doc. Ing. Pavel Kordík, Ph.D.

In the recommendation systems, we are currently focusing research on a few open problems that have deep theoretical underpinnings, but whose solutions also have very concrete practical applications. We are exploring the utilization of deep neural networks to reduce the cold start problem of recommender systems, the design of transformers to predict shopping baskets. In the field of general machine learning, we focus on reinforced learning to optimize longer-term metrics such as user satisfaction, on transfer learning methods to incorporate new referral databases, and on using AutoML to optimize the architecture and hyperparameters of recommender systems.

doc. Ing. Pavel Kordík, Ph.D.

doc. Ing. Ivan Šimeček, Ph.D.

Department of Applied Mathematics

This thesis provides a review of modern neural networks for the representation of text and typo corrections with emphasis on transformer-based architectures and pre-trained models. A benchmarking framework is proposed and used to evaluate several models in both their pre-trained versions and after fine-tuning. The evaluations contain several metrics assessing both performance and inference cost on a corpus containing both natural sentences and code snippets, GitHub Typo Corpus. Since the NeuSpell library for spelling correction operates solely on a token-based level, a space correction algorithm was proposed to approximate the spacing after prediction.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

Ing. David Pešek

Department of Software Engineering

Startup formation poses difficulties for technology-oriented student founders who are primarily trained in software development but often lack practical skills in refining a product. They are typically taught to select a topic and complete it, rather than engaging in the iterative process of product development that involves pivoting and launching small. Additionally, these founders must transition from being engineers to assuming roles such as HR managers, team leads, and product owners. This thesis investigated how to systemize early-stage startup development for student entrepreneurs. Existing frameworks, such as Lean Startup, Design Sprint, Agile, and Lean Canvas, were analyzed using criteria including their relevance for B2B contexts, their applicability to technological innovations, and the aspects of startup formation they do not address. Interviews with experienced startup founders and industry experts were also conducted to identify additional gaps and areas for improvement. Based on this analysis, the Lean and Execute startup formation framework was developed, specifically tailored to the needs of tech-oriented student founders operating in the European context. The framework emphasized a focused approach centered on understanding user and client needs, combined with iterative improvements, guiding founders to start with a minimal viable product (MVP) aligned with validated user requirements, rather than pursuing unvalidated solutions. The framework also provided structured recommendations on team formation, funding strategies, and product development. Validation was performed through its application to the authors own startup using a qualitative case study approach. The study evaluated effective aspects of the framework, identifying which elements were implemented successfully and why they were beneficial. It also examined limitations of the framework, noting what it failed to address and the resulting consequences, as well as the implications of deviations from the framework. This case study demonstrated the frameworks effectiveness in reducing uncertainty, streamlining decision-making processes, and aligning development efforts with market demands. The findings underscored the value of combining established methodologies, such as the Lean Startup approach and Lean Canvas, with insights from experienced professionals to address the specific challenges faced by student entrepreneurs. This thesis advanced the study of startup formation by offering a practical and tailored solution in the form of a Notion guide. The guide incorporates insights from the Background chapter, synthesizing the analyzed frameworks and interview findings. It provides a visual walkthrough of startup formation up to the launch phase, focusing in each stage on what the founder needs to learn. Based on this knowledge, it outlines specific tasks for execution within their startup.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

Dr. rer. nat. Rodrigo Augusto da Silva Alves

Department of Applied Mathematics

This thesis is concerned with improving next basket recommendation by taking cold items into account. To overcome this problem, two architectures were introduced. Content-aware recurrent model for Next basket recommendation (CARMEN ), which uses a recurrent architecture to learn the basket sequence and Recency-aware, Feature-enhanced Next basket recommendation model (RAFEN ), which seeks to understand the inter- and intra-basket relationships between the items. Both methods were comprehensively tested on two realworld datasets in three scenarios. The results found that the proposed methods perform exceptionally well in the cold start scenario i.e. recommending cold items as a part of the next basket. CARMEN outperformed the baselines by up to 1750% in this given scenario.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

Ing. Miroslav Čepek, Ph.D.

Department of Applied Mathematics

This thesis explores the integration of NeuroEvolution of Augmenting Topologies (NEAT) with Quality-Diversity (QD) algorithms, specifically the MAP-Elites framework, to evolve diverse and high-quality neural network architectures for continuous control tasks. Leveraging the hardware-accelerated TensorNEAT library, several variants of MAP-Elites were implemented, including standard MAP-Elites with evolving topologies, a variant incorporating speciation, and hybrid approaches using Evolution Strategies (MEES and MEMES). These algorithms were benchmarked against a fixed-topology baseline in simulated environments (Walker2d, Hopper, Ant). The results demonstrate that evolving topologies within MAP-Elites achieves performance comparable to fixed-topology methods while offering the potential for structural adaptation. Furthermore, the experiments highlight the necessity of GPU acceleration for efficient neuroevolution and reveal the limitations of ES-based methods under strict time constraints.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

Ing. Stanislav Kuznetsov, Ph.D.

Department of Software Engineering

This thesis describes the design and implementation of an application for scheduling and sending of personalized content. It gives theoretical background into the topics of content personalization, notifications, and explains how we can improve the problem of alert fatigue by carefully scheduling user notifications. It analyzes use cases and existing solutions, as well as technologies and requirements for this notification scheduling orchestrator. The software is then designed and implemented as an open-source library. It is thoroughly tested and a demo integration is created, acting as a first step in its future real-world usage.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

prof. Ing. RNDr. Martin Holeňa, CSc.

Department of Applied Mathematics

Spoken dialogue systems face a fundamental trade-off: cascaded pipelines (ASRLLMTTS) achieve strong semantic reasoning but accumulate latency and discard prosodic information, while end-to-end speech-to-speech models preserve naturalness but underperform pretrained text language models on reasoning tasks. This thesis investigates whether these limitations can be overcome by architectural decouplingrouting an "inner monologue" text stream through a pretrained language model while maintaining a separate audio pathway for prosody and timing. We propose the Decoupled Semantic-Acoustic Architecture (DSAA), a dual-LLM extension of the Moshi speech-to-speech framework by integrating a pretrained Qwen 2.5 3B text model alongside Moshi's audio temporal transformer. The modalities are fused through late concatenation with learnable gating, preserving information from both streams. To maintain computational feasibility, we freeze the transformer backbones and train only fusion interfaces, enabling rapid experimentation with alternative pretrained models. Through 28 training configurations across two experimental phases, we achieved measurable improvements on phonetic and grammatical benchmarks (sWUGGY: 0.818 vs 0.739 baseline, sBLIMP: 0.589 vs 0.561, sSC: 0.641 vs 0.589) while maintaining acoustic fidelity and inference latency. However, representational analysis using Centered Kernel Alignment reveals extreme text dominance (CKA = 0.97), manifesting as 3.3× worse turn-taking performance compared to baseline. We identify two contributing factors: frozen backbones forcing all cross-modal alignment into small fusion layers, and training data dominated by read-speech monologues rather than conversational dialogue. The work demonstrates that semantic-acoustic decoupling is architecturally viable, provides a modular framework for future research, and publicly releases a novel Classroom Q\A evaluation dataset. The validated contributions and identified integration challenges offer both practical tools and diagnostic insights for advancing multimodal spoken dialogue research.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

Mgr. Jan Spěvák, Ph.D.

Department of Applied Mathematics

This thesis focuses on precipitation nowcasting from weather radar images using neural networks, particularly diffusion-based generative models. The task is formulated as predicting a sequence of future radar frames from recent observations. Through a set of ablation experiments, the effects of regularization and conditioning are analyzed. The final model is benchmarked against DeepMinds DGMR GAN baseline. Evaluation metrics include MSE, MAE, PCC, SSIM, threshold-based categorical metrics (e.g., CSI), and spatial skill measured by FSS. The proposed diffusion model produces smooth and stable forecasts with lower average errors, while DGMR tends to better preserve and localize stronger precipitation features.

Thesis on DSpace