程玉胜 钱坤 王一宾 赵大卫

摘 要：已有的多标签懒惰学习算法（IMLLA）在利用近邻标签时因仅考虑了近邻标签相关性信息，而忽略相似度的影响，这可能会使算法的鲁棒性有所降低。针对这个问题，引入萤火虫方法，将相似度信息与标签信息相结合，提出一种融合萤火虫方法的多标签懒惰学习算法（FFMLLA）。首先，利用Minkowski距离来度量样本间相似度，从而找到近邻点;然后，结合标签近邻点和萤火虫方法对标签计数向量进行改进;最后，使用奇异值分解（SVD）与核极限学习机（ELM）进行线性分类。该算法同时考虑了标签信息与相似度信息从而提高了鲁棒性。实验结果表明，所提算法较其他的多标签学习算法有一定优势，并使用统计假设检验与稳定性分析进一步说明所提出算法的合理性与有效性。

关键词：多标签学习;萤火虫方法;标签相关性;多标签懒惰学习算法;极限学习机

中图分类号：TP181

文献标志码：A

Abstract： The existing Improved Multilabel Lazy Learning Approach （IMLLA） has the problem that the influence of similarity information is ignored with only the neighbor label correlation information considered when the neighbor labels were used， which may reduce the robustness of the approach. To solve this problem， with firefly method introduced and the combination of similarity information with label information， a Multilabel Lazy Learning Approach based on FireFly method （FFMLLA） was proposed. Firstly， Minkowski distance was used to measure the similarity between samples to find the neighbor point. Secondly， the label count vector was improved by combining the neighbor point and firefly method. Finally， Singular Value Decomposition （SVD） and kernel Extreme Learning Machine （ELM） were used to realize linear classification. The robustness of the approach was improved due to considering both label information and similarity information. The experimental results demonstrate that the proposed approach improves the classification performance to a great extent compared to other multilabel learning approaches. And the statistical hypothesis testing and stability analysis are used to further illustrate the rationality and effectiveness of the proposed approach.

英文關键词Key words： multilabel learning; firefly method; label correlation; Improved Multilabel Lazy Learning Approach （IMLLA）; Extreme Learning Machine （ELM）

0 引言

多标签学习[1]是一种应用非常广泛的学习范式，是机器学习研究的重要热点之一。传统的单标签学习，每个对象只与单个标签相关联;然而，真实世界中的对象往往具有多义性，比如一篇文章可能属于军事、体育、运动等多个主题[2]。

多标签学习作为处理具有丰富语义真实世界对象的学习框架之一，且其研究成果已经广泛应用到文本分类[3]、基因工程[4]、图像识别[5-6]、Web数据挖掘[7]和视频自动标注[8]等多个领域。对此许多学者提出了针对多标签分类的学习算法，例如BR（Binary Relevance）算法、LP（Label Power）算法[9]等，它们通过增加分类器个数或者标签的种类来解决多标签问题，但在一定程度上影响了分类器效率。……