The Neural Network for Online Learning Task Without Manual Feature Extraction

Abstract

The article is devoted to the problem of feature extraction in online learning tasks. In many cases, the proper feature extraction is very time-consuming. Currently, in some cases, this problem is successfully solved by deep neural networks. However, deep models are computationally expensive and so hardly applicable for online learning tasks which require frequent updating of the model. This paper proposes the lightweight neural net architecture that can be learned in online mode and doesn’t require complex handcrafted features. The small sample processing time distinguishes the proposed model from more complex deep neural networks. The architecture and learning process of the proposed model are discussed in detail. The special attention is paid to fast software implementation. On benchmarks, we show that developed implementation processes one sample several times faster than implementations on the base of deep learning frameworks. The conducted experiments on CTR prediction task show that the proposed neural net with raw features gives the same performance as the logistic regression model with handcrafted features. For a clear description of the proposed architecture, we use the metagraph approach.