Jianshan Sun

Sentiment classification: The contribution of ensemble learning

With the rapid development of information technologies, user-generated contents can be conveniently posted online. While individuals, businesses, and governments are interested in evaluating the sentiments behind this content, there are no consistent conclusions on which sentiment classification technologies are best. Recent studies suggest that ensemble learning methods may have potential applicability in sentiment classification. In this study, we conduct a comparative assessment of the performance of three popular ensemble methods (Bagging, Boosting, and Random Subspace) based on five base learners (Naive Bayes, Maximum Entropy, Decision Tree, K Nearest Neighbor, and Support Vector Machine) for sentiment classification. Moreover, ten public sentiment analysis datasets were investigated to verify the effectiveness of ensemble learning for sentiment analysis. Based on a total of 1200 comparative group experiments, empirical results reveal that ensemble methods substantially improve the performance of individual base learners for sentiment classification. Among the three ensemble methods, Random Subspace has the better comparative results, although it was seldom discussed in the literature. These results illustrate that ensemble learning methods can be used as a viable method for sentiment classification.

Leveraging Content and Connections for Scientific Article Recommendation in Social Computing Contexts

Rapid proliferation of information technologies has generated a great volume of information that makes scientific information searching more challenging. Personalized recommendation is a widely used technique to help researchers find relevant information. Researchers involved in a social computing context generate abundant content and form heterogeneous connections. Existing article recommendation techniques fail to perform a deep analysis of this information. This research proposes a novel approach to recommend scientific articles to researchers by leveraging content and connections. In this approach, we first analyze the semantic content of the article by keyword similarity calculation and then extract online users’ connections to support article voting and finally employ a two-stage recommendation process to suggest relevant articles. The proposed method has been implemented in ScholarMate (www.scholarmate.com), an online research social network platform. Two experiments are conducted and the evaluation results indicate that the proposed method is more effective than the baseline methods.

A personalized information recommendation system for R&D project opportunity finding in big data contexts

With the rapid proliferation of online information, how to find useful information, such as suitable jobs, appropriate experts, and proper projects, is really an important problem. Recommendation technique, as one of emerging tools to deal with information overload and information asymmetry, is critically important for providing personalized online information services. With the increase of R&D investment in government and industry, such as high-tech companies and advanced manufacturing enterprises, more and more R&D project information are launched in public websites for cooperation. When the number of online information and users is extremely huge, how to effectively recommend R&D project opportunities to related researchers and practitioners is a challenging and complex task. In this paper, a novel two-stage method is proposed for R&D project opportunity recommendation. An information filtering method is first offered to identity proper R&D projects as a candidate set. Then, an information aggregation model with various constraints is suggested to recommend appropriate R&D projects for applicants. The proposed method has been implemented in an online research community - ScholarMate (www.scholarmate.com). An online user study has been conducted and the evaluation results exhibit that the proposed method is more effective than existing ones.

A Novel Approach for Personalized Article Recommendation in Online Scientific Communities

Rapid proliferation of information technologies has generated sheer volume of information which makes scientific research related information searching more challenging. Personalized recommendation is the widely adopted technique to recommend relevant documents to researchers. Current methods are suffering from mismatch problem and match irrelevance problem and fail to generate highly related results. To overcome these problems, we propose a novel approach to recommend articles to the researchers. In our approach we integrate three types of similarity measures: keyword similarity, journal similarity, and author similarity to measure the relevance of the articles to researchers. The keyword similarity is used to generate candidate list of articles, and the journal similarity and author similarity are used to select most suitable articles from the candidate list. The integrated similarity measure is used to rank the articles based on their relevance. The proposed method is implemented in Scholar Mate (www.scholarmate.com), the online research social network platform. The evaluation results exhibit that proposed method is more effective than existing ones.

Collaborative matrix factorization mechanism for group recommendation in big data-based library systems

Academic groups are designed specifically for researchers. A group recommendation procedure is essential to support scholars’ research-based social activities. However, group recommendation methods are rarely applied in online libraries and they often suffer from scalability problem in big data context. The purpose of this paper is to facilitate academic group activities in big data-based library systems by recommending satisfying articles for academic groups.,The authors propose a collaborative matrix factorization (CoMF) mechanism and implement paralleled CoMF under Hadoop framework. Its rationale is collaboratively decomposing researcher-article interaction matrix and group-article interaction matrix. Furthermore, three extended models of CoMF are proposed.,Empirical studies on CiteULike data set demonstrate that CoMF and three variants outperform baseline algorithms in terms of accuracy and robustness. The scalability evaluation of paralleled CoMF shows its potential value in scholarly big data environment.,The proposed methods fill the gap of group-article recommendation in online libraries domain. The proposed methods have enriched the group recommendation methods by considering the interaction effects between groups and members. The proposed methods are the first attempt to implement group recommendation methods in big data contexts.,The proposed methods can improve group activity effectiveness and information shareability in academic groups, which are beneficial to membership retention and enhance the service quality of online library systems. Furthermore, the proposed methods are applicable to big data contexts and make library system services more efficient.,The proposed methods have potential value to improve scientific collaboration and research innovation.,The proposed CoMF method is a novel group recommendation method based on the collaboratively decomposition of researcher-article matrix and group-article matrix. The process indirectly reflects the interaction between groups and members, which accords with actual library environments and provides an interpretable recommendation result.

A crowdsourcing-based topic model for service matchmaking in Internet of Things

Abstract The Internet of Things (IoT) provide intelligence for the communication between people and physical objects. An important and critical issue in the IoT service applications is how to match the suitable IoT services with service requests. To solve this problem, researchers use semantic modeling methods to make service matching. Semantic modeling methods in IoT extract meta-data from text using rule-based approaches or machine learning techniques often suffer from the scalability and sparseness since text provided by sensors is short and unstructured. In recent years, topic modeling has been used in IoT service matchmaking. However, most topic modeling methods do not perform well in IoT service matchmaking since the text is too short. In order to address the issues, this paper proposes a new topic modeling method to extract topic signatures provided by intelligent devices. The method extends the classical knowledge representation framework and improves the qualities of service information extraction, and this process is able to improve the effectiveness of service matchmaking in IoT service. The framework incorporates human cognition to improve the effectiveness of the algorithm and make the algorithm more robust in heterogeneous systems in the IoT. The usefulness of the method is illustrated via experiments using real datasets.

Diversified Recommendation Incorporating Item Content Information Based on MOEA/D

There has been an increasing awareness that accuracy is not the only criteria in the evaluation of recommender systems. Additional properties such as diversity, novelty and interpretability are playing more important roles in increasing satisfaction of users when interacting with the recommender systems. However, designing a recommendation algorithm that optimizes the abovementioned properties simultaneously is hard since these objectives are conflicting. In this paper, we propose a multi-objective evolutionary algorithm based on decomposition to recommend diversified recommendation lists to each user. Notably, the item content information are taken into account when devising the diversity objective function, which makes the recommendation lists highly explainable. Experimental results on the movie dataset demonstrate that the proposed algorithm can generate a more diversified and novel recommendation, without sacrificing the accuracy significantly.

Online Social Network as a Powerful Tool to Identify Experts for Emergency Management

Widespread use of social network has changed people’s daily life as well as the way of emergency management.As a useful tool for information dissemination, communication and collaboration, social network plays an important role in the process of emergency management: mitigation, preparedness, communications, response and recovery. Emergency problem solving nowadays often need experts from various domain areas such as medicine, nuclear, chemistry, information technology and so on. It’s difficult and costly for local emergency management databases to be well prepared since emergency disasters are small probability events. This research captures the advantage of social network to tackle such issues. Expertise on online social network is available and trustworthy with supervision of crowds. Therefore, we propose a method though the integration of social position analysis and expertise level analysis to profile online individuals. Social position analysis captures the importance or prominence of an individual in social network, and expertise level analysis measures one’s expertise relevance and expertise level. Expert finding on online social can also capture one’s interest in the specific emergency disaster. We also give an empirical analysis in a selected small group. Outperformed individuals can be identified. More work is needed to validate the method on the specified social network in the future.

Group recommendation based on a bidirectional tensor factorization model

Capturing the preference of virtual groups that consist of a set of users with diversified preference helps recommend targeted products or services in social network platform. Existing strategies for capturing group preference are to directly aggregate individual preferences. Such methods model the preference formation of a group as a unidirectional procedure without considering the influence of the group on individual’s interest. In the context of social group, however, the preference formation is a bidirectional procedure because group preference and individual interest are interrelated. In addition, the influence of group on individuals is usually distinct among users. To address these issues, this paper models the group recommendation problem as a bidirectional procedure and proposes a Bidirectional Tensor Factorization model for Group Recommendation (BTF-GR) to capture the interaction between individual’s intrinsic interest and group influence. A Bayesian personalized ranking technique is employed to learn parameters of the proposed BTF-GR model. Empirical studies on two real-world data sets demonstrate that the proposed model outperforms the baseline algorithms such as matrix factorization for implicit feedback and Bayesian personalized ranking.

A hybrid approach for article recommendation in research social networks

With the prevalence of research social networks, determining effective methods for recommending scientific articles to online scholars has become a challenging and complex task. Current studies on article recommendation works are focused on digital libraries and reference sharing websites while studies on research social networking websites have seldom been conducted. Existing content-based approaches or collaborative filtering approaches suffer from the problem of data sparsity. The quality information of articles has been largely ignored in previous studies, thus raising the need for a unified recommendation framework. We propose a hybrid approach to combine relevance, connectivity and quality to recommend scientific articles. The effectiveness of the proposed framework and methods is verified using a user study on a real research social network website. The results demonstrate that our proposed methods outperform baseline methods.