Semi-supervised classification refers to a type ofpattern classification problem involving both labeledand unlabeled data, where the number of labeledinstances is often significantly smaller comparedto the number of unlabeled ones. Althoughthere exist several semi-supervised classifiers withhigh performance over different tasks, most of themare complex models that do not allow explainingthe obtained outcome, thus behaving like blackboxes. In this paper, we perform a critical analysisof the interpretability of state-of-the-art semisupervisedclassification approaches. In addition,we present a self-labeling grey-box classifier thatuses a black-box to estimate the missing class labelsand an interpretable white-box to make theactual predictions. The main contribution of thismodel relies on its transparency while also beingable to outperform most state-of-the-art semisupervisedclassifiers.
|Title of host publication||Proceedings of the 2nd Workshop on Explainable Artificial Intelligence|
|Subtitle of host publication||27th International Joint Conference on Artificial Intelligence and the 23rd European Conference on Artificial Intelligence|
|Editors||David W. Aha, Trevor Darrell, Patrick Doherty, Daniele Magazzeni|
|Place of Publication||Stockholm, Sweden|
|Publication status||Published - 19 Jul 2018|
Grau, I., Sengupta, D., Lorenzo, M. M. G., & Nowé, A. (2018). Interpretable self-labeling semi-supervised classifier. In D. W. Aha, T. Darrell, P. Doherty, & D. Magazzeni (Eds.), Proceedings of the 2nd Workshop on Explainable Artificial Intelligence: 27th International Joint Conference on Artificial Intelligence and the 23rd European Conference on Artificial Intelligence (pp. 52-57). http://home.earthlink.net/~dwaha/research/meetings/faim18-xai/