Prediction reliability of deep learning based systems allows users to confirm if the prediction is reliable to real applications, which is the key to the success of recommendation systems. Current research on deep learning based recommendation systems focused on estimating model reliability, but seemed to be missing data contributions to the prediction reliability. This paper proposed a novel framework to estimate the prediction reliability for deep learning-based methods through combining the model reliability and the data reliability. The proposed framework has been validated in a case study based on named entity recognition (NER) that is from an Intuit recommendation task. It employed two NER datasets: WNUT and GMB to examine detailed performance of the proposed framework, where, specifically, we proposed a novel evaluation metric to comprehensively evaluate the performance. Experimental results demonstrated that, compared to model reliability only, combining data reliability with model reliability will significantly improve performance, as well as enhance the prediction interpretability.

Analyzing micro-array data faces many challenges such as high dimension, low sample size and sparse data. Feature selection is a technique to select more relevant features to implement dimension reduction to mitigate these challenges. In this paper, we propose a novel framework of feature selection based on recurrent neural network (RNN) to select a subset of features. Specifically, the proposed framework has been applied to select features from micro-array data for cell classification. We implement four feature selection models with different architectures of recurrent neural networks under the proposed framework, where these architectures include gated recurrent unit (GRU), long short-term memory (LSTM), RNN and bi-directional LSTM. The advantages of the proposed framework is demonstrated via real-world micro-array datasets.

In this paper, we address the issue of how to enhance the generalization performance of convolutional neural networks (CNN) in the early learning stage for image classification. This is motivated by real-time applications that require the generalization performance of CNN to be satisfactory within limited training time. In order to achieve this, a novel hierarchical transfer CNN framework is proposed. It consists of a group of shallow CNNs and a cloud CNN, where the shallow CNNs are trained firstly and then the first layers of the trained shallow CNNs are used to initialize the first layer of the cloud CNN. This method will boost the generalization performance of the cloud CNN significantly, especially during the early stage of training. Experiments using CIFAR-10 and ImageNet datasets are performed to examine the proposed method. Results demonstrate the improvement of testing accuracy is 12% on average and as much as 20% for the CIFAR-10 case while 5% testing accuracy improvement for the ImageNet case during the early stage of learning. It is also shown that universal improvements of testing accuracy are obtained across different settings of dropout and number of shallow CNNs.

Rumor detection becomes increasingly important in social media. The effects of rumor propagation are dreadful in case of time-critical events, for example, during natural disasters. In this paper, we proposed a multiple time-series data analysis model to detect rumors on Twitter. Instead of checking the contents of the tweets, the proposed method only uses temporal properties of the tweets. As a result, the computational complexity measured by the training time and prediction time has been reduced significantly, which allows quick detection of rumors. Experimental results show that the proposed model combined with Gaussian Naive Bayes classifier achieved a high precision score of 94%.