融合多源信息的专家学术专长语义匹配方法

通过评议文档与专家库的专家学术专长匹配,可以输出领域相关的候选专家列表,是同行评议中专家遴选和专家推荐的重要参考依据.针对学术专长匹配存在语义鸿沟、无法反映专家和评议文档多源信息间语义关联的问题,首先对专家信息和评议文档的多源信息进行语义特征抽取,融合多类特征进行表示学习,利用卷积神经网络设计专家特征抽取器ExpFeat和评议特征抽取器RevFeat,采用词嵌入方法和注意力机制对专家专长标签、评议文档关键词、学科分类树语义特征进行抽取和融合,生成具有“小同行”特征的专家和评议文档语义特征向量表示,解决多类信息源间不同学术分类标准造成的语义差异,反映内在语义联系,利用低维稠密向量表达语义信息,降低匹配复杂度.然后,根据专家语义特征表示和评议文档语义特征表示进行学术专长语义匹配,将专家和评议文档特征向量映射到相同语义空间,计算向量间余弦相似度衡量语义相似性,引入负例专家进行模型训练,通过softmax函数计算最大化正例专家概率优化特征提取器参数,进一步提升语义差异的捕捉能力,解决专家信息和评议文档之间的语义鸿沟问题,提升专家匹配效果.最后,在开源的论文评审数据集和项目评审数据集上进行了实验...     

同行评议专家网络评价系统的研究与开发

同行评议专家网络评价是根据专家近几年评审过的项目,结合项目立项资助情况和专家的评审结果,对专家的评审质量和能力进行的网上评估,是专家评审结果的科学性和可信度的重要保障。文章提出了影响同行评议专家综合评价的五因素,并基于此模型,开发了同行评议专家网络评价子系统。     

基于图卷积网络的重大事件趋势预测

针对基于公开新闻数据的重大事件趋势预测研究在文本特征提取上语义理解不足问题,提出基于图卷积网络的重大事件趋势预测方法.利用结构化事件信息构建以事件为中心的语义关联图;按时间片粒度分割出局部事件语义关联图,利用预训练词向量对节点文本初始化特征表示;将带有节点特征的局部事件语义关联图输入图卷积网络,聚合节点邻域信息计算图特征表示并输入到分类器,训练预测模型并输出预测结果.以朝鲜核行为趋势预测问题验证所提出方法的有效性,对测试集(2017年3月～2018年3月)预测准确率达到76.92％,全时间段预测准确率为89.58％.预测结果表明,该方法增强了对文本的语义理解,预测精度优于基于专家知识构建特征的方法.     

关联项目增强的多兴趣序列推荐方法

现有基于多兴趣框架的序列推荐方法仅从用户近期交互序列中学习得到用户多兴趣表示,忽略了数据集中项目间的关联信息。针对这一问题,提出了一种关联项目增强的多兴趣序列推荐方法IAMIRec（item associations aware multi-interest sequential recommendation method）。首先通过数据集中用户交互序列计算得到项目关联集合和对应的项目关联矩阵,然后根据项目关联矩阵通过多头自注意力机制建模用户的近期交互序列,最后使用多兴趣框架学习得到用户的多个兴趣向量并进行top-N推荐。在三个数据集上对该方法进行了测试与分析,IAMIRec在recall、NDCG（normalized discounted cumulative gain）和hit rate指标上的表现均优于相关方法。实验结果说明 IAMIRec可以实现更优的推荐性能,也表明引入项目关联信息可以有效增强用户的多兴趣表示。     

关联文本信息的网络表示学习推荐算法

通过基于随机游走的网络表示学习算法得到节点的低维嵌入向量,进而将其应用于推荐系统是推荐领域很流行的研究方向.针对当前基于随机游走的网络表示学习算法仅着重考虑了网络结构特性而忽略文本信息的问题,提出一种关联文本信息的网络表示学习推荐算法.首先在随机游走阶段,考虑到了节点文本间的相似度,联合结构和文本信息对下一游走节点进行筛选;然后在网络表示学习部分融合文本信息,引入注意力矩阵,对文本信息矩阵中的向量进行加权表示;最后将生成的节点向量应用于推荐系统.在实验部分,将所提算法与常见的3种算法在两个数据集上进行对比分析,并对所提算法进行了参数敏感性分析.实验结果表明所提算法在AUC评价指标上的性能优于另外3种算法,可见该算法在个性化推荐中的有效性.     

融合多层次交互注意力的引文推荐算法

学术引文推荐是指通过论文间的匹配关系为查询论文提供深度匹配的引文文献列表，提高学者科研工作效率.现有方法主要基于短文本匹配(如关键词、标题等),缺乏对论文结构和整体语义的表示能力，导致检索结果语义相关性差.本文从长文本的深层次数据特征出发，提出一种基于层次化交互注意力匹配的引文推荐算法.基于深度神经网络构建单词、句子、文章的层次化表示框架，提升长文本的结构化表示能力；使用内部注意力机制增强学术论文的内部语义表示；使用交互注意力机制挖掘引文间细粒度匹配特征.在计算机、自然语言处理、医学等学术文献数据集上进行实验验证，提出的方法在ACC和F1等指标均优于短文本匹配模型，结果表明层次化交互注意力能获得更好的引文匹配效果.     

多空间交互协同过滤推荐

大数据时代,由于信息过载,用户很难从海量数据中寻找出感兴趣的内容,个性化推荐系统的诞生极好地解决了这个问题.协同过滤算法被广泛应用于个性化推荐领域,但由于模型的限制,推荐效果未能得到进一步提升.现有的基于协同过滤模型的改进方法大多都是通过引入表示学习方法来得到更好的用户表示向量和项目表示向量,或通过改进用户项目匹配函数来提升推荐能力,但此类工作都致力于从单个交互提取用户-项目交互信息.文中提出了一种多空间交互协同过滤推荐算法,将用户向量和项目向量映射到多空间,从多角度做用户-项目交互,使用两层注意力机制聚合最终的用户表示向量和项目表示向量,以进行评分预测.在公开的真实数据集上,多空间交互协同过滤模型(MSICF)与多个基线模型进行了对比实验,MSICF模型的评估优于对比的基线方法.     

知识增强的图神经网络序列推荐模型

现有基于图神经网络的序列推荐模型大多仅关注用户与项目交互的结构性信息,序列偏好的学习仅涉及项目交互顺序,缺乏项目自身的内容信息,并且未有效利用用户信息及挖掘项目之间更深层的语义关系。提出一种知识增强的图神经网络序列推荐模型KGGNN,引入知识图谱,并结合用户交互数据构建协同知识图谱,学习得到项目语义关联辅助信息以及用户关联辅助信息。将交互序列构建成有向序列图,利用门控图神经网络以及用户关联辅助信息学习序列中项目节点的结构性信息。通过注意力机制组合项目向量作为全局序列偏好,将最近交互的项目作为当前兴趣偏好,融合两者形成最终序列偏好,并结合项目语义关联辅助信息进行模型预测。在Amazon-Book、Last-FM、Yelp2018这3个公开数据集上的实验结果表明,辅助信息能有效提升序列推荐的准确性,该模型在命中率（HIT@K）和归一化折损累计增益（NDCG@K）2个指标上相较于GRU4Rec、NARM、SASRec等模型均有显著提升。当评估指标K值选取10时,与KGSR模型相比,其HIT@10指标在3个数据集上分别提升12.9%、4.5%、6.9%,NDCG@10指标在3个数据集上分别提升...     

超边约束的异质超网络表示学习方法

与普通网络相比，超网络具有复杂的元组关系（超边），然而现有的大多数网络表示学习方法并不能捕获元组关系。针对上述问题，提出一种超边约束的异质超网络表示学习方法（HRHC）。首先，引入一种结合团扩展和星型扩展的方法，从而将异质超网络转换为异质网络；其次，引入感知节点语义相关性的元路径游走方法捕获异质节点之间的语义关系；最后，通过超边约束机制捕获节点之间的元组关系，从而获得高质量的节点表示向量。在3个真实世界的超网络数据集上的实验结果表明，对于链接预测任务，所提方法在drug、GPS和MovieLens数据集上都取得了较好的结果；对于超网络重建任务，当超边重建比率大于0.6时，所提方法在drug数据集上的准确性（ACC）优于次优的Hyper2vec(biased 2nd order random walks in Hyper-networks)，同时所提方法在GPS数据集上的ACC超过其他基线方法中次优的基于关联图的超边超边约束的异质超网络表示学习方法（HRHC-关联图）15.6个百分点。     

基于图注意力网络表示学习的协同过滤推荐算法

针对传统的基于模型的协同过滤推荐算法未能有效利用用户与项目的属性信息以及用户之间与项目之间的关系结构信息, 本文提出一种基于图注意力网络表示学习的协同过滤推荐算法. 该算法使用知识图谱表示节点的属性特征信息和节点间的关系结构信息, 并在用户和项目的同质网络上进行节点的图注意力网络表示学习, 得到用户和项目的网络嵌入特征表示, 最后构建融合网络嵌入信息的神经矩阵分解模型获得推荐结果. 本文在Movielens数据集上与相关算法进行对比实验, 实验证明该算法能优化模型的推荐性能, 提高推荐的召回率HR@K和归一化折损累计增益NDCG@K.     

A semantic social network-based expert recommender system

This research work presents a framework to build a hybrid expert recommendation system that integrates the characteristics of content-based recommendation algorithms into a social network-based collaborative filtering system. The proposed method aims at improving the accuracy of recommendation prediction by considering the social aspect of experts’ behaviors. For this purpose, content-based profiles of experts are first constructed by crawling online resources. A semantic kernel is built by using the background knowledge derived from Wikipedia repository. The semantic kernel is employed to enrich the experts’ profiles. Experts’ social communities are detected by applying the social network analysis and using factors such as experience, background, knowledge level, and personal preferences. By this way, hidden social relationships can be discovered among individuals. Identifying communities is used for determining a particular member’s value according to the general pattern behavior of the community that the individual belongs to. Representative members of a community are then identified using the eigenvector centrality measure. Finally, a recommendation is made to relate an information item, for which a user is seeking an expert, to the representatives of the most relevant community. Such a semantic social network-based expert recommendation system can provide benefits to both experts and users if one looks at the recommendation from two perspectives. From the user’s perspective, she/he is provided with a group of experts who can help the user with her/his information needs. From the expert’s perspective she/he has been assigned to work on relevant information items that fall under her/his expertise and interests.     

Review Expert Collaborative Recommendation Algorithm Based on Topic Relationship

The project review information plays an important role in the recommendation of review experts. In this paper, we aim to determine review expert's rating by using the historical rating records and the final decision results on the previous projects, and by means of some rules, we construct a rating matrix for projects and experts. For the data sparseness problem of the rating matrix and the “cold start” problem of new expert recommendation, we assume that those projects/experts with similar topics have similar feature vectors and propose a review expert collaborative recommendation algorithm based on topic relationship. Firstly, we obtain topics of projects/experts based on latent Dirichlet allocation (LDA) model, and build the topic relationship network of projects/experts. Then, through the topic relationship between projects/experts, we find a neighbor collection which shares the largest similarity with target project/expert, and integrate the collection into the collaborative filtering recommendation algorithm based on matrix factorization. Finally, by learning the rating matrix to get feature vectors of the projects and experts, we can predict the ratings that a target project will give candidate review experts, and thus achieve the review expert recommendation. Experiments on real data set show that the proposed method could predict the review expert rating more effectively, and improve the recommendation effect of review experts.      

Developing cognitive task analysis and the importance of socio-cognitive competence/insight for professional practice

Accelerating the cognitive expertise of professionals is a critical challenge for many organizations. This paper reports a collaborative, longitudinal, academic practitioner project which aimed to elicit, document, and accelerate the cognitive expertise of engineering professionals working with the manufacture and management of petroleum additives. Twenty-five engineering experts were trained by three academic psychologists to use applied cognitive task analysis (ACTA) interview techniques to document the cognition of their expert peers. Results had high face validity for practitioners who elicited hot/sensory-based cognition, a number of perceptual skills and mental models, highlighting undocumented context specific expertise. We conclude from a peer review of findings, combined with experienced CTA analysts that ACTA techniques can be advanced in context by the explicit recognition and development of socio-cognitive competence/insight.     

Temporal expert finding through generalized time topic modeling

This paper addresses the problem of semantics-based temporal expert finding, which means identifying a person with given expertise for different time periods. For example, many real world applications like reviewer matching for papers and finding hot topics in newswire articles need to consider time dynamics. Intuitively there will be different reviewers and reporters for different topics during different time periods. Traditional approaches used graph-based link structure by using keywords based matching and ignored semantic information, while topic modeling considered semantics-based information without conferences influence (richer text semantics and relationships between authors) and time information simultaneously. Consequently they result in not finding appropriate experts for different time periods. We propose a novel Temporal-Expert-Topic (TET) approach based on Semantics and Temporal Information based Expert Search (STMS) for temporal expert finding, which simultaneously models conferences influence and time information. Consequently, topics (semantically related probabilistic clusters of words) occurrence and correlations change over time, while the meaning of a particular topic almost remains unchanged. By using Bayes Theorem we can obtain topically related experts for different time periods and show how experts’ interests and relationships change over time. Experimental results on scientific literature dataset show that the proposed generalized time topic modeling approach significantly outperformed the non-generalized time topic modeling approaches, due to simultaneously capturing conferences influence with time information.     

融合主题信息和卷积神经网络的混合推荐算法

针对传统的协同过滤算法中数据稀疏和推荐结果不准确的问题,提出了一种基于隐狄利克雷分布（LDA）与卷积神经网络（CNN）的概率矩阵分解推荐模型（LCPMF）,该模型综合考虑项目评论文档的主题信息与深层语义信息。首先,分别使用LDA主题模型和文本CNN对项目评论文档建模;然后,获取项目评论文档的显著潜在低维主题信息及全局深层语义信息,从而捕获项目文档的多层次特征表示;最后,将得到的用户和多层次的项目特征融合到概率矩阵分解（PMF）模型中,产生预测评分进行推荐。在真实数据集Movielens 1M、Movielens 10M与Amazon上,将LCPMF与经典的PMF、协同深度学习（CDL）、卷积矩阵因子分解模型（ConvMF）模型进行对比。实验结果表明,相较PMF、CDL、ConvMF模型,所提推荐模型LCPMF的均方根误差（RMSE）和平均绝对误差（MAE）在Movielens 1M数据集上分别降低了6.03%和5.38%、5.12%和4.03%、1.46%和2.00%,在Movielens 10M数据集上分别降低了5.35%和5.67%、2.50%和3.64%、1.75%和1.74%,在Amazon数据集上分别降低17.71%和23.63%、14.92%和17.47%、3.51%和4.87%,验证了所提模型在推荐系统中的可行性与有效性。     

基于本体的项目和领域专家匹配原型系统*

设计和实现了一个基于本体的项目和领域专家匹配原型系统。首先在集成基于点和基于边的语义相似度计算方法的基础上,提出一种计算两棵树型概念结构中两个概念节点之间以及两棵树型概念结构之间的语义相似度的方法;然后以项目文档和领域专家文档匹配为实例,按照相似度大小排序,为一个项目选择合适的评审专家。测试结果表明,所提出的方法及实现的系统原型可以在较大程度上拟合人工判断,获得较高的准确率和召回率,为评审项目时选择合适的领域专家提供了一个新的思路。     

融合高低层语义信息的自然语言句子匹配方法

针对目前自然语言句子匹配方法未能融合公共语义信息以及难以捕获深层次语义信息的问题,提出了一种融合高低层语义信息的自然语言句子匹配算法。该算法首先采用预训练词向量GloVe和字符级别的词向量获取句子P和句子Q的词嵌入表示;其次,使用双向LSTM(long-short term memory)对P和Q进行编码,再初步融合P和Q获取低层语义信息;然后,计算P与Q的双向注意力,拼接得到语义表征,再计算其自注意力获取高层语义信息;最后,采用启发式的融合函数将低层语义信息和高层语义信息进行融合,得到最终的语义表征,使用卷积神经网络预测得到答案。在文本蕴涵识别和释义识别两个任务上评估了该模型。在文本蕴涵识别任务SNLI数据集和释义识别任务Quora数据集上进行了实验,实验结果表明该算法在SNLI测试集上的准确率为87.1%,在Quora测试集上的准确率为86.8%,验证了算法在自然语言句子匹配任务上的有效性。     

An intelligent decision support approach for reviewer assignment in R&D project selection

In the process of Research and Development (R&D) project selection, experts play an important role because their opinions are the foundation on which to judge the potential value of a project. How to assign the most appropriate experts to review project proposals might greatly affect the quality of project selection, which in turn could affect the return on investment of the funding organization. However, in many funding organizations, current approaches to assigning reviewers are still based on simply matching the discipline area of the reviewers with that of the proposal, which could result in poor quality of project selection and poor future financial return. Additionally, these approaches might make it difficult to balance resources and resolve conflicts of interests between reviewers and applicants. Therefore, to overcome these problems, there is an urgent need for a systematic approach to support and automate the reviewer assignment process. This research aims at proposing an intelligent decision support approach for reviewer assignment and developing an Assignment Decision Support System (ADSS). In this approach, heuristic knowledge of expert assignment and techniques of operations research are integrated. The approach uses decision models to determine the best solution of reviewer assignment that maximizes the total expertise level of the reviewers assigned to proposals. It also balances the distribution of proposals at different grades and solves conflicts of interests between reviewers and applicants. Its application in the National Natural Science Foundation of China (NSFC) and the computational results of its effectiveness and efficiency are also described.     

利用门控网络构建用户动态兴趣的序列推荐模型

在推荐系统中,现有多数序列推荐方法将用户行为视为一个时间有序的序列进行用户兴趣建模,用户兴趣的动态变化导致模型难以从用户行为序列中捕捉准确的用户兴趣信息。针对该问题,同时考虑到项目间成对的共现模式应作为交互信息的补充,提出利用门控网络构建用户动态兴趣的序列推荐模型DCGN。使用门控线性单元捕获交互序列中的用户兴趣,利用带有注意力权重的门控循环网络学习用户的动态兴趣。在此基础上,对用户交互项目间的共现模式进行建模,与用户兴趣信息以及用户信息进行融合后输入深度神经网络,得到最终推荐结果。在ML100K、Amazon 5-Elect、Retailrocket 3个公开数据集上进行实验,使用精确率、归一化折损累积增益和命中率进行性能评估,结果表明,DCGN模型较NARM、GRU4Rec、NLR等主流序列推荐模型性能更优,其归一化折损增益和精确率在Retailrocket数据集上平均提升1.9%和1.22%,在Amazon 5-Elect数据集上平均提升0.82%和1.05%,在ML100K数据集上平均提升0.36%和0.31%。     

基于层次注意力机制神经网络模型的虚假评论识别

针对虚假评论识别任务中传统离散模型难以捕捉到整个评论文本的全局语义信息的问题，提出了一种基于层次注意力机制的神经网络模型。首先，采用不同的神经网络模型对评论文本的篇章结构进行建模，探讨哪种神经网络模型能够获得最好的篇章表示；然后，基于用户视图和产品视图的两种注意力机制对评论文本进行建模，用户视图关注评论文本中用户的偏好，而产品视图关注评论文本中产品的特征；最后，将两个视图学习的评论表示拼接以作为预测虚假评论的最终表示。以准确率作为评估指标，在Yelp数据集上进行了实验。实验结果表明，所提出的层次注意力机制的神经网络模型表现最好，其准确率超出了传统离散模型和现有的神经网络基准模型1至4个百分点。