Tematy projektów WSHOP -- zima 2026/2027

• : Possibility of extending it to master thesis
• : Quick project
• : Linked to international scientific project
• : Linked to JU-internal scientific project

• Student:
• Namespace in the wiki: FIXME
• The goal of the project:
• Technology:
• Description:
• Links:
  • 

: Possibility of extending it to master thesis

• Student:
• Namespace in the wiki: agenticmlops
• The goal of the project: surveying and prototyping agentic solutions for MLops with the special focus an emerging methods and tools such as XAI, NeuroSymbolic AI(NeSy), resource constrained ml (rcml), edge ai
• Technology: agentic ai, ml ops, xai, nesy, rcml
• Description: 1) survey and understand state of the art for mlops, 2) identify agentic platforms integration for mlops, 3) look for dedicated agenticml ops platforms, 4) study opportunities and needs for xai, nesy, rcml, 5) prototype your own solution/s. The project gives an opportunity for a longer term cooperation with GEIST, spring WSHOP, master thesis, research project, etc
• Links:
  • 

• Student:
• Namespace in the wiki: affective_game
• The goal of the project: Creation of game with affective feedback loop
• Technology: Eyetracking, biosignals (EDA)
• Description: The task aims to integrate signals from various sources to create a coherent, interactive virtual environment. The task involves developing the specifications, interfaces, and mechanisms required to log gameplay data for subsequent analysis. Control mechanisms driven by gaze (eye-tracking data), with force feedback dynamically activated by an EDA (electrodermal activity) signal.
• Links:
  • 

• Student:
• Namespace in the wiki: bir3-analysis-w27
• The goal of the project: Multimodal data study
• Technology: Eyetracking, biosignals (EDA, EKG), game events
• Description: Multimodal analysis of data collected during two editions of the BIRAFFE3 experiment (2025, 2026). The objective is to integrate results from a series of measurements comprising eye-tracking data, biosignals (ECG, EDA), logs from the research-specific game events, and questionnaires describing the psychological context of the data.
• Links:
  • 

• Student:
• Namespace in the wiki: eyetracking-CBM
• The goal of the project: Advanced analysis of eye-tracking signals
• Technology: Eyetracking
• Description: Advanced analysis of eye-tracking signals within the context of the BIRAFFE3 experiment. A project conducted in collaboration with the Jagiellonian University Brain Research Centre.
• Links:
  • 

• : Possibility of extending it to master thesis
• Student: Filip Foltyn
• Namespace in the wiki: cfbd
• The goal of the project: The goal of the project is to enhance a prototype of a counterfactual-ready time-series classifier for ECG data.
• Technology: Python, PyTorch, NumPy, pandas
• Description: The Back2Normal classifier decomposes an input signal into a normal component and a residual component. Classification is performed using the residual. Counterfactuals are generated by attenuating the residual, producing a signal that remains similar to the original but is closer to the learned normal pattern. A working implementation is available to the project team, but it requires the following extensions:
  1. an architecture that processes signal slices instead of the entire signal;
  2. an improved NormalPredictor that separates the normal and residual components;
  3. support for multi-label classification.

• Links:
  • Paper describing the method

• : Possibility of extending it to master thesis
• Student: Szymon Sołtys
• Namespace in the wiki: rcfbd
• The goal of the project: The goal of the project is to develop a counterfactual-ready classifier for evolving data streams, designed to generate counterfactual explanations that remain valid under anticipated concept drift.
• Technology: Python, PyTorch, NumPy, pandas, scikit-learn
• Datasets: SEA, Hyperplane
• Description: Counterfactual explanations are typically generated for a classifier at a particular point in time. In evolving environments, however, the data distribution and the classifier's decision boundary may gradually change. As a consequence, a counterfactual that is valid when it is generated may become invalid after even a relatively small concept drift.
• The project will investigate a different approach: instead of making only the counterfactual generator robust to such changes, the classifier itself will be designed and trained to support the generation of temporally robust counterfactuals. The starting point will be the Counterfactuals by Design idea, in which properties required from counterfactual explanations are incorporated directly into classifier design and its learning objective. The project should develop and experimentally evaluate a classifier in which additional constraints imposed during training shape the probability space around the decision boundary. These constraints should encourage the existence of regions in which generated counterfactuals remain valid despite bounded future changes in the data distribution and the resulting movement of the decision boundary. In particular, the project will investigate:
  1. how concept drift can be formally described and bounded in the SEA and Hyperplane data streams;
  2. how the expected range or direction of future drift can be incorporated into the classifier's training objective;
  3. how additional loss terms or architectural constraints can shape the probability distribution around the decision boundary;
  4. how to generate counterfactuals that do not merely cross the current decision boundary, but reach regions expected to remain valid under future bounded drift;
  5. how to quantify the robustness of a counterfactual as the range of future drift under which its target prediction remains unchanged;
  6. the trade-off between predictive performance, counterfactual proximity and temporal robustness.
• An optional research direction will be a Dual Decision Boundary Classifier. Instead of using exactly the same boundary for prediction and counterfactual generation, the model could learn two related boundaries:
  1. a classification boundary, optimized primarily for predictive performance;
  2. a counterfactual robustness boundary, defining a safer region that a counterfactual should reach in order to remain valid under anticipated future drift.
• Under this formulation, a counterfactual would not be considered sufficient immediately after crossing the current classification boundary. It should cross the robustness boundary and therefore enter a region providing a safety margin against expected movement of the classification boundary.
• Experiments will be performed on synthetic evolving data streams, initially using SEA and Hyperplane generators. Their controlled drift mechanisms will make it possible to systematically study when a counterfactual becomes invalid and whether the proposed classifier can generate explanations that remain valid for longer periods of stream evolution.
• The final evaluation should compare a standard classifier and counterfactual generation approach with the proposed Counterfactuals-by-Design model. Apart from standard predictive and counterfactual quality measures, the evaluation should introduce measures of counterfactual temporal robustness, e.g., how long or under how much drift a generated counterfactual remains valid.
• Expected outcome: A prototype of a drift-aware, counterfactual-ready classifier and an experimental evaluation demonstrating the relationship between classifier design, concept drift and the temporal robustness of generated counterfactual explanations.

• Links:
  • Paper describing the original Counterfactuals by Design method

• : Possibility of extending it to master thesis
• Student: Aleksanda Mes, Łucja Obrusik
• Namespace in the wiki: tobiipm
• The goal of the project: The goal of the project is to implement an additional Tobii-Pytracker analyzer that uses process mining to study how users investigate images during predefined tasks.
• Technology: Python, Tobii-Pytracker, PM4Py, pandas
• Description: Tobii-Pytracker is a Python framework for running eye-tracking experiments, collecting multimodal data, and analysing gaze, fixations, saccades, scanpaths, and attention to image regions. The project will extend it with an analyzer based on PM4Py, a Python process-mining toolkit. Image regions will be treated as activities and transitions between them as process events, allowing common image-inspection patterns to be discovered and compared.

• Links:
  • Tobii-Pytracker
  • PM4Py

• Student: Bartosz Kosmala i Maria Rostworowska
• Namespace in the wiki: dc
• The goal of the project: The goal of the project is to prepare an anomaly-detection competition using the CoinDefect6k dataset created with the industrial partner Proditec.
  • Note: Possible forks of this project:
    • Enhancing original dataset with masks indicating real anomaly locationo – combination of automatic/manual anomaly detection.
    • Creating labelling system with TobiiPytracker that will create multimodal set: images/masks/eye-gaze/voice (for a challenge to NeurIPS)
• Technology: Python, Codabench or Kaggle, Docker
• Description: CoinDefect6k contains images of normal and defective coins acquired using an industrial inspection system. The project involves selecting a competition platform, defining the dataset split and evaluation protocol, preparing a scoring procedure and baseline, and configuring a pilot competition for coin-defect detection.
• Links:
  • CoinDefect6k dataset
  • Codabench
  • Kaggle competitions

• Student: Maria Tyrpa, Magdalena Kiszka
• Namespace in the wiki: cmano_labels
• The goal of the project: Design a system for labelling textual samples.
• Technology: Python, Streamlit or FastAPI, pandas, SQLite
• Description: The XAI-FUNGI dataset contains transcripts from a study on the comprehensibility of XAI methods. In many cases, participants disagreed with, questioned, or misunderstood the explanations. The project has two goals:
  1. manually label part of the data to identify different types of contestation and create training data for ML-based methods;
  2. build an online interface that allows multiple users to label larger numbers of transcript fragments.

• Links:
  • XAI-FUNGI dataset
  • Labelling demo repository -- more details and instructioons

• : Possibility of extending it to master thesis
• Student: Michał Kobylski
• Namespace in the wiki: cmano_jev
• The goal of the project: The goal of the project is to test JEV-9B in an existing conversational interface for hyperparameter optimization.
• Technology: Python, PyTorch, Hugging Face Transformers, Pydantic or JSON Schema
• Description: Hyperparameter optimization is computationally intensive and requires knowledge of model parameters and their effect on model quality. An existing system uses an LLM to translate between ML concepts and user-oriented requirements, such as performance, labelling effort, time, and energy consumption. JSON is used as an intermediate representation, but translating user intent into valid JSON is not always reliable. JEV-9B will be tested as an alternative because it returns probabilities over predefined answers, making it suitable for structured decisions.

The possible architecture is presented below. The important thing: the system has to be as lightweight as possible, no queries to expertnal services, only small language models that runs locally. How JEV can be used here

For example user says:

“I want to decrease labelling effort by 20%.”

JEV can decide:

Which parameter is referenced?

review_budget
training_fraction
both
ambiguous

What operation is requested?

decrease
increase
set
restrict to range

What does the value mean?

relative percentage
percentage points
absolute value
target proportion

What is the reference?

current configuration
best configuration
selected configuration
unspecified

• Links:
  • AutoTrust JEV-9B

• : Possibility of extending it to master thesis
• Student: Kacper Mielniczek, Aleksander Pugowski
• Namespace in the wiki: cmano_interface
• The goal of the project: The system combines image labelling, hyperparameter optimization, model evaluation, and explanation. The goal is to design a common interface that allows users to communicate with these components in natural language.
• Technology: Python, FastAPI, Pydantic, JSON or JSON Schema
• Description: The project involves designing a REST API that enables communication between the system components using JSON or another structured representation. The interface will connect the anomaly-detection system with an LLM-based, or possibly JEV-based, conversational component responsible for communication between the user and the ML system.

• Links:
  • Demoo GUI App and more instructions
  • Additional system for conversaitonal management designed for the Auto-labelling component. Access on request, only for students involved in the project: PEER Cont Management

• : Possibility of extending it to master thesis
• Student: Natalia Żuchowska, Tomasz Fela
• Namespace in the wiki: xaifungiFIN
• The goal of the project: Celem pracy jest zebranie + miejscami uspójnienei metodologii z poprzedniego semestru i pomoc przy pisaniu artykułu naukowego podsumowującego te analizy i wnioski. W szczególnosci odnośnie do analizy dynamiki wywiadów.
• Technology: Python, text analysis,
• Description: Zbiór XAI-FUNGI zawiera transkrypcje wywaidów z ekspertami, studentami informatykami i studentami studiów nieinformatycznych. Każdy transkrypt jest plikiem CSV, z wydzielonymi fragmentami w których uczestnik analizuje na głos wybrany element charaktecryzujacy model do klasyfikacji grzybów. Celem jest zbudownaie klasyfikatora, który będzie staral się określić do jakiej grupy należy dany użytkownik, bazujać na tych transkrypcjach.
• Links:
  • Wyniki z poprzedniego semestru: xaifungiA, xaifungiB, xaifungiC, xaifungiD
  • XAI-FUNGI
  • Robimy forka tego repozytorium i pracujemy na forku!: Starting point GitHub

• : Possibility of extending it to master thesis
• Student: Piotr Klocek, Benedykt Okramus,
• Namespace in the wiki: sgg
• The goal of the project: Develop and evaluate a pipeline that automatically generates small knowledge graphs describing objects depicted in images and the relations between them
• Technology: Python, computer vision, vision-language models, scene graph generation, RDF/Semantic Web, RDFLib
• Description: Object detection models such as YOLO can identify objects depicted in an image, but their output is usually limited to a set of labels and bounding boxes. In this project, we want to go one step further and generate a small graph that also describes relations between the detected objects, for example, a person riding a horse, a person holding a sword, or a tree behind a building. The project will focus primarily on cultural heritage images, such as digitized paintings or photographs, but methods and datasets from other domains may also be used.
  The work should include: (a) reviewing and testing existing scene graph generation or vision-language models, (b) selecting or preparing a small collection of images for evaluation, (c) transforming generated objects and relations into an RDF knowledge graph, (d) linking selected concepts to external knowledge bases or controlled vocabularies, and (e) evaluating the quality of the generated graphs. Depending on the selected scope, the project may also explore pixel-level grounding, validation with SHACL rules, provenance and confidence representation, comparison of several models, or a simple interface for presenting and correcting generated graphs.
• Links:
  • Scene Graph Generation: A Comprehensive Survey
  • Neural Motifs
  • From Pixels to Graphs: Open-Vocabulary Scene Graph Generation with Vision-Language Models + Pix2Grp: code and pretrained models
  • Open-set Panoptic Scene Graph Generation + OpenPSG: Panoptic Scene Graph Generation
  • Visual Genome
  • Synthetic Visual Genome

• : Possibility of extending it to master thesis
• Student:
• Namespace in the wiki: emb
• The goal of the project: Develop an improved knowledge graph embedding for recommendations in a CIDOC-CRM-based cultural heritage knowledge graph.
• Technology: Python, RDF, SPARQL, PyTorch/PyKEEN, knowledge graph embeddings, recommendation systems
• Description: Our recommendation pipeline combines knowledge graph embeddings, nearest-neighbour search, and SPARQL-based filtering. In the first prototype, ComplEx performed better than TransE, but entities considered relevant by experts were still often too distant in the embedding space. The project will investigate models better suited to complex, event-centric graph structures, such as RotatE, PairRE, CompGCN, or NodePiece, as well as relation weighting and selected graph transformations. Experiments will initially use small, representative subgraphs, with the most promising solutions subsequently evaluated on the full graph.
  Expected deliverables: a reproducible experimental pipeline, an expert-based evaluation dataset, a comparison of selected approaches in terms of recommendation quality and computational cost, and an improved embedding configuration integrated with the existing recommendation pipeline.
• Links:
  • Our paper: https://link.springer.com/chapter/10.1007/978-3-032-29924-6_37 + repo: https://gitlab.geist.re/pro/chex-recs
  • Analysis from previous WSHOP: graphrecs

• : Possibility of extending it to master thesis
• Student: Adam Lisik, Kacper Kula
• Namespace in the wiki: iconclass2
• The goal of the project: Continuation from previous semester, probably by the same people, so the details are obsolete here
• Links:
  • Previous project: iconclass

• : Possibility of extending it to master thesis
• Student: Jakub Budzyński, [są wolne miejsca]
• Namespace in the wiki: polykgch
• The goal of the project: Develop and evaluate a benchmark dataset for representing multiple, potentially conflicting narratives in cultural heritage knowledge graphs
• Technology: Python, RDF/Semantic Web, data integration, NLP/LLMs, SPARQL
• Description: Existing knowledge graph benchmarks typically assume that one consistent representation of an entity or event can be constructed. This project will instead focus on cases in which museums, researchers, communities, or other sources provide different interpretations of the same cultural heritage object or event. Students will identify suitable open datasets, connect descriptions referring to the same entities, extract and annotate claims, and model their provenance and relationships such as agreement, contradiction, or refinement. The resulting dataset should include benchmark tasks and evaluation criteria for selected problems, such as claim extraction, perspective identification, or polyvocal question answering. There is no need for next polyvocal ontology, we have to develop some datasets!
• Links:
  • Some datasets:
    • https://collectie.wereldculturen.nl/ Nationaal Museum van Wereldculturen or other Colonial Collections / Provenance Research
    • Europeana Migration https://www.europeana.eu/en/themes/migration Why did person X migrate? (1) institutional narrative (2) personal narrative (3) family narrative (4) national narrative
    • Wikipedia + Talk Pages
  • One idea:
    • 50-100 objects (colonial objects, etnographic artefacts) from e.g., Wereldmuseum, Rijksmuseum, British Museum, Smithsonian
    • Get several descriptions for each: museum description, provenance report, scholarly article (if exists), Wikipedia article, any others?
    • Build the graph, including Entity, Claim, Perspective, Evidence, Conflict

• : Possibility of extending it to master thesis
• Student: Igor Zamojski, [są jeszcze wolne miejsca]
• Namespace in the wiki: highxai
• The goal of the project: Develop a method for transforming feature-level explanations of machine learning predictions into concise, higher-level explanations based on a formal domain ontology
• Technology: Python, machine learning, XAI (SHAP, others?), RDF/Semantic Web, ontologies, SPARQL
• Description: XAI methods such as SHAP explain predictions through the contributions of individual features, which may be difficult to interpret when many features are involved. This project will use the NHANES dataset and existing mappings of its variables to biomedical ontologies to generate higher-level explanations, for example, metabolic factors instead of a list of laboratory measurements. The main task is to develop a method that automatically selects an appropriate level of ontological abstraction while controlling information loss and preserving fidelity to the original explanation. The method will be compared with standard SHAP explanations and fixed feature groups.
• Links:
  • NHANES datasets and documentation
  • NHANES-metadata/ontology-mappings – mappings to Experimental Factor Ontology (EFO), National Cancer Institute Thesaurus (NCIT), Food Ontology (FOODON)
  • On the multiple roles of ontologies in explanations for neuro-symbolic AI

• The goal of the project: Experiment with different methods of GPU acceleration for ontology reasoning
• Technology: OWL ontologies, python, optional PyTorch or JAX, GPU/CUDA.
• Description:

Ontology reasoners (the engines behind Knowledge Graphs and the Semantic Web) usually run sequentially on CPUs. Because they rely on complex symbolic logic, scaling them up is notoriously difficult. Modern GPUs excel at massive parallel math—can we leverage GPU hardware to make logical reasoning lightning fast?

In this project you will investigate whether core reasoning tasks can be accelerated by converting logical rules into GPU-friendly operations (like matrix multiplication or Boolean circuit evaluations) and benchmark your results against standard CPU reasoners. KLAY may be explored as a backend for layerised, GPU-friendly evaluation of such compiled circuits. The implementation should compare a CPU baseline with one or more GPU approaches. Evaluation should measure correctness, runtime, throughput for increasing batch sizes, GPU memory use, and preprocessing cost. The project should identify which reasoning tasks benefit from GPU acceleration and when conventional CPU reasoners remain more efficient.

• Links:
  • To Neuro-Symbolic Classification and Beyond by Compiling Description Logic Ontologies to Probabilistic Circuits
  • Neuro-symbolic learning over OWL 2 DL via consequence-based compilation to differentiable circuits
  • KLay: Accelerating Arithmetic Circuits for Neurosymbolic AI

• The goal of the project: Implement and evaluate a neurosymbolic pipeline using DeepLog to combine neural spatial perception with logic rules for automated room classification in architectural layouts.
• Technology: Python, PyTorch, DeepLog
• Description:

Spatial enrichment in architectural models (such as 2D floor plans and 3D BIM models) requires combining noisy visual perception with strict domain rules. Pure neural networks excel at extracting spatial features (e.g., room size, geometry, or fixture presence) but lack topological guarantees, whereas pure expert systems fail when facing architectural variations or missing features.

In this project, you will build a DeepLog pipeline to classify room types (e.g., bedroom, bathroom, living room) by integrating neural predicates with a symbolic knowledge base. The results may serve both as an improvement in the room classification task, but also as a benchmark and use case of Neurosymbolic systems.

• Links:

• DeepLog: Official Tutorial and Documentation

• DeepLog: A Software Framework for Modular Neurosymbolic AI

• Automatic Room Type Classification using Machine Learning for Two-Dimensional Residential Building Plans

• Comparing Machine Learning and Rule-Based Inferencing for Semantic Enrichment of BIM Models

• The goal of the project: Implement and evaluate an interactive LLM-based system that converts a heterogeneous knowledge graph and its ontology into user-defined networks for network analysis.
• Technology: Python, RDF/OWL, SPARQL, an LLM API, and an interactive visualization framework such as Streamlit, Dash, or Cytoscape.js
• Description:

Knowledge graphs represent entities, relations, attributes, class hierarchies, and semantic constraints. One way of using the information is for network analysis under different conditions. For example, a bibliographic KG can produce researcher–publication, co-authorship, citation, institution–researcher, or researcher–topic networks.

In this project, you will build an interactive system that allows users to describe a desired network in natural language, such as: “Create a weighted co-authorship network for the 18th century.” An LLM will translate the request into a structured network specification containing node types, relations, filters, projection rules, directions, and weights.

The system will validate this specification against the KG ontology, generate and execute a SPARQL query, retrieve the relevant subgraph, and deterministically construct the requested network. The result will be visualized interactively and analysed using measures such as centrality, shortest paths, density, and community structure.The interface should make the transformation transparent by showing the selected ontology elements, generated query, construction rule, supporting KG triples, and network statistics.

• Links:
  • Networks and knowledge graphs
  • Preview version of our ChexGraph
  • CHExNet an example of network manually created from KG in which we conducted some analysis

• The goal of the project: Investigate and compare different ways of representing narratives using knowledge graphs.
• Technology: : Python, RDF/OWL, SPARQL
• Description:

Narratives can be represented in knowledge graphs in different ways. A KG may model a story as a sequence of events, a network of characters and relationships, a causal structure, a temporal chain, a set of narrative roles, or a combination of these representations. More complex models may also include perspectives, uncertainty, themes, sub-events, and different levels of narrative organization. In this project, you will compare different proposed semantic-web models for narratives. You will classify them and reflect on each of their advantages and disadvantages. If there is enough time you can evaluate their interactions with LLMs

• Links:
  • Paper Structured Representation for Narratives
  • Paper From Nodes to Narratives: A Knowledge Graph-based Storytelling Approach