Feature selection for classifying multi-labeled past events
Feature selection for classifying multi-labeled past events
复制标题
DOI：
10.1007/s00799-020-00293-5
复制
发表时间：
2020-09
影响因子：
1.5
通讯作者：
Yasunobu Sumikawa;Ryohei Ikejiri
中科院分区：
文献类型：
--
作者：
Yasunobu Sumikawa;Ryohei Ikejiri
研究和分析过去的事件可以提供许多好处。虽然以前已经研究过事件分类，但它通常只为一个事件分配一个事件类别。在这项研究中，我们专注于过去的事件，这是一个更普遍和更具挑战性的问题，比以前的研究中所接近的多标签分类。我们使用一系列不同的特征和分类器将事件分类为13种不同的类型，这些特征和分类器在一个数据集上训练，每个类别至少有50篇标记的新闻文章。我们已经证实，使用所有特征来训练分类器具有统计意义，并提高了所有微观和宏观平均值，多标签准确度，平均精度@5，接收器工作特征曲线下的面积和基于示例的损失函数。
The study and analysis of past events can provide numerous benefits. While event categorization has been previously studied, it usually assigned only one event category to an event. In this study, we focus on multi-label classification for past events, which is a more general and challenging problem than those approached in previous studies. We categorize events into thirteen different types using a range of diverse features and classifiers trained on a dataset that has at least 50 labeled news articles for each category. We have confirmed that using all the features to train classifiers has statistical significance and improves all micro- and macro-average, multi-label accuracy, average precision@5, area under the receiver operating characteristic curve and example-based loss functions.