SM03: Evaluation of Feature Selection and Weighting methods for topical Website Multi-class Classification

The repository is related to a website classification research, named: "Evaluation of Feature Selection and Weighting methods for topical Website Multi-class Classification" The main focus of the study is a comprehensive evaluation of state-of-the-art term weighting models, in the context of business website classification. The models are decomposed into their local and global components and recombined into 32 hybrid models, representing all viable variations, beyond what was initially considered by the original authors. The results showed that multi-class classification performances can be significantly improved if recently proposed global weighting components of Inverse Gravity Moment and Inverse Class Space Density Frequency, are combined with less addressed, but highly effective, local functions, like square root Term Frequency and Glasgow. In addition, filter-model feature selection functions, based on information theory, are empirically evaluated together with web page selection functions for website representation construction. The repository provides: + content analysis and other statistics on used datasets: WebKB's 7-Sector 1997 and WebKB 7-Sector 2018 Reports generated during three stages of experiments: + Feature selection function evaluation + 32 hybrid term weighting models evaluation + Weg page selection functions evaluation Note: the content snippets are removed from the experiment reports, in order to comply to the copyrights of source websites. Hence many folders in the reports remained empty. An experiment report directory, normally contains the following: + Subdirectories for each fold of cross validation 5-fold[0-5] directory_readme.txt -- description of contained files dt_test_results.xlsx -- classification results, after aggregated from k-folds log.txt -- Log output generated by imbWBI Console Tool note.txt -- Notes on the experiment In fold subdirectories: + Corpus -- subdirectory, contains reports of selected features and processed corpus note.txt -- provides description of the experiment setup本仓库关联一项网站分类研究，研究名称为《面向主题式网站多分类的特征选择与加权方法评估》（Evaluation of Feature Selection and Weighting methods for topical Website Multi-class Classification）。 本研究的核心是在商业网站分类场景下，对当前主流的词项加权（term weighting）模型开展全面评估。研究将模型拆解为局部与全局组件，并重组为32种混合模型，覆盖了初始研究作者未考虑的所有可行变体。实验结果表明，若将近期提出的逆重力矩（Inverse Gravity Moment）与逆类别空间密度频率（Inverse Class Space Density Frequency）全局加权组件，与此前较少被关注但效果优异的局部函数（如平方根词频（square root Term Frequency）与格拉斯哥（Glasgow）函数）相结合，可显著提升多分类任务的性能。此外，本研究还基于信息论对过滤式特征选择函数展开实证评估，并结合网页选择函数用于构建网站表征。 本仓库提供以下内容： + 所用数据集的内容分析与其他统计信息：WebKB 1997年7类别数据集与WebKB 2018年7类别数据集 实验三阶段生成的报告如下： + 特征选择函数评估报告 + 32种混合词项加权模型评估报告 + 网页选择函数评估报告 请注意，为遵守源网站的版权要求，实验报告中已移除内容片段，因此报告中的多数文件夹均为空文件夹。 标准的实验报告目录通常包含以下内容： + 对应5折交叉验证（cross validation）各折的子目录（5-fold[0-5]） + directory_readme.txt：所含文件说明文档 + dt_test_results.xlsx：经k折结果聚合后的分类结果文件 + log.txt：由imbWBI控制台工具生成的日志输出文件 + note.txt：实验相关说明文档 在各折子目录中： + Corpus（语料库）子目录：存放所选特征与处理后的语料库报告 + note.txt：实验设置说明文档