Grading Rules 2026/2027
The final grade is based on two components, for a total of at most 100 points:
| Component | Maximum | Minimum required |
|---|---|---|
| Semestral project | 70 points | 35 points (to be admitted to the exam) |
| Oral exam | 30 points | 15 points (to be approved) |
| Total | 100 points |
- Semestral project: an individual project. Detailed guidelines are given in the Semester Project section below.
- Oral exam: can be taken only after reaching at least 35 points in the semestral project.
Exam
Format
- You will receive two topics for discussion:
- a question related to your semestral project;
- a topic based on the content covered in lectures and tutorials.
- You will have 15 minutes to prepare.
- The oral exam follows right after the preparation period.
Points
| Part | Maximum |
|---|---|
| Questions related to your semestral project | 15 points |
| Questions related to lecture topics | 15 points |
| Total (at least 15 points to pass) | 30 points |
Exam Topics
TBA. The lecture topics for the final exam will be announced later.
Evaluation Scale
Your grade is determined by the sum of points from the semestral project and the exam (at most 100 points). The evaluation scale follows The Study and Examination Code for Students of CTU in Prague.
| Grade | Points | Wording |
|---|---|---|
| A | 90 – 100 | excellent |
| B | 80 – 89 | very good |
| C | 70 – 79 | good |
| D | 60 – 69 | satisfactory |
| E | 50 – 59 | sufficient |
| F | less than 50 | failed |
Semester Project
The goal of the semester project is to carry out scientific or applied research that extends an existing personalised machine learning method, and to present the results as a short scientific paper. The main focus is on recommender systems, although the same ideas can be applied in many other fields.
Your project must start from a paper published in the research track (full or short paper) of one of the following conferences in 2024, 2025 or 2026: RecSys, KDD, CIKM, WWW (The Web Conference) or another A* machine learning conference. You are expected to extend this work, for example by improving its accuracy, efficiency or applicability, or by applying it to a new setting. The extension should be original and clearly evaluated against the base paper.
The topic is open: any topic related to personalised machine learning is allowed. Projects based on the following directions are still very welcome:
- Language models, for example using embeddings from large language models or sentence transformers to represent users and items, or using instruction-tuned models directly to rank, generate or re-rank items.
- Agentic approaches, where one or more LLM-based agents take part in the recommendation process, for example by planning, using tools or simulating users.
- Modern architectures, such as transformers, state-space models, linear recurrent models, diffusion models or contrastive learning models.
- Explaining these models, for example by analysing what attention or embeddings capture, or by generating explanations for the recommendations.
The paper must follow the ACM template (a prepared template is available in the shared Overleaf project) and be 4 to 8 pages long, excluding references.
Evaluation Criteria
Students will be assessed based on the quality of their scientific paper, the novelty and significance of their research extension, and their ability to communicate their findings effectively.
Important Dates
Each milestone is discussed in the tutorial of the given week. The deadline to submit it is always the Sunday before that tutorial (23:59).
| Tutorial | Deadline | Milestone | Points |
|---|---|---|---|
| 05.10.2026 | 04.10.2026 | Project topic (registered in the spreadsheet) | 5 |
| 12.10.2026 | 11.10.2026 | Related works, motivation, novelty and baselines | 5 |
| 26.10.2026 | 25.10.2026 | Proposed methodology and evaluation | 5 |
| 23.11.2026 | 22.11.2026 | First experiments (at least the baselines) | 5 |
| 30.11.2026 | – | Project tutorial | – |
| 14.12.2026 | – | Poster presentation | – |
| – | 30.12.2026 | Paper submission (together with the poster) | 50 |
Topic Registration and Submission
- Topic: add the topic of your project here by Sunday 04.10.2026. Each student must have a different topic, so it’s first come, first served.
- Submissions: every student gets their own paper in a shared Overleaf project (ACM template), named with their KOS login. The link will be shared with you by email. For each deadline, update your paper there with the current version.
- Poster: must be printed and presented during the project presentation on 14.12.2026, as in academic conferences. See an example poster. Professional printing is not required; you can, for example, print it on a few A4 pages.
Important:
The shared Overleaf project is for submissions only, not for working. Write your paper in your own Overleaf project (or any other application) and copy it to the shared project only when a milestone is done.
Topic Suggestions
The following ideas are only examples; you can also propose your own topic, as long as it fits the requirements above. In every case, start from a recent paper and extend it. Your topic will be discussed with and approved by the teacher on 05.10.2026.
Language models
- Compare general-purpose and instruction-based text embedding models as item representations, and study how the embedding size affects recommendation quality.
- Use language model embeddings to recommend new items with no interaction history (cold start), and compare them with ID-based models.
- Initialise or align a recommendation model with LLM embeddings, and compare frozen, fine-tuned and partially frozen setups.
- Use an open instruction-tuned LLM to generate user profiles or item knowledge, and measure how much it improves a standard recommender.
- Replace item IDs with short codes learned from text embeddings (semantic IDs), and study their effect on new and rarely seen items.
Agentic approaches
- Measure how faithfully LLM agents simulate real users, and whether a simulator ranks recommendation models in the same order as offline evaluation.
- Build a user simulator for conversational recommendation that does not leak the target item, and measure how it changes the reported success of recommenders.
- Analyse a tool-using recommender agent to find how much of its quality comes from the LLM planner and how much from the recommendation models it calls.
- Reduce the cost of LLM agents that represent users and items, for example with embedding-based memory, and report quality against the number of LLM calls.
Modern architectures
- Compare state-space or linear recurrent models with transformers on long user histories in terms of accuracy, speed and memory.
- Re-evaluate a recent diffusion recommender against carefully tuned simple baselines, over several random seeds.
- Speed up diffusion recommenders with fewer sampling steps, and study the trade-off between accuracy and latency.
- Guide a diffusion recommender with item text instead of only item IDs, and study the balance between accuracy and diversity.
- Study whether graph structure is really needed in contrastive recommenders, and how the loss temperature shapes the learned embeddings.
- Improve contrastive recommenders with harder negative examples chosen by popularity or by language model similarity, and measure the effect on popularity bias.
Explainability
- Train a sparse autoencoder on item embeddings of a modern recommender and describe the learned features using item information.
- Evaluate whether LLM-generated explanations reflect the real reasons for the recommendations, for example by removing the explained evidence and checking whether the ranking changes.
- Visualise and measure what the attention weights of a transformer recommender capture, compared with item similarity and co-occurrence.