Preference elicitation with argument-based machine learning
master's thesis
Abstract
We have developed a novel method for determining people's preferences based on their explanations of visual data. To this end, we have extended the existing framework for argument-based machine learning (ABML), which includes argument-based rule learning and an interactive knowledge refinement loop, with a recommendation engine and a pipeline based on convolutional neural networks to obtain interpretable data from images. We have developed an interactive application inspired by ABML to determine users' dating preferences. To enable a user to argue and explain his preferences based on image data, we introduced a novel approach where the user explains his preferences by drawing rectangles to select a part of the image he likes or dislikes. The ABML knowledge refinement loop allows the user to focus on the most critical parts of the current knowledge base and helps the user to adequately explain selected relevant examples - in our case, images.
We have shown experimentally that the new approach to preference elicitation
allows successful preference elicitation when it comes to dating. All users
found the final selection of images useful, and the selection of images that the
user is likely to prefer gradually improved during the interaction. The
identified preferences of each user of the application are presented as a
rule-based model that helps to quickly find images according to the user's
taste. We have shown that this rule model is easy to interpret. All participants
found that most of the rules in the final model matched their preferences.
The beauty of our approach to preference elicitation is that, at least in principle, we can address any domain that can be represented by images, where people can explain which parts of the image they like or dislike, provided that it is possible to obtain meaningful attributes from images.
Keywords
knowledge acquisition;preference elicitation;argument-based machine learning;convolutional neural networks;weakly supervised object localization;transfer learning;online dating;computer science;computer and information science;master's thesis;
Data
| Language: | English |
|---|---|
| Year of publishing: | 2020 |
| Typology: | 2.09 - Master's Thesis |
| Organization: | UL FRI - Faculty of Computer and Information Science |
| Publisher: | [J. Grabnar] |
| UDC: | 004.85(043.2) |
| COBISS: |
40090883
|
| Views: | 822 |
| Downloads: | 168 |
| Average score: | 0 (0 votes) |
| Metadata: |
|
Other data
| Secondary language: | Slovenian |
|---|---|
| Secondary title: | Ugotavljanje preferenc z uporabo argumentiranega strojnega učenja |
| Secondary abstract: | V magistrskem delu smo razvili novo metodo za določanje uporabnikovih preferenc, ki temelji na analizi slik in uporabnikovih argumentov o slikah. V ta namen smo razširili obstoječe ogrodje za argumentirano strojno učenje (ABML), ki vključuje argumentirano učenje pravil in interaktivno zanko za zajemanje znanja. Zanka vključuje priporočilni sistem in cevovod, ki temelji na konvolucijskih nevronskih mrežah, s katerimi iz slik dobimo atribute, ki jih je mogoče interpretirati. Za namen določitve preferenc uporabnikov smo razvili interaktivni vmesnik, ki sloni na metodi ABML. Uporabniku smo omogočili, da lahko svoje preference utemelji in pojasni glede na atribute na sliki. Za ta namen smo uporabili nov pristop, pri katerem lahko uporabnik svoje preference razloži z označevanjem delov slike. S pravokotnikom lahko izbere del slike, ki mu je všeč ali ne. Zanka za zajemanje znanja ABML uporabniku omogoča, da se osredotoči na najbolj kritične dele trenutne baze znanja in mu pomaga, da ustrezno razloži izbrane slike. S tremi udeleženci smo izvedli poizkuse v domeni slik ljudi in tako pokazali, da nov pristop omogoča uspešno pridobivanje uporabnikovih preferenc. Vsem udeležencem se je zdel končni izbor slik uporaben. Število všečkov slik, ki jih je predlagala metoda, se je postopoma izboljševalo. Končne preference vsakega uporabnika aplikacije so predstavljene kot model, ki temelji na pravilih in pomaga hitro najti slike, ki ustrezajo uporabniku. Pokazali smo, da je ta model pravil enostavno interpretirati. Vsi udeleženci so ugotovili, da se večina pravil v končnem modelu ujema z njihovimi preferencami. Prednosti obravnavanega pristopa pri ugotavljanju preferenc je v širini spektra, ki ga pokriva. Načeloma je z njim mogoče nasloviti katero koli domeno s slikami. Da je uporaba tega pristopa smiselna, morajo biti uporabniki zmožni pojasniti, kateri deli slike so jim všeč ali ne, hkrati pa mora biti uresničen pogoj, da je iz slik mogoče pridobiti smiselne atribute. |
| Secondary keywords: | zajemanje znanja;ugotavljanje preference;argumentirano strojno učenje;konvolucijske nevronske mreže;šibko nadzorovana lokalizacija objektov;preneseno učenje;spletno iskanje partnerjev;računalništvo;računalništvo in informatika;magisteriji; |
| Type (COBISS): | Master's thesis/paper |
| Study programme: | 1000471 |
| Embargo end date (OpenAIRE): | 1970-01-01 |
| Thesis comment: | Univ. v Ljubljani, Fak. za računalništvo in informatiko |
| Pages: | XII, 116 str. |
| ID: | 12168684 |
Recommended works:
Preference elicitation with argument-based machine learning
2020,
master's thesis
Superposition and compression of deep neutral networks
2020,
master's thesis
Recovery of superquadric parameters from depth images using deep learning
2020,
master's thesis
Avtomatsko postavljanje ločil v surovem tekstu
2020,
magistrsko delo
Analysis of methods for few-shot semantic segmentation
2021,
master's thesis