Cross-domain Aspect-Based Sentiment Analysis (ABSA) leverages unsupervised domain adaptation techniques to transfer knowledge from a source domain, which is rich in labeled data, to the target domain which lacks labeled data. Many recent studies have attempted to address this issue by generating a large amount of labeled target domain data, and the domain adaptive model DA2LM has achieved state-of-the-art results. However, training this model requires the use of target domain data, which should be annotated with pseudo labels. Therefore, it is important to generate high-quality pseudo labels effectively. Furthermore, when a substantial amount of labeled data for the target domain is obtained using the generative model, it becomes essential to train an effective model to predict the label of test data from the target domain. In this work, we propose a novel Cross-Domain Data Augmentation approach based on Contrastive Learning, named CLCDDA. This approach combines the Multiple Kernel Maximum Mean Discrepancy and enhanced contrastive learning to obtain high-quality target domain data with pseudo labels, which is used to improve the performance of the generative model. In addition, contrastive learning is introduced to improve the performance of the sequence labeling model. The experimental results on four benchmarks show that CLCDDA significantly outperforms previous approaches in both cross-domain End2End ABSA and AE tasks.

As a branch of psychology, personality plays an important role in distinguishing individuals in the society. The existing personality prediction models need to be further improved in precision and generalization. Recently, deep neural network (DNN) models are being applied to personality prediction tasks to obtain promising results. However, only extracting the semantic features of text through deep learning is very limited to improve the performance of the model. We propose a BERT-based Model for Personality Prediction named BSAM to extract semantic features and use the statistical information of corpus as external features. In this model, we concatenate the output of BERT with the statistical information and use bidirectional long short term memory networks (Bi-LSTM), bidirectional gated recurrent unit (Bi-GRU) and improved convolutional neural networks (CNN) to extract deep semantic features. We also compare the results with benchmark models on social media datasets and test the effectiveness of statistical features. The experimental results show that our model can effectively improve the classification performance of the five dimensions of the Big-Five personality.

在教育领域中,命名实体识别在机器自动提问和智能问答等相关任务中都有应用.传统的中文命名实体识别模型需要改变网络结构来融入字和词信息,增加了网络结构的复杂度.另一方面,教育领域中的数据对实体边界的识别要十分精确,传统方法未能融入位置信息,对实体边界的识别能力较差.针对以上的问题,该文使用改进的向量表示层,在向量表示层中融合字、词和位置信息,能够更好地界定实体边界和提高实体识别的准确率,使用BiGRU和CRF分别作为模型的序列建模层和标注层进行中文命名实体识别.该文在Resume数据集和教育数据集(Edu)上进行了实验,F1值分别为95.20％ 和95.08％.实验结果表明,该文方法对比基线模型提升了模型的训练速度和实体识别的准确性.

信息技术的发展带动了教学模式的变革,文章通过问卷调查收集了133个样本数据,对大规模在线教学与课堂教学混合式交互教学模式的教学效果进行了深入探索.采用非参数检验和回归分析的方法,对混合学习模式下学生自主学习能力和倾向选择的在线学习平台与学业成就间的关系进行了差异分析.定量研究发现,自学能力强的学生比自学能力弱的学生高2.9分,倾向于选择编程实践平台的学生比倾向于选择概念解释测试平台的学生高3.3分.回归模型的复样本确定系数为0.466,这说明其具有良好的泛化能力,从而保证了模型能够广泛应用于实际的计算机教育场景中.

It is vital for educators to teach learners in accordance with their aptitude, which can be useful to help learners reach their full potential. Educators have been taking the Myers-Briggs Type Indicator (MBTI) as a powerful tool to understand the differences in students’ learning styles, adopting appropriate teaching strategies to accommodate the learning styles of different types of students can effectively prevent students from being tired of studying. It is a problem worthy of research to recognize the students’ personality traits with technological means. Therefore, we propose a method to recognize learners’ MBTI from videos, which can be applied in the course learning and practice stages of software engineering education. We propose a novel approach to recognize the MBTI personality traits of learners from videos. Personality and emotion unconsciously affect facial expression, the speaking style in social contexts. However, in the current literature, there is no publicly available source of images dataset labeled with the MBTI personality scale; nearly all the available data are text. In this paper, we use two datasets: images extracted from ChaLearn First Impressions dataset and the Myer-Briggs Personality Type Dataset from Kaggle for our training tasks. Furthermore, we take plentiful text data labeled with the MBTI personality scale as the source domain and image data as the target domain for borrowing knowledge from the source domain to facilitate the learning task in a target domain. By adopting feature transfer, a bridge is built between the source domain and the target domain. We perform experiments on the transfer task and evaluate the effectiveness of this approach, the results of this study can assist educators in regards to the identification of learners’ MBTI personality types in a new way.

Pair programming has been widely used in programming experiment teaching in programming courses. One of the important factors affecting the successful completion of pair programming is the timing of periodic role switching. We organized an experiment in the course of Python Programming for 102 freshmen who did not major in computer science. By comparing the accuracy of code submitted by students in the online judge system, we evaluated the influence of pairing programming on students’ programming ability under three different periodic role switching intervals of 15 min, 20 min and 30 min, and collected students’ perception towards pairing programming under different modes. We also made a standard to judge the normalization of code, to study the influence of pair programming on the normalization of code written by students. The results show that when the periodic role switching interval is 30 min, pairing programming is helpful for students to solve difficult problems, and it has a positive impact on the solution of subsequent problems after experiencing the process of solving difficult problems. When the periodic role switching interval is 20 min, students have a positive attitude towards pair programming. Therefore, the best switching interval can be set between 20 min and 30 min. However, in terms of code normalization, there is no significant relationship between the standard degree of student code and the switching interval of pair programming. We gave some explanations for this in this paper.

The problem of low-rank matrix completion has attracted massive attention in various engineering fields such as machine learning, image processing and video denoising. Assuming the data are low-rank, missing entries can be estimated in matrices by matrix completion algorithms, and the best approximation which fixes the constraints is produced. However, most matrix completion methods usually result in unsatisfactory reconstruction accuracy with adding non-Gaussian noise. This paper proposes a new robust algorithm for matrix completion considering some traditional methods which can enhance the robustness and avoid fitting. This paper can identify the location of outliers and replace them with approximate data, reducing the influence of outliers to the result, and enhancing the accuracy of the reconstruction. Simulated data and real data show that the algorithm is good at robustness and accuracy when the data sets are polluted by the non-Gaussian noise.

Abstract Nowadays, a great deal of educational data has been produced by E-learning system and MOOC. Educational data is important for Teaching and research. These educational data can be classified many kinds of feature, such as demographic features, social features and behavioral features. And which feature is the most import for student’s performance? In this paper, In this paper, we use some common data mining technologies including Naïve Bayesian(NB), Artificial Neural Network(ANN), Support Vector Machine(SVM) and Decision Tree Classifier(DT) to predict students’ performance, and try to find out the influence of characteristics on students’ academic performance. From the conclusion, we can see that SVM technique outperform others, and Behavioral Features have good effect on students’ performance.

Link prediction, which aims at recommending potential links between network nodes, is an important step to understand and study the characteristics of social networks. With the development of social networks, many networks contain rich node attributes. This paper focuses on using both network structure and node attributes to predict links. Based on the assumption that two nodes in the network may be connected because they are close in the network, or may be linked for they have similar attributes, a new random walk model for link prediction by combining network structure and node attributes is proposed. First, two different graphs and transition matrices are created for new iteration rule. Second, the model is simplified for calculation and then a fast approximation algorithm is presented. The experiment on two standard datasets reveals that this method has better performance compared with other similar methods. Meanwhile, the effect of the probability of particle walking on different graphs is analyzed and it shows that node attributes can promote the prediction ability effectively.

教育数据挖掘是指开发、研究和应用计算机技术来发掘存在于大规模教育数据背后的潜在模式.本文着重梳理数据挖掘技术在学生学习成绩预测领域内应用的文献,重点介绍自2016年以来的主要研究成果,对其应用的典型算法进行详细阐述,并分析不同算法的应用范围与局限性.最后总结出学生成绩预测的主要流程及其未来发展方向.

Heterogeneous information networks (HIN) contain different types of nodes and edges. Predicting the connection between nodes in HIN is a non-trivial problem. Meta paths connect multiple types of nodes through a set of relationships, and are used to describe the different semantics of connections between different types of nodes in HIN. Although several similarity measures based on meta paths have been proposed, the challenge of how to use the measures under different paths to predict links remains open. Besides, the attribute information of nodes and edges in HIN can also be used for link prediction. In this paper, we propose a framework that combines similarity measures of meta path with other attribute information, and formulate a supervised learning task to find the optimal parameters. Experiments on a real data set show that the method has good performance in the problem of link prediction.

Teachers tend to ask students submit their assignments online not only in online courses but also face to face courses. The phenomena of plagiarism is becoming more and more serious due to the ease with which resources can be found on the Internet also, especially in a computer programming course. This paper aims to develop a robust automated detection technology of code plagiarism towards programming course. After analyzing and summarized state of art of code plagiarism technology, a more robust detection technology is developed by combining word2vec with Word mover's distance (WMD) similarity metric in the paper. We consider the different plagiarism methods when students commit their program source code. Then we collect more than 20 thousands code submissions in our introductory C++ programming course for non-major students and check whether it is a plagiarized code manually. In the process, we examine how our proposed method compare with two other main algorithms and their suitability for different plagiarism characteristics. The results obtained on the dataset indicate that our approach is well suited for detect different types of code plagiarism. We conclude that incorporating WMD similarity metric is crucial for improved effective and adaptability.

Experiences of complex search tasks are important in social interaction and in problem solving. Considering the high importance of complex search experiences, many search experience management systems (SEMSs) have been introduced. Like any other life experience, complex search experiences should maintain 3 types of global coherence: temporal, causal and thematic coherence. However, to the best of our knowledge, none of the available SEMSs were designed to support all the 3 types of global coherence. In this paper, we introduce a coherence-oriented complex search experience management method named TimeTree. By organizing queries and clicks of a complex search task as a relative chronological source-tracking tree (RCST), TimeTree manages to support all the 3 types of global coherence. We describe a user study to evaluate TimeTree in 2 typical types of complex search task. The subjective evaluation results, the expert evaluation results, and the objective evaluation results all suggest that TimeTree can help maintain temporal, causal and thematic coherence for complex search experiences.

Outliers in student activities record will cause misjudgment on student grading in online courses, especially in a SPOC course due to increased student flexibility. These outliers increased model complexity dramatically while grading students. Inspired by the process of human knowledge construction, Self-paced Learning Model which starts with an easy sample and gradually incorporates complex samples into the objective function optimization give potential to deal with high complexity. However, traditional Self-paced Learning Models use dichotomous weights to classify samples into easy ones or complex ones only. The accuracy of the model could be decreased due to the change of sample boundary caused by outliers. To this end, an identification refinement learning model is proposed and then is applied to online SPOC student grading experiments. The matrix factorization and classification experiments verify that the algorithm can improve the robustness and accuracy of the model.

Since complex search tasks are usually divided into subtasks, providing subtask-oriented query recommendations is an effective way to support complex search tasks. Currently, most subtask-oriented query recommendation methods extract subtasks from plain form search logs consisting of only queries and clicks, providing limited clues to identify subtasks. Meanwhile, for several decades, the Computer Human Interface (CHI)/Human Computer Interaction (HCI) communities have been working on new complex search tools for the purpose of supporting rich user interactions beyond just queries and clicks, and thus providing rich form search logs with more clues for subtask identification. In this paper, we researched the provision of subtask-oriented query recommendations by extracting thematic experiences from the rich form search logs of complex search tasks logged in a proposed visual data structure. We introduce the tree structure of the visual data structure and propose a visual-based subtask identification method based on the visual data structure. We then introduce a personalized PageRank-based method to recommend queries by ranking nodes on the network from the identified subtasks. We evaluated the proposed methods in experiments consisting of informative and tentative search tasks.

随着大规模社会网络的发展,链接预测成为了一个重要的研究课题。研究了在社会网络中融合节点属性信息进行链接预测,在传统的社会-属性网络图模型的基础上,将节点属性的类别这一重要参量加入到网络构建中。基于此,提出了一系列为网络中不同类型的连边分配边权重的方法,最后通过随机游走的方法进行网络链接的预测。实验表明,所提链接预测方法相比同类方法有明显的效果提升。

In this paper, we detail our efforts at designing a study to evaluate the quality of peer-grading for a SPOC course. Peer grading is an common used strategy to evaluate students learning performance in MOOC/SPOC courses. The strategy assign an author's work to his classmates who take the responsibility to be reviewers of the author's work. Both reviewer and reviewee are classmates who come from same course. Low quality peer grading is still a challenge due to the level of reviewers and the relationship between reviewer and reviewee. We propose a regression model to represent the correlation of student performance on subjective question and objective question. Then a measurement is presented to assess the quality of peer grading of subjective assignments. Addressing the high missing rate of score record of objective questions, three matrix completion algorithms are introduced in order to make the record complete. Experiments show that the validity of the proposed evaluation method. Instructors can leverage peer grading in their classes and to deal with homework evaluation more easily.

The massive online open course ( MOOC) is mainly dependent on video,the lecturer can not directly access to the student′s state of listening. However in the face-to-face course,the lecturer can timely access to the feedback,and improve the teaching process through the student′s facial expression. In order to solve this problem,the student′sfacial expression identification was introducedinto MOOC,7 kinds of expressions are defined,and a dynamic expression recognition algorithm is explored. Firstly,the face images are ex-tracted from the video captured by camera at 5 frames intervals, and then its are processed by FACE + + to select facial feature points. Secondly,the selected several feature modes of the important parts of face studied by the SVM are convert to the nodes of the classification tree. Finally,the facial expressions are recognized by the combination of feature modes of the important face parts. Exper-iment shows that the algorithm does not be limited to recognize the 6 kinds of traditional facial expression recognition,it has better effect on the MOOC.

社区发现一直是社会网络研究中的热点内容。但是当前社区发现算法更加关注用户与用户之间的链接关系,而对社会网络中用户生成内容(user generated contents,UGC)大数据研究较少。用户生成内容是Web2.0的特点,也是社会网络平台吸引用户的重要原因之一,对社区的形成起着重要作用。提出了一种新的社区发现算法,能够综合利用用户与用户之间的链接关系以及用户生成内容来确定用户的社区划分。该算法用LDA(latent Dirichlet allocation)算法分析用户生成内容中主要的内容形式——文本信息,同时通过谱分析方法分析用户与用户之间的链接关系,并有机结合以发现网络的社区结构。通过分析科学网的真实数据,证明了所提算法能够有效综合利用用户生成内容与用户链接关系,使社区发现的结果更加客观准确。