Evolving multi-label fuzzy classifier

Multi-label classification has attracted much attention in the machine learning community to address problem of assigning single samples more than one (not necessarily non-overlapping) class at same time. We propose an evolving multi-label fuzzy classifier (EFC-ML) which is able self-adapt and self-evolve its structure with new incoming incremental, single-pass manner. It based on a multi-output Takagi–Sugeno type architecture, where for each separate consequent hyper-plane defined, yields flexibility partially approximating respective classes binary [0,1]-regression context. The procedure embeds locally weighted incremental correlation-based algorithm combined (conventional) recursive fuzzily least squares Lasso-based regularization. Locality important avoid out-masking effect labels or rules; part ensures that interrelations between labels, specific well-known property improved performance, are preserved properly; regularization reduces curse dimensionality effects case higher number inputs. Antecedent achieved by product-space clustering conducted all together, rule base (opposed related techniques such as one-versus-rest chaining, achieving multiple different bases, per class), allowing compact knowledge view thus enabling better interpretable insights. Furthermore, our approach comes online active (AL) strategy updating just (smaller) selected samples, turn makes applicable scarcely labelled streams applications, annotation effort typically expensive. three essential concepts: novelty content antecedent space, uncertainty due ambiguity (output) space parameter instability reduction, these combination upper-allowed selection budget (which could be predefined user). Our was evaluated several data sets from MULAN repository showed significantly accuracy average precision trend lines compared (evolving) chaining concepts. A significant result that, AL method, 90% reduction used updates had little accumulated full update most set cases.