Project Guidelines
This document defines what counts as an acceptable semester project in MST0052, which methods are in scope, and what every project must include. Read it before you commit to a project idea. If you are unsure whether your idea fits, come and talk to me early — in class, at office hours, or by email. Earlier is cheaper than later.
Submission format, grading, the 20 November 2026 hand-in deadline, and the oral exam are covered in the Assessment Guidelines. This page is about what you work on, not how it is graded.
Core requirement
Your project must use machine learning to solve a predictive problem.
Concretely, this means:
- There is a target you are predicting — a numeric value (regression) or a class label (classification).
- You train one or more models on historical data and evaluate them on held-out data using a principled validation strategy.
- You produce a reproducible pipeline from raw data to evaluated predictions.
Exploratory analyses, descriptive dashboards, and pure clustering studies do not by themselves satisfy this requirement. Unsupervised methods are welcome as part of a predictive project — for example as preprocessing, feature engineering, or segmentation — but there must be a predictive target at the core of the work.
Methods in scope
The course covers, and your project may use, any of the following:
| Category | Methods |
|---|---|
| Linear models | OLS regression, ridge, lasso |
| Classification | Logistic regression, k-NN, naive Bayes |
| Trees and ensembles | Decision trees, random forests |
| Boosting | Gradient boosting, XGBoost |
| Kernel methods | Support vector machines |
| Unsupervised (as support) | PCA, k-means, hierarchical clustering |
| Workflow | scikit-learn pipelines, cross-validation, hyperparameter tuning |
You are welcome to go beyond the course within this family — for example, calibration techniques, stacking ensembles, probabilistic regression, quantile regression, survival models, conformal prediction, or advanced feature-engineering methods. Just remember the oral exam: every method in your report is something you must be able to explain without notes.
Methods out of scope: no deep learning
Your project must not use deep learning as its predictive model.
By deep learning we mean, concretely:
- Multi-layer neural networks trained from scratch (beyond the illustrative one-hidden-layer network built in the Lecture 14 notebook)
- Convolutional neural networks (CNNs), recurrent neural networks (RNNs, LSTMs, GRUs), Transformers
- Fine-tuning large pretrained deep models (BERT, GPT, ResNet, Vision Transformers, etc.)
- Using an LLM as the predictor (e.g. prompting GPT to classify text) — the predictive contribution must come from classical ML you train yourself
- AutoML or neural-architecture-search tools whose selected final model is a deep network
Why this restriction?
Deep learning has its own dedicated course later in the programme. MST0052 is about the statistical foundations and classical methods that sit beneath deep learning — the vocabulary you will need to reason about any ML model, deep or not. Classical methods also remain the right choice for the majority of tabular business problems, which are the most common cases you will meet in practice.
Working inside this constraint forces you to focus on problem framing, feature engineering, validation, and interpretation These skills transfer when you take the deep-learning course afterwards.
Borderline cases
- A small neural network shown alongside classical models as a comparison point: acceptable only as an illustrative comparison. It cannot be your chosen or headline model, and the report’s methodological core must remain classical ML.
- Pretrained embeddings (e.g. BERT, sentence-transformers, image embeddings) used as input features for a classical model: check with the instructor first. Allowed only when the predictive model itself is classical ML and this is not the main methodological contribution of the project.
- Off-the-shelf libraries that wrap deep models (e.g.
fastai, Hugging Face pipelines used for prediction): out of scope for the main model. - Reinforcement learning, generative modelling: out of scope for this course.
If in doubt, ask — before you invest weeks in an off-scope approach.
Problem types that fit the course
Good project problems typically take one of these shapes:
- Regression — predicting a numeric outcome (house prices, expected sales, credit loss given default, delivery time)
- Binary classification — yes/no outcomes (churn, default, readmission, spam)
- Multi-class classification — one of several categories (product category from description, customer segment)
- Ranking or recommendation framed as scoring each candidate with a classical model
Problems you should not pick without first discussing them with the instructor: image classification from raw pixels, NLP tasks where the state of the art is a Transformer, and anything that genuinely requires a large pretrained model to be competitive.
Data
- You may use public datasets (UCI, Kaggle, OpenML, government open data, Eurostat, SSB, Norges Bank, Hugging Face Datasets, API pulls) or data you collect yourself.
- Do not use proprietary or confidential data that you are not permitted to share with the instructor and examiner.
- Check licensing and cite your source in the report.
- Respect privacy and GDPR — do not use personal data you are not authorised to process.
- Size guidance: aim for something large enough to do honest cross-validation (hundreds to hundreds-of-thousands of rows is a healthy range) but small enough to iterate quickly on your laptop. Massive datasets are not required — a thoughtful small-data project beats a sloppy big-data one.
- Using a dataset attached to a live or past Kaggle competition is fine, but your analysis must be your own. Do not copy a public notebook.
Required project elements
Every project, regardless of topic, should include:
- A clearly stated prediction problem and why it is worth solving.
- A description of the data — source, size, target, features, preprocessing.
- At least one honest baseline (e.g. a mean/majority-class predictor and a simple linear or logistic model). This is the bar your more sophisticated models must beat.
- At least two contrasting modelling families compared fairly on the same resampling rule (e.g. linear + tree-based, or tree-based + SVM).
- A reproducible validation strategy appropriate to the data — k-fold for i.i.d. data, time-series split for temporal data, grouped splits when rows are not independent.
- Honest evaluation: at least one appropriate metric, a sense of the variability across folds, and a discussion of where the model fails.
- A brief reflection on interpretability, fairness, and limitations.
- A reproducible workflow: pinned dependencies, set seeds, and a pipeline rather than scattered ad-hoc steps. You do not hand in your code, but your report must describe the workflow clearly enough that someone could reproduce it.
Scope and ambition
You can calibrate the ambition of your project. Any of the following make a project more interesting:
- A novel or unusual dataset that you collected, scraped, or merged yourself
- Careful feature engineering grounded in domain knowledge
- Thoughtful handling of a hard problem: class imbalance, concept drift, missing-not-at-random data, small samples, or serious leakage risks
- Rigorous validation: nested cross-validation, proper time-series splits, probability calibration, conformal prediction intervals
- A thoughtful comparison across the model families from the course, with reasons for why one wins
- Interpretability work: SHAP values, partial dependence, permutation importance, surrogate models
- A responsible-ML angle: fairness analysis across subgroups, robustness checks, error-slice analysis
You do not need to do all of these. Depth in one or two dimensions beats shallow breadth across all of them.
Check-in and approval
- Recommended: have a dataset and a preliminary problem statement by around Lecture 3 — a suggestion to keep you on pace, not a deadline.
- Consider signing up for the optional Lecture 8 project showcase — 5-minute slots for student presentations, with feedback from the instructor and ideas from your classmates. Signup is in Canvas and closes 6 October.
- Rehearse at the Lecture 15 project workshop.
- Submit the final PDF report by 20 November 2026 at 12:00. Code is not submitted.
- Optional: put your code on GitHub and link it in your report. Encouraged as good practice and a portfolio piece — it is not graded.
- At any point you are uncertain whether your idea is in scope, talk to me — in class, at office hours, by email, or by asking for a meeting. There is no single formal scope-check deadline; the responsibility to ask is yours, and the cheaper-to-ask window is now, not late October.
Example project ideas (illustrative, not prescriptive)
These sketches are meant to show the shape of a good project. Do not copy them — pick something you are genuinely interested in.
- Predicting next-month customer churn for a telecom dataset using gradient boosting, with SHAP-based analysis of the main drivers.
- Forecasting weekly retail sales per store from historical sales and promotions, using time-series-aware cross-validation and an ensemble of lasso and gradient boosting.
- Classifying news articles by topic using TF-IDF features and an SVM, with a fairness check across source outlets.
- Predicting apartment rental prices in Oslo from scraped listings, comparing ridge, random forest, and gradient boosting, with calibration and interpretability analyses.
- Early warning of student drop-out from administrative records, with careful handling of class imbalance and a thoughtful discussion of fairness.
A good project is one you are interested enough in to defend in detail on the oral exam.