course before starting the thesis, with the components undertaken while the thesis is being planned.

This course is important preparation, but it is not the thesis itself. Do not add its 5 ECTS to the 30 ECTS thesis and call the thesis 35 ECTS. The current CSEE-DSY structure still identifies the thesis itself as 30 ECTS.

Do not resurrect the old DATA.810 seminar automatically

Older Tampere curricula contained DATA.810 Master’s Thesis Seminar, Data Science & Statistical Data Analytics, 5 ECTS. That historic information is still discoverable online. It is unsafe to treat it as a current 2026–2027 CSEE Data Science requirement merely because the course title matches the field.

The current CSEE-DSY structure and current ITC.CEE.800 evidence used for this guide do not establish historic DATA.810 as the governing current seminar course. Tampere’s general non-technical thesis rules still require seminar participation, an oral presentation and acting as opponent. Therefore follow your current Sisu/Moodle implementation for the exact seminar registration and schedule, but do not import the old five-credit course unless it actually appears in your current study plan.

Build the topic around a research question, not a dataset

A common weak starting point is “I have this dataset, what can I predict?” A stronger thesis starts with a question about a phenomenon, system or decision and then asks what evidence is required. The dataset should be suitable for answering the question, not the reason the question exists.

For predictive work, define the target, unit of analysis, evaluation horizon and intended use. For descriptive work, define which patterns matter and why. For causal work, be explicit that prediction and association do not establish causality; causal interpretation requires assumptions and a design capable of supporting it. For computational-method work, define the baseline and the dimension on which the new method is supposed to improve.

Data splitting and leakage can determine whether the result is valid

In machine-learning and statistical prediction, the train/validation/test design is part of the research method. Leakage can occur when information from the test period or target indirectly enters preprocessing, feature construction, imputation, normalisation, hyperparameter selection or model choice. A visually impressive score can therefore be invalid even when the code runs correctly.

Choose a split that reflects the data-generating structure. Random row splits may be inappropriate for time series, repeated observations from the same participant, multiple images from the same subject, spatially clustered data, or records from the same organisation. Grouped, temporal or nested evaluation may be required. Document the split before final model selection whenever possible.

If you compare models, tune them under the same evaluation logic. Do not give the proposed method extensive tuning while leaving the baseline at defaults and then interpret the comparison as scientific evidence.

Metrics must match the research question

Accuracy is not universally meaningful. For imbalanced classification, precision, recall, F1, area-under-curve measures, calibration or class-specific error may matter. For regression, MAE, RMSE, calibration and domain-specific error thresholds answer different questions. For ranking or recommendation, ranking metrics and offline evaluation may still fail to represent real user outcomes.

Choose the primary metric before examining the final test result where practical. Explain why it corresponds to the thesis question. Secondary metrics can reveal trade-offs. If a model improves average accuracy while making rare but serious errors worse, report that rather than hiding it behind one aggregate score.

Uncertainty, robustness and statistical inference

Data Science is not only about finding the highest-performing model. Report uncertainty where it affects interpretation. Confidence intervals, bootstrap distributions, repeated cross-validation, posterior uncertainty, sensitivity analysis or other approaches may be appropriate depending on the method.

If several models are very close, avoid claiming superiority from trivial numerical differences without considering variability. If your conclusion changes when a seed, threshold, preprocessing choice or reasonable sample restriction changes, that instability is itself part of the result.

For statistical testing, predefine the hypotheses and avoid treating a small p-value as proof of practical importance. Report effect sizes and uncertainty. Multiple comparisons, optional stopping and repeated model searching can make nominal significance misleading.

Reproducibility is part of the evidence

A computational thesis should make it possible to understand how raw data became the reported result. Record data provenance and version, filtering rules, preprocessing, feature construction, train/test logic, package and environment versions, random seeds and analysis configuration when these materially affect the outcome.

Use version control for analysis code. Keep generated results separate from source code. A run log can include commit, dataset version, configuration and output path. If proprietary or sensitive data cannot be shared, reproducibility can still be improved with a data dictionary, synthetic example, pseudocode, documented transformations and scripts that would run inside the authorised environment.

Reproducibility does not mean publishing confidential data. It means making the research process sufficiently transparent that the evidence can be examined and, where possible, repeated.

Data quality comes before model sophistication

Before fitting complex models, inspect missingness, duplicates, impossible values, measurement changes, selection bias and target-definition problems. If a label is noisy or inconsistent, more model complexity may only fit the noise more efficiently.

Document how observations enter the dataset. A company database may contain only customers who reached a certain stage of a process; a health dataset may reflect who was measured rather than the whole target population; an online platform dataset may omit deleted or inaccessible activity. These mechanisms affect what population the conclusion can represent.

If preprocessing removes a large fraction of observations, report the attrition and justify it. Do not silently discard difficult cases to improve model performance.

Also separate data availability from target-population validity. A large dataset can still be systematically unrepresentative. Describe who or what could enter the data, who is missing, and whether the collection mechanism changes over time. If the thesis evaluates a model on one organisation, platform, country or time period, state that boundary explicitly before generalising to a wider population.

When labels are created by human judgement, heuristics or another model, document the labelling process and assess its reliability where possible. Label disagreement can be substantive information rather than mere noise. If a proxy target is used because the true outcome is unavailable, explain what the proxy captures and what it does not.

Fairness, privacy and sensitive variables

Data Science can affect people even when the thesis looks technically focused. If the data include personal information, discuss the controller arrangement and processing plan with the supervisor before collection or transfer. Tampere’s data-protection instructions require appropriate purpose definition, minimisation, privacy information and secure processing.

Fairness is context dependent. Comparing error rates across groups can be useful, but a fairness metric does not by itself answer whether a deployment is ethically acceptable. Consider whether group labels are appropriate to process, whether the sample supports subgroup conclusions, and whether the model could amplify an existing selection or measurement bias.

Do not upload personal, confidential or company data to external AI services merely to speed up coding or analysis.

Treat personal data governance as part of the research design, not as a final administrative check. Decide who can access raw data, where identifiers are stored, when pseudonymisation or anonymisation occurs, and what outputs can safely be published. Aggregate statistics can sometimes reveal individuals in small groups, so disclosure risk should also be considered at reporting stage.

AI tools in a Data Science thesis

Tampere allows AI use within its current study guidance, but the student remains responsible for the academic work, analysis and correctness. If AI is used for coding, debugging, drafting, translation, idea generation or another permitted task, follow the current acknowledgement instructions and the supervisor’s guidance.

AI-generated code must be tested and reviewed. AI-generated citations must be independently checked. AI-generated interpretation is not a substitute for your own analysis. Do not provide confidential data, unpublished company code, participant data or credentials to an external service without an appropriate basis and approved processing environment.

If the AI tool itself is part of the research object, document model/service version, relevant settings and known limits so later readers understand the experimental conditions.

Literature review for Data Science

The literature review should justify both the substantive problem and the methodological choices. A thesis on prediction in healthcare, for example, needs literature on the domain problem as well as literature on modelling and validation. A thesis on a new algorithm needs comparison with relevant methodological work, not only papers that used the same dataset.

Use systematic information searching from ITC.CEE.800 to build a defensible evidence base. Keep records of databases, important search strings and selection logic where the review has a systematic component. Distinguish peer-reviewed evidence from software documentation, blog posts, benchmark leaderboards and vendor claims.

Supervision and research plan

Tampere appoints one or two supervisors, with one primary supervisor. Agree the research question, scope, method, data access and timetable early. If the thesis is commissioned by a company, the external contact can support the practical work, but Tampere’s academic supervision and examination requirements remain controlling.

The research plan should state the question, related work, data, inclusion/exclusion rules, method, evaluation design, ethics/data-protection issues, expected limitations and timeline. For computational work, include how the data and code will be stored, how experiments will be tracked and what must be frozen before the final evaluation.

A useful milestone is a “method freeze”: by that point the target, split, primary metric, baseline and main analysis should be defined. Later changes can still be justified, but they should be documented rather than silently adjusted to improve results.

Seminar, presentation and opposition

Tampere’s general non-technical master’s-thesis process includes participation in a thesis seminar, oral presentation and acting as opponent for another student’s work. These activities are part of the academic process even though this guide does not assign them the historic DATA.810 credit structure.

Use the current implementation shown in Sisu and Moodle. Confirm when you need to present, what counts as opposition, and whether there are programme-specific milestones. Do not assume an old timetable remains active simply because an archived course page is still searchable.

Writing the thesis

Write methods before final results are known. This makes it easier to identify hidden researcher degrees of freedom and missing information. For computational work, the method section should be detailed enough to explain data construction, preprocessing, model specification, training/tuning procedure and evaluation.

Keep results and interpretation distinguishable. Report failed analyses when they materially affect the conclusion. Discuss threats to validity such as dataset shift, small subgroup samples, leakage risk, model-selection bias, missing data, unmeasured confounding and limited external validity.

A common structure is introduction and questions, background/related work, data and methods, results, discussion, limitations and conclusions. Topic-specific structure and supervisor instructions take priority over any generic template.

Maturity test

The master’s-thesis process includes a maturity test. When the required Finnish/Swedish proficiency has already been demonstrated as specified by Tampere, the thesis abstract normally functions as the maturity test and only its content is assessed. When language checking is still required, the current electronic EXAM route applies.

When the maturity test is not used to demonstrate Finnish or Swedish proficiency, it is written in the thesis language. The maturity test cannot be replaced by recognition of prior learning.

Turnitin, Trepo and final submission

Obtain the primary supervisor’s permission before final submission. The thesis goes through Turnitin originality checking before final deposit in Trepo, Tampere University’s institutional repository. Turnitin and Trepo are different stages: one supports originality review, the other is the repository and permanent archive route.

You need attending registration for examination and thesis credits. After Trepo deposit, the Library normally provides an access address that the student forwards to the examiners. The permanent archival file should be PDF/A.

Check the final file carefully before deposit. Once a thesis is formally approved, it is final and cannot simply be replaced by a revised new submission.

Examination, grading and response rights

For non-technical master’s theses, the dean approves the thesis based on a joint statement and grade proposal from two examiners. At least one examiner must be employed by Tampere University, and both must have at least a master-level degree. The general thesis scale is 0–5.

When the thesis abstract serves as the maturity test, the normal combined assessment period is 21 days. When a separate electronic maturity exam with language checking is required, it is 28 days. Summer schedules can be longer. These periods are assessment windows, not the time required to research and write the thesis.

The proposed grade and statement are sent to the student. A written response can be submitted within seven days if the student disagrees. After grading, the assessment decision can be appealed in writing to the Faculty Council within 14 days of access to the result and assessment criteria information.

Graduation planning

Do not plan graduation solely from the 21/28-day thesis assessment clock. Other credits, Sisu study-plan approvals, maturity completion and graduation processing must also be ready. Check the current graduation schedule before choosing a final submission date.

A safe final-stage sequence is: supervisor approval, final originality check, Trepo PDF/A deposit, examiner access, maturity route, assessment, possible response period, grade registration and then graduation application/process according to the current schedule.

Practical Data Science workflow

1. Confirm that your study right uses CSEE-DSY, not the separate TIM-DSY object.
2. Confirm the 30 ECTS thesis and current ITC.CEE.800 requirements in Sisu.
3. Turn the practical problem into a narrow research question.
4. Define data provenance, population, unit of analysis and target before modelling.
5. Design train/validation/test or statistical analysis logic to prevent leakage.
6. Fix primary metrics, baselines and main evaluation logic before the final test.
7. Plan privacy, ethics and any company restrictions before data collection or transfer.
8. Keep version-controlled code, environment information and an experiment log.
9. Complete the current seminar, presentation and opposition implementation shown in Sisu/Moodle; do not import historic DATA.810 automatically.
10. Write and revise alongside the analysis.
11. Complete maturity, Turnitin, Trepo PDF/A deposit, examination and graduation steps.

The strongest Data Science thesis is not necessarily the one with the most complex model. It is the one where the question, data, assumptions, method and evaluation fit together tightly enough that another reader can understand exactly what the evidence does and does not support.