COMMTRUST: A MULTI-DIMENSIONAL TRUST MODEL FOR E-COMMERCE APPLICATIONS

International Journal on Natural Language Computing (IJNLC) Vol. 5, No.5, October 2016
DOI: 10.5121/ijnlc.2016.5504 43
COMMTRUST: A MULTI-DIMENSIONAL TRUST
MODEL FOR E-COMMERCE APPLICATIONS
M. Divya1
, Y. Sagar2
1
M.Tech (Software Engineering) Student, VNR Vignana Jyothi Institute of Engineering
& Technology, Hyderabad, Telangana, India, 500090.
2
Associate professor, CSE, VNR Vignana Jyothi Institute of Engineering & Technology,
Hyderabad, Telangana, India, 500090.
ABSTRACT
E-Commerce applications use reputation-based trust models based on the feedback comments and ratings
gathered. The “all better Reputation” problem for the sellers has become very huge because a buyer
facing problem to choose truthful sellers. This paper proposes a new model “CommTrust” to valuate trust
by mining feedback comments that uses buyer comments to calculate reputation scores using multi-
dimensional trust model. An algorithm is proposed to mine feedback comments for dimension weights,
ratings, which combine methods of topic modeling, natural language processing and opinion mining.
This model has been experimenting with the dataset which includes various user level feedback comments
that are obtained on various products. It also finds various multi-dimensional features and their ratings
using Gibbs-sampling that generates various categories for feedback and assigns trust score for each
dimension under each product level.
KEYWORDS
E-Commerce, Feedback mining, Trust score, Topic modeling, Reputation-based trust score
1. INTRODUCTION
Accurate trust evaluation plays a vital role in e-commerce systems. A reputation system [2]
is
implemented to get superior service deals in e-commerce systems like Amazon, eBay etc. To
allocate rank for sellers, feedback ratings are calculated which are given by the buyers. The “all
better reputation” [2]
problem is an issue for all sellers where the feedback rating is above 99%
positive on average [2]
. Such strong positive bias is not helpful to the buyers to select a right seller
or product. At the Amazon, the system uses detailed seller ratings for sellers (DSRs) on four
conditions, i.e. item, shipping, communication and cost. In DSRs we find strong positive bias
even though there is a little problem with product or delivery. The One potential negative rating is
the chance for the absence of negative ratings at electronic commerce websites, it attracts the
buyer who provides the negative feedback about the items and it harms their own reputation [2]
in
purchasing sites.
Buyers express some disappointment and negative opinions about the product in feedback
comments. Example: a buyer may have liked an item, packaging, and the overall transaction, but
the delivery would have been postponed. For this situation, the buyer may tend to score more 4
on a 5-star scale and comment on the postponement in the content field. In order to overcome the
above-mentioned “all better reputation” [2]
problem, a Comment based multi-dimensional
(CommTrust) is proposed for the trust valuation model accomplished by mining e-commerce
comments. CommTrust, trust profile is calculated for a seller that incorporates dimension
reputation scores, weights and overall trust scores. Thus, trust profiles for sellers are made by
mining feedback comments.

44
In CommTrust, access that unites dependency relation analysis [3, 4]
and lexicon based opinion
mining techniques are proposed to extract feature opinion expressions from feedback comments.
Furthermore, based on dependency relation analysis and Latent Dirichlet Allocation (LDA) topic
modeling methods [5, 18]
an algorithm is proposed to cluster feature expressions into the
dimensions and calculate total dimension weights and ratings, called Lexical-LDA [5]
. Therefore,
the reputation profiles in CommTrust contain dimension reputation scores, weights and complete
trust scores for ranking sellers.
2. RELATED WORK
The work focus on three major areas: 1) Computing approach to trust, mainly reputation based
trust valuation; 2) Analyzing feedback comment in e-commerce application and usually mining
opinions on product analysis and another form of free text documents; and 3) opinion mining and
summarization.
2.1. Computing Trust Valuation
The positive trust score aspect of the Amazon reputation system is well documented [1, 6]
. No valid
solutions have been reported. As proposed in [6]
, to observe feedback comments to get seller
reputation score below the balanced ratio, where feedback comments do not create the positive
rating which allows negative rating for a transaction. Complete trust scores for seller rating on
transactions farther aggregated. In this, our focus is on extracting dimensions from buyer
feedback comments and these dimension ratings are calculated to find a trust score for
dimensions.
2.2. Analying Feedback Commenets
In [13]
the e-commerce application, there have been different learning’s on analysis feedback
comments, even though an inclusive trust valuation is not their focus. The focal point is on the
sentiment classification [7, 20]
of feedback comments. It concludes that feedback comments are
audible by evaluating them as a trail. Omitted conditions for comments are assumed negative,
these methods are made from an aspect rating [15, 16]
are used to allocate feedback comments may
be positive or negative. The approach enhanced to encapsulate feedback comments. It aims to sort
out the considerate comments that do not present in actual feedback. It aims at developing “rated
aspect summary” [8]
given by Amazon feedback comments. The numerical developing model is
based on regression about a complete rating.
2.3. Opinion Mining and Summarization
The main performance is relevant to opinion mining and sentiment analysis [9, 10]
on free text
documents. Aspect Opinion mining on item review is the existing work. In a product description
and the opinion towards them are extracted. By choosing and re-constructing sentences
according to the extracted characteristics are summarized. Feedback mining and summarization is
the mission of generating sentiment summary [17]
that consists of sentences from feedback which
arrest the buyer’s opinion. Feedback summarization is interested in features or aspects on which
customers have opinions. It also concludes either the opinions are positive or negative. This
creates it from classic content summarization. A comprehensive overview is presented. Most
existing works on survey mining and summarization [11, 19]
concentrate on item reviews. For
example, [14]
concentrated on mining and summarizing ratings by extracting opinion sentences
with regard to the product features.

45
3. COMMTRUST: MULTI-DIMENSIONAL TRUST VALUATION FOR
COMMENTS
In electronic commerce application, feedback comments are the source in which users state their
opinions about the product honestly in a text box. Feedback comment analysis is done on the
various e-commerce sites announce that, even if the buyer gives positive comment he/she still
gives comments of mixed opinions about the product. For example, the buyer leaves a comment
as “Worst response, will not purchase again”. So the buyer has a negative opinion towards the
customer service and delivery of the product and gave a complete positive feedback score for the
purchase. Therefore, a comment based trust valuation is multi-dimensional. The terms are used
for opinion and rating correspondently to express their positive, negative and neutral polarity
toward entities that expressed in natural language text.
3.1. Commtrust Model
The commtrust framework Figure 1 Shows, Feedback comments are extracted based on opinion
expressions and their association ratings. Dimension trust and weights are calculated using cluster
form expression into dimensions which accumulate the complete trust score.
Figure 1: The CommTrust framework
The below equation 1 is used to compute trust score and weights for overall trust score evaluation
.
Equation 1: A complete trust score is weighted for a seller is accumulated using dimension trust
score.
= ∑ ∗ (1)
Where = and = ℎ dimensioned where ( = 1. . ).
The below equation 2 is used to compute dimensions trust scores.
Equation 2: Given positive (+1) and negative (-1) ratings towards dimension i, = | | =
+1 ∀ = −1}| , the trust score for d is:
=
| #$ % }|% &⁄ ∗
%
(2)
The above equation is called m-estimated [12]
. = [0.1] and [0.5] which represents a constant
trend for truth valuation. In equation 2, is a hyper parameter which may be in peusedo counts -
1/2 ∗ for the positive and negative. The further genuine considerations are required to review
the real, constant trust score of 0.5, which represents the higher value of . By proposing the
previous delivery use the super-parameter m, importantly, the modification may decrease the
positive preference in ratings, supremely although a finite number of negative and positive ratings
[2]
.

46
4. MINING FEEDBACK COMMENTS FOR RANKING
In this section we present the dependency relation analysis for each feedback comment that helps
in the generation of trust score for seller products. This also presents an algorithm for LDA which
is used to calculate dimension weights and rating.
4.1. Typed Dependency Relation Analysis For Extracting An Expression And
Rating
The typed dependency relation analysis [4]
tool currently refined in natural language processing
(NLP) and it is used to interpret grammatical errors in sentences. With typed dependency relation
parsing, a set of dependency relation represented [4]
by a sentence between a couple of words in
the type of (dependent, head), heads are given as content words and other similar words as turn
on the heads as shown in Figure. 2. Whenever a comment indicates an opinion pointing to
dimensions, hence opinion words and dimension words must form some dependency relations.
Words are additionally commented on their parts of speech tags functioning as an adjective
(ADJ), adverb (AVB), noun (NN) and verb (VB). The dimension expressions pointing to head
terms by ratings are analyzed by distinguishing the prior polarity changes terms through an
opinion of a user’s lexicon SentiWordNet. The previous polarity of the words in SentiWordNet
consists of Positive, neutral and negative and that compare to the ratings of +1, 0 and -1.
Figure 2: Typed dependency relation analysis
4.2. Clustering Dimensions
The Lexical-LDA algorithm is proposed to cluster expressions into semantically called
dimensions. In the topic-modeling technique, it assumes the file as input by using term matrix, for
effective clustering Lexical-LDA that allows shallow lexical knowledge in dependency relations
for the topic modeling.
Lexical knowledge makes use of two types of supervise clustering dimension expressions that are
helpful in the generation of appropriate clusters.
• Comments are small, hence re-occurrence of a head condition are not exact instructive.
Rather, re-occurrence of dimension expressions a pone consideration to the same change
across comments is used, and it possibly considers other relevant terms for dimension
expressions.
• As recognized in few conditions to the similar condition of e-commerce purchases are
commented n number of times in feedback comments.

47
Under this topic modeling, clustering complication is formulated as follows: the distribution of
topics generates dimension expressions for the equal change term or negation of a change term.
The distribution of head terms generates each and every topic successively. The above confess to
adapting the structured dependency relation illustration from the dependency relation parser for
clustering. Dependency relations will be input for lexical-LDA for dimension expression in the
form of (head, modifier) couples, or their denial like (quick, shipping) or (bad, seller).
4.3. Lexical LDA-Evaluation
In the feedback comments set of informal language, expressions used. Before processing is
performed, and then spelling correction is applied. For example: let us consider “thankx” is
replaced by “thanks”. Then, the Stanford dependency parser was utilized to generate the
dependency relation representation of the comments and dimension expressions were abstracted.
Lexical- LDA algorithm is applied to cluster dimension expression into dimensions, then finally
after the computing trust score for seller’s figure 3.
Figure 3: Mining feedback comments
Algorithm 1: Variational inference algorithm for the LDA
Input: A No of cases (
Corpus with ) files, * terms in a file +
Output: A Model parameter: ,, ., /
∅ 1
2
∶= 1 4⁄ for all and
51 ∶= 61 + * 4⁄ for all
repeat
for = 1 to *
for = 1 to (
∅ 1
7%
≔ ,19 exp (:;5 1
7
<)
normalize ∅ 1
7%
to sum to 1
57%
≔ 6+∑=
∅7%
until convergence
The varitional inference method in above-shown algorithm 1, 5 and ∅ are starting points. The
pseudo code is understandable with every iteration in varitional interference requires O;(* +
1)(<, in the file number of iterations are necessary required for each and every file on the order of
words in the file. Approxmently produced total operations are *&
(.
5. EXPERIMENTATION AND RESULTS
The model is experimented in Net Beans IDE with MySQL environment. We have taken 1000
users feedback comments extracted from the Amazon for MP3 player products. These feedback
comments are based on the item, shipping, communication, and cost. The DSRs are used to rate
seller, that helps the customer to buy standard products. In the following figure 4 & 5, dataset
information is exported with three buttons, the first button is to browse the user datasets and then
click on the view button to view the dataset information and these datasets are extracted and load
the data into the database and click on the next button this direct to a dependency analysis page.

48
Figure 4: Dataset Information
Figure 5: Extracted Data
Figure 6: Dependency Relation Analysis

49
The figure 6 shows the dependency analysis page. In this user feedback comments are viewed
because in each comment dependency relation identifies the parts of speech tags for each
category like noun, verb, adjective, and adverb etc. For example: as shown below before POS
tagging comment like: good shipping, a great deal and after POS tagging comment like:
good/ADJ shipping/NN, great/ADJ deal/NN and then click on the pre-process button, this ends
with dependency relation.
Figure 7: Dimension Expression Rating
The figure 7 shows dimension expression ratings pointing to head terms are analyzing by
distinguishing the prior polarity of change terms through an opinion of a user’s lexicon
SentiWordNet. The pervious polarity of the words in SentiWordNet Consist of positive, neutral
and negative and that compare to the ratings of +1, 0 and -1. Here, the +1 rating is given to the
positive feedback comments, the 0 rating is given to the neutral feedback comment like a semi-
positive and semi-negative and -1 rating is given to the negative feedback. Then click on the
process button.
Figure 8: Gibbs Sampling
The figure 8 shows LDA window it consists of four buttons they are Gibbs sampling, product
selection, dimensions and process. If we click on Gibbs sampling it is a generative model for

50
LDA process and the dimensions are divided into four expressions such as item, shipping,
communication and cost. This is shown in figure 9. In this process we have to select a product
after that click on the dimension button and we can view dimension words. Now click the process
button.
Figure 9: Dimension Expressions
Next, the Lexical-LDA algorithm is used to cluster aspect expression and these dimension
expressions are the input for LDA. In this process figure 10 shows a weight and trust score
button. It is used to calculate the dimension trust scores and weights for each product and click on
the select product and evaluation button then we can view the dimension scores for the product as
shown in the figure.
Figure 10: Weights and Trust score

51
Figure 11: LDA Clustering
The Figure 11 shows the main LDA process, it is carried out in cluster formation for products, in
the below screenshot it shows the dimension trust score for different products and these
dimensions are clustered, it is called overall trust score evaluation. Now click on the seller trust
profile.
The following graphical representation figure 12 shows the trust score for dimension expressions
like item, shipping, communication and cost for different mp3 player products.
Figure 12: Trust score Dimensions

52
Figure 13: Over all Trust score Evaluation
The figure 13 shows comparison between the products with respect to scores. These scores are
assigned for each product now the buyer can choose the trustworthy seller based on the overall
trust score.
6. CONCLUSION
The “reputation system” problem is well known on popular websites like Amazon eBay etc. High
reputation scores cannot rank sellers effectively so the customers are misguided to select genuine
and trustable sellers. As observed that the buyers give their negative opinions in free text
feedback comments fields, although they provide higher ratings. In this paper, we presented a
multi-dimensional trust valuation model for calculating comprehensive trust profiles for sellers.
The trust valuation model also includes an effective algorithm that computes dimension trust
scores and dimension weights by extracting feature opinion expressions from feedback comments
and clustering them into dimensions. By combining the NLP (natural language processing) with
opinion mining can evaluate the trustworthy sellers in the e-commerce application. All inclusive
experiments on feedback comments for Amazon sellers determine that our technique figures out
trust score in an impressive way and rank sellers.
REFERNCES
[1] Xiuzhen Zhang, Lishan cui, and Yan Wang, “Computing Multi-Dimensional Trust by Mining E-
Commerce Feedback Comments,” IEEE Transaction on Knowledge and Data engineering, Vol: 26
No:7, Year 2014.
[2] P. Resnick, K. Kuwabara, R. Zeckhauser, and E. Friedman, “Reputation Systems: Facilitating Trust in
Internet Interactions,” Communications of the ACM, Vol: 43, pp. 45-48, 2000.
[3] M. DeMarneffe, B. MacCartney, and C. Manning, “Generating typed dependency parses from phrase
structure parses,” in proc. LREC, vol: 6, pp. 449-454, 2006.
[4] M. De Marneffe and C. Manning, “The stand ford typed dependencies representation,” in proc. The
workshop on Cross-Framework and cross-Domain parser Evaluation, 2008.
[5] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet Allocation,” the journal of machine
Learning research, vol: 3, pp. 993-1022, 2003
[6] J. O’Donovan, B. Smyth, V. Evrim, and D. Mcleod, “Extracting and Visualizing trust relationships
from online auction feedback comments,” in proc. IJCAI, pp.2826-2831, 2007
[7] M. Gamon, “Sentiment Classification on customer feedback data: noisy data, large feature vectors,
and the role of linguistic analysis,” in proc. The 20th Int. Conf. On Computational Linguistics, 2004.

53
[8] Y. Lu, C. Zhai, and N. Sunaresan, “Rated aspect summarization of short comments,” in proc. The
18th Int. Conf. On WWW, 2009.
[9] B. Pang and L. Lee, “opinion mining and sentiment analysis,” Found Trends Inf. Retr, Vol:2 No. 1-2,
pp. 1-135, 2008.
[10] B. Lui, Sentiment analysis and opinion mining, Morgan & Claypool Publishers, 2012. [11] M. Hu and
B. Lui, “Mining and Summarizing customer reviews,” in proc. The fourth Int. Conf. On KDD,
pp.168-177, 2004.
[12] K. Karplus, “Evaluating regularies for estimating distributions of amino acids,” in proc. The third Int.
Conf. On Intelligent Systems for Molecular Viology, Vol: 3, pp. 188-196, 1995.
[13] Y. Hijikata, H. Ohno, Y. Kusumura, and S. Nishida, “Social summarization of text feedback for
online auctions and interactive presentation of the summary,” Knowledge-Based Systems, vol. 20, no.
6, pp. 527–541, 2007.
[14] M. Hu and B. Liu, “Mining and summarizing customer reviews,” in Proc. the fourth Int. Conf. on
KDD, 2004, pp. 168–177.
[15] H. Wang, Y. Lu, and C. Zhai, “Latent aspect rating analysis without aspect keyword supervision,” in
Proc. the 17th ACM SIGKDD international conference on Knowledge discovery and data mining,
2011, pp. 618–626.
[16] Wang, H., Lu, Y., & Zhai, C. (2010, July). Latent aspect rating analysis on review text data: a rating
regression approach. In Proceedings of the 16th ACM SIGKDD international conference on
Knowledge discovery and data mining (pp. 783-792). ACM.
[17] I. Titov and R. T. McDonald, “A joint model of text and aspect ratings for sentiment summarization.”
in Proc. ACL, 2008, pp. 308–316.
[18] C. Lin and Y. He, “Joint sentiment/topic model for sentiment analysis,” in Proc. the 18th ACM
conference
[19] A. Fahrni and M. Klenner, “Old wine or warm beer: Targetspecific sentiment analysis of adjectives,”
in Proc. of the Symposium on Affective Language in Human and Machine, AISB, 2008, pp. 60–63.
[20] J. Blitzer, M. Dredze, and F. Pereira, “Biographies, bollywood, boom-boxes and blenders: Domain
adaptation for sentiment classification,” in ACL, vol. 7, 2007, pp. 440–447.

COMMTRUST: A MULTI-DIMENSIONAL TRUST MODEL FOR E-COMMERCE APPLICATIONS

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to COMMTRUST: A MULTI-DIMENSIONAL TRUST MODEL FOR E-COMMERCE APPLICATIONS (20)

Recently uploaded (20)

COMMTRUST: A MULTI-DIMENSIONAL TRUST MODEL FOR E-COMMERCE APPLICATIONS