We use cookies to give you the best experience possible. By continuing we’ll assume you’re on board with our cookie policy


Many real-world Web excavation undertakings need to detect subjectsinteractively, which means the users are likely to interfere the subject find and choice procedures by showing their penchants. In this paper, a new algorithm based on Latent Dirichlet Allocation ( LDA ) is proposed for synergistic subject development pattern sensing. To slake those subjects non interested or related, it allows the users to add supervised information by seting the posterior topic-word distributions at the terminal of each loop, which may act upon the illation procedure of the following loop. A model is designed to integrate different sorts of supervised information. Experiments are conducted both on English and Chinese principal and the consequences show that the extracted subjects gaining control meaningful subjects, and the synergistic information can assist to happen better subjects more expeditiously.

Keywords- Web excavation ; probabilistic subject theoretical accounts ; topic evolutionary forms

LDA-based Approach for Interactive Webmining TOPICS SPECIFICALLY FOR YOU

I. Introduction

The Web has become a really convenient and effectual communicating channel for people to portion their cognition, show their sentiment, promote their merchandises, or even educate each others, by printing textual informations through a Web browser interface. However, the Web users besides have to confront the defeat of information flood. For illustration, an active forum may bring forth more than one hundred posts/articles which cover over 10s of subjects in a individual twenty-four hours. Some of the hot posts/articles may be interested by 100s of users, and they join the treatment by answering to those original paperss and/or raising some more related subjects. These subjects so germinate over clip reiterating the above process and bring forth more subsequent textual informations.

Mining meaningful and utile information ( such as explainable subjects and subject development forms ) from textual information is believed to be helpful to get the better of information overflow. One of the excavation undertakings is to happen those interesting temporal forms in the text informations to see what themes they have and how these subjects vary along clip line. The typical applications may include papers mold, sentiment excavation, subject cleavage, information extraction and drumhead [ 1,2,3 ] .

Probabilistic subject theoretical account has gain much research involvements since the debut of probabilistic Latent Semantic Analysis ( pLSA ) , Latent Dirichlet Allocation ( LDA ) and other discrepancies. Basically, all these theoretical accounts are based upon the premises that paperss are mixtures of subjects, and each subject is a chance distribution over words. This household of subject theoretical accounts are calledproductive theoretical accountsbecause they specify probabilistic processs by which paperss can be generated. To bring forth a new papers, a distribution over subjects is chosen foremost. Then, to obtain each word in that papers, one chooses a subject at random harmonizing to this distribution, and draws a word from that subject.

It is non hard to happen that in many real-world subject sensing undertakings, the procedure of the subject sensing is frequentlysynergistic, which means the users are likely to interfere the ground procedure by showing their penchants. For illustration, the users may wish to execute multiple loops on the forum informations set to see what are most often discussed during the past two month. Or, narrow his/her involvements in the subsequent excavation procedure on several specific subjects found in the old tallies.

If we straight use these theoretical accounts ( e.g. the naA?ve LDA ) to mine the subject theoretical accounts and evolutionary forms, we frequently faced with the undermentioned troubles.

( 1 ) The algorithm such as naA?ve LDA is frequently run in afire-and-gomanner. However, the users are ever in demand of the ability topathcertain subject along clip line, to see how this subject evolves over clip.

( 2 ) the result of the LDA theoretical account is frequently non easy to construe, and farther attempts are required to deduce the concealed explainable significance behind those word-distributions of the subjects, and may publish new mining operations thenceforth.

In this paper, we proposed an algorithm and package model for synergistic subject development form sensing which is based on LDA. The proposed algorithm, iOLDA, is similar to LDA for its unsupervised subject constellating procedure ; but unlike LDA, it runs in an synergistic manner which allows the users to add supervised information by modifying the posterior distribution of each loop through a unvarying model, trusting to steer the subject sensing procedure of the following loop.

The remainder of the paper is organized as follows. Section 2 provides an overview on bing probabilistic subject mold and development sensing methods. Section 3 introduces an on-line version of LDA theoretical account. In Section 4 we describe in item our synergistic Latent Dirichlet Allocation attack. We give a simple extension to LDA theoretical account to enable the users to set the topic-word distributions interactively, which may act upon the illation procedure of the following loop by slaking those subjects non interested or related. Section 5 describes the experiments conducted on two different dataset to formalize the effectivity of the proposed methods. Section 6 concludes the paper and discusses future work.

II. Related Plants

Probabilistic subject mold is a new attack being successfully applied to detect and foretell the implicit in construction of textual informations. Many probabilistic subject theoretical accounts have been proposed these old ages, such as PLSI, LDA, CTM etc [ 2 ] . They are statistical productive theoretical accounts that relates paperss and words through latent variables which represent the subjects. In this paper, we follow the LDA algorithm as the basic probabilistic theoretical account. For a comprehensive study of these theoretical accounts, delight mention to [ 2 ] .

Topic development is the procedure of alterations in all signifiers of subject over clip, and subject development excavation is the survey of discoverying evolutionary subject. Similar to the construct of biological development, subjects can besides holdsubject traitsandsubject mutants. Informally, in the context of probabilistic theoretical account, by topic traits we mean the stableness or similarity of the topic-word distributions between back-to-back clip slides. By topic mutants we mean the divergency or difference between them.

Recently, there are some research works on the mold of subject evolutionary forms. Reference [ 4 ] is likely the most similar one to our work introduced in this paper. It presents a solution which is based on pre-discretized clip analysis that paperss are divided by clip into discretaized clip pieces, and for each piece, a separate subject theoretical account is developed. Our method is distinguished by its ability of incorporating the user ‘s counsel into an iterative procedure, and enable the user to detect and tracking subject evolutionary forms interactively. Other recent work has been concerned with temporal paperss utilizing a aggregation of incremental and efficient algorithms [ 5, 6 ] .

III. Modeling Development of Subjects

In this paper, we will concentrate on how to incorporate personal synergistic activities into the procedure of probabilistic subject sensing and trailing, in order to assist the Internet users to choose the interesting subjects of their ain. This can be divided into the undermentioned two inquiries: ( 1 ) how to capture the evolutionary form of the subjects ; and ( 2 ) how to utilize the user ‘s supervised information to steer the hunt or recommendation services for a more focussed hunt. In the following sub-section, we will discourse an on-line version of LDA to capture the temporal features of the subjects. In subdivision IV, we will turn to the synergistic LDA theoretical account which incorporates the user ‘s feedback as prejudice on distribution priors.

A.iOLDA Topic Model

To capture the evolutionary form of the subjects, we developed an synergistic on-line version of the Latent Dirichlet Allocation algorithm ( iOLDA ) . In this paper, we follow the notations of [ 4 ] . Figure 1 ( a ) illustrates the chief thought of iOLDA. The extension of the synergistic excavation to LDA will be discussed subsequently in Section IV.

Comparing to the inactive LDA, iOLDA takes a text watercourse as its input, instead than a whole inactive text principal, and happen the concealed subjects within that watercourse. For illustration, in the context of Internet forum, a text watercourse is formed by the stations or articles published in each individual different channels. Each watercourseSecondis partitioned intoT= { 0,1,2, … T, … }go uping clip slides, called watercourse sectionsSt, following a skiding window theoretical account. The Windowss size are fixed but the coarseness can change from a twenty-four hours, a hebdomad, to a month, depending on user picks and/or the mean volume of the channel. For simpleness ground, we do non see the intersection between different channels ( i.e. different watercourses ) in this paper.

Within each watercourse sectionSt, the paperss are exchangeable, i.e. , we do non see the publication clip order of the paperss within each watercourse section. This simplifies the procedure of detecting the subjects within the scope of the watercourse section. In this paper, we make usage of LDA algorithm for each watercourse section with a simple extension to the coevals of topic-word distribution prior, except the first loop, i.e. LDA0 in Figure 1 ( a ) , which is a naA?ve LDA procedure [ 2 ] .

The end product of iOLDA is really a sequence of LDA ‘s subject in go uping order of theThyminewatercourse sections. The subject theoretical account of clipTconsists of two parts which is similar to the end product of a inactive LDA: ( 1 ) the estimation of the topic-word distribution,foot; and ( 2 ) the estimation of the document-topic distribution,qt.

In order to gaugefootandqt, we use Gibbs sampling, a signifier of Markov concatenation Monte Carlo, proposed by Grif? Thursdaies and Steyvers in [ 1 ] , because it is easy to implement and is besides an efficient method for subject pull outing from big principal. The Gibbs sampling process considers each word token in the principal in bend and estimates the chance of delegating the current word item to each subject, conditioned on the subject assignments to all other word items. Because of the limited infinite, the estimations are given below. For more inside informations about Gibbs Sampling process, delight mention to [ 2 ] .

whereHundredweightandCDTare matrices of counts with dimensionsTungstentenThymineandCalciferoltenThymineseverally ;CwjWTcontains the figure of times wordtungstenis assigned to topicAJ, non including the current instanceAIandCdjDTcontains the figure of times topicJis assigned to some word token in papersvitamin D, non including the current caseI.Tungstenis the figure of distinguishable words andThymineis topic Numberss.

Besides the two distributions, the found subjects are besides correlated by semantic distances between each other along the clip line, which can be used to stand for the subject development forms of the current text watercourse ( e.g. forum channels ) . Here we use the Kullback-Leibler divergency ( or K-L distance, information addition, comparative information ) [ 7 ] to mensurate the subject traits and mutants, quantitatively.

For chance distributions P and Q of a distinct random variable, the K-L divergency of Q from P is defined to be

Here P and Q are the back-to-back topic-word distributions obtained from the LDAi, where i=0,1,2, … . Because the Kullback-Leibler divergency is non symmetric, we compute both DKL ( P||Q ) and DKL ( Q||P ) and use their norm as a step of the semantic distance between two back-to-back subjects.

With DKL, we can detect fresh subject by look intoing if the distance between a freshly found subject and its predecessor is larger than a threshold, where a subject mutant occurs. In this manner, the subjects found can be partitioned along clip line to meaningful subjects.

IV. Mining the Patterns Interatively with iOLDA

In this subdivision, we will discourse the inquiry of how to utilize the user ‘s supervised information to steer the hunt or recommendation services. An synergistic theoretical account, iOLDA, is developed to integrate the user ‘s penchant and/or feedback as prejudice on distribution priors of the following loop.

A.Different Interaction Policies

After the first interaction of iOLDA, the subjects generated by LDA0 represent the chief topical contents in S0. This consequence can be used to lend to the prior of the following loop of LDA1, because the paperss in the same text watercourse portion similar semantic background. From the statistical point of position, this anterior service as a prejudice on the following watercourse section, and carries the subject traits to the following loop. However, the following watercourse section may present new items, and the sizes of the two sections possibly excessively different, or the users may stipulate their penchants. Then, a mapdegree Fahrenheitis introduced for smoothing, standardization, and other intents. We have:

prior-ft=Adegree Fahrenheit( buttocks-ft-1) ( 3 )

The different ways to find the mapdegree Fahrenheitcan be used to carry through different interaction policies, which may includes:

1 )Corpus degree smoothing: A to take the principal degree influences into considerations. For illustration, to smooth the difference of the item set size of the back-to-back principal, we have:

prior-ft=A A buttocks-ft-1A wsmoothing ( 4 )

where wsmoothing can be a smoothing weight vector, such as a individual exponential smoothing weight as defined in ( 5 ) .

2 )Word degree filtering: to leverage word degree prior cognition to filtrate the influence of words, e.g. the common sense words. For illustration, we can utilize a TFIDF filter to extinguish the noise of common sense words, by diminish their anterior chances or put them to zero.

prior-ft=A A buttocks-ft-1wfiltering ( 6 )

where wfiltering is a filtering weight vector which set the nominal chance to zero, if their TFIDF weight is less than a predefined threshold.

3 )User degree penchant: to encode the user ‘s penchants as prejudice on the anterior distribution. For illustration, the user can explicitly sepecify an uninterested item set and put the corresponding anterior chance to zero.

All the different policies can be implemented under an unvarying model, which treat the alteration of the priors as different services which portion the same interface.

B.iOLDA Algorithm

All the different interaction policies as discussed in the old sub-section have been realized under a unvarying model of maps, as illustrated in Figure 1 ( B ) . The symbolplus( ‘+ ‘ ) indicates a wages map to a specific item ‘s prior of the topic-word distribution. The symbolsubtraction( ‘- ‘ ) indicates a punishment map to either a specific item ‘s anterior or to a subject as a whole. The reward/penalty maps can be designed to increase/decrease the corresponding item ‘s happening in the principal straight, or the item ‘s probabilistic values in the subject theoretical account indirectly. The symbolten( ‘X ‘ ) indicates other maps on the posterior distribution of the old loop.

Through this unvarying model, alterations to the anterior can be implemented easy and the most of import thing is that it besides makes the combination of different interaction policies really flexible.

So far, we can sum up the overview of the proposed synergistic online LDA algorithm as in Algorithm 1.

The algorithm takes as input the current text watercourse sectionSt, the subject figureK, the clip castT, the Dirichlet hyperparametersaandB, and the user interaction mapdegree Fahrenheit, severally. The end product of the algorithm contains the appraisals and the distances of subject distributions along each subject subject.

V. Experiments

For the experiment, we have prepared two datasets, an English version of the NIPS dataset [ 8 ] , and a Chinese version of TIANYA forum [ 9 ] dataset. NIPS dataset consists of the full text of the 14 old ages ( 1988-2001 ) of proceedings of the Neural Information Processing Systems ( NIPS ) Conferences, which is stored in 1,958 text files in 14 booklets ( one for a twelvemonth ) , and occupied about 36MB. Because this dataset portion a common background of machine acquisition and computational neuroscience, and besides used by many related plants for experiment or rating intents, it is chosen to demo iOLDA ‘s abilities to capture the evolutionary form of the subjects.

TIANYA dataset consists of the full text of Chinese articles for 5 month ( from August to December, 2008 ) , which is collected by a Web sycophant. The original information is more than 1.4GB. After the preprocessing of de-tagging and halt words, the trial information set is about 111MB, in 43,731 text files. Because the forum is really active, the subjects of this dataset covers a batch of spheres including political relations, economic system, instruction, amusement, literature, societal events, and others. All the experiments are run on a 2.4G Duo CPU laptop with 2GB memory. We used theMalletJava library [ 10 ] for the execution of the naA?ve LDA.

We besides tested the effectivity of the different interaction policies by synergistic excavation on the TIANYA dataset. Table III gives some consequences of proving the word degree filtrating mechanism. The left column ( in bright grey ) and the right column of each subject contains topic words before and after the word degree filtering, severally. Notice that words with lower TFIDF ( in dark grey ) weight are non in the top list.

iOLDA discovered many interesting subject evolutionary forms. From Table II, it is really easy to see that subject 5 revealsthe 2008 Chinese milk dirtand topic 7 revealsthe Beijing Olympic games. We noticed that word set in subject 5 changed really small, which is coherent with the fact that the milk dirt were hot subject along that period of clip. Another interesting thing is that from Octerbor, the word ( which meansfigure onein Chinese ) started to be on list. Further probe to the principal shows that the corresponding paperss are all about China ‘s gold mental figure werefigure onein Olympic game 2008.

Similarly, Table IV shows some consequences of proving the user degree penchant maps. The left column ( in bright grey ) , the right column and the mid column of each subject contains topic words of two back-to-back loops and the penchant keywords or maps. Symbol + and – are reward keyword set and punishment keyword set, severally. All the affected words are in dark grey ( For brevity ground, we merely display the top 13 words and omitted the remainder ) . The rewarded keywords are in white and the penalized words are in bold. By comparing the left and right columns of the same subject, we found that by encoding user penchant as a prejudice into the illation distribution prior, the subjects discovered from the principal can be closer to those contents that are interested by the users.

VI. Decisions

In this paper, we presented an algorithm and package model for synergistic subject development form sensing which is based on LDA. The proposed algorithm, iOLDA, is fundamentally a simple extension of LDA, but runs in an synergistic manner which allows the users to add supervised information by modifying the posterior distribution of each loop. We have developed three classs of interaction policies, viz. , corpus degree smoothing, word degree filtering, and user degree penchant, to leverage the heuristic and supervised information to steer the subject sensing procedure. We besides designed and realized an unvarying model for this semi-supervised illation procedure. Experiments utilizing both English text and Chinese Internet forum text are conducted to formalize the effectivity of iOLDA.

The proposed method is still a preliminary work towards integrating sphere cognition and user supervised information into probabilistic subject mold. It can be extended in many waies. First, how to automatically larn and adaptively adjust the parametric quantities of different interaction maps can be really of import for this method to be used in larger principal. We are besides interested in integrating a conceptual hierarchy into the interaction excavation procedure, so that the users can utilize sphere cognition at a larger coarseness. Although the public presentation is non a critical concern of this paper, we have to develop advanced indexing mechanism and parallel algorithms if we are traveling to utilize this method for larger principal.


The writers would wish to thank Dafu Shan, Yi Han and Zheng Liang for helpful treatments and informations readyings. We are besides like to thankMalletundertaking for an first-class execution of LDA algorithm. The work in this paper is supported by National Natural Science Foundation of China under the grant 60873204 and 60933005. It is besides partly supported by the National 242 plan under the grant 2009A90.


[ 1 ] T. Griffiths and M. Steyvers, “ Finding scientific subjects, ” In Proceedings of the National Academy of Sci-ences of the United States of America, 2004.

[ 2 ] M. Steyvers and T. Griffiths, “ Probabilistic subject theoretical accounts, ” In T. Landauer, D. Mcnamara, S. Dennis, and W. Kintsch, editors, Latent Semantic Analysis: A Road to Meaning. Lau-rence Erlbaum, 2005.

[ 3 ] J. Boyd-Graber, J. Chang, S. Gerrish, C. Wang, and D. Blei, “ Reading tea foliages: How worlds inter-pret subject theoretical accounts, ” In Neural Information ProcessingSystems ( NIPS2009 ) , 2009.L. AlSumait, D. Barbara, and C. Domeniconi, “ OnlineLDA: adaptative subject theoretical accounts for mining text streamswith applications to topic sensing and trailing. InICDM, 2008.

[ 4 ] Qiaozhu Mei and ChengXiang Zhai, “ Detecting evolutionary subject forms from text: an geographic expedition oftemporal text excavation ” , In SIGKDD, pages 198-207, New York, NY, USA, 2005. ACM.

[ 5 ] D. Blei and J. La erty. Dynamic subject theoretical accounts. InICML, 2006.

[ 6 ] T. Cover and J. Thomas, Elements of Information Theory, Wiley & A ; Sons, New York, 1991. Second edition, 2006.

[ 7 ] Available at the NIPS Online Repository. hypertext transfer protocol: //nips.djvuzone.org/txt.html

[ 8 ] Available at the TIANYA forum Website, hypertext transfer protocol: //www.tianya.cn/new/publicforum/ArticlesList.asp? strItem=free.

[ 9 ] McCallum, Andrew Kachites.A “ MALLET: A Machine Learning for Language Toolkit. ” hypertext transfer protocol: //mallet.cs.umass.edu. 2002.

Share this Post!

Send a Comment

Your email address will not be published.