# Empirical benchmark for recent University of Tartu Data Science master’s theses

**Prepared:** 19 October 2026  
**Programme selector:** exact registry field `curriculum = msc_data_science`  
**Defence years:** 2022–2026

## Decision result

A realistic UT Data Science / Data Science and Artificial Intelligence master’s thesis is usually **one bounded research problem with a frozen evidence base, a defensible baseline or prior solution, one primary outcome, and an evaluation that remains useful even if the complex method does not win**. It does not need to be a production platform or a collection of several equally large contributions.

The official ICS guideline treats the master’s thesis as a research paper and permits theoretical, applied, or combined work.

For applied/software work it explicitly asks how the solution compares with previous solutions, whether and how well it worked, measurable results, justified technical choices, and documented architecture.[2]

The empirical corpus confirms that software systems, data pipelines, datasets, evaluation frameworks, observational studies, and optimization studies all occur alongside conventional model-development theses.[1]

The current ICS programme page uses the name “Data Science and Artificial Intelligence.”[3]

## Corpus and acquisition

### Observed

- The UT Institute of Computer Science Graduation Theses Registry exposes annual public records with an explicit curriculum field, titles, authors, supervisors, abstracts, keywords, persistent UUIDs, and public thesis-file links.[1]
- Exact filtering produced **133 registry records**: **22** (2022), **23** (2023), **24** (2024), **29** (2025), and **35** (2026).
- All **133 public files** were downloaded and size/SHA-256 verified locally: **752,355,436 bytes (0.752 GB / 0.701 GiB)**.
- **124** files meet the deterministic full-length-public-copy rule; **9** are 3–7-page short/public/restricted copies. The latter classification describes the public file, not the privately submitted thesis.

### Calculated document-length benchmark

For the **124 full-length public copies**:

- PDF pages: median **51.5**, interquartile range **43.0–60.0**, observed range **29–113**.
- Extracted words: median **11495**, interquartile range **8661–14738**, observed range **4464–30088**.

These PDF counts include front matter and appendices. They are therefore compatible with, but not equivalent to, the ICS guideline’s approximate **40–50 pages excluding appendices**.[2]

## Abstract-level contribution benchmark

English abstracts were hand-coded for **133** exact-curriculum records (2022, 2023, 2024, 2025, 2026 complete). The cross-period categories below are conservative normalizations over two compatible but non-identical codebooks; native period results remain in the JSON evidence.

### Dominant contribution

- Model or method development/evaluation: **64 (48.1%)**
- Software or data system: **32 (24.1%)**
- Observational, descriptive, or qualitative analysis: **21 (15.8%)**
- Optimization or decision support: **11 (8.3%)**
- Data resource or benchmark creation: **4 (3.0%)**
- Human-subject intervention: **1 (0.8%)**

### Evaluation and scope signals

- An explicit comparator was reported in **89/133 abstracts (66.9%)**. “Not reported” is not proof that the full thesis had no comparison.
- A concrete output—model, method, dataset, report, software, workflow, or decision-support artifact—was coded in **128/133 (96.2%)**. This broad category must not be read as “deployed software.”
- A bounded setting, population, dataset, domain, prototype, or simulation was coded in **133/133 (100.0%)**.
- An explicit limitation, trade-off, method failure, small-sample warning, or null result was reported in **58/133 (43.6%)**.

## What reasonable scope looks like

### 1. One primary task can sustain the thesis

A 2026 harbor study compared six forecasting model families at fixed 10–120-minute horizons for one operational target.[4]

A 2025 calibration study evaluated a fixed calibration problem, while the earlier temperature-scaling thesis treated calibration behavior itself as the central contribution.[11]

The lesson is not “use six models”; it is **freeze the target, horizon, data cut, and primary metric before adding model breadth**.

### 2. Engineering-heavy work is valid when the evaluation is the thesis

A schema-constrained knowledge-graph thesis built a decoupled evaluation framework, began with a small annotated document sample, and tested transfer rather than claiming a universal extraction system.[5]

A general-AI-tutor thesis centered on an evaluation framework and pilot validation rather than requiring a full educational intervention.[6]

The corpus also contains operational data-quality and drift studies grounded in explicit comparison or validation.[9][12]

### 3. Small or local studies can be defensible if the inference boundary is honest

A smartphone inertial-sensing thesis evaluated feedback for a defined strength-exercise setting.[7]

A personalised-feedback pilot used a small human sample and reported mostly non-significant results rather than inflating the claim.[8]

A narrow dataset or pilot is acceptable evidence for a narrow conclusion; it is not evidence for national, population, or cross-domain generalization.

### 4. Negative results remain useful results

The corpus includes complex methods that did not outperform simpler approaches, explicit null findings, and operational constraints. A thesis should therefore define a **minimum defensible result**—benchmark, calibrated error analysis, dataset, evaluation protocol, or validated negative finding—before promising a new model or production intervention.

### 5. Privacy-preserving or rule-first designs are legitimate

The privacy-preserving text-synthesis thesis evaluated a bounded framework rather than treating privacy as a disclaimer.[10]

For private finance records, agent logs, health records, or public-record conflicts, the proposal should specify data minimization, publication boundaries, independent gold labels, and abstention before model sophistication.

## Corpus-calibrated proposal gate

Every shortlisted topic should now state:

1. **One primary research question** and one central claim.
2. **One frozen data cut or cohort**, with independent-unit counts rather than raw rows/events/frames alone.
3. **At least one simple baseline or previous solution** plus the proposed method.
4. **One primary metric** and a leakage-safe split (temporal, source, participant, merchant, route, or session-level where appropriate).
5. **One minimum defensible result** that survives a null or negative model result.
6. **A publication/privacy/rights boundary** for any private, personal, monitored, or legally sensitive data.
7. **A scope cut list** naming attractive components that are explicitly not primary contributions.

Software, dashboards, deployments, and existing pipelines count as research infrastructure unless their quality, reliability, effect, or comparative performance is independently evaluated.

## Implications for Kristjan’s project ranking

The empirical benchmark supports four concrete ranking corrections:

- **Reward bounded evaluation over platform breadth.** A smaller benchmark with valid labels and holdouts outranks a larger system with unclear outcomes.
- **Lower data-readiness scores when labels are only rule-derived or still prospective.** Raw logs, pages, frames, snapshots, or transactions do not by themselves establish supervised-data readiness.
- **Treat software/data-system topics as programme-plausible, not second-class, when they have a CS research question and comparative evaluation.**
- **Penalize over-scoping through research-question clarity, evaluation feasibility, and finishability.** Do not add a separate prestige bonus for more components.

These rules motivate narrowed titles such as precipitation calibration instead of a universal forecast combiner; verified-completion evaluation instead of an entire agent reliability stack; calibrated receipt matching instead of general multimodal finance intelligence; and source-aware vacancy-removal modelling instead of combined lifecycle plus personalised ranking.

## Evidence files

- `evidence/ut-ds-thesis-corpus-2022-2026.json` — normalized authoritative corpus
- `evidence/ut-ds-thesis-corpus-provenance.json` — retrieval timestamps and source hashes
- `evidence/ut-ds-theses-pdf-manifest.json` — size and SHA-256 for every public file
- `evidence/ut-ds-thesis-document-benchmark.json` — page/word metrics and short-copy classification
- `evidence/ut-ds-benchmark-2022-2023.json` — record-level abstract coding
- `evidence/ut-ds-benchmark-2024-2025.json` — record-level abstract coding
- `evidence/ut-ds-benchmark-2026.json` — record-level abstract coding when present
- `evidence/ut-ds-benchmark-aggregate.json` — normalized cross-period benchmark
- `evidence/ut-corpus-ranking-review.json` — candidate-by-candidate independent recalibration

## Interpretation boundaries

- Registry membership means the thesis is in the defended-thesis registry for the selected curriculum; it does **not** reveal grade, examiner judgment, excellence, or supervisor endorsement.
- Abstract coding is analyst interpretation by one reviewer and was not independently double-coded.
- Missing abstract detail means “not reported in the abstract,” not “absent from the thesis.”
- The defended corpus has survivorship and reporting bias; it does not show rejected or abandoned proposals.
- Similarity to a prior thesis does not establish novelty, ethics approval, data access, or supervisor fit.

## Sources

[1] https://thesis.cs.ut.ee — UT ICS Graduation Theses Registry
[2] https://cs.ut.ee/sites/default/files/2026-08/2026_02_03_CSI_Guidelines_for_preparing_and_grading_of_graduation_theses.docx.pdf — ICS graduation-thesis guidelines
[3] https://cs.ut.ee/en/content/masters-studies — UT ICS master’s studies
[4] https://thesis.cs.ut.ee/e856e557-9e17-4a8a-8b8a-c886f605cb27 — Short-Term Harbor Water Flow Forecasting
[5] https://thesis.cs.ut.ee/103c320d-bd31-414a-8972-160a04514695 — KG extraction evaluation framework
[6] https://thesis.cs.ut.ee/d9f7f130-5124-4156-bcbe-13467bc31922 — General AI tutor evaluation framework
[7] https://thesis.cs.ut.ee/b0c8864d-5b69-43d8-9cd2-a8fa33bb6253 — Smartphone inertial strength exercise feedback
[8] https://thesis.cs.ut.ee/3a51227f-bec6-42cd-8041-854a525d4afe — LLM personalised feedback pilot
[9] https://thesis.cs.ut.ee/656faca3-e202-4384-8fa0-cf8b5562dc69 — Call-centre data drift thesis
[10] https://thesis.cs.ut.ee/f5508af2-13ed-4658-9f53-4948c40c64bb — Privacy-preserving text synthesis framework
[11] https://thesis.cs.ut.ee/a980bc0d-f4b7-4b86-9df9-4a8b24951807 — Overconfidence and temperature scaling
[12] https://thesis.cs.ut.ee/41c9911a-9649-45af-aaf9-e206647a8f21 — Automated data-quality rule detection