Open Topics

Text Mining:

Explore novel interaction techniques - Chatbots for Data Analysis

Chatbots are becoming more prevalent and are actively used by many companies. They offer a voice or text interface to interact with a computer. An example of a chatbot is amazon’s alexa, which can tell the time when asked.
The goal of this project is to find new possible ways to interact with an exploratory data analysis tool. Developing new interaction techniques would allow the user to explore and understand the data in a new fashion. For example, it could be possible to have a chat window next to a scatterplot that enables the user to enter queries such as: ‘show me the average’, which would then be reflected in the scatterplot.

  • Learn about natural language processing
  • Understand and compare interaction techniques
  • Develop a ‘conversation’ with a data analysis tool

Prerequisites: VIS, HCI

Contact: Torsten Möller

Access to Justice

The challenge of "Access to Justice (A2J)" is to make laws and court decisions accessible to lawyers and laymen. A lot of legal information is available online. However, this data is often spread among different hard to use databases that are mostly aimed at experts. 
While it will be unlikely that you will solve this issue with a semester project, an interesting piece to start with is to better understand the needs of different users (within the user group of lawyers as well as the user group of laymen) and to transform them into mockups of a possible website for the retrieval of the appropriate information.

Your results can build upon some efforts from legal scholars (in particular Paul Eberstaller and his efforts at https://risplus.at/) with whom we are collaborating with.

Milestones:

  • conduct interviews with several potential users of said information retrieval interfaces
  • conduct a literature search for interfaces trying to achieve similar tasks
  • summarize your findings through a requirement analysis, including context analysis, task analysis, and personas
  • build low-fidelity prototypes and gather feedback on them
  • build high-fidelity prototypes and gather feedback on them
  • improve your high-fidelity prototype based on these feedbacks

some helpful references:

contact: Torsten Möller
in collaboration with: Paul Eberstaller, Nikolaus Forgo (for advice on users and use case); Evangelos Milios (for advice on information retrieval and text analysis)

pre-req: HCI [NLP would be nice, but not required]
 

Machine Learning:

Interpreting Deep Clustering Results

Deep embedded clustering also called deep clustering is a growing field that combines ideas from clustering and deep learning. The integration of these techniques makes it possible to learn features automatically from the data to increase clustering performance. Current deep clustering methods are hard to interpret,  making it difficult to understand how a clustering result was reached. 

The goal of this project is to develop an interactive visualization tool, e.g. a web based application, for exploring the predictions of deep clustering algorithms and helping to understand their decision making process. 
The student is expected to do a literature review of existing visualization techniques developed for (supervised) deep learning, e.g. feature visualizations, that could be applicable to interpreting unsupervised deep clustering algorithms. The identified methods should then be applied (if necessary adapted) to and compared for existing deep clustering algorithms.

Some research questions of interest that should be considered during the project would be:  How suitable are existing visualization techniques to interpret deep clustering results? How do the different parts of the multi-objective loss of deep clustering techniques relate to each other? Considering multiple clustering models, e.g. K-Means vs DBSCAN, how do the neural network visualizations differ for each of them?

Students working on this project need basic background knowledge in machine learning (e.g. Foundations of Data Analysis), visualisation (e.g. Visualisation and Visual Data Analysis), solid programming skills in Python, and desirably some background with PyTorch, deep learning and some visualization framework, like d3.

Prerequisites: VIS, FDA

Contact: Aleksandar Doknic, Torsten Möller, in collaboration with Claudia Plant & Lukas Miklautz

Visually exploring neural networks

We have a collection of 100,000 different neural networks from the Tensorflow Playground . The core goal of this project is to create a visual interface to understand some of the basic properties of neural networks. Enabling a user to explore should help answer questions like the relationship of number of neurons and number of hidden layers, the impact of batch size, activation functions and other parameters on the quality of the network. Your tasks include:

  • fast prototyping with Tableau
  • getting familiar with the data set
  • querying neural network users on what parameters they want to explore (requirement analysis)
  • development of low-fi and high-fi prototypes

 

Prerequisites: VIS, FDA

Contact: Torsten Möller

Visually Analyzing the Fault Tolerance of Deep Neural Networks

The main objective is to design and implement a good and efficient way of visually investigating the resilience of deep neural networks against silent data corruption (bit flips) based on given empirical measurements. There are many possible causes for such faults (e.g., cosmic radiation, increasing density in chips, lower voltage which implies lower signal charge, etc.), and their "incidence" is expected to increase with current trends in chip architecture.

Starting point for the project is a given example data set which contains information about the relationship between single bit flips across various locations of a certain neural network (which layer, which neuron, which weight, which position within the floating-point representation of a real number, etc.) and the resulting accuracy of the network.

The task is to develop a tool which supports answering various questions about the influence of a bit flip on the resulting accuracy.

Examples for interesting questions are the following:

  • (empirical) distribution of the influence of a bit flip on the resulting accuracy over the positions in the floating-point representation
  • (empirical) distribution of the influence of a bit flip on the resulting accuracy over the layers in the network architecture
  • (empirical) distribution of the influence of a bit flip on the resulting accuracy over the weights in a given layer in the network architecture

In order to answer these questions, an iterative design process is required to

  • start with a requirement analysis (task & data analysis)
  • low-fi prototypes
  • high-fi prototypes
  • refinement
  • constant evaluation of the visual analysis tool.

The data set, the problem setting and the details of the requirements are provided by Prof. Gansterer, the supervision in visual analysis aspects is provided by Prof. Möller.

Prerequisites: VIS, FDA

Contact: W. Gansterer | Torsten Möller

Library for visualization of slices

This is primarily a programming project. Slicing methods are a novel way of visualizing multi-dimensional data. However, there is no publicly-available library for R or Python that makes it easy to use these visualization techniques. The goal of these projects is to develop such a library. Students should have knowledge of Javascript and either R or Python.

 

Prerequisites: VIS
Programming languages: Javascript and (R or Python)

Contact:Torsten Möller|Thomas Torsney-Weir

Visualization for optimization problems

Can visualization beat traditional (offline) optimization problems? The goal of this project is to see how well visually guided optimization can compete with traditional optimization algorithms. Students will develop a visualization system to find optimum configurations of black box (i.e. unknown) algorithms from a contest.

Prerequisites: VIS, Mathematical Modeling
Programming languages: Javascript, (R or Python), and C++

Contact: Torsten Möller|Thomas Torsney-Weir

Semantic understanding of Charts

Research areas: Machine Learning, HCI - Human Data Interaction

This project aims to automatically understand charts (= data visualizations) and translate their meaning into natural language text. This will be done using deep learning. Neural nets draw bounding boxes around objects and label these objects. After detecting all possible objects another network creates a sentence describing the scene in the image by using the object labels. Examples of deep neural networks for image descriptions and chart extraction can be found here: 

A resulting sentence would sound something like this: ‘black and white dog jumps over bar.’ 

The goal of this project is to use such an approach and apply it to different types of charts. We would have a neural network detecting the objects in a plot and another describing the objects by creating a sentence.

  • Learn about state of the art Machine learning
  • Use machine learning libraries such as tensor flow
  • Find/Aggregate datasets

Prerequisites: FDA, possibly VIS

Contact: Torsten Möller, Laura Koesten

Clustering in high noise astronomical data

Area: Machine Learning

With the second Gaia data release astronomers have been flooded with data. One interesting research question which Gaia helps answering is concerning open clusters (OC). OCs are groups of stars born in the same place and time which are regarded as the building blocks of galaxies. Gaia provides precise positions and velocities of  1.6 billions stars which are the features used to extract these open clusters. However, OCs constitute only a small fraction of the full data set and are embedded in a sea of field stars which we consider noise. Hence, the results of density based clustering techniques depend strongly on small changes in the algorithms hyper-parameters. The goal of this project is to survey current techniques which can help to better extract these OCs from the Gaia catalog and potentially design new algorithms which better suit this task.

  • Unsupervised learning

  • Density based clustering

  • Big Data

Prerequisites: FDA

Contact: Sebastian Ratzenböck

Exploratory data analysis in Gaia

Area: VIS/Machine Learning

Astronomical discoveries depend on the quality of the data. Therefore, quality criteria are introduced to filter out bad data points. Within the Gaia data set there are multiple metrics which provide information about the quality of a single entry in the tabular data set (e.g. a star). The goal of this project is to analyze current quality criteria and their effects on the data features and optimally find better suited filter solutions.

  • Visualize the effects of different data filters
  • Unsupervised learning
  • Big data

Prerequisites: VIS, FDA

Contact: Sebastian Ratzenböck 

Anomaly Detection in High Energy Physics with Variational Autoencoders (in cooperation with CERN)

At CERN's Large Hadron Collider (LHC), researchers are searching for new unobserved physics phenomena that could convey missing pieces of today's understanding of the universe. For more than 40 years, many theories for new particles have been put forward and the LHC's data was probed for their evidence. This led to many discoveries, most notably that of the Higgs particle. However, to this date, numerous questions about the nature of matter remain unanswered. Thus, CERN is exploring the use of Machine Learning (ML) in its quest to shed light on those rare unknown phenomena, called anomalies.

Inside the LHC, 1 billion proton-proton collisions are produced every second. The collisions result in new particles which are registered by detectors. Information from the detectors is sent through a data-processing pipeline, where it is denoised, filtered and interpreted. On that resulting data, ML techniques can be applied to search for patterns hinting at anomalous phenomena.

Variational Autoencoding A promising approach to tackle this challenge are unsupervised ML techniques, where no prior theory about the anomaly is needed. One prominent unsupervised ML algorithm is the Variational Autoencoder, which learns to compress input to a much smaller dimensionality called the latent space from where it reconstructs the input. This compression-decompression flow can be used to flag anomalies, whenever the reconstruction does not resemble the corresponding input.

Project Description: The goal of this project is to explore a variational autoencoder applied to simulated LHC collision events. Concretely the four tasks at hand are:

  • Train a Variational Autoencoder on a simulation of anomalous particles (e.g. the Randall-Sundrum Graviton)
  • Explore functional space, especially through a visual analysis
  • Hyperparameters (loss function weighting, dimensionality of latent space, learning rate, etc.)
  • Architecture (number of layers, size of layers and of convolutional filters, pooling, etc.)
  • Analyze latent space (Can we find an interpretation?)

Programming: Python, Tensorflow
Prerequisites: VIS, FDA

Contact: Torsten Möller

Understanding anomalies in high-dimensional spaces

The goal of this project is to visually explain how datasets with varying number of features (or dimensions) per data point behave differently. We understand that high-dimensional spaces behave differently than low-dimensional spaces, but it is difficult to develop an intuition of what these differences are and how they affect datasets.

Visualization could help us intuitively understand what happens to the distances, local geometry, projections, etc. when data points have 3, 10 or 50 dimensions. In other words: Show how the dimensionality of the data affects mathematical operations (dist(X,Y), PCA(X), etc.).

1. Create synthetic datasets to play around with

2. Validate with real datasets with different numbers of dimensions

3. Validate your approach with a user study

Optional: How does data dimensionality affect ML models?

Prerequisites: FDA, optional: VIS

Contact: Aleksandar Doknic, Torsten Möller 

 

Data Visualization/Story Telling:

Identifying charts on climate change or COVID-19


Research area: Data Visualisation

Description: Data visualisations, such as charts, are often used to communicate data about climate change and the COVID-19 pandemic, both in research and in popular news sources. Which charts are commonly used to communicate data about these topics? This project tests and modifies an existing algorithm developed to extract charts from articles in journals and popular news sources in order to inform the creation of a larger dataset of charts which are used to communicate data about climate change and COVID-19.

Tasks:

Prerequisites:
FDA
Programming languages: any

Contact: Torsten Möller, Laura Koesten, Kathleen Gregory

Understanding COVID-19 data

Research area: Data Visualisation, HCI / Human Data Interaction

Description: Data visualisations, such as charts, are used frequently to communicate data about COVID-19, both in research and in popular news sources. In this project we investigate the types of questions that are frequently asked during the COVID-19 pandemic and how charts are used to answer them. We will do this by collecting commonly asked questions and conducting a qualitative study about how people answer these questions for themselves using COVID data visualisations.
Tasks:

  • Collect a sample dataset of COVID related questions (from online resources)
  • Design a study aiming to investigate people’s sensemaking practices

Prerequisites:

  • FDA
  • possibly VIS

Contact: Laura Koesten (+ Kathleen Gregory)

Communicating complex vaccination data

Research area: Data Vis / Storytelling

Description: 

The question is simple: How many people are protected against COVID-19? But the data get's rather complex rather fast. We have people that have been ill before, but some of them have had COVID-19 too long ago and are not protected anymore. Some of them had it without knowing about it. Some are vaccinated, but lost their protection (since they didn't have a booster shot), etc. Are you able to explain these complex connections to

  • a vis expert
  • an epidemeologist
  • a doctor
  • a politician
  • your relatives?

Would you use the same way of explaining it each time or would you choose a different way? Why? The goal of this project is to iterate and find optimal approaches of communicating this data and compare what works and what doesn't.

Prerequisites: HCI + Vis

Contact: Torsten Moeller

HCI - Human Computer Interaction:

Interfaces and influences for fruit fly larvae brains

The goal of this project is to develop interfaces and visualization to answer questions related to the connections and similarities between different anatomical structures of the fluit fly larvae brains. For example, how do neurons on the left and right hemishpheres compare? Or, how does a particular neuron change as the organism grow? To answer these questions we need effective interfaces and visualizations.

Prerequisites:

  • Programming languages: Knowledge of or willingness to learn TypeScript and Go

Contact: Thomas Torsney-Weir, Torsten Möller

Sliders for decision making

Research area: Human Computer Interaction, Data Science, Interfaces

Description: Sliders on interfaces provide a range to select an input value. Sliders can restrict users to entering valid values by only offering a valid range, or they can be used to support multi-criteria decision making. In this project we aim to compare different types of sliders for decision making. This includes triangular, binary and single, sliders as well as “scented widgets”, which are embedded visualizations to facilitate navigation in information spaces.

(See for instance https://dl.acm.org/doi/pdf/10.1145/3240167.3240185)

Tasks:

  • Creating interfaces using different slider types, develop simple alternatives of slider components
  • Design an online user study (including task design, recruitment, usability evaluation)
  • Analyse quantitative and qualitative data from the user study

Prerequisites:

  • HCI, possibly FDA
  • Programming languages: Python or R

Contact: Laura Koesten, Torsten Möller

HDI - Human Data Interaction:

How do people understand charts?

Research areas: HCI - Human Data Interaction, Data Visualization

Textual descriptions of charts are relevant for a variety of application and research areas.

In this project we will create a crowdsourcing study to collect a dataset of charts annotated with a description of their key messages as perceived by the readers of the charts. The data will consist of images (charts) and free text interpretations of the charts. We will analyse the resulting descriptions qualitatively and visualise the results in an interactive manner.

  • Qualitative (content analysis) and quantitative analysis of text and image data
  • Apply NLP techniques to cluster and analyse free text data

Prerequisites: FDA, VIS

Contact: Laura Koesten

Explaining descriptive statistics

With Anscombe's Quartet [1] it was demonstrated quite figuratively that summary statistics can be very misleading or, at least, hard to interpret. Just recently, this example has become quite playful with the Dinozaur Dozen [2]. However, there are a number of statistical measures, that don't have an easy (visual) explanation. One of them is Krippendorf's alpha [3], a very common measure in the social science for measuring the agreement between subjective coders (as in labeling text or documents). The challenge of this project will be to:

  • understand the measure
  • develop simple alternatives
  • develop different visual representations that "bring this measure to life", i.e. make it easy(er) to understand

pre-req: Vis

[1] Anscombe, F. J. (1973). "Graphs in Statistical Analysis". American Statistician. 27 (1): 17–21. doi:10.1080/00031305.1973.10478966. JSTOR 2682899, see also en.wikipedia.org/wiki/Anscombe%27s_quartet

[2] Justin Matejka, George Fitzmaurice (2017), "Same Stats, Different Graphs: Generating Datasets with Varied Appearance and Identical Statistics through Simulated Annealing," ACM SIGCHI Conference on Human Factors in Computing Systems. see also www.autodesk.com/research/publications/same-stats-different-graphs

[3] Krippendorff, Klaus (1970). Estimating the reliability, systematic error, and random error of interval data. Educational and Psychological Measurement, 30 (1), 61–70. see also en.wikipedia.org/wiki/Krippendorff%27s_alpha

Common data or spreadsheet fears

Research area: Human Data Interaction

Description: We are increasingly exposed to data in different aspects of our lives, be that in an ever growing range of professions reliant on data analysis, or in our private lives exposing us to data about us, our activities or using data to inform our decisions. However, many people still do not feel comfortable engaging with a spreadsheet, nor do they have the skills to perform more complex types of data analysis. In this project we aim to conduct a qualitative study to better understand people’s preconceptions by observing them interacting with a spreadsheet and discussing their experiences.

Tasks:

  • Design a mixed method study 
  • Recruit respondents 
  • Qualitative data analysis

Prerequisites:

  • FDA
  • possibly VIS and HCI

Contact: Laura Koesten

Data Science:

Crowdsourcing dataset summaries

Research area: Data Science, Human Computation

Text is more accessible than metadata when describing what a dataset is about and how it should be used. In previous studies we used crowdsourcing to generate data summaries and understand what good summaries look like:

https://www.sciencedirect.com/science/article/pii/S1071581918306153

In this project, the aim is to improve on this method to iterate over the summaries written by the crowd and create a larger dataset of summaries. This can be a useful resource, for instance to train Machine Learning algorithms to create dataset summaries automatically.

Tasks: 

  • Learn crowdsourcing as a method
  • Create a dataset of crowdsourced summaries
  • Analyse text data 

Prerequisites: FDA, HTML and basic Javascript, basic Python, possibly familiarity with APIs 

Contact: Laura Koesten

 

Image Processing:

Neuron signature visualization

Neuroscience is focused on understanding how the brain functions both structurally as well as functionally. This is accomplished by imaging brains using a variety of different methods including EM, confocal, LM, etc. The challenge is to identify neurons across these images to understand how the brain develops throughout an organism's life. The aim of this master's thesis is to push the state of the art in searching for neurons through exploration and evaluation of different methods of exploring similarities between neurons.


Prerequisites: SIP

Contact: Thomas Torsney-Veir | Torsten Möller

Visualization-Supported Comparison of Image Segmentation Metrics

Area: Visualization, Image Processing

Segmentation algorithms, which assign labels to each element in a 2D/3D image, need to be evaluated regarding their performance on a given dataset. The quality of an algorithm is typically determined by comparing its result to a manually labelled image. Many metrics can be used to compute a single number representing the similarity of two such segmentation results, all with specific advantages and disadvantages. The goal in this project is to:

  • Research the segmentation metrics in use in the literature.
  • Create a tool that calculates multiple segmentation quality metrics on an image.
  • With the help of this tool, analyze how the single segmentation metrics perform in detecting specific kinds of errors in the segmentation results, as well as correlations between the metrics.


Prerequisites: SIP, VIS

Contact: Bernhard Fröhler | Torsten Möller

Usability Evaluation of Open Source Volume Analysis Software

open_iA enables users to perform general and specialized visual analysis and processing of volumetric datasets (such as from a computed tomography device). Since it has been developed mainly as a basis for research prototypes, the user interface so far was not developed with usability as first concern.

The goals of this project are:

  • To evaluate the usability of its general capabilities, and optionally of its advanced visual analysis tools. This could for example happen through usability interviews, or user studies comparing it to other (open source and commercially available) solutions.
  • To find innovative ways of overcoming the problems found in the evaluation.
  • Depending on time and interest, to implement some or all of these improvements.

Prerequisites: finished the Signal and Image Processing & the Human Computer Interaction class

Contact: Bernhard Fröhler 

Improving ground truth for CNNs through Uncertainty and a Human-in-the-loop

Convolutional neural networks (CNNs) have become one of the most used tools for image segmentation in any application domain. CNNs, however, require a lot of training data. Especially in volumetric datasets (i.e. 3D images such as from a Computed Tomography device), where the size of a single dataset is typically no less than 1000³ = 1 billion voxels, it is already infeasible to fully manually segment just one of these. Other segmentation algorithms (or CNNs trained for slightly different application domains or input data) can be used to help in creating the ground truth, but their results need to be checked. The goal of this project is to

  • Design and implement a hybrid segmentation method, based on

    • A CNN performing segmentation of an input volume, and
    • Gathering user input for iterative refinement

  • For gathering the user input, a tool should implemented which visualizes the results of the CNN, and allows a user to provide input for future trainings of the CNN
  • The tool should ideally provide guidance on regions to check via uncertainty metrics from the CNN
  • Evaluate the implemented method in comparison to state-of-the-art methods in e.g. material science (datasets and reference algorithms will be provided)

Prerequisites: finished the Signal and Image Processing & the Machine Learning class; attending the Visualization and/or Human Computer Interaction class before also would be beneficial

Contact: Bernhard Fröhler 

Ensemble methods in image segmentation

An image segmentation algorithm labels a pixel. While no segmentation algorithm is always correct, the idea is to work with many different segmentation algorithm that each create a label for a pixel. We call this an ensemble. The idea of this project is to explore how to best combine these different ensemble members to "always" create the right label for the pixel (to explore 'the wisdom of the crowd').
In image segmentation, several methods are known of how to combine a given collection of segmentation results. For example voting methods might label a pixel according to the majority of labels for that pixel in the collection. However, such a vote can be ambiguous, therefore additional rules might be required to arrive at a definitive labeling.

Goal:

  • Gather and/or define a set of useful rules to combine image segmentation results. Furthermore, define a pipeline containing these rules, such that the usage of the rules is depending on the parameterization. A simple example: The pipeline could be based on the majority voting rule, combined with intelligent rules for handling the case of ambiguous pixels, for example through considering the neighborhood of the pixel or the uncertainty of the single segmentation results (if probabilistic segmentation algorithms are used).
  • Explore the parameter space of this generalized pipeline. Set up a framework to “learn” suitable parameters for this pipeline. Test your pipeline on several different datasets and try to come up with optimal parameters. Refine your pipeline until it can produce results at least close to the state of the art algorithms for segmentation such images.
  • Once a set of optimal parameters for some limited number of datasets are established, perform experiments on whether those parameters learned for the generalized combination pipeline are transferable to the processing of new datasets, i.e. other than those the parameters were learned with.

Milestones:

  • Definition of a parameterized, rule-based pipeline for (specific) image analysis tasks.
  • Evaluation of the pipeline and refinement of its parameters on a limited number of datasets
  • Application of the pipeline and the found parameters on a broader range of datasets

Prerequisites: VIS, SIP

Contact: Torsten Möller | Bernhard Fröhler

Smart Image Filter Preview

The analysis of large images (2D or 3D), requires applying filters like smoothing or denoising. Finding the most suitable parameters for a given analysis task through a trial-and-error approach can be  time-consuming. The goal of this project is to develop a tool for a smart preview over the possible outcome of some image processing filters for different parameters for a small region of the image; the outcome of different parameterizations could for example be presented in a matrix; the tool should also be evaluated regarding usability.

Prerequisites: SIP, HCI

Programming languages: Python, C++


Contact: Bernhard Fröhler | Torsten Möller

Theory of Vis:

iTuner: Touch interfaces for high-D visualization

 

In order to understand simulations, machine learning algorithms, and geometric objects we need to interact with them. This is difficult to perform with something like a mouse which only has 2 axes of movement. Multitouch interfaces let us develop novel interactions for multi-dimensional data. The goals of this project are:

  • Develop a touch-screen interface for navigating high-dimensional spaces.
  • User interface designed for a tablet (ipad) to be used in concert with a larger screen such as a monitor or television.

Prerequisites: VIS, HCI

Contact: Torsten Möller|Thomas Torsney-Weir