ComEx Miner: Expert Mining in Virtual Communities

The utilization of Web 2.0 as a platform to comprehend the arduous task of expert identification is an upcoming trend. An open problem is to assess the level of expertise objectively in the web 2.0 communities formed. We propose the “ComEx Miner System” that realizes Expert Mining in Virtual Communities, as a solution for this by quantifying the degree of agreement between the sentiment of blog and respective comments received and finally ranking the blogs with the intention to mine the expert, the one with the highest rank score. In the proposed paradigm, it is the conformity & proximity of sentimental orientation of community member’s blog & comments received on it, which is used to rank the blogs and mine the expert on the basis of the blog ranks evaluated. The effectiveness of the paradigm is demonstrated giving a partial view of the phenomenon. The initial results show that it is a motivating technique.


INTRODUCTION
Expert identification is an intricate task because experts and their expertise are rare, expensive, unevenly disseminated, hard to qualify, continuously varying, unstable in level, and often culturally isolated and oversubscribed.The expert seekers behavior further complicates this, as they typically have improperly articulated requirements, are ignorant of expert's performance history, and are not well equipped to differentiate between a good and a bad expert.Web 2.0 [1] is an evolution from passive viewing of information to interactive creation of user generated data by the collaboration of users on the Web.The proliferation of Web-enabled devices, including desktops, laptops, tablets, and mobile phones, enables people to communicate, participate and collaborate with each other in various Web communities, viz., forums, social networks, blogs.Thus, evidently the Internet now forms the basis for the constitution of virtual communities.According to the definition of Howard Rheingold in [2], virtual communities are social aggregations that emerge from the Net when enough people carry on public discussions long enough, with sufficient human feeling, to form webs of personal relationships in cyberspace.
Thus, the expected alliance of these active areas of research, namely, Expert Identification & Web 2.0, fills the gaps that exist in the diversified Web.In response to the identified need to better exploit the knowledge capital accumulated on the Web 2.0 as a place for people to seek and share expertise, the operative challenge is to mine experts in virtual communities.Expert identification in virtual communities is noteworthy for the following reasons [3].Firstly, virtual communities are knowledge pools where members communicate, participate and collaborate to gather knowledge.Intuitively, we tend to have more confidence on an expert's text.Secondly, virtual communities allow interaction of novices with experts, which otherwise in real world is tedious and expensive.
Instigated by the challenge to find experts in the virtual communities, we propose a Community Expert Mining system called the ComEx Miner system, where, firstly we build an interest similarity group, an online community which is a virtual space where people who are interested in a specific topic gather and discuss in depth a variety of sub-topics related to the topic using blogs.We further propose to mine the sentiment of the each group member's blog along with the sentiment of their respective comments.This is based on the intuition that the blogger and the commenter talk about the same topic or product, treated as feature for opinion orientation identification and if the blog's sentiment about a topic/product matches with the commenter's sentiment about the topic/product this implies that blogger's knowledge about the topic/ product is acceptable as people agree to what has been talked about in the blog.This degree of acceptance matching would then help to rank the blog and mine the expert with highest blog rank.

The main components of the ComEx Miner are:
Interest Mining Module: This module puts forward an algorithm for Interest Group construction by uncovering shared interest relationships between people, based on their blog document entries.The key point of constructing this Collaborative Interest Group is the calculations of interest similarity relations and application of the K-means clustering technique to cluster researchers with similar interests into the same group.

 Expert Mining Module:
The ranking of member's blog within the built group is done on the basis of the score obtained by conjoining the blog & average comment orientation.This helps to identify the expert, the one with the highest ranking blog.The module is further divided into the following sub-modules: www.ijacsa.thesai.org  Sentiment Mining Module: The goal of this module is to perform sentiment analysis of the group member's blogs and the comments received on the respective blog.It gives the strength of the blogs and strength of their respective comments. Blog Ranking: Once the blog strength and comment strength has been determined; this module ranks the blogs by calculating the blog score, a metric which combines the respective blog's strength to the average comment strength.The blogs are then ranked as per the blog score and the expert is identified as the one with the highest rank.The paper is organized into 4 sections.Section 2 highlights the related and background work pertinent to the research carried.Section 3 illustrates the proposed ComEx Miner System expounding the methodology used to mine the expert from a virtual community, followed by section 4 which demonstrates the results and analysis of proposed paradigm with the help of sample data.Finally, the conclusion lists out the key contributions of the research work presented.

II. RELATED WORK
We seek guidance from people who are familiar with the choices we face, who have been helpful in the past, whose perspectives we value, or who are recognized experts [4].Expert finding addresses the task of identifying the right person with the appropriate skills and knowledge [5].There are various approaches related to expert identification & expertise search available in literature.Bogers et al. [6] used two methods: content based expert finding using academic papers and expert finding using social citation network between the documents and authors for finding experts.Breslin et al. [7] introduced a concept of re-using and linking of existing vocabularies in the semantic web, which can be used to link people based on their common interest.They described that a framework made by the combination of popular ontologies FOAF, SIOC, SKOS could allow one to locate an expert in a particular field of interest.Metze et al. [8] proposed a system to provide exchange of information by determining experts who can answer a given question.They provided a prototype expert finding system which enables individual within a large organization to search for an expert in certain area.Schall and Dustdar [9] addressed the problem of expertise mining based on performed interactions between people.Their approach comprised of two steps: Firstly, of offline analysis of human interaction considering tagged interaction links.Secondly, composition of ranking scores based on performance.Huh et al. [10] presented a grid enabled framework of expertise search (GREFES) engine, which uses online communities as sources for expert on various topics.They also suggested an open data structure SNML (Social Network Markup Language) for sharing community data.Smirnova and Balog [11] have argued that in real world, the notion of best expert depends on the individual performing the search.They proposed a user oriented model that incorporates user-dependent factor.It is based on the assumption that the user's preferences for an expert is balanced between the time needed to contact the expert and the knowledge value gained after .Li et al. [12] describe a method for finding expert through rules and taxonomies.They have proposed a combination of RDF FOAF facts and RuleML FOAF rules.
Punnarut and Sriharee [13] have introduced a method for finding expertise research using data mining and skill classification ontology.Zhang et al. [14] utilize an online community to find the people who may have expertise for answering a particular question.They analyze the experts by considering interactions of the people in questioning and answering the questions.Tang et al. [15] propose an expertise search system that analyses information from a web community.They use ontology to determine the correlation between information collected from different sources.
In the research presented in the paper, we intend to mine the experts in an online community which is a virtual space where people who are interested in a specific topic gather and discuss in depth a variety of sub-topics related to the topic using blogs.The conformity & proximity of sentimentally orientation of community member's blog & comments received is then used to rank the blogs and mine the expert on the basis of the blog ranks evaluated.The next section furnishes the details of the proposed paradigm.

III. THE PROPOSED COMEX MINER SYSTEM
In general, an expert is someone who possesses a high level of knowledge in a particular area.This entails that experts are reliable sources of relevant resources and information.An open problem thus arises to assess the level of expertise objectively.We propose the "ComEx Miner System" that realizes Expert mining in virtual communities, as a solution for this by quantifying the degree of agreement between the sentiment of blog and respective comments and finally ranking the blogs with the intention to mine the expert, the one with the highest rank score.
Figure 1 shows the architectural overview of ComEx Miner System proposed in this research.www.ijacsa.thesai.orgThe following sub-sections expound the details of the ComEx Miner:

A. Interest Mining Module
In this module, we focus on the problem of discovering people who have particular interests.The Interest Group construction algorithm is based on interest similarity, which can cluster researchers with similar interests into the same group and facilitate collaborative work.
The following sub-sections expound the details of the Collaborative Interest Group construction [4]: 1) Interest Vector: Each researcher writes blog entries according to his or her interest.The interest vector of the researcher, V i , is represented as a bag-of-words with frequently used words being assigned high weights.The interest vector is calculated by the equation described below: and; And where s ik means the strength of interest in word w k ; ef i (w k ) means the number of entries containing w k in researchers i's site; uf(w k ) means the number of researchers who use w k ; and N u means the number of researchers.www.ijacsa.thesai.org2) Interest Similarity Score: A similarity score represents how similar the interests of a pair of researchers are.If researcher i and j have similar interests, their interest vectors should be similar.Thus, we calculate the similarity score between them, R ij , using the cosine similarity of V i and V j as described below.
All elements of V i and V j are positive and thus the range of R ij is 0 to 1.

3) Collaborative
Interest Group Construction: Construction of an interest group is done to cluster the researchers with similar interests into the same group and facilitate collaborative work.Collaborative Interest Group Construction is done by using the technique of K-means clustering algorithm [16] where K is a user-specified parameter and it refers to the total number of clusters required.
Each point is then assigned to the closest centroid, and each collection of points assigned to a centroid is a cluster.The centroid of each cluster is then updated based on the points assigned to the cluster.We keep repeating this procedure again and again and update steps until no point changes clusters, or equivalently, until the centroids remain the same.

a) Finding total number of clusters, denoted by K:
The value of K is found out by first forming the researcher groups.Total number of researcher groups formed is equal to the total number of researchers and researchers belonging to a particular group can carry out the co-operative work among themselves.Each group will have its respective threshold value which will decide the membership of a particular researcher in that group.T i denotes the threshold for group i and is found out by averaging all the similarity scores corresponding to researcher i.
Membership criteria: group other some to belongs j researcher else, Now, once all the researcher groups have been formed, then the value of K is equivalent to the minimum number of groups required to cover all the data points.

b)
Assigning Points to the Closest Centroid: To assign a point to the closest centroid, we need a proximity measure that quantifies the notion of 'closest' for the specific data under consideration.We use the proximity measure as the distance between any two researchers, denoted by d ij and is given as (5) where d ij denotes the distance between researchers i and j R ij denotes the similarity score between researchers i and j.

c)
Centroids and Objective Functions: The next step is to re-compute the centroid of each cluster, since the centroid can vary, depending on the proximity measure for the data and the goal of clustering.
Once the virtual collaborative interest similarity group is put together, the next step is to identify the expert from this group.To realize this task, the sentiment of each group member's blog along with the sentiment of their respective comments is analyzed for opinion strengths.As mentioned previously the degree of acceptance matching would then help to rank the blog and mine the expert with highest blog rank.

B. Expert Mining Module
The expert mining module is divided into three submodules; namely, the Data Repository module which collects the web pages from the member's blog & comments, cleans them and then stores them in the repository, the Sentiment Mining Engine that receives these cleaned web pages from the repository and then provides orientation strengths of blogs & respective comments by extracting opinion features and opinion words and the Blog ranking module which finally We calculate the interest similarity score R ij between researchers i and j using the cosine similarity of V i and V j

Collaborative Interest Group Construction
We construct the collaborative interest group by using the technique of K-means clustering algorithm.It consists of two basic steps as follows:  We find the total number of clusters, denoted by K with the help of researcher groups so formed. And then we assign points to the closest centroid by taking the proximity measure as the distance between two researchers.
Basic K-means algorithm 1: Select K points as initial centroids.2: repeat 3: Form K clusters by assigning each point to its closest centroid.

4:
Re-compute the centroid of each cluster.5: until centroids do not change.www.ijacsa.thesai.orgranks the blogs on the basis of combined orientation strength of blog & comments to mine the expert as the one with the highest ranking blog.
The details of each of these sub-modules are given in the sections below.
1) Data Repository: This sub-module deals with collecting the web pages and storing them in the repository.Firstly the web crawler periodically crawls the member's blog and respective comments to collect them as web pages.Thereafter, these pages are cleaned up to remove the HTML tags and then are organized properly to be stored in the " Repository".
2) The Sentiment Mining Module: This sub-module deals with providing the actual orientation strenghts of both member's blog & comments received on it.The Sentiment Mining Module receives the web pages from the repository, i.e., if there are k members in a group then k blogs and their respective n comments will be processed to finally calculate the opinion strengths using the following three steps:a) Feature Extraction: This is most basic and crucial step for providing orientation strength by identifying those features that the bloggers & commenters have expressed their opinion on.Such features are known as Opinion Features.We make use of both the data mining and NLP Techniques to perform the task of feature extraction.We extract the opinion features with the help of POS Tagging and Preprocessing techniques.

 POS Tagging (Part of Speech Tagging)
POS Tagging is done to find out the features of the product that have been written about.As we know, features are usually noun or noun phrases in the review sentences.Therefore, we use NL Processor linguistic Parser [17] to parse each text, to split texts into sentences and to produce POS Tag for each word (whether the word is a noun, verb, adjective etc.) NL Processor generates XML output and deals only with explicit features, which are the features that occur explicitly as nouns or noun phrases.Each sentence is then saved in a Database along with the POS Tag information of each word in the sentence.

 Pre-Processing
In this sub-step, a transaction file is created which consists of pre-processed noun/noun-phrases of the sentences in the database.Here pre-processing includes the deletion of stop words, stemming and fuzzy matching.

b)
Opinion Direction Identification: In this step, we find out the opinion direction using the opinion features extracted in the previous step.To find the opinion direction, we will first extract the opinion words in the text and then find out their orientation strengths.It includes the following substeps:

 Opinion Words Extraction
In this sub-step, we extract the opinion words from the text given by the member's in their respective blog & by the commenter's in their comments on that blog.Opinion words are the words that people use to express their opinion (either positive, negative or neutral) on the features extracted in the previous steps.In our work, we are considering the opinion words as the combination of the adjectives along with their adverbs.We have called them collectively as an Adjective-Group (AG).Although, we can compute the sentiment of a certain texts based on the semantic orientation of the adjectives, but including adverbs is imperative.This is primarily because there are some adverbs in linguistics (such as "not") which are very essential to be taken into consideration as they would completely change the meaning of the adjective which may otherwise have conveyed a positive or a negative orientation.For example; One user says, "This is a good book" and; Other says, "This is not a good book" Here, if we had not considered the adverb "not", then both the sentences would have given positive review.On the contrary, first sentence gives the positive review and the second sentence gives the negative review.Further, the strength of the sentiment cannot be measured by merely considering adjectives alone as the opinion words.In other words, an adjective cannot alone convey the intensity of the sentiment with respect to the document in question.Therefore, we take into consideration the adverb strength which modify the adjective; in turn modifying the sentiment strength.

 Opinion Words Orientation
In this sub-step, we find out the orientation strength of the opinion word.As our opinion word consists of adjective + adverb, therefore to find out the orientation of the opinion word, we first find out the polarity of the adjective in the opinion word and then identify the strength of its corresponding adverb in the opinion word which modifies the adjective.Finally, the product of the adjective polarity and the adverb strength gives us the strength (orientation) of the opinion word.The details for finding adjective polarity, calculating adverb strength and deducing the final opinion word strength are as follows: a) Adjective Polarity Here, we will identify the semantic orientation for each of the adjective.As we know, words that have a desirable state (e.g.good, great) have a positive orientation, while words that have an undesirable state (e.g.bad, nasty) have a negative orientation.In general, adjectives share the same orientations as their synonym and opposite orientations as their antonyms.Using this idea, we propose a simple and effective method by making use of the adjective synonym set & antonym set in WordNet [18] to predict the semantic orientation of adjectives.Thus, our method is to use a set of seed adjectives whose orientations we know, & then grow this set by searching in the WordNet.The complete procedure for predicting adjective polarity is given below: Procedure "determine_ polarity" takes the target adjective whose orientation needs to be determined and the adjective seed list as the inputs.

Note:
1) For those adjectives that Word Net cannot recognize, they are discarded as they may not be valid words.2) For those that we cannot find orientations, they will also be removed from the opinion words list and the user will be notified for attention.3) If the user feels that the word is an opinion word and knows its sentiment, he/she can update the seed list.4) For the case that the synonyms/antonyms of an adjective have different known semantic orientations, we use the first found orientation as the orientation for the given adjective.b) Adverb Strength We collect all the adverbs which are used to modify the adjectives from English lexicon.Based on the different emotional intensity expressed by the adverb, we mark the negative adverbs with a negative score and other positive adverbs with different score in different sentiment level.The score is ranging from -1 to +1 and a higher score expresses a stronger sentiment.For example, we consider that the adverb "extremely" has higher strength than "more" does, but lower than that of "most".Consequently, "most" is marked with 0.9, "extremely" with +0.7, and "more" with +0.3.Negative adverbs, such as "not", "never", "hardly", "seldom", are marked with a negative score accordingly.

c) Opinion Word Strength
It is calculated by the product of adjective polarity i.e.P(adj i ) and the adverb strength i.e.S(adv i ) and is given by the following formula: ) S(adv ) P(adj ) S(OW i i i    (6) where, S(OW i ) represents the sentiment of i th opinion word , P(adj i ) represents the polarity of i th adjective and S(adv i ) represents the strength of i th adverb.The value of P(adj i ) is either -1 or +1 and the value of S(adv i ) ranges from -1 to +1.Therefore, the strength of each opinion word i.e., S(OW i ) will also lie in the range of -1 to +1.

Note:
Sometimes, there is no adverb in the opinion word, so the S(adv) is set as a default value 0.5 .When there is no adjective in the opinion word, then the P(adj) is set as +1.

d)
Blog & Comment Orientation Strength: After extracting all the opinion words from the blog and finding their respective strength, the overall strength of a Blog B is calculated by averaging the strength of opinion words as shown below: For each sentence in the review database If (it contains a product feature, extract all the Adjective-Group i.e. adjectives and their adverbs as opinion words) For each feature in the sentence The nearby adjective and adverb is recorded as its effective opinion (which modifies the noun / noun phrase which is a product feature)   [3] where; |OW(B)| denotes the size of the set of opinion words extracted from the blog and S(OW i ) denotes the sentiment strength of i th opinion word.As the overall strength of the blog is calculated by averaging the strength of the opinion words, therefore the strength of the review i.e.S(B) will also lie in the range of -1 to +1; where, S(B) = -1 indicates a strong negative opinion, S(B) = +1 indicates a strong positive opinion and S(B) = 0 indicates a neutral opinion.
Similar to blog orientation, comment orientation, S (C), of each comment received on a particular blog is determined.Once the orientation of every comment is known, the average comment orientation, Avg.S (C), is calculated (dividing the total comment orientation by the no. of comments).
3) Blog Ranking Module: This module takes as input the blog orientation strength and the average comment orientation strength for each member to compute the blog score.
Blog Score = Blog Orientation Strength* Avg.Comment Orientation Strength The member's blogs are are then ranked on the basis of this computed blog score.Finally, the expert is identified as the one with the highest ranking blog score.

IV. ILLUSTRATION
To clearly illustrate the use and effectiveness of the proposed system, a case study is presented to describe a typical scenario and examine the result of each module of the approach.

A. Interest Mining Module:
To demonstrate the Interest mining module we directly take the sample data calculations of interest vector and interest similarity from [3], where there are 5 researchers viz.i, j, k, n & m.Therefore, N u = 5 and there are 5 entries in each of the researcher's blog site.The following table I shows the blog entries of each of the Researcher i, j, k, n & m.
The key point of constructing this Collaborative Interest Group is the calculations of interest similarity relations and application of the K-means clustering technique to cluster researchers with similar interests into the same group.As all the elements of both the vectors taken at a time to calculate the similarity score are positive, thus the range of similarity score is between0 to 1.

This indicates that:
The value of 1 means that the 2 researchers have exactly similar interests and; The value of 0 means that the 2 researchers do not have any similar interests at all.Therefore, we can say that: The researchers n & m have almost similar interests (as R nm = 0.9169, approx 1 ) The researchers k & n have similar interests to a very great extent (as R kn = 0.8786) The researchers "k & m" and "i & m" have quite a lot similar interests (as R km = 0.8140 and R im = 0.8064)  The researchers "j & k" and "j & n" have quite less similar interests (as R jk = 0.6688 and R jn = 0.6132) a) Collaborative Interest Group Construction We construct the collaborative interest group by using the technique of K-means clustering algorithm with the help of two basic steps.We first construct the researcher groups by finding the membership of each of the researcher using the formula defined in equation ( 4).This step would give us the total number of clusters required, denoted by K.And then we assign points to the closest centroid by taking the proximity measure as the distance between two researchers using the formula defined in equation (5).

1) Membership for group i
Step 1: Calculate the threshold for this group i.e.T i = [1 + 0.7063 + 0.7110 + 0.7502 + 0.8064] = 0. 79478 Step 2: Deciding the members for group i As we can see, R ii >T i and R im >T i , therefore Researcher i and Researcher m belong to group i.We find Membership for group j, group k, group n, group m in a similar way and the following Researcher Groups are formed with their respective members:

Group i Group j Group k
Group n Group m CONSTRUCTION OF CLUSTERS

1) Total number of clusters
Now as we know total number of clusters i.e.K is equivalent to the minimum number of groups required to cover all the data points.Therefore, K=3.In other words, we can say that there are total three number of clusters required with the centroid as i, j, and n respectively.1 st cluster 2 nd cluster 3 rd cluster

2) Assigning points to the closest Centroid
In this step we assign points (researcher m and k) to the closest centroid by taking the proximity measure as the distance between two researchers.Therefore using the formula defined in equation ( 5), we calculate the distance of these two researchers with each of the above researchers:

B. Expert Mining module:
As described in section III, the expert mining module is divided into three sub-modules, namely the data respository; sentiment mining & blog ranking modules, here we demonstrate them & examine their effectiveness.We consider a group with 4 members and analyze their blogs & cooments received on them.Our final task is to determine the expert from this group of 4 members.

Comments:
1) Yes there's very little boot space.
3) There are only 3 colours available.4) I think the ride is quite good.5) Instability at high speeds is a major drawback.6) According to me, the seats are very comfortable.7) Ground clearance is a very common issue in India!The Sentiment mining module works in the following manner: www.ijacsa.thesai.orgWe manually mark the strengths of a few frequently used adverbs with values ranging from -1 to +1 based on our intuitions.We consider the most frequently used adverbs (for our illustration) along with their strength as below in table III:- Opinion Strength Calculations: The strength of each opinion word is given by the formula defined in equation (7) Opinion Words (for blog): 1. boring -1 * +0.5 = -0.1) Jerky drive.
3) Its an okay okay buy.  1) Yes, very low maintenance required.
2) The ride is not very smooth.
3) Affordable price .Just right for middle class families.4) Streamlined shape.5) Cooling is not good especially for Delhi summers.6) Its a good buy.Opinion Words (for Blog): 1. never big 1* -0.9= -0.9 2. pretty reasonable 1 *-0.3=-0.-0.0005 4 Thus, comparing all the blog strengths, according to our approach, the highest blog score is for blog 3 and therefore the Expert is blogger 3!

Limitations:
1) It covers comments only written in English.
2) No abbreviations or acronyms can be accounted for.
3) It does not cover interrogative sentences.4) A negative adjective and a negative adverb convert into a positive opinion word.5) A positive adjective and a negative adverb also convert into a negative opinion word.6) The method has no way of detecting and dealing with emoticons.

V. CONCLUSION
We proposed a novel ComEx Miner System for mining experts in virtual communities.This work is exploratory in nature and the prototype evaluated is a preliminary prototype.The major contributions of this research are: The size of this car is never big and this makes its price pretty reasonable and affordable.It does not demand any maintenance and its performance and safety are also amazing.Not much of car service is required.The cooling is very effective and this car is not very smooth on hilly terrains.www.ijacsa.thesai.orgi. Constructing a collaborative interest group known as the virtual community which will cluster researchers with similar interests in a same group and thereby facilitate collaborative work.
ii. Accessing the expertise from the virtual community using sentiment analysis of each group member's blog & comments received on it.Their combined orientation strength determined the blog score which enabled to rank the blogs and identify the expert as the one with the highest blog rank.The practice result proves that this algorithm has the characteristics of highly effective group arranging and identifying expert.This study is just one step in this direction.Due to the complex nature of framework, it is impossible to consider and incorporate all the factors that could have an impact on the effectiveness and efficiency of this system. .More research needs to be done in order to validate or invalidate these findings, using larger samples.

7 )
Better cars are available in the market.Now we calculate the S (B 4 ) & Avg.S (C 4 ) to compute blog score 4

Input :
Member's blog and comments on each blogOutput: Expert, one with the highest ranking blog Rank the blogs by their Blog Score .www.ijacsa.thesai.orgAdverb strength helps in assessing whether a document gives a perfect positive opinion, strong positive opinion, a slight positive opinion or a less positive opinion.

TABLE I .
SAMPLE BLOG ENTRIES OF 5 RESEARCHERS

TABLE III .
ADVERB STRENGTHS

TABLE IV .
BLOG RANKING