Search

Depression is a common mental disease, and it is committed to promote the research of depression assessment based on physiological signals. By collecting the resting state EEG data of depressive disorder, we collected the resting state EEG data of 14 patients with depression and 17 normal people. Through the analysis, the number of troughs of each person's data was statistically analyzed, combined with the convolution neural network model. The accuracy rate of some data is 94.88% by the statistical trough number method, and the accuracy rate of the remaining part of the test data set is 85.0% through the convolution neural network model, and the final fusion accuracy rate is 86.88%. The experimental results show that the combination of statistical trough number and convolution neural network can distinguish depression patients better in resting state EEG data.

Knowledge Tracing (KT) refers to the problem of predicting learners’ future potential performance given their past learning history in e-learning systems. In order to better trace the learners’ knowledge, KT tasks have become increasingly complicated recently, and various factors related to learning (such as skill, exercise, hint, etc.) have been incorporated into the modeling of KS of the learner, which renders it inadequate for the traditional KT definition to formalize these tasks. Therefore, this paper first gives a more general formal definition of KT tasks, and then proposes an Extensible Deep Knowledge Tracing model for multiple learning factors based on this general definition, named EDKT. EDKT can integrate various different learning factors by extending or ablating factors in two plug-ins on the basis of minor modifications. To demonstrate the effectiveness of the proposed model, we conduct extensive experiments on three real-world benchmark datasets, and the results show that EDKT comprehensively outperforms the state-of-the-art KT models on predicting future learner responses.

Deep Learning based Knowledge Tracing (DLKT) has been shown to outperform other methods due to its strong representational ability. However, DLKT models usually exist a common flaw that all learners share the same model with identical network parameters and hyper-parameters. The drawback of doing so is that the learned knowledge state for each learner is only affected by the specific learning sequence, but less reflect the personalized learning style for each learner. To tackle this problem, we proposes a novel framework, called Adaptive Deep Knowledge Tracing (ADKT), to directly introduce personalization into DLKT. The ADKT framework tries to retrain an adaptive model for each learner based on a pre-trained DLKT model and trace the knowledge states individually for each learner. To verify the effectiveness of ADKT, we further develop the ADKVMN (Adaptive Dynamic Key-Value Memory Network) model by combining the ADKT and the classic DKVMN, which has been widely referred as a state-of-the-art DLKT model. With extensive experiments on two popular benchmark datasets, including the ASSISTments2009 and ASSISTments2015 datasets, we empirically show that ADKVMN has superior predictive performance than DKVMN.

With the advancement of big data and education technology, MOOCs (Massive Online Open Courses) has become a popular education model in online education community. A large number of online courses with diverse disciplinary background are offered freely to global learners. When a learner finishes a series of courses, it is very important to effectively and efficiently recommend the most relevant courses to study next. Traditional item-based recommendation methods are all pointwise approaches where models bias towards estimating the precise rating or relevance score of each item. It would be better to model this problem from a pairwise learning perspective which is more close to the ranking nature of course recommendation. In this paper, we combine item-based collaborative filtering and Bayesian Personalized Ranking for course recommendation problem. We theoretically derive the optimization schema based on Bayesian Personalized Ranking and develop a novel neural network model, called Bayesian Personalized Ranking Network (BPRN), which learns pairwise course preference for each user given her historically enrolled courses. With extensive experiments on a large-scale MOOCs enrollment dataset from XuetangX, we empirically demonstrate that our BPRN framework performs better than state-of-the-art item-based course recommendation methods.

Distant supervision (DS) has the advantage of automatically generating large amounts of labelled training data and has been widely used for relation extraction. However, there are usually many wrong labels in the automatically labelled data in distant supervision (Riedel, Yao, & McCallum, 2010). This paper presents a novel method to reduce the wrong labels. The proposed method uses the semantic Jaccard with word embedding to measure the semantic similarity between the relation phrase in the knowledge base and the dependency phrases between two entities in a sentence to filter the wrong labels. In the process of reducing wrong labels, the semantic Jaccard algorithm selects a core dependency phrase to represent the candidate relation in a sentence, which can capture features for relation classification and avoid the negative impact from irrelevant term sequences that previous neural network models of relation extraction often suffer. In the process of relation classification, the core dependency phrases are also used as the input of a convolutional neural network (CNN) for relation classification. The experimental results show that compared with the methods using original DS data, the methods using filtered DS data performed much better in relation extraction. It indicates that the semantic similarity based method is effective in reducing wrong labels. The relation extraction performance of the CNN model using the core dependency phrases as input is the best of all, which indicates that using the core dependency phrases as input of CNN is enough to capture the features for relation classification and could avoid negative impact from irrelevant terms.

The exponential growth of Massive Open Online Courses (MOOCs) provides increasing learning options along with distinguishing difficulties in choosing courses or finding learning paths for learners. To improve online learning resource utilization, a knowledge graph, MOOC-KG, collecting and formulating the information of online courses on major platforms was constructed. The aim of MOOC-KG is to not only provide users with a new MOOC resource organization but also supply an easy approach for learning resource using among several platforms. In our work, an acknowledged educational ontology was adopted to model the knowledge of online learning resources and the web information extraction methods were developed to collect the metadata on major platforms. The entity disambiguation and conflict resolution methods were adopted to integrate the resources among different platforms. As a result, MOOC-KG is the biggest MOOC-related knowledge graph, which represents and stores 28591 instances with their corresponding relations, including 4 platforms, 604 universities, 18671 teachers and 9312 courses until now. Related valuable data analyses illustrate the fundamental information and reveal explorable parts of MOOC-KG. We have released the related project on GitHub to facilitates further improvement and extensive use.

During the past five years, tons of economic 360 VR cameras (e.g., Ricoh Theta, Samsumg Gear360, LG 360, Insta 360) are sold in the market. While 360 VR videos become ubiquitous very soon, 360 VR video standardization is still under discussion in the digital industry, and more concrete efforts are desired to accelerate its standardization and applications. Though ERP has been widely used for projection and packing layout while encoding 360 VR videos, it has severe projection distortion near poles. In this paper, we introduce a new format for encoding and storing 360 VR videos using hybrid cylindrical projection after thoroughly analyzing the problems with ERP. We show that our new hybrid format can minimize stretching distortion and generate well balanced pixel distribution in the resulting projection.

With the application of electronic medical records in medical field, more and more people are paying attention to how to use these data efficiently. In this paper, the BiLSTM-CRF model is applied to Chinese electronic medical records to recognize related named entities in these records. For the characteristics of Chinese electronic medical records, firstly, the one-hot vector of each word is obtained in units of sentences. Secondly, map one-hot vector to a low-dimensional dense word vector. Thirdly, word vector is used as the input of the BiLSTM layer to achieve automatic extraction of sentence features. Finally, the CRF layer performs sequence-level labeling of sentences. In addition, drug dictionary and post-correction rules are added to correct the segmentation error of entity boundary, to improve recognition accuracy of related named entities. The F1 value of this method on a given test data set is 87.68%.

Word embeddings have been used as popular features in various Natural Language Processing(NLP) tasks. To overcome the coverage problem of statistics, compositional model is proposed, which embeds basic units of a language, and compose structures of higher hierarchy, like idiom, phrase, and named entity. In that case, selecting the right level of basic-unit embedding to represent semantics of higher hierarchy unit is crucial. This paper investigates this problem by Chinese phrase representation task, in which language characters and words are viewed as basic units. We define a phrase representation evaluation tasks by utilizing Wikipedia. We propose four intuitionistic composing methods from basic embedding to higher level representation, and investigate the performance of the two basic units. Empirical results show that with all composing methods, word embedding out performs character embedding on both tasks, which indicates that word level is more suitable for composing semantic representation.

Recently there has been an increasing research interest in short text such as news headline. Due to the inherent sparsity of short text, the current text classification methods perform badly when applied to the classification of news headlines. To overcome this problem, a novel method which enhances the semantic representation of headlines is proposed in this paper. Firstly, we add some keywords extracted from the most similar news to expand the word features. Secondly, we use the corpus in news domain to pre-train the word embedding so as to enhance the word representation. Moreover, Fasttext classifier, which uses a liner method to classify text with fast speed and high accuracy, is adopted for news headline classification. On the task for Chinese news headline categorization in NLPCC2017, the proposed method achieved 83.1% of the F-measure, which got the first rank in 33 teams.

MOOCs have attracted a large number of learners with different education background all over the world. Despite its increasing popularity, MOOCs still suffer from the problem of high drop-out rate. One important reason may be due to the difficulty in understanding learning demand and user interests. To helper users find the most suitable courses, personalized course recommendation technology has become a hot research topic in e-learning and data mining community. One of the keys to the success of personalized course recommendation is a good user modeling method. Previous works in course recommendation often focus on developing user modeling methodology which learns latent user interests from historic learning data. Recently, interactive course recommendation has become more and more popular. In this paradigm, recommender systems can directly query user interests through survey tables or questionnaires and thus the learned interests may be more accurate. In this paper, we study the user interest acquisition problem based on the interactive course recommendation framework (ICRF). Under this framework, we systematically discuss different settings on querying user interests. To reduce performance-cost score, we propose the ICRF user interest acquisition algorithm that combines representative sampling and interest propagation algorithm to acquire user interests in a cost-effective way. With extensive experiments on real-world MOOC course enrollment datasets, we empirically demonstrate that our selective acquisition strategy is very effective and it can reduce the performance-cost score by 30.25% compared to the traditional aggressive acquisition strategies.

Relation extraction is very useful for many applications and has attracted much attention. The dominant prior methods for relation extraction were supervised methods which are relation-specific and limited by the availability of annotated training data. In this paper, we propose a method using hierarchical clustering to extract unbounded relations without relying on training data. The relation among entities in a sentence depends on the terms associated with the entities. Terms on the expandPath capture the relations between the entities. Given a relation, though an expandPath may have more than one dependency phrase, only the core dependency phrase describes the specific relation between the subject and the object. Our method uses heuristic rules to select the core dependency phrases and clusters entity pairs according to the similarity of the core dependency phrases in order to avoid irrelevant information and capture the semantics of the relation between entities more precisely. At last, our method automatically labels the relation clusters on basis of the semantics of core dependency phrases. The experimental results show that our method can cluster entity pairs which have the same relations more accurately and generate appropriate labels for the relations.

In order to be considered as Linked Data, the datasets on the web must be linked to other datasets. Current studies on dataset interlinking prediction researches do not distinguish the type of links, which are of less help for real application scenarios, as dataset publishers still do not know what kinds of RDF links can be established and furthermore how to configure the data linking algorithms. In this paper, we focus on predicting the possible links between datasets with the most important RDF link type, owl:sameAs. Since the goal is to discriminate between linked dataset pairs against not-linked ones, we formulate the link prediction problem as a classification problem. We adopt Random Forest as the basic classifier to incorporate features of the scores output by unsupervised predictors, and apply the bagging technique to combine multiple forests to reduce variance and improve the accuracy. Experiments show we can improve the prediction performance by about 10% in AUROC.

Due to the outstanding role of owl:sameAs as the most widely used linking predicate, the problem of identifying potential Linked Data datasets for sameAs interlinking was studied in this paper. The problem was regarded as a Recommender systems problem, so serveral classical collaborative filtering techniques were employed. The user-item matrix was constructed with rating values defined depending on the number of owl:sameAs RDF links between datasets from Linked Open Data Cloud 2014 dump. The similarity measure is a key for memory-based collaborative filtering methods, a novel dataset semantic similarity measure was proposed based on the vocabulary information extracted from datasets. We conducted experiments to evaluate the accuracy of both the predicted ratings and recommended datasets lists of these recommenders. The experiments demonstrated that our customized recommenders out-performed the original ones with a great deal, and achieved much better metrics in both evaluations.