基於鏈結與相似度對網路化資料分群的主動式學習

我們提出了一個主動式學習的方法來解決網路化資料的分類問題。在網路的領域，當資料間彼此有鏈結時他們可能會有相似的屬性。近年來，一些集體分類的方法利用鏈結的資訊來改善網路化資料的分類方法已經被提出。但集體分類的方法需要不斷迭代來推論出鄰居特徵，那將會花費大量的時間在訓練集體分類器上。因此，我們提出了一個主動式學習的方法來減少訓練的時間而且準確度較集體分類方法更佳。我們主要的想法是利用分群來聚集有鏈結以及有相似屬性的資料以增加訓練的效率。我們使用區域性網路分群的方法來對資料分群，接著在每個群獨立訓練分類器。在實驗部分，我們使用真實世界的網路資料來評量我們的方法以及顯示我們的方法勝過其他的基準線方法，包含了使用集體分類方法的ALFNET。

關鍵字

分群；網路化資料；主動式學習；鏈結；相似度

並列摘要

This paper presents a novel active learning approach to solve the classification of networked data. In network domain, the instances may have similar attributes when they are connected for each other. Recently, some collective classification methods used the link informations to solve the classification of networked data have been proposed. But collective classification need to infer the neighbor features iteratively, that will spend a lot of time learning a collective classifier. Hence, we present an active learning approach that reduce the learning time and the accuracy also be better than collective classification methods. Our main idea is to increase the learning efficiency by clustering the data and aggregate the instances which has links and similar attributes. We use community-based network clustering method to cluster the data, and then train the local classifiers independently for each cluster. In the experiments, we evaluate our approach on real-world networks and show that our method outperforms several baselines, including the ALFNET [6] which using the collective classifiers.

並列關鍵字

clustering ； networked data ； active learning ； link ； similarity

參考文獻

[1] Saar-Tsechansky, M. and Provost, F. Active sampling for class probability estimation and ranking. Machine Learning, 54(2):153178, 2004.

Eliassi-Rad, T. Collective classiﬁcation in network data. AI Magazine,

29(3):93106, 2008.

Proceedings of the Workshop on Statistical Relational Learning at the

17th National Conference on Artiﬁcial Intelligence, 2000.

國際替代計量

基於鏈結與相似度對網路化資料分群的主動式學習

未授權

主題瀏覽