Design of efficient and scalable algorithms for label ranking problems
Juan Carlos Alfaro Jiménez · RUIdeRA - Institutional University Repository (University of Castilla-La Mancha) · 2023
Preference learning is an area of research that has gained significant importance in the field of artificial intelligence, mainly due to the growing complexity and realism of many domains. The representation and processing of preferences has emerged as a critical solving paradigm in this context. To acquire preferences, automatic learning, discovery, and adaptation methods are necessary, and their implications extend to domains such as electronic commerce, recommender systems, and personalized medicine. One critical problem in preference learning is the induction of preference models from empirical data. Two major techniques exist in the literature to model preferences: learning from utility functions and learning from preference relations. The last strategy is more closely related to traditional machine learning problems such as classification and regression. However, it involves predicting complex structures, such as rankings, rather than single values. This thesis focuses on the learning from preference relations scenario known as label ranking. The objective is to map instances to a weak order of the class labels. While extensive research has been conducted on learning from total orders using specific probability distributions and aggregation methods, less attention has been given to weak orders, where some class labels may be tied. This thesis proposes methods for this particular learning framework, referred to as the partial label ranking problem. The algorithms designed in this research are interesting in domains such as personalized recommendations for online streaming services, where weak orders can help capture more nuanced user preferences. For example, a user may not have a clear preference between two items but may prefer them over a third item, which a total order may miss. The preference model can provide users with more personalized and accurate recommendations by incorporating weak orders into the output. A benchmark of datasets is utilized to evaluate the performance of the proposed algorithms. These datasets are generated by transforming standard classification problems into partial label ranking learning scenarios. Additionally, a comprehensive statistical analysis procedure is conducted to ensure the adequacy of the methods.