An Embedded-Based Weighted Feature Selection Algorithm for Classifying Web Document

Table 1

The usage of machine learning algorithms for document classification—a summary.

Reference	Method	Summary
[30]	DragPushing	(i) Proposes kNN optimization which automatically balances the data points evenly across all the classes to avoid model misfits.; (ii) They have set the value of to 7 for their experiment.; (iii) Three datasets are used Reuter-21578, industry sector, and TDT-5.
[31]	Prototype selection	(i) Much faster than [30].; (ii) Prototype selection recommends the most portable prototypes for training purposes.; (iii) [31] eliminates most of the data points in training to increase speed.
[32]	ForesTexter	(i) The Gini index can easily predict the skewness in the majority class and creates many subtrees to balance the data points.; (ii) Can work faster than [30, 31].; (iii) [32] combines both feature subspace selection and splitting criterion to create multiple subtrees to balance the data.
[33]	Resampling	(i) Handles the imbalance problem better than [32] by performing resampling.; (ii) Instance weighting enables one to assign few weights for the imbalanced class so that the end performance (in terms of accuracy) is balanced.; (iii) They validated the proposed method with SVM classifier.
[34].	Topic modelling	(i) Instead of balancing the data points across all the classes, this method uses the topic model to construct new data points in each class to create a complete dataset.; (ii) This method considers more data points than [30, 32, 33] because of topic modeling which can construct new data points.
[35]	Bag of concepts	(i) Aims to reduce the dimensions of the document matrix representation.; (ii) Instead of recommending data points from the data set (such as [31]), the bag of concepts groups one or more data points into topics.; (iii) Bag of concepts solves many problems in the traditional bag of words models such as high dimensionality and sparsity issues.
[36]	Ontology-based deep learning model	(i) Enhances the problem of [35] by not considering the relationships among the documents.; (ii) The features and their relationships are extracted based on deep learning.; (iii) The ontology enhancement proposed in [36] helps to reduce the high dimensions.; (iv) This method consumes more time in training the samples.
Proposed	Weighted feature selection	(i) Resolves the imbalance problem by assigning weights to the most important features.; (ii) Three classifiers are used, namely, kNN, SVM, and Naïve Bayes.; (iii) [35] fails to detect the relationships among the documents. In this paper, the proposed system detects the relationship and considers it for classification.