Multi-feature hierarchical topic models for human behavior recognition

Human behavior recognition is one important task of image processing and surveillance system. One main challenge of human behavior recognition is how to effectively model behaviors on condition of unconstrained videos due to tremendous variations from camera motion,background clutter,object appearance and so on. In this paper,we propose two novel Multi-Feature Hierarchical Latent Dirichlet Allocation models for human behavior recognition by extending the bag-of-word topic models such as the Latent Dirichlet Allocation model and the Multi-Modal Latent Dirichlet Allocation model. The two proposed models with three hierarchies including low-level visual features,feature topics,and behavior topics can effectively fuse two different types of features including motion and static visual features,avoid detecting or tracking the motion objects,and improve the recognition performance even if the features are extracted with a great amount of noise. Finally,we adopt the variational EM algorithm to learn the parameters of these models. Experiments on the YouTube dataset demonstrate the effectiveness of our proposed models.     

Unsupervised Activity Perception in Crowded and Complicated Scenes Using Hierarchical Bayesian Models

We propose a novel unsupervised learning framework to model activities and interactions in crowded and complicated scenes. Hierarchical Bayesian models are used to connect three elements in visual surveillance: low-level visual features, simple "atomic" activities, and interactions. Atomic activities are modeled as distributions over low-level visual features, and multi-agent interactions are modeled as distributions over atomic activities. These models are learnt in an unsupervised way. Given a long video sequence, moving pixels are clustered into different atomic activities and short video clips are clustered into different interactions. In this paper, we propose three hierarchical Bayesian models, Latent Dirichlet Allocation (LDA) mixture model, Hierarchical Dirichlet Process (HDP) mixture model, and Dual Hierarchical Dirichlet Processes (Dual-HDP) model. They advance existing language models, such as LDA [1] and HDP [2]. Our data sets are challenging video sequences from crowded traffic scenes and train station scenes with many kinds of activities co-occurring. Without tracking and human labeling effort, our framework completes many challenging visual surveillance tasks of board interest such as: (1) discovering typical atomic activities and interactions; (2) segmenting long video sequences into different interactions; (3) segmenting motions into different activities; (4) detecting abnormality; and (5) supporting high-level queries on activities and interactions.     

Unsupervised Learning of Human Action Categories Using Spatial-Temporal Words

We present a novel unsupervised learning method for human action categories. A video sequence is represented as a collection of spatial-temporal words by extracting space-time interest points. The algorithm automatically learns the probability distributions of the spatial-temporal words and the intermediate topics corresponding to human action categories. This is achieved by using latent topic models such as the probabilistic Latent Semantic Analysis (pLSA) model and Latent Dirichlet Allocation (LDA). Our approach can handle noisy feature points arisen from dynamic background and moving cameras due to the application of the probabilistic models. Given a novel video sequence, the algorithm can categorize and localize the human action(s) contained in the video. We test our algorithm on three challenging datasets: the KTH human motion dataset, the Weizmann human action dataset, and a recent dataset of figure skating actions. Our results reflect the promise of such a simple approach. In addition, our algorithm can recognize and localize multiple actions in long and complex video sequences containing multiple motions.     

基于LDA特征选择的文本聚类

特征选择在文本聚类中起着至关重要的作用,将产生式模型Latent Dirichlet Allocation(LDA)引入基于K-means算法的文本聚类中,通过提取特征与隐含主题的关系进行特征选择。在第2届中文倾向性分析评测的语料上的实验结果表明,当选择2%的特征时,相对于单词贡献度(TC,Term Contribution)方法的纯度和F值分别提高了0.15和0.16,相对于LDA直接得到文本与主题的关系的实验结果的纯度和F值分别提高了0.14和0.13。     

集成FM的短视频喜好率预测模型

短视频喜好率预测往往面临着用户及广告的数量巨大且训练数据集高维、稀疏等问题，从而导致预测准确度下降。针对这些问题提出了基于LDA-GBDT-FM的短视频喜好率预测模型，该模型利用隐狄利克雷分配模型（LDA）对原始数据集基于主题分割，利用梯度提升决策树（GBDT）对不同主题的子训练集提取连续型特征的高影响力特征，将其与离散特征合并来训练因子分解机（FM）模型，最后有效组合子模型，进而预测短视频的喜好率。实验基于Bytedance公司的数据集，实验结果表明，提出的LDA-GBDT-FM模型相较于LDA-FM、FM和LR在预测指标上分别提高了3.0%、5.7%和8.5%。     

Latent topic model for audio retrieval

Latent topic model such as Latent Dirichlet Allocation (LDA) has been designed for text processing and has also demonstrated success in the task of audio related processing. The main idea behind LDA assumes that the words of each document arise from a mixture of topics, each of which is a multinomial distribution over the vocabulary. When applying the original LDA to process continuous data, the word-like unit need be first generated by vector quantization (VQ). This data discretization usually results in information loss. To overcome this shortage, this paper introduces a new topic model named Gaussian-LDA for audio retrieval. In the proposed model, we consider continuous emission probability, Gaussian instead of multinomial distribution. This new topic model skips the vector quantization and directly models each topic as a Gaussian distribution over audio features. It avoids discretization by this way and integrates the procedure of clustering. The experiments of audio retrieval demonstrate that Gaussian-LDA achieves better performance than other compared methods.     

基于LDA话题演化研究方法综述

现实生活中不断有新话题的产生和旧话题的衰减,同时话题的内容也会随着时间发生变化。自动探测话题随时间的演化越来越受到人们的关注。Latent Dirichlet Allocation模型是近年提出的概率话题模型,已经在话题演化领域得到较为广泛的应用。该文提出了话题演化的两个方面 内容演化和强度演化,总结了基于LDA话题模型的话题演化方法,根据引入时间的不同方式将目前的研究方法分为三类 将时间信息结合到LDA模型、对文本集合后离散和先离散方法。在详细叙述这三种方法的基础上,针对时间粒度、是否在线等多个特征进行了对比,并且简要描述了目前广泛应用的话题演化评测方法。文章最后分析了目前存在的挑战,并且对该研究方向进行了展望。     

融合主题与语言模型的蒙古文信息检索方法研究*

摘 要: 为了从日益丰富的蒙古文信息中快速准确地检索用户需求的主题信息,提出了一种融合主题模型LDA与语言模型的方法。该方法首先对蒙古文文本建立一元和二元语言模型,得到文本的语言概率分布;然后基于LDA建立主题模型,利用吉普斯抽样方法计算模型的参数,挖掘得到文档隐含的主题概率分布;最后,计算出文档主题分布与语言分布的线性组合概率分布,以此分布来计算文档主题与查询关键词之间的相似度,返回与查询关键词主题最相关的文档。语言模型充分利用蒙古文语法特征,而主题模型LDA又具有良好的潜在语义挖掘及主题发现的泛化学习能力,从而结合两种方法更好的实现蒙古文文档的主题语义检索,提高检索准确性。实验结果表明,融合LDA模型与语言模型的方法相比单一模型体现主题语义方面取得了较好的效果。     

MTRF:融合空间信息的主题模型

针对主题模型中词汇独立性和主题独立性假设忽略了视觉词汇间空间关系的问题,提出了一种融合了视觉词汇空间信息的主题模型,称为马尔可夫主题随机场(MTRF),并且提出了主题在图像处理中的表现形式为对象的组成部件。根据相邻视觉词汇以很大概率产生于同一主题的特点,该算法在产生主题的过程中,通过视觉词汇间是否产生于同一主题,来判断主题产生于马尔可夫随机场(MRF),还是产生于多项式分布。同时,从理论和实验两方面论证了主题并非对象的实例,而是以中层特征的形式表达对象的各个组成部件。与隐狄利克雷分配(LDA)相比,MTRF在Caltech101上的平均准确率提高了3.91%;在VOC2007数据集上的平均精度均值(mAP)提高了2.03%;此外,MTRF更准确地为视觉词汇分配了主题,能产生更有效表达对象的组成部件的中层特征。实验结果表明,MTRF有效地利用了空间信息,提高了模型的准确率。     

Discriminative sequential association latent dirichlet allocation for visual recognition

Graphical models have been employed in a wide variety of computer vision tasks. Assignments of latent variables in typical models usually suffer the confused explanation in sampling way. In this paper we present discriminative sequential association Latent Dirichlet Allocation, a novel statistical model for the task of visual recognition, and especially focus on the case of few training examples. By introducing the switching variables and formulating the direct discriminative analysis, the sequential associations are considered as priori to establish a relevance determination mechanism to obtain the reasonable assignments of latent variables and avoid the invalid labeling oscillations. We demonstrate the power of our model on two common-used datasets, and the experiment results show that our model can achieve better performances with efficient convergence and give well interpretations of specific topic assignments at the same time.     

MB-HL模型的微博主题挖掘研究*

为了解决传统的文本主题模型对微博主题挖掘准确率低及不考虑主题之间关联的问题,针对中文微博语料本身的特点,分析LDA和HMM模型优缺点,提出了微博主题挖掘模型MB-HL(Microblog-Hidden Markov Model Latent Dirichlet Allocation)。该模型用逐条微博作为处理单元,建立分布主题-词语矩阵并进行优化,通过LDA模型对微博用户不同的行为建模并提取特征,利用HMM模型强大的时序状态建模能力弥补LDA在主题相关性上的不足,采用Gibbs采样进行推理求解。在真实的新浪微博数据上对比实验表明MB-HL模型能提高近9%主题关键词的准确度,并能有效地发现主题之间的关联关系。     

社交网络用户发布模式和兴趣预测研究

数十亿用户通过在社交网络服务上发布照片和文本来分享他们的想法。他们对各种主题感兴趣,通常有不同的情感倾向和发布活动。提出了一个模型来表征社交网络用户的发布活动,以预测用户的兴趣。应用LDA来构建用户发帖的典型模式模型,以将用户的发贴行为表示为发帖模式的概率分布。从发布模式结果中提取出用户行为特征,并与从用户点赞的主页中提取的语言特征结合,构建兴趣预测模型。实验结果显示,使用从用户的发布行为中提取出的用户行为特征可以提高预测的准确性。     

基于狄利克雷多项分配模型的多源文本主题挖掘模型

随着文本数据来源渠道越来越丰富,面向多源文本数据进行主题挖掘已成为文本挖掘领域的研究重点。由于传统主题模型主要面向单源文本数据建模,直接应用于多源文本数据有较多的限制。针对该问题提出了基于狄利克雷多项分配（DMA）模型的多源文本主题挖掘模型——多源狄利克雷多项分配模型（MSDMA）。通过考虑主题在不同数据源的词分布的差异性,结合DMA模型的非参聚类性质,模型主要解决了如下三个问题：1）能够学习出同一个主题在不同数据源中特有的词分布形式;2）通过数据源之间共享主题空间和词项空间,使得数据源间可进行主题知识互补,提升对高噪声、低信息量的数据源的主题发现效果;3）能自主学习出每个数据源内的主题数量,不需要事先给定主题个数。最后通过在模拟数据集和真实数据集的实验结果表明,所提模型比传统主题模型能更有效地对多源数据进行主题信息挖掘。     

主题模型在短文本上的应用研究

针对短文本上以LDA为主的传统主题模型易受特征稀疏、噪声以及冗余影响的问题,首先梳理了文本特征表示法的变化以及短文本上主题模型的发展现状,并系统地总结了LDA模型和狄利克雷多项混合模型(DMM)各自的生成过程和相应的吉布斯采样参数推导。关于主题模型最优主题数,选取常见的4种优化指标进行了详细的对比说明。最后分析了近2年主题模型的扩展研究和其在网络舆情上的简单应用,并以此指明了未来主题模型的研究方向和侧重点。     

面向动态主题数的话题演化分析

话题演化用于自动分析话题变化趋势,具有较高的应用和研究价值。ILDA(Infinite Latent Dirichlet Allocation)模型在LDA(Latent Dirichlet Allocation)模型的基础上增加了狄利克雷过程,除了能获取隐变量,更重要的是能完成超参的动态更新和主题数的变动。而已有的话题演化研究中,话题的主题数需要事先指定且无法变动,基于ILDA模型的方法则可以针对性地解决该问题。构建的话题演化分析系统可实现如下功能:各周期内按不同主题分类、相邻周期间的主题进行关联、按时间顺序计算子话题强度。实验显示,基于ILDA模型的参数动态更新符合实际需求,话题演化分析过程完善可行。     

LDA模型参数有偏估计方法

LDA(Latent Dirichlet Allocation)等基于隐含topic的模型在离散数据处理中的应用逐渐增多.然而LDA使用Dirichlet分布作为隐含topic的分布函数,未能很好表示各topic之间相互关系.目前常见改进方法是通过DAG(Directed Acyclic Graph)图或对数正态分布等其他分布函数表达topic之间的关系.本文通过参数有偏估计的方法,考虑topic混合过程中词项上的重叠关系,改变topic内部词项分布,最终改进LDA模型性能.在回顾一些基础内容后,重点介绍参数有偏估计及简化计算方法.最后通过LDA模型在信息检索中的实验验证这种改进的有效性,并初步分析模型参数选用规律.     

Harnessing global expertise: A comparative study of expertise profiling methods for online communities

Building expertise profiles in global online communities is a critical step in leveraging the range of expertise available in the global knowledge economy. In this paper we introduce a three-stage framework that automatically generates expertise profiles of online community members. In the first two stages, document-topic relevance and user-document association are estimated for calculating users’ expertise levels on individual topics. We empirically compare two state-of-the-art information retrieval techniques, the vector space model and the language model, with a Latent Dirichlet Allocation (LDA) based model for computing document-topic relevance as well as the direct and indirect association models for computing user-document association. In the third stage we test whether a filtering strategy can improve the performance of expert profiling. Our experimental results using two real datasets provide useful insights on how to select the best models for profiling users’ expertise in online communities that can work across a range of global communities.     

基于LDA模型的文本分类研究

针对传统的降维算法在处理高维和大规模的文本分类时存在的局限性,提出了一种基于LDA模型的文本分类算法,在判别模型SVM框架中,应用LDA概率增长模型,对文档集进行主题建模,在文档集的隐含主题-文本矩阵上训练SVM,构造文本分类器。参数推理采用Gibbs抽样,将每个文本表示为固定隐含主题集上的概率分布。应用贝叶斯统计理论中的标准方法,确定最优主题数T。在语料库上进行的分类实验表明,与文本表示采用VSM结合SVM,LSI结合SVM相比,具有较好的分类效果。     

基于跨语言神经主题模型的汉越新闻话题发现方法

针对汉越跨语言新闻话题发现任务中汉越平行语料稀缺,训练高质量的双语词嵌入较为困难,而且新闻文本一般较长导致双语词嵌入的方法难以很好地表征文本的问题,提出一种基于跨语言神经主题模型（CL-NTM）的汉越新闻话题发现方法,利用新闻的主题信息对新闻文本进行表征,将双语语义对齐转化为双语主题对齐任务。首先,针对汉语和越南语分别训练基于变分自编码器的神经主题模型,从而得到单语的主题抽象表征;然后,利用小规模的平行语料将双语主题映射到同一语义空间;最后,使用K-means方法对双语主题表征进行聚类,从而发现新闻事件簇的话题。实验结果表明,所提方法相较于面向中英文的隐狄利克雷分配主题改进模型（ICE-LDA）在Macro-F1值与主题一致性上分别提升了4个百分点与7个百分点,可见所提方法可有效提升新闻话题的聚类效果与话题可解释性。     

Classification and prediction of whereabouts patterns from the Reality Mining dataset

Classification and prediction of users’ whereabouts patterns is important for many emerging ubiquitous computing applications. Latent Dirichlet Allocation (LDA) is a powerful mechanism to extract recurrent behaviors and high-level patterns (called topics) from mobility data in an unsupervised manner. One drawback of LDA is that it is difficult to give meaningful and usable labels to the extracted topics. We present a methodology to automatically classify the topic with meaningful labels so as to support their use in applications. We also present a topic prediction mechanism to infer user’s future whereabouts on the basis of the extracted topics. Both these two mechanisms are tested and evaluated using the Reality Mining dataset consisting of a large set of continuous data on human behavior.