The Role Of Hubness in High-dimensional Data Analysis

Machine learning in intrinsically high-dimensional data is known to be challenging and this is usually referred to as the curse of dimensionality. Designing machine learning methods that perform well in many dimensions is critical, since highdimensional data arises often in practical applications and typical examples include textual, image and multimedia feature representations, as well as time series and biomedical data. The hubness phenomenon [1] has recently come into focus as an important aspect of the curse of dimensionality that affects many instance-based machine learning systems. With increasing dimensionality, the distribution of instance relevance within the models tends to become longtailed. A small number of hub points dominates the analysis and influences a disproportionate number of system predictions. Most remaining points are rarely or never retrieved in relevance queries, resulting in an information loss. High data hubness has been linked to poor system performance in many data domains. The dissertation [2] proposes several novel hubness-aware machine learning algorithms to improve the effectiveness of machine learning in intrinsically high-dimensional data. The proposed