Currently, significant technological advances are being made in the field of artificial intelligence. In particular, machine learning is transforming the way we process data and extract useful knowledge in areas such as healthcare, industry, and sports. In the sports context, analyzing sensor and video data allows you to model performance, detect patterns, and anticipate risks. However, analysis alone does not decide what to do next in training. Coaches still have to select, sequence, and adapt exercises under time and operational constraints. This is where recommendation systems add value: they transform analytical signals and historical preferences into suggestions tailored to the athletes’ needs and the context of the session. Indoor team sports, such as futsal or basketball, are dynamic and strategic games where cooperation between team members is crucial to success. Basic characteristics such as speed of play, precise coordination, and tactical strategy give these sports their unique character. In this context, there is a clear gap: specific recommendation solutions for indoor sports at the moment are rare. This dissertation proposes and evaluates recommender systems to support training planning, capable of suggesting personalized exercises and even complete training plans. The system integrates three complementary components: (i) matrix factorization based on implicit feedback, (ii) temporal co-occurrence modeling sequences of exercises within a session, and (iii) a content-based component using cosine similarity. The scores are combined in a hybrid model, with weights optimized during validation and with eligibility and context filters to ensure practical recommendations. We developed a pipeline covering data preparation for exercises and contextual information, model generation, hyperparameter optimization, and hybrid weight optimization. Finally, we evaluate the model considering two sports modalities and two temporal protocols, Hold-Out and Next-Step, reporting Recall@N, NDCG@N, Coverage@N, Diversity@N, HR@N, and MRR@N. The performance evaluation and results demonstrate the feasibility of recommender systems to support training planning in indoor team sports, paving the way for more personalized training plans and progressive integration with contextual data in future work.

Music streaming services heavily rely upon recommender systems to acquire, engage, and retain users. One notable component of these services are playlists, which can be dynamically generated in a sequential manner based on the user's feedback during a listening session. Online learning to rank approaches have recently been shown effective at leveraging such feedback to learn users' preferences in the space of song features. Nevertheless, these approaches can suffer from slow convergence as a result of their random exploration component and get stuck in local minima as a result of their session-agnostic exploitation component. To overcome these limitations, we propose a novel online learning to rank approach which efficiently explores the space of candidate recommendation models by restricting itself to the orthogonal complement of the subspace of previous underperforming exploration directions. Moreover, to help overcome local minima, we propose a session-aware exploitation component which adaptively leverages the current best model during model updates. Our thorough evaluation using simulated listening sessions from Last.fm demonstrates substantial improvements over state-of-the-art approaches regarding early-stage performance and overall long-term convergence.

In e-commerce applications, buyers are overwhelmed by the number of products due to the high depth of assortments. They may be interested in receiving recommendations to assist with their purchasing decisions. However, many recommendation engines perform poorly in the absence of community data and contextual data. This thesis examines a hybrid matrix factorisation model, LightFM, representing users and items as linear combinations of their content features’ latent factors. The model embedding item features displays superior user and item cold-start performance. The results demonstrate the importance of selectively embedding contextual data in the presence of cold-start.