Master and Bachelor Theses

2022

Student Name: 
Chenkai Wang

Project Duration:
Fall Semester, 2022

Type:
Master thesis

Abstract:
The fairness issue has been an urgent topic, with the popularity of mobile applications equipped with recommender systems. In this paper, we investigated the fairness problems from user’s point of view through a survey. We prudently designed the survey questions and collected 1,300 responses in total. After filtering, 630 samples composed our Dec2Dataset. (Since the dataset is finally confirmed in December.) We embraced the common constructs of Perceived Ease Of Use and Perceived Usefulness. Meanwhile, we introduced the concept of Perceived Fairness to explore the relationships among these three principal constructs. We applied structural equation modeling methods to validate the model structure from collected data, proving our hypotheses’ consistency. Through analysis with statistical reliability and validity, we confirmed that (1) Perceived Ease Of Use has a significantly positive influence on Perceived Fairness; (2) Perceived Fairness positively influences Perceived Usefulness; (3) these three constructs have both direct and indirect positive effects on Behavioral Intentions. This could bring a new perspective to researchers about the fairness problems in two-sided recommender systems.

Thesis:

Student Name: 
Mohamed Aziz Dhouib

Project Duration:
Spring Semester, 2022

Type:
Bachelor thesis

Abstract:
Never in the history of mankind have we been as safe from diseases and injuries as today. However, one domain has not seen much improvement and breakthroughs despite having been studied for over a century now: The care and treatment of Alzheimer’s patients. To this day, no cure has been found, and the available treatments are slim to none, while all being palliative in nature. Taking care of an Alzheimer’s patient eats up a lot of time. They require near-constant attention, and caregivers are often unable to work or live a normal life because of this. Some decide to hire a professional nurse to assist with these basic tasks, which often end up being extremely expensive. While professional care is absolutely mandatory at the most advanced stages of the disease, the earliest stages require nothing more than supervision, which can easily be digitalized. To address this challenge, this project prototypes a solution using AR and artificial intelligence, to help Alzheimer’s patients in their day-to-day lives to accomplish the most basic tasks by themselves, making them a bit closer to living like healthy people.

Thesis:

 

2021

Student Name: 
Timothée Duran

Project Duration:
Fall Semester, 2021

Type:
Master thesis

Abstract:
Mobility is one of the many challenges blind or visually impaired people face. This thesis started with the hypothesis that technologies such as LiDAR could replace white canes by providing a more effective obstacle detection tool. However, through literature review, SMEs, and ethnographic interviews with blind and visually impaired people, we refuted the hypothesis and quickly pivoted the project. The result is an application allowing users to explore simplified indoor or outdoor venues maps by moving their finger on the screen. Precise haptic feedback, sounds, and audio descriptions are given depending on the element under the finger. By doing so, users can understand any map even when having no vision, allowing them to gain autonomy and confidence in their everyday lives. Talks are taking place to open-source the project and transforming it into a framework for use in existing apps such as audio-GPS or even public transport apps to display maps of stations.

Thesis:

Student Name: 
Junze Li

Project Duration:
Fall Semester, 2021

Type:
Master thesis

Abstract:
Humor is a magnetic component in daily human-human communication. Teaching machines to recognize and understand humor is beneficial to human-computer interaction systems. Humor is a subjective feeling, which means that the text considered humorous may be offensive to others. However, most state-of-the-art models designed for humor recognition are based on large-scale deep learning models, falling short of interpretability. Very little work exists to inspect whether and why detected humor is rooted in offensive social biases.

In this work, we design an approach to inspect and mitigate gender biases in humor recognition systems using the influence function, which is a type of example-level interpretable method to compute the influence of each training sample on the prediction for each test sample. By conducting a series of experiments, we found that the influence function is helpful for interpreting the behaviors of humor recognition models and can detect gender-biased training samples. Furthermore, since high-quality labeled humor data is limited, we also tested our approach in the low-resource regime. Finally, we found that gender-swapping is a promising method to
debias the models and datasets for humor recognition.

Overall, our work is the first step towards interpreting and debiasing methods for humor recognition. Our approach to detecting gender bias in humor recognition could also be transferred to detecting other types of social biases.

Thesis:

Student Name: 
Chun-Hung Yeh

Project Duration:
Fall Semester, 2021

Type:
Master thesis

Abstract:
In the modern era, people more often suffer from severe emotional distress, which may result from the risks such as relationship difficulties, financial strain, or chronic medical illness. Being unable to recover from it may potentially lead to self-destructive behaviors or even suicide. Although therapeutic consultations are available to assist people in distress, most of them are required to be synchronous and face-to-face with therapists. Moreover, in recent years, more and more people may choose to use text-based platforms for their mental health support. Therefore, the task of empathetic conversation agents has been a popular research topic aiming at generating syntactically and emotionally appropriate responses following dialog contexts.

In this work, we develop multi-turn empathetic dialog models which can not only recognize human emotions but also rely on the well-developed cognitive framework to assess expressed empathy in texts. The experiments reveal that our models can evidently generate empathetic responses in accordance with the context of speaker utterances. Lastly, we benchmark them with the baselines of two previous works based on automatic evaluation and human assessment respectively.

Thesis:

Student Name: 
Costanza Volpini

Project Duration:
Spring Semester, 2021

Type:
Master thesis

Faculty Supervisor:
Dr. Pearl Pu

Industry Supervisor:
Sarah Cheng,
Logitech Europe S.A.

Abstract:
The demand for web pages has been growing steadily in recent years; consequently, the web teams at Logitech have grown a lot. However, due to difficulties in scaling with the increased demand, teams often had to leave out some good design resulting in a design debt. The latter comes in many forms, such as accessibility requirements not being met or inconsistencies across web pages. This thesis aims to explore design systems as a solution to enable more effective collaboration and to ensure consistency across pages without compromising designers’ creative flexibility. A design system is defined as the set of interconnected patterns and shared practices for efficiently and consistently delivering digital products. Moreover, a design system also represents a centralized resource that will be referred as the single source of truth by all the web teams, therefore enabling a better collaboration. A large part of the research focused on what value a design system would bring to the organization. Many qualitative methods, e.g. interviews and usability testing, have been used to build our system iteratively and to assess how much each version was aligned with the users’ need. Constructive positive feedback has been received, and next steps, that will be tackled after this thesis, have been identified. Lastly, Apollo’s Logitech web design system confirmed that different teams can collaborate thanks to a shared vision.

Thesis:

 

2020

Student Name: 
Thanuditha Ruchiranga Wickramasinghe

Project Duration:
Fall Semester, 2020

Type:
Master thesis

Faculty Supervisor:
Dr. Pearl Pu

Industry Supervisor:
Sydney Bovet,
Logitech Europe S.A.

Abstract:
The Logitech G HUB software provides a Lua scripting environment for users to perform advanced customizations on Logitech Gaming gear. This feature requires the user to have some familiarity with programming in Lua language and a lengthy reference manual has to be referred to learn the Lua API functions provided by G HUB. This project aims to explore how visual programming can be utilized in this setup to make this scripting environment in G HUB easy to use, more intuitive and more approachable, for users without any programming experience. The project follows a User Centered Design process where progress is made in iterations, carrying out user tests at each step to verify the design decisions.

To start with, two types of candidate visual programming setups were evaluated. One setup used a notion of blocks that fit together to create meaningful programs. The other setup used a notion of a graph of nodes connected to each other using lines that specify the flow of execution and the flow of data. Low fidelity prototypes of the two setups were developed and tested with users. Based on user input, it was decided to proceed with the blocks-based setup. Thereafter, a high-fidelity visual programming prototype that functioned fully integrated with the G HUB software was developed. A final round of user tests was carried out to discover any further usability issues in the prototype to be addressed in future iterations. A majority of the participants of the high fidelity prototype user test stated with confidence that they would prefer to use the prototype rather than write code in text to perform device customizations.

Thesis:

 

2017

Student Name: 
Yumeng Hou

Project Duration:
Spring Semester, 2017

Type:
Master thesis

Abstract:
Food is vital to human life. Recent studies indicate that there is an influence between food and emotional wellbeing, but existing tools seldom provide insights on it. To fill this gap, we propose to build digital profiles on users’ food log data with emotion records. Information visualization is used to enable users to visually explore their food journeys while revealing the correlation between personal nutrition intake and emotions. We designed a two-level visual representation featuring self-designed glyphs and novel storytelling layouts for nutrition-emotion visualization. The entire solution is implemented as a web-based interactive interface enriched with dynamics. Evaluations verify that our designs boast pleasant aesthetics and achieve more effective storytelling due to improved visual intuition.

Thesis: