Most supervised methods for relation extraction (RE) involve time-consuming human annotation. Distant supervision for RE is an efficient method to obtain large corpora that contains thousands of instances and various relations. However, the existing approaches rely heavily on knowledge bases (e.g., Freebase), thereby introducing data noise. Various relations and noisy labeling instances make the issue difficult to solve. In this study, we propose a model based on a piecewise convolution neural network with adversarial training. Inspired by generative adversarial networks, we adopt a heuristic algorithm to identify noisy datasets and apply adversarial training to RE. Experiments on the extended dataset of SemEval-2010 Task 8 show that our model can obtain more accurate training data for RE and significantly outperforms several competitive baseline models. Our model has an $F_{\mathrm{1}}$ score of 89.61%.

In this paper, we propose a novel Residuals-Based Deep Least Squares Support Vector Machine (RBD-LSSVM). In the RBD-LSSVM, multiple LSSVMs are sequentially connected. The second LSSVM uses the fitting residuals of the first LSSVM as input time series, and the third LSSVM trains the residuals of the second, and so on. The original time series is the input of the first LSSVM. Additionally, to obtain the best hyper-parameters for the RBD-LSSVM, we propose a model validation method based on redundancy test using Omni-Directional Correlation Function (ODCF). This method is based on the fact when a model is appropriate for a given time series, there should be no information or correlation in the residuals. We propose the use of ODCF as a statistic to detect nonlinear correlation between two random variables. Thus, we can select hyper-parameters without encountering overfitting, which cannot be avoided by only cross validation using the validation set. We conducted experiments on two time series: annual sunspot number series and monthly Total Column Ozone (TCO) series in New Delhi. Analysis of the prediction results and comparisons with recent and past studies demonstrate the promising performance of the proposed RBD-LSSVM approach with redundancy test based model selection method for modeling and predicting nonlinear time series.

Traffic prediction plays an integral role in telecommunication network planning and network optimization. In this paper, we investigate the traffic forecasting for data services in 3G mobile networks. Although the Box-Jenkins model has been proven to be appropriate for voice traffic (since the arrival of calls follows a Poisson distribution), it has been demonstrated that the Internet traffic exhibits statistical self-similarity and has to be modeled using the Fractional AutoRegressive Integrated Moving Average (FARIMA) process. However, a few studies have concluded that the FARIMA process may fail in modeling the Internet traffic. To this end, we conducted experiments on the modeling of benchmark Internet traffic and found that the FARIMA process fails because of the significant multifractal characteristic inherent in the traffic series. Thereafter, we investigate the traffic series of data services in a 3G mobile network from a province in China. Rich multifractal spectra are found in this series. Based on this observation, an integrated method combining the AutoRegressive Moving Average (ARMA) and FARIMA processes is applied. The obtained experimental results verify the effectiveness of the integrated prediction method.