Ing. Petr Kasalický

Ing. Petr Kasalický

doc. Ing. Pavel Kordík, Ph.D.

Department of Applied Mathematics

This work describes the NVIDIA Merlin application framework and all its parts in detail and how it can be used when implementing scalable data preprocessing pipeline, designing a model architecture, training a model, or generating predictions for recommendation system. Merlin is presented by NVIDIA as easy-to-use and highly scalable framework capable of handling real-world data and workloads. This thesis verifies claims made by NVIDIA and reviews Merlin from developers perspective. We describe the general approaches used in recommendation systems and then discuss the optimizations NVIDIA implemented. Next we provide a guide in how to use some of the Merlin modules. We have managed to use Merlin components to implement a movie recommendation system that is producing relevant results.

Thesis on DSpace

Ing. Petr Kasalický

Ing. Karel Klouda, Ph.D.

Department of Applied Mathematics

This master thesis aims to survey content-based recommendation systems based on image data, state-of-the-art convolutional neural networks for feature extraction from images, and possibilities of producing recommendations based on more than one image of a single item, so-called multiple instance learning. Subsequently, new content-based recommendation methods that use one and more images of a single item and incorporate interactions from users behavior to improve the recommendations were described. Several prototypes were implemented based on the state-of-the-art convolutional neural networks, and compared in offline tests in their ability to extract important features for recommendations on cold start problem. Finally, four of the models were compared in the online A/B test against each other. The results showed that the incorporation of user interactions and more images into the recommender system improved the measured click-through rate metric.

Thesis on DSpace

Ing. Petr Kasalický

Rodrigo Augusto da Silva Alves, Ph.D.

Department of Applied Mathematics

This thesis provides a comprehensive study of recommender systems, with a specific focus on multi-stage recommender systems. The work investigates the complex nature of such systems and examines the motivations behind the use of a four-stage pipeline. The study covers various retrieval algorithms and approaches employed in the ranking stage. A tailored multi-stage recommender system for the Dressipi dataset is designed and implemented, with a range of retrieval models evaluated to create a robust retrieval stage. The state-of-the-art implementation of gradient-boosted decision trees is employed for ranking, with the best results obtained by ensembling multiple approaches. The importance of utilizing various features is examined, and the pointwise and listwise ranking approaches are compared. The performance evaluation demonstrates that utilizing a multi-stage pipeline for recommendation outperforms using single recommendation algorithms or their ensembles. Finally, future improvements and work are discussed.

Thesis on DSpace