@article{oai:ipsj.ixsq.nii.ac.jp:00184919,
 author = {Yuya, Yoshikawa and Tomoharu, Iwata and Yuya, Yoshikawa and Tomoharu, Iwata},
 issue = {3},
 journal = {情報処理学会論文誌数理モデル化と応用（TOM）},
 month = {Dec},
 note = {In various machine learning problems, bag-of-words (BoW) representation, i.e., a multiset of features, is widely used as a simple and general data representation. Deep learning is successfully used in many areas. However, with BoW data, deep learning models are often outperformed by kernel methods such as support vector machines (SVMs), where each sample is simply transformed into a fixed-length count vector for the input. In this paper, we propose a deep learning model for BoW data. Based on the idea introduced in the framework of SVMs that has achieved a better performance in BoW count vector inputs, the proposed model assigns each feature to a latent vector, and each sample is represented by a distribution of the latent vectors of features contained in the sample. To transform the distribution efficiently and nonparametrically to the inputs of deep learning, we integrate kernel mean embeddings and a random Fourier feature algorithm. Our experiments verify the effectiveness of the proposed model on BoW document datasets. Because the proposed model is a general framework for BoW data, it can be applied directly to various supervised and unsupervised learning tasks. Moreover, because the proposed model can be combined with existing deep learning models, it further extends the potential applications of deep learning., In various machine learning problems, bag-of-words (BoW) representation, i.e., a multiset of features, is widely used as a simple and general data representation. Deep learning is successfully used in many areas. However, with BoW data, deep learning models are often outperformed by kernel methods such as support vector machines (SVMs), where each sample is simply transformed into a fixed-length count vector for the input. In this paper, we propose a deep learning model for BoW data. Based on the idea introduced in the framework of SVMs that has achieved a better performance in BoW count vector inputs, the proposed model assigns each feature to a latent vector, and each sample is represented by a distribution of the latent vectors of features contained in the sample. To transform the distribution efficiently and nonparametrically to the inputs of deep learning, we integrate kernel mean embeddings and a random Fourier feature algorithm. Our experiments verify the effectiveness of the proposed model on BoW document datasets. Because the proposed model is a general framework for BoW data, it can be applied directly to various supervised and unsupervised learning tasks. Moreover, because the proposed model can be combined with existing deep learning models, it further extends the potential applications of deep learning.},
 pages = {32--38},
 title = {Randomized Kernel Mean Networks for Bag-of-Words Data},
 volume = {10},
 year = {2017}
}