Massive Open Online Courses (MOOCs) belong to a new cloud-based service in education that suffers from low completion rates. Effective pre-learning intervention services, such as recommending courses with a high probability of completion or filtering courses with a very low probability of completion, will encourage students to spend more time and energy on proper courses, thus can reduce the dropout ratio. In practice, intervention services are introduced when students are predicted to drop out. However, existing methods concentrate on analyzing students’ learning actions and predicting final dropout after a period of enrollment, which are insufficient in preventing students from enrolling in unsuitable courses and withdrawing mid-way. This paper presents a neural network-based Explainable Self-supervised Model (ESM) to predict MOOC dropout before enrollment. Specifically, the student's learning actions on an unenrolled course are estimated using previous logs by the neural network. And then, the action's contribution to the completion of a course is calculated in a similar way. Therefore, the probability of completion for an unenrolled course is predicted by aggregating the learning actions and their contribution to the completion. To train the neural network, a self-supervised training strategy is proposed, where enrolled courses in the training data are randomly selected as validation in each epoch. The ESM outperforms existing methods in terms of prediction accuracy and efficiency. The average increment of Area Under the ROC Curve (AUC) and F-score (F1) in the two MOOCs datasets, XuetangX and KDDCUP, are 8.3% and 0.6%, respectively. Furthermore, the two pre-learning intervention services named courses recommendation and courses filtration are proposed. When courses are recommended, the completion rate increased from 22% to 60% in XuetangX, and from 27% to 45% in KDDCUP. By filtering courses predicted with low completion probability, 40% wasted time in uncompleted courses will be saved in XuetangX.

Due to the increasing popularity of smart mobile devices, the amount of mobile data communications has led to explosive growth of data traffic in cellular networks. Cellular networks have to face the challenge of huge communication traffic. Offloading data traffic through opportunistic communication among smart mobile devices is a promising solution to partially solve this problem since there is almost no monetary cost for it. Large amount of smart mobile devices can communicate each other using Bluetooth or WIFI Direct in short communication range and they can form an opportunistic mobile social network. The opportunistic communications among smart mobile devices can effectively reduce the amount of cellular data traffic. However, mobile users take a long time to obtain useful data. In order to reduce data communication latency, this paper proposes three data dissemination protocols named RRDP(Request-Reply Dissemination Protocol), RDP(Random Dissemination Protocol) and LDP(LRU Dissemination Protocol) respectively. The three proposed protocols are based on opportunistic sharing policy. Extensive NS-2 simulation results show that (1) on the campus situation, the user's access delay of RDP is 56.4% less than the RRDP and LDP is 44.8% less than RRDP. (2) in the vehicular environment, the user's access delay of RDP is 32.5% less than the RRDP and LDP is 28.1% less than RRDP. RDP is the best protocol.

Biological frequent patterns usually correspond to the important function (or structure) in biological sequences. Along with the rapid growth of biological sequences, it is significant to find frequent patterns over a large bio-sequence efficiently. However, most of existing algorithms need to produce lots of short patterns or projected databases, which influence the efficiency badly and also increase the cost of space. Graphics processing units (GPUs) embracing many core computing devices, have been extensively applied to accelerate computation performance in many areas. In order to meet the demand of biologists, we redefine the frequent pattern problem with length constraints for finding frequent patterns. We present pruning optimization method for the serial algorithm (POSA), and based on this technique, we propose a parallel algorithm (POPA) which not only reduces the time complexity with a low space cost but also obtains better performance on CUDA. To validate the presented algorithms, we implemented the algorithms on multiple-core CPU and various GPU devices. Also, CUDA optimization techniques are applied to speed up calculation in the paper. Finally, experimental results show that compared with the serial algorithm on CPU with six cores, POSA achieves 1.2~4.5 speedup, and POPA gains 3~20 speedup.

The repetitions in biological sequence analysis are of great biological significance. Finding the repetitions has been a hot topic in gene projects naturally. In recent years, graphics processing unit (GPU) has been far exceeded the CPU in terms of computing capability and memory bandwidth, especially CUDA dramatically increases in computing performance by harnessing the power of the GPUs. This paper proposes efficient parallel algorithms on CUDA to accelerate finding PTRs which is redefined as LPRs based on the SUA Index. The proposed parallel algorithms have been utilized with the parallel primitives offered by Thrust library and the effective parallel bit compression technology based on division to achieve better acceleration. Optimization techniques include CUDA streams technology are also realized to reduce transmission latency. Experimental results show that the proposed parallel algorithms are faster than the benchmark with 1.6∼5.4 speedup.

Approximate string matching using the k-difference technique has been widely applied to many fields such as pattern recognition and computational biology. Data dependency exists in the traditional sequential algorithm. Therefore, it is hard to design a parallel algorithm for approximate string matching with k differences. This paper presents a technique to eliminate data dependency. Based on this technique, this paper also presents a parallel algorithm which can calculate the elements in the same row of the edit distance matrix in parallel by eliminating data dependency. The algorithm has high parallelism, but requires synchronization. To validate the proposed algorithm, it is implemented on GPU and multiple-core CPUs. Moreover, the CUDA optimization techniques are also presented in the paper. Finally, experimental results show that, compared with the traditional sequential algorithm on CPU with twenty-four cores, the proposed parallel algorithm achieves speedup of 7-42 on GPU.

Solving the eigenvalues of matrices is an open problem which is often related to scientific computation. With the increasing of the order of matrices, traditional sequential algorithms are unable to meet the needs for the calculation time. Although people can use cluster systems in a short time to solve the eigenvalues of large-scale matrices, it will bring an increase in equipment costs and power consumption. This paper proposes a parallel algorithm named Jacobi on gpu which is implemented by CUDA (Computer Unified Device Architecture) on GPU (Graphic Process Unit) to solve the eigenvalues of symmetric matrices. In our experimental environment, we have Intel Core i5-760 quad-core CPU, NVIDIA GeForce GTX460 card, and Win7 64-bit operating system. When the size of matrix is 10240×10240, the number of iterations is 10000 times, the speedup ratio is 13.71. As the size of matrices increase, the speedup ratio increases correspondingly. Moreover, as the number of iterations increases, the speedup ratio is very stable. When the size of matrix is 8192×8192, the number of iterations are 1000, 2000, 4000, 8000 and 16000 respectively, the standard deviation of the speedup ratio is 0.1161. The experimental results show that the Jacobi on gpu algorithm can save more running time than traditional sequential algorithms and the speedup ratio is 3.02~13.71. Therefore, the computing time of traditional sequential algorithms to solve the eigenvalues of matrices is reduced significantly.