基于多视角学习和注意力的特征选择算法

Feature selection algorithm based on multi-view learning and attention

从海量且包含噪声的数据中挑选出关键性的特征,即特征选择,一直是机器学习任务中的重要挑战.鉴于此,提出了基于多视角表征学习和注意力机制的特征选择方法.首先,该算法通过多实例生成器为每个样本生成实例包,同时采用特殊的填补方式来保证特征位置不变性.其次,多视角表征模块从多个视角挖掘特征的自身信息及交互信息,并利用注意力机制模块为这些表征计算贡献度权重,最后,分类网络利用权重化表征进行分类.实验结果表明:该模型可以为每个标签挑选出最具代表性的特征群,并在不同类型数据集上取得了性能提升.

Feature selection plays an important role in many machine learning tasks,which is always a major challenge for selecting the key features from big and noisy data.Therefore,this paper proposes a feature selection method based on multi-view representation learning and attention mechanism.Firstly,Multi-instance generator of the algorithm generates instance bags for each sample,and uses a special padding method to ensure the invariance of feature positions.Then,the multi-view representation module mines the feature’s own information and interaction information from multiple perspectives.Next,the attention mechanism module calculates the contribution weights for these representations.Finally,the classification network uses the weighted representations for classification tasks.The experimental results show that the proposed method can select the most representative feature group for each label and improve the accuracy on different datasets with diverse types.