|
|
Human-computer interaction model for multi-objective recommender systems
Machala, Patrik ; Peška, Ladislav (advisor) ; Lokoč, Jakub (referee)
One of the most developing research fields of information retrieval are recommender systems. They typically try to recommend a few of the most relevant or most suitable items to users from all the candidates when the number of canidates can be in the or- der of thousands or millions. However, it turns out that relevance alone is not enough. Therefore, this work focuses on multi-objective recommender systems using the beyond- relevance objectives. The aim of the thesis is to find out new knowledge about this specific type of recommendation, especially in the connection with the field of HCI, i.e. user and computer interaction that has not been explored much so far. The software output of the work is a web application and a modified recommender system. These two components were used in a user study, where, among other things, we investigated whether users were willing to explicitly set the parameters for a multi- objective recommender system by assigning weights to each of the objectives, compared different variants of metrics for these objectives, mechanisms for assigning weights and different level of detail of texts and visualization of the explanations of the recommen- dations. The results of our experiment show that users perceive the benefit of setting weights for objectives to improve recommendations....
|
|
|
Generating synthetic data for an assembly of police lineups
Dokoupil, Patrik ; Peška, Ladislav (advisor)
Eyewitness identification plays an important role during criminal proceedings and may lead to prosecution and conviction of a suspect. One of the methods of eyewitness identification is a police photo lineup when a collection of photographs is presented to the witness in order to identify the perpetrator of the crime. In the lineup, there is typically at most one photograph (typically exactly one) of the suspect and the remaining photographs are the so-called fillers, i.e. photographs of innocent people. Positive identification of the suspect by the witness may result in charge or conviction of the suspect. Assembly of the lineup is a challenging and tedious problem, because the wrong selection of the fillers may end up in a biased lineup, where the suspect will stand out from the fillers and would be easily identifiable even by a highly uncertain witness. The reason why it is tedious is due to the fact that this process is still done manually or only semi-automatically. This thesis tries to solve both issues by proposing a model that will be capable of generating synthetic data, together with an application that will allow users to obtain the fillers for a given suspect's photograph. 1
|
|
|
E-commerce platform designed for continuous optimization and personalization
Dräxler, Peter ; Peška, Ladislav (advisor) ; Bernhauer, David (referee)
The ability of an e-commerce company to collect, synthesize, and utilize data can be vital to its survival. The degree to which a company has the ability to gather data about sources of conversion is proportional to their ability to allocate advertising / im- plementation budget effectively. Measurement and optimization of user flow on a website is equally important as it can deliver a measurable increase in user loyalty, revenue, and profits. Similarly, it can reveal that numerous improvements, while appearing effective on paper, may not perform as well in real-world scenarios. This thesis focuses on the construction of an e-commerce platform centered around data collection and analysis. We use the data to conduct continuous, real-time experi- ments on the platform using the contextual multi-armed bandits algorithm. We develop a recommendation system based on collaborative filtering and set up an experiment to evaluate it's real-world performance. The platform can be easily connected to a business intelligence dashboard to allow data exploration, to support its management in making informed tactical & strategic decisions. The platform is developed with Blazor, which is an emerging technology that en- ables usage on C# code in the browser with compilation to WebAssembly as opposed to JavaScript. We describe...
|
|
|
Sparse Approximate Inverse for Enhanced Scalability in Recommender Systems
Spišák, Martin ; Peška, Ladislav (advisor) ; Vančura, Vojtěch (referee)
In theory, the linear autoencoder EASE is one of the most capable collaborative filtering recommenders for large item domains with sparse user-item feedback. However, the model's weights are determined by the inverse of a matrix of dimension equal to the item set size. This inverse matrix is generally dense, and for large item sets, the computed weight matrix might be too large to store in memory during inference. Consequently, scaling the model beyond tens of thousands of items quickly becomes very expensive. We propose a modification of EASE called SANSA to alleviate the issue. SANSA approximates the weights of EASE with prescribed density via an end-to-end sparse training procedure. To find a method capable of computing the sparse approximation efficiently, we investigate approaches for constructing sparse approximate inverse precon- ditioners. We select a method fitting for very large SPD problems with general sparsity patterns. The training procedure is robust and finds a good approximation of EASE even on datasets with dense item relations. Moreover, as the number of items in datasets grows, SANSA achieves unparalleled efficiency, even compared to EASE's previous state- of-the-art modification focused on scalability. Consequently, SANSA effortlessly scales the concept of EASE to millions of items. 1
|
|
|
Hybrid recommender systems for books domain
Varga, Ondřej ; Peška, Ladislav (advisor) ; Dokoupil, Patrik (referee)
The bachelor thesis deals with the topic of recommender systems, which are especially important in e-commerce field. The main goal of the thesis was to implement a recom- mender system that would cover the needs of recommending in the book domain (on an e-shop dealing with book sales). The focus of the work is the implemented recommender system that handles the data available for the book domain, however, it is designed more generally to be deployable as Recommendation-as-a-Service. The system includes both the recommendation algorithms themselves (collaborative, content-based and hybrid), as well as support for the different phases of the recommendation lifecycle, recommendation performance monitoring and easy administration through an interactive web interface. We then evaluated the recommendation algorithms, with experiments showing that collaborative methods, in particular ALS matrix factorization and ELSA, perform better with respect to relevance metrics. However, hybrid approaches and content-based methods may have advantages with respect to beyond accuracy metrics, especially Coverage and Novelty. 1
|
|
|
Similarity Models for Content-based Video Retrieval
Veselý, Patrik ; Peška, Ladislav (advisor) ; Sixtová, Ivana (referee)
Multimedia retrieval is increasingly important with the skyrocketing multimedia vol- umes produced every day. Therefore many image and video retrieval tools are being developed utilising visual similarity modelling algorithms for similar image retrieval or various visualisations. As such, the quality of the similarity modelling is crucial for these tools. This thesis explores diverse similarity models, their agreement with human percep- tion of similarity and possible improvements of these models. The examined similarity models consisted of colour-based, SIFT-based, and DNN-based models. For the purpose of model evaluation, a user study was conducted to create a dataset of relative image similarity comprising both generic images as well as two compact domains. In this study, the participants were asked to state which of the candidate images was more similar to the query image. The collected data showed the superiority of DNN-based models compared to other evaluated variants. Nonetheless, all similarity models performed significantly better than a random guess. In order to further enhance the performance of the simi- larity models, we fine-tuned the best-performing model (W2VV++) with the collected dataset and achieved significant improvement in some areas. 1
|
|
|
Portfolio performance evaluation
Simonov, Dmitry ; Zavoral, Filip (advisor) ; Peška, Ladislav (referee)
The goal of this work is to develop a tool for tracking and analysis of investment portfolios, with support for automatic financial data retrieval and highly customizable chart creation. In this thesis, we research existing investment tracking tools, comparing them based on their supported features. We also analyze different methodologies of eval- uating performance of an investment. Following this, we develop a web application using ASP.NET and React.js allowing the user to import, track, and chart their investments. Financial data for tracked securities is automatically fetched from various data sources using a handcrafted library, which can be used for retrieval of any kind of data from dif- ferent sources, while supporting easy addition or replacement of such sources. Finally, we evaluate the implemented application and suggest possible extensions and improvements. 1
|
|
|
Configurable point rasterization for large scatterplots
Kulichová, Tereza ; Hoksza, David (advisor) ; Peška, Ladislav (referee)
Scattermore is a simple R package used for scatterplot visualizations. Before its reinvention, it gained popularity with the cytometric community because of its functionality and speed. The new version of scattermore offers a highly customizable API. Again, it is much faster than the R standard plot function because some parts of the code are implemented in C language. Plotting points or lines, combining the data in various ways, and avoiding overplotting simultaneously are possible. Except for that, the conducted analysis offers potential speed optimizations regarding cache utility and par- allelization. 1
|
|
|
Fairness in group recommender systems
Maleček, Ladislav ; Peška, Ladislav (advisor) ; Lokoč, Jakub (referee)
The goal of this thesis is to explore the area of group recommender systems with an emphasis on fairness. In the core part of our thesis, we have created a novel aggregation method called Exactly Proportional Fuzzy D'Hondt's Aggrega- tion that works on top of single-user recommender systems. We have evaluated it on five datasets, in three different recommendation scenarios, and with two different types of artificially created groups. The proposed algorithm performed favorably with respect to several fairness metrics while maintaining a reasonable utility of the recommendation. Furthermore, we have created a set of tools to simplify the evaluation pipeline of group recommender systems. The main parts of the pipeline are a dataset downloader, matrix factorizer, and synthetic group creation scripts. We believe these tools may contribute towards more reproducible research in the group recommender systems domain. 1
|
|
|
Visual Explanations in Music Recommender Systems
Savčinský, Richard ; Peška, Ladislav (advisor) ; Petříček, Tomáš (referee)
Music recommendations from industry-leading algorithms are a product of a hybrid system combining multiple techniques. However, in the end, the user is simply left without additional information why a certain song is present in the result. One way to improve the experience for the user is to provide so-called visual explanations. For that purpose, in this thesis, we designed and proposed various forms of visual explanations for the recommended data from the Spotify API. The main goal was to highlight important hidden relationships between familiar and new music, used by Spotify but also to utilize the actual audio features for the construction of our own recommender system. We developed a modern mobile application that allows users to explore and interact with the visualizations of their own music tastes and also provides tools to customize the experience.
|
guest ::
RSS feed.