Automatic Clustering Based Feature Combination for Latent Topic Based Classification

[[abstract]]在本研究中, 我們示範了一個基於自動分群為基礎的特徵合併程序, 這個合併程序藉由自動分群進行特徵轉換, 並使用合併程序產生新的特徵. 這個新的特徵處理流程相較於傳統的分類來說,展現了更好的分類效能. 在本研究的印證實驗中, 我們依據各個原始單獨特徵本身的資料分佈,提出了偵測最佳群數的估計公式, 依此方式進行自動分群. 這個估計最佳群數的公式是依據分群狀態的多項特性, 如群內密度與群間距離, 輔以特別的比率調和計算, 用以比較各種群聚分布下的優劣. 然後依此結果進行後續的特徵合併與隱性主題分析, 最後得到較佳的分群結果. 根據本研究的實驗結果, 我們發現本論文所提出的處理流程, 相較於傳統的分類器效能有著更好的接受器操作特性(ROC). In this work, we demonstrate a feature combination process based on automatic clustering. This procedure delivers a better classification result than the traditional classification models. In our experiment, we detect the best cluster number for each feature based on the distribution off feature data with a specially designed cluster number decision formula called HCV function. This function estimates the best intra- and inter- cluster relationship with a harmonic based mean calculation. With HCV, we successfully decrease the number of parameters needed in our feature combination process in respect to our previous method. According to the experiment, we found a better ROC performance than traditional classifier and other configurations with our design