Ranking Importance Based Information on the World Wide Web

Identifying useful features for classification and forecast tasks from a ranked data is highly difficult and challenging. By ranking user popularity ratings from normalised area histograms, a method of feature selection for ranked data inspired from the law of vital few is proposed. We propose that the attributes that are most stable against the variations in classes have their usefulness in a forecasting task, while the attributes that are most unstable between inter-class samples but most stable within intra-class samples have their usefulness in classification tasks. The performance of the proposed method is demonstrated through a realistic example of web-content data from Yahoo! research repository: the user rating of web pages. The attributes in the data when ranked based on their importance in a year show distinct characteristics of performance in the tasks of popularity forecast and classification.