Active learning for Web accessibility evaluation

To evaluate the accessibility level of a website, we need to obtain the accessibility evaluation results of the pages in this website. Due to the massive number of pages in a website and possible involvement of human inspection for conformance checking, directly evaluating all the pages is prohibitively expensive. In practice, we usually select a representative sample for accessibility evaluation of the whole site. This makes the evaluation results heavily dependent on the pages selected. Undersampling may lead to a large bias in evaluation. But oversampling will incur high evaluation expense. To address this issue, this paper proposes a semi-supervised machine learning method, called active-prediction, to obtain the accessibility evaluation results for all pages in a site. Active-prediction casts the website accessibility evaluation into a prediction problem by building learning models for each checkpoint in evaluation and consequently avoids the expensive cost in human inspection. To achieve a higher prediction accuracy with only a small number of training data, active-prediction exploits active learning techniques to select the most informative pages to train the models. Experimental results show that the active-prediction could achieve a high accuracy on predicting the accessibility results and better reflect the accessibility level of the websites than the existing methods.