针对以往流行度预测方法未利用演化模式之间的差异和忽略预测时效性的问题,提出了一种基于演化模式的推特话题流行度预测方法.首先,基于K?SC算法对大量历史话题的流行度序列进行聚类,并得到6类演化模式;然后,使用各类演化模式下的历史话题数据分别训练全连接网络(FCN)作为预测模型;最后,为选择待预测话题的预测模型,提出幅度对齐的动态时间规整(AADTW)算法来计算待预测话题的已知流行度序列与各演化模式的相似度,并选取相似度最高的演化模式的预测模型进行流行度预测.在根据已知前20 h的流行度预测后5 h的流行度的任务中,与差分整合移动平均自回归(ARIMA)方法以及使用单一的全连接网络进行预测的方法相比,所提方法的预测结果的平均绝对百分比误差(MAPE)分别降低了58.2％和31.0％.实验结果表明,基于演化模式得到的模型群相较于单一模型能更加准确地预测推特话题流行度.

To quickly identify bursty events that are emerging and developing in their early stages is important for our emergency response and public security. Daily news and social media are two major channels for people to contact with the world, thus become the main sources for the bursty event detection. However, recent works either use daily news only which is authoritative and well-organized, but easily out of date, or use social media only which is real-time and abundant, but contain a lot of noise. In this paper, to construct an efficient and effective bursty event detection system, we propose to combine the data from daily news and social media channels. Firstly, bursty features are extracted from social media and initially grouped into bursty events. Then, the data from two channels are aligned by using supervised learning at the level of events. Finally, we use the news to verify the detected result from social media by alignment algorithms. Experimental results show that our framework outperforms baselines.

本文提出一种基于突发词聚类的跨媒体突发事件检测方法。根据事件分析,发现微博具有文本丰富、用户活跃度高、在突发事件检测中具有速度快且高效的特点,但是由于微博文本长度较短,内容过于随意,使得事件发现的结果不够精确。新闻作为官方媒体,其真实性和权威性较高,内容比较规范,事件发现较为准确,但因为新闻数量较少,对于突发事件检测任务来说,时效性较低。现有的方法只针对一种媒体的数据进行挖掘,无法规避掉该媒体的数据所固有的缺点。本文提出一种方法,将微博和新闻2种媒体的数据进行融合,在满足突发事件检测的时效性的同时,提升了突发事件检测的准确率。

用户评论往往同时包含多个方面、多种情感,如何正确判断一条评论中不同方面的情感倾向性是方面情感分析的难点.文中提出基于词嵌入与记忆网络的方面情感分类.通过在记忆网络的不同模块引入方面词向量,加强方面词的语义信息,指导注意力机制捕捉方面相关的上下文信息,提升方面情感分类效果.在SemEval 2014任务4的短文本英文评论数据集和文中标注的长文本中文新闻数据集上实验表明,文中方法分类效果较好,在记忆网络框架下引入方面词嵌入信息是有效的.

随着互联网发展的日益深入，网络已成为当今社会信息传播的主要媒介之一。网络舆情形成迅速，对社会影响巨大，加强网络舆情的监测和分析，值得引起社会各界的高度重视。天玑舆情监测系统是专门针对网络舆情监测的工作要求和特点而打造的，不仅采用了专业化的搜索引擎技术，还融入了更加智能的数据挖掘技术，建立一个以日为周期的网络舆情监测平台，同时配上以周或者以月为基础的舆情分析报告，从而提供了一个便捷、科学、可操作的舆情工作平台。

Information retrieval (IR) is finding material (usually documents) of an unstructured nature (usually text) that satisfies an information need from within large collections (usually stored on computers) [1]. The most common and popular information retrieval application is web search engine such as Google[2], Baidu[3], Bing[4] and Sogou[5]. These application will return top-N best retrieval result to users. Information Retrieval systems for the Web, i.e., web search engines, are mainly devoted to finding relevant web documents in response to a user’s query[6]. Current retrieval systems performance well in phrase-level retrieval tasks which provide simple fact and entitycentric needs. Complex Answer Retrieval Track is a new track in 2017, which requests a more complex and longer retrieval result to answer a query. It focuses on developing systems that are capable of answering complex information needs by collating relevant information from an entire corpus. Given an article stub Q, retrieval for each of its sections Hi, a ranking of relevant entity-passage tuples (E, P). Tow tasks are offered: passage ranking and entity ranking. This paper introduces an algorithm and a system for passage ranking. The retrieval queries are outlines which consist of titles and section titles of articles. The retrieval collection consists of paragraphs which are come from Wikipedia articles. We use the BM25 algorithm and develop a system to retrieval the top-100 most relevant paragraphs.

We present a method for sentiment classification based on sentiment-specific word embedding (SSWE).Word embedding is the distributed vector representation of a word with fixed length in real topological space.Algorithms for learning word embedding,like word2vec,obtain this representation from large un-annotated corpus,without considering sentiment information.We make sentiment improvement for the initial word embedding and get the sentiment-specific word embedding that contains both syntactic and sentiment information.Then text representations are built based on sentiment-specific word embeddings.Sentiment polarities of texts are obtained through machine learning approaches.Experiments show that the presented algorithm performs better than sentiment classification method based on texts modeling by word,N-gram and word embeddings from word2vec.

近些年来,在线社交网络出现了爆炸式的增长.微博、脸谱网、推特等社交网站已经被越来越多的人熟知和使用.这些社交网站变得越来越流行,汇聚了全球数以亿计的用户.在社交网络中,用户可以根据自己的喜好,自由关注感兴趣的其他用户,组建自己的社交圈.同时,用户还可以轻松地发布消息以及转发其他用户的消息,极大地加速了消息的产生和传播过程.因此,对于社交网络的研究,主要集中于用户网络的结构特性研究和网络中的传播动力学研究两方面.

Recently, Deep Convolutional Neural Networks (CNNs) have been widely applied to sentiment analysis of short texts. Naturally, word embedding techniques are used to learn continuous word representations for constructing sentence matrix as input to CNN. As for sentiment analysis of customer reviews, we argue that it is problematic to learn a single representation for a word while ignoring sentiment information and the discussed aspects. In this poster, we propose a novel word embedding model to learn sentimental word embedding given specific aspects by modeling both sentiment and syntactic context under the specific aspects. We apply our method as input to CNN for sentiment analysis in multiple domains. Experiments show that the CNN based on the proposed model can consistently achieve superior performance compared to CNN based on traditional word embedding method.

In this paper, we consider load balancing and maintenance of distributed similarity search systems using locality sensitive hashing (LSH) in a structured peer-to-peer (P2P) network based on Distributed Hashing Table (DHT). LSH has been proven efficient in K-Nearest Neighbor (KNN) search in high dimensions. Recently, a number of schemes have been proposed to implement LSH over DHT-based P2P systems to process distributed similarity searches. We provide an efficient structure using virtual nodes to manage the multi-dimensional LSH bucket space in DHT peers and the maintenance algorithm, which improves load balancing in comparison with state-of-the-art techniques such as the virtual node algorithm. Here, we demonstrate effectiveness of the proposed method by experiments.

网络是目前最重要的信息传播渠道,其自由性和丰富性使得信息迅速传播。挖掘网络中的热点主题对政府政策的制定、企业经营决策的调整可以提供强有力的支持,并能够满足网民对热点主题的关注需求。主题数量的庞大使得主题热度值的计算尤为重要,该文分析热度的形成原因,基于因果模型并采用面板数据,给出一种较为客观可行的主题热度计算模型。该模型使用易于获取的数据进行计算,给出较为客观的热度度量,进而便于不同主题、不同日期间的热度对比。在此基础上,通过对热度变化规律的考察,提出一种基于多峰高斯曲线拟合热度变化进行主题热度预测的思路。

复杂网络的模式与演化分析具有重要的研究和应用价值,链路预测问题是其中一个研究热点.当前学者们提出了很多基于局部信息的相似性指标和链路预测算法,但是在应用于真实网络尤其是社交网络时,随着对网络宏观的演化模式与微观的链接生成机制的深入研究,链路预测算法的准确性仍有很大的提升空间.本文在现有算法基础上考虑网络同质性,分析局部结构内部的关联模式,将局部共同邻居集合根据全局最短路径信息进行建模,提出一种改进的链路预测方法:局部差异融合算法.该算法不仅保持了链接与节点之间的相似性的密切相关,而且反映了共同邻居集合内部的差异性.在各种真实网络数据集上的实验证实了本文提出算法的有效性.

排序是信息检索、数据挖掘以及社会网络分析的基础工作之一。在线社交网络和社会媒体的快速发展积累了大量的图数据——由表示实体的节点和表示实体间关系的连边构成。图数据中节点之间连接关系复杂,通常缺少显式的全序结构,使得图排序在图数据分析中显得尤为重要。图排序算法主要包括2大类,面向节点中心度的图排序算法和面向节点集合多样性的图排序算法。与传统的图排序不同,多样性图排序考虑排序和聚类的融合,体现为节点集合对网络整体的覆盖程度。近年来,多样性图排序得到了广泛的关注,取得了一系列研究进展,研究成果成功应用到了搜索结果排序、文档自动摘要、信息推荐系统和影响最大化等诸多场景中。文章评述了多样性图排序的研究现状及主要进展,将现有的多样性图排序方法按照研究思路的不同分为边际效益最大化、竞争随机游走、聚类与排序互增强3类,分别评述了每类方法的优势和不足。最后指出,设计有效的评价指标和标准测试集、克服多样性图排序面临的精度和速度的矛盾等是多样性图排序未来的研究重点。

Understanding the link patterns of networks is important for many tasks in link mining such as link prediction, friend recommendation, community detection and network evolution models. Recently, more and more models based on directed networks are present to formulate the underlying networks in various applications. However, most existing studies of link analysis are focused only in the undirected settings. In this paper, a theory of link formation named as Local Relative Position Theory (LRP) is proposed by analyzing the local structures to reveal the microscopic organizing principles of directed networks. Then a corresponding link prediction algorithm is proposed combing with existing microscopic mechanisms. Extensive experiments were applied on real world directed networks and the results testified the effectiveness of our link formation theory.

Large scale near-duplicate video detection is very desirable for web video processing, especially the computational efficiency is essential for practical applications. In this paper, we present a computationally efficient algorithm based on multi-layer video content analysis. Local features are extracted from key frames of videos and indexed by an novel adaptive locality sensitive hashing scheme. By learning several parameters, fast retrieval in the new hashing structure is performed without high dimensional distance computations and achieves better real-time retrieving performance compared with other state-of-the-art approaches. Then a descriptor filtering method and a two-level matching scheme is performed to generate a relevance score for detection. Experiments on near-duplicate video detection tasks including various transformed videos demonstrate the efficiency gains of the proposed algorithm.

P2P-VoD systems have gained tremendous popularity in recent years. Existing research is mostly based on the traditional theoretical assumptions. However, what is the performance of these assumptions is not known in the real world. It is particularly valuable to understand how the real-world system works. This paper presents a com-prehensive measurement and analysis on CoolFish, a real deployed P2P-VoD system, and proposes some statistical models different from the traditional observations on user behavior:the access pattern does not match Poisson distri-bution;session time does not have positive correlation with movie popularity;jump frequency does not have a negative correlation with movie popularity as assumed in previous studies. This paper analyzes the reasons for these results.

One challenge of link prediction in online social networks is the large scale of many such networks. The measures used by existing work lack a computational consideration in the large scale setting. We propose the notion of social distance in a multi-dimensional form to measure the closeness among a group of people in Microblogs. We proposed a fast hashing approach called Locality-sensitive Social Distance Hashing (LSDH), which works in an unsupervised setup and performs approximate near neighbor search without high-dimensional distance computation. Experiments were applied over a Twitter dataset and the preliminary results testified the effectiveness of LSDH in predicting the likelihood of future associations between people.

In order to achieve real-time and efficient access to the data of Twitter,two different methods based on Twitter List API and Lookup API were presented after analyzing the shortcomings of traditional collection methods.By classi-fying users,this method can precisely control the frequency of calling API.A series of experiments on over 260,000 users and over 6 million messages were carried out,and the results show that the combination of the two methods can be efficiently used to collect Twitter data in real-time.