为了解决就医过程中医疗资源短缺和患者时间不充裕、行程不便的问题,提出了结合外部知识的基于记忆网络的知识感知医疗 对话生成模型(memory networks based knowledge-aware medical dialogue generation model,MKMed).该模型首先通过利用精确字匹配的方法在对话历史中进行实体追踪;随后在外部实体知识数据库里设计2阶段的实体预测,筛选出可能出现在回复中的医疗实体及对应知识,其中2阶段实体预测分别利用计算共现矩阵和余弦相似度的方法;模型接着用记忆网络来存储知识和对话历史的信息;最后整合记忆网络存储的信息,并使用注意力机制以及循环神经网络生成回复.在带有外部知识的大规模医疗对话数据集KaMed上进行了相关实验,该数据集为收集自在线平台的真实数据.实验结果表明提出的模型生成的回复在流畅性、多样性、正确性和专业性等方面均显著优于大部分基准模型.证明了合理引入外部知识的医疗对话模型能产生成更有医疗价值的回复.

The generation models such as Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN) have been demonstrated to be of high effectiveness in standard collaborative filtering applications. However, the conventional VAE can't capture the data distribution well when the data is sparse or the auxiliary information is added, resulting in low recommendation accuracy. In this paper, we propose a novel VAE-GAN-based collaborative filtering (CF) framework, named CF-VAE-GAN, to provide higher accuracy in recommendation. First, auxiliary information such as user comments and item multimedia features are added to VAE. And then, we use the discriminator of GAN to improve the reconstruction objective of VAE. Finally, we design two CF-VAE-GAN models of users and items, respectively. Empirical results indicate that our method outperforms state-of-the-art methods in terms of Recall and NDCG.

During a two-day strategic workshop in February 2018, 22 information retrieval researchers met to discuss the future challenges and opportunities within the field. The outcome is a list of potential research directions, project ideas, and challenges. This report describes themajor conclusionswe have obtained during the workshop. A key result is that we need to open our mind to embrace a broader IR field by rethink the definition of information, retrieval, user, system, and evaluation of IR. By providing detailed discussions on these topics, this report is expected to inspire our IR researchers in both academia and industry, and help the future growth of the IR research community.

Although people can use Web search engines to explore scenic spots for traveling , they often find it very difficult to discover the sighting sites which match their personalized need well .Tour recommendation systems can be used to solve the issue .A good tour recommendation system should be able to provide personalized recommendation and take the time and cost factors into account . Furthermore ,our investigation shows that often a user u will consider the distance between her∕his habitual residence and the tour destination when she∕he makes her∕his travel plan .It is because that the travel distance reflects the effect of time and cost indirectly .Therefore ,we propose a distance-aware tour recommendation model ,named GeoPMF (geographical probabilistic matrix factorization) , which is developed based on the Bayesian model and PMF (probabilistic matrix factorization) .The main idea of GeoPMF is that for each user we try to get a most preferred travel distance span by mining her past tour records .Then we use it as a kind of weight factors added into the traditional PMF model .Experiments on travel data of Ctrip show that ,our new method can decrease RMSE (root mean square error) nearly 10% compared with some baseline methods .And when compared with the traditional PMF model ,the average decline on RMSE is nearly 3 .5% in virtue of the distance factor .

Current binary sentiment classification has been focusing on improving the performance of classification, while the imbalance of sentiment data set in practical applications, which means the number of samples in one category is several folds of that of another category, is neglected. Most study on sentiment classification has been done on the balanced data, so these methods perform well on balanced data, while are unable to maintain the same performance on unbalanced data set. This paper proposed a method for unbalanced sentiment classification that combines unbalanced classification method and ensemble learning technique. Both algorithm and data set are considered to enhance the classification performance of imbalance sentiment data set. Under the framework of ensemble learning, this hybrid method integrates three different methods: under-sampling, bootstrap re-sampling and random feature selection to process the data set. Experiments on the unbalanced data set prove that this ensemble method can improve the classification performance of unbalanced sentiment data set.

当前,检索结果多样化作为一种提升用户满意度的有效方法已成为Web和数据库检索、文本摘要及推荐系统等领域的研究热点之一.但已有研究工作大都只考虑语义多样化策略.而实际上,多样化是一个非常复杂的优化问题,还需考虑许多其他的策略,如新颖性、质量、价值等.众所周知,Web是一个动态的信息空间,用户的查询需求也随时间不断演化,只有在一个特定的时间模式下,检索系统才能返回满意的结果.故该文提出一种新的结合语义和时效性两个维度的查询结果多样化方法.该文首先给出了多维度查询结果多样化框架的通用定义.然后,对于给定的查询,探讨了如何基于文档、词和查询频率来计算其时效性意图的概率分布.之后,提出一种新的针对时效性多样化的评价方法.最后,构建了针对多维度多样化问题的真实数据集,并通过实验证明该文提出的方法,不管是在传统的多样化评价指标上,还是在该文提出的时效性多样化指标上,性能都超过了当前主流的基准方法.

随着社会化媒体的兴起,信息资源的数量呈现爆炸式增长,如何在海量的信息中帮助用户发现有用的知识成为亟需解决的问题.社会化推荐方法作为一种有效的信息过滤技术,由于能够结合社会网络的特点,模拟现实社会中的推荐过程,在分析用户历史行为的基础上,主动向用户推荐满足他们兴趣和需求的信息,受到了研究者们的广泛关注.但目前已有的方法大都只从用户间社会关系的角度出发,仅认为相互信任的朋友间具有相似的兴趣爱好,而忽略了推荐对象间的关联关系对推荐结果产生的影响.针对以上存在的问题,文中从推荐对象间关联关系的角度出发,假设具有关联关系的推荐对象更容易受到同一用户的关注,并进而在已有的社会化推荐算法的基础上,提出了一种结合推荐对象间关联关系进行推荐的算法.算法使用共享的潜在特征空间对目标函数的求解过程进行约束,使其在考虑用户间社会关系的同时,也考虑到推荐对象间关联关系所起到的重要作用.实验结果表明,与主流的推荐算法相比,文中所提出的方法在分类准确率和评分误差等多种评价指标上都取得了更好的结果.

The formation of target wave in the network of Hodgkin-Huxley neuron with four variables is investigated by inputing a distribuled current. A stimulus current (I1) is input in to a local square area, and another stimulus current (I2) is input into the nodes of the network thus the distributed current is depicted. The development of target wave is measured by changing the coupling intensity, the size (number of the controlled neurons) of the local area into which current I1 is input, the gradient current (?I=I1?I2). It is found that higher gradient current (?I) is necessary to induce target wave when local area into which current I1 is input is smaller in size and the coupling intensity is higher in value. Finally, the potential mechanisms of the stimulus current and target wave formation are discussed in brief. Eextensive numerical results confirm that the developed target wave is robust to a certain channel noise.

Flickr groups are self-organized user communities revolving around a common interest and of massive popularity. In recent years, the explosive growth in the number of the group makes it difficult for Flickr users to find relevant groups they are really interested in. In this paper, we focus on automatically recommending groups to users. Many existing works have utilized only ternary relationship such as users-tags-groups to generate their recommendations. In our work, we show that ternary relationship is insufficient to provide accurate recommendations. Instead, we represent the quaternary relationship among users, tags, image clusters and groups as a 4-order tensor and further employ the Higher-Order Singular Value Decomposition to reduce the dimensionality of the 4-order tensor; accordingly the group recommendation problem is casted as a latent semantic analysis problem between users and groups. Experiments on the dataset crawled from Flickr and comparisons with the ternary relationship model demonstrate the effectiveness of the proposed approach. Copyright © 2013 Binary Information Press.

In this paper, we investigate the evolution and transition of collective electric activities of neurons in a ring network, induced by inhomogeneous distribution of ion channels. The local kinetics is measured by Morris-Lecar under voltage coupling type. In the numerical studies, the effect of inhomogeneous distribution of ion channels is simulated by changing the conductance in ion channels embedded in the membrane, and the potential mechanism is discussed. The effect of diversity of conductance between calcium and potassium ions on the activating of the adjacent neurons, and the dependence of developed travelling wave on the coupling intensity, are investigated in detail. The activating and waking up the nonexcitable or quiescent neurons with type I and type II excitability, are investigated, respectively. The numerical results confirm that the adjacent neurons are activated and the stable travelling wave is developed in the ring network of neurons when the conductance of calcium ions is increased beyond a certain threshold or the conductance of potassium ions is reduced below another threshold; while the propagation of the travelling wave could be slowed down or suppressed when the conductance of calcium ions is reduced or the conductance of potassium ion is increased. The development or emergence of travelling wave and propagation are greatly dependent on the increase of conductance of calcium ions and the decrease of potassium conductance.

In recent years, the bipartite graph model is becoming popular because of its simplicity and efficiency in recommender systems. However, this model may be ineffective due to the data sparsity and scalability problems. Clustering techniques are effective methods to alleviate these two problems. In this paper we propose a novel method denoted as Multi-US_BG which makes recommendation on the user-item bipartite graph using multiple user subgroups. This method first uses SVD to decompose the rating matrix to get user feature vectors, then utilizes a fuzzy c-means clustering algorithm to cluster users into multiple subgroups. Finally it integrates subgroups with the recommendation method on the user-item bipartite graph. Experimental results on MovieLens show that our method can improve the top-N recommendation performance in Precision, Recall and F1-measure in comparison with the pure recommendation method on the bipartite graph (Pure_BG). Copyright © 2013 Binary Information Press.

The development of spiral wave in a two-dimensional square array due to partial ion channel block (Potassium, Sodium) is investigated, the dynamics of the node is described by Hodgkin-Huxley neuron and these neurons are coupled with nearest neighbor connection. The parameter ratio x(Na) (and x(K)), which defines the ratio of working ion channel number of sodium (potassium) to the total ion channel number of sodium (and potassium), is used to measure the shift conductance induced by channel block. The distribution of statistical variable R in the two-parameter phase space (parameter ratio vs. poisoning area) is extensively calculated to mark the parameter region for transition of spiral wave induced by partial ion channel block, the area with smaller factors of synchronization R is associated the parameter region that spiral wave keeps alive and robust to the channel poisoning. Spiral wave keeps alive when the poisoned area (potassium or sodium) and degree of intoxication are small, distinct transition (death, several spiral waves coexist or multi-arm spiral wave emergence) occurs under moderate ratio x(Na) (and x(K)) when the size of blocked area exceeds certain thresholds. Breakup of spiral wave occurs and multi-arm of spiral waves are observed when the channel noise is considered.

Forcing currents are generated by using two types of constant currents to stimulate different areas of a medium. The external current is mapped into transmembrane current and stable gradient force is induced to drive the stable rotating spiral wave. The medium becomes homogeneous when the spiral wave is removed and the time series of sampled membrane potentials becomes stable value without considering noise. The scheme is confirmed to be effective in the presence of Gaussian white noise.

为了帮助用户高效地组织和检索图像资源,多数图像分享站点允许用户为图像添加标签.图像标签推荐系统旨在提供一组标签候选项来方便用户完成添加标签的过程.以往的图像标签推荐方法往往利用标签间的共现信息进行标签推荐.但是,由于忽略了图像的视觉内容信息和被推荐标签之间的多样性,以往方法的推荐结果常存在标签歧义和标签冗余的问题.为了解决上述问题,文中提出了一种新的图像标签推荐方法,该方法综合考虑了被推荐标签的相关性和多样性.首先,利用视觉语言模型,该方法分别计算标签与图像的相关性和标签之间的视觉距离.然后,基于上述计算,给出一个贪心搜索算法来找到能合理地平衡相关性和多样性的标签集合,将该集合作为最终的推荐.在Flickr数据集上的实验结果表明,该方法在准确率、主题覆盖率和F1测度上均优于目前的代表性方法.

While searching through the Internet, users are often confronted with a great number of results. Facing this situation, many efforts have been devoted to search result visualization. Thereinto, the tag cloud has been demonstrated as an effective method. Unlike the previous work, we propose a new tag cloud scheme named Visual Tag Cloud (VTC) which jointly provides visual and textual tags. Compared with the traditional tag cloud, VTC is created in two stages, tag selection and image selection. In the tag selection stage, we present a series of strategies for ranking and selecting tags. In the image selection stage, we employ the Affinity Propagation (AP) algorithm to select exemplar images for each tag selected in the first stage. Then we evaluate our tag selection algorithms under a user navigation model as well as general metrics. The result shows that VTC is able to specify and deliver users' query intentions more precisely as well as summarize and navigate query results efficiently. Copyright © 2013 Binary Information Press.

Complete synchronization could be reached between some chaotic and/or hyperchaotic systems under linear coupling.More generally,the conditional Lyapunov exponents are often calculated to confirm the stability of synchronization and reliability of linear controllers.In this paper,detailed proof and measurement of the reliability of linear controllers are given by constructing a Lyapunov function in the exponential form.It is confirmed that two hyperchaotic systems can reach complete synchronization when two linear controllers are imposed on the driven system unidirectionally and the unknown parameters in the driving systems are estimated completely.Finally,it gives the general guidance to reach complete synchronization under linear coupling for other chaotic and hyperchaotic systems with unknown parameters.

Document clustering plays an increasing significant role with the exponential growth of documents on the Web. Many approaches have been proposed to solve clustering problem and work well on static data sets. However, most of them create clusters only based on the text semantic distance between the elements without considering their temporal and spatial relationship. In addition, almost all existing methods aim to find all topic groups present in a news collection. But it is extremely difficult to coverage all topic groups exactly. In this paper, we propose a novel spatiotemporal clustering approach for Web news. We first utilize the temporal burst characteristic of topics to automatically and quickly predict an optimal number of groups to be clustered. Then, we extend the traditional distance measure to be a novel function which utilizes the temporal, spatial and content information of the documents. Finally, we outline an algorithmic on the basis of above two steps. We collect the real Web data set and demonstrate through a series of experiments that our method significantly outperforms the baseline clustering methods. © 2012 by Binary Information Press.

网络作弊检测是搜索引擎的重要挑战之一,该文提出基于遗传规划的集成学习方法 （简记为GPENL）来检测网络作弊。该方法首先通过欠抽样技术从原训练集中抽样得到t个不同的训练集;然后使用c个不同的分类算法对t个训练集进行训练得到t*c个基分类器;最后利用遗传规划得到t*c个基分类器的集成方式。新方法不仅将欠抽样技术和集成学习融合起来提高非平衡数据集的分类性能,还能方便地集成不同类型的基分类器。在WEBSPAM-UK2006数据集上所做的实验表明无论是同态集成还是异态集成,GPENL均能提高分类的性能,且异态集成比同态集成更加有效;GPENL比AdaBoost、Bagging、RandomForest、多数投票集成、EDKC算法和基于Prediction Spamicity的方法取得更高的F-度量值。

Josephson jtmction is an important superconducting device,values of the critical parameters within the Josephson junction model are mapped with different orders of magnitude.As a result,the unknown parameters with smaller order of magnitude can not be detected with high precision by using the original scheme of adaptive synchronization.Based on the Lyapunov stability theory,appropriate controller and parameter observers with gain coefficients are approached analytically,a scheme of scale conversion is proposed to estimate the unknown parameters with different orders of magnitude.By defining the statistical error function,the areas of synchronization and non-synchronization are demarcated.It is found that the area of synchronization becomes larger when two smaller unknown parameters are amplified simultaneously. In this way,precision of the identified results is increased greatly,and the results confirm that the scheme is successful to identify the unknown parameters with small order of magnitude.