TT-RNNPool3D: A Tensor-Based High-Order RNNPool Model for Mobile Edge Consumer Applications

With the development of the new generation of information technologies such as the Internet-of-Things, Big Data, and High-Performance Computing, mobile edge devices are widely used in almost every aspect of our lives to provide personalized services. One key challenge how to efficiently process the multi-attribute data is essential to analyze the users’ information including the users’ consumption habits and the foundation to provide high-quality services as well. Although the existing models such as the Recurrent Neural Network (RNN) are used to process and analyze multi-attribute data, there are still many challenges about multi-attribute data processing. In this paper, a tensor-aided high-order pooling operator based on RNN (RNNPool), namely the TT-RNNPool3D model, using tensor as the multi-attribute data representation tool is proposed. The TT-RNNPool3D model uses tensor-train (TT) decomposition to improve the accuracy by decomposing multi-attribute data into multiple tensor-cores for parallel processing. This improves TT-RNNPool3D’s efficiency by reducing the number of parameters, which is suitable for mobile edge devices in processing multi-attribute data. To demonstrate the performances of the TT-RNNPool3D model such as the accuracy and efficiency, experiments are carried out on three different datasets.