Sei sulla pagina 1di 9

International Journal of Advanced Science and Technology

Vol. 28, No. 11, (2019), pp. 462-470

Sentiment Quantification Approaches For Customer Reviews Of Amazon


Products
Smita Suresh Daniel 1, Ani Thomas 2, Neelam Sahu3, Aradhana 4

1,3Dept of Computer Science ,Dr. C.V. Raman University Kota ,C.G ,India
2Dept of Information Technology , BIT Durg , Bhilai, C.G, India
4Dept of Information Technology , SSEC , Bhilai, C.G, India

Abstract
Quantifiying various aspects of a product is important in understanding the sentiments of reviewers for
a product. Our research helps to correctly estimate the prevalence of various sentiment classes for each
aspect in a review accurately.

Keywords: Quantification, sentiment Analysis.

1. Introduction

Sentiment Analysis is an important source of decision making which is used to extract, identify and
quantify product reviews based on its various aspects . However, the growing scale of data demands
automatic data analysis techniques. We focused on issues and problems faced by the customers on
buying Amazon product online and measured its sentiments for various aspects by Machine learning
and various Quantification learning techniques.
The final goal of our studies is not just estimating the class label (e.g., Positive, Negative, or Neutral) of
individual reviews, but estimating the relative frequency or prevalence of the different classes in the
dataset accurately despite having a classifier with poor accuracy.
Performing “classify and count” approach has been shown to yield suboptimal quantification accuracy,
this has established quantification as a task of its own. Various Quantification techniques has
convincingly shown that it should be tackled as a task of its own, using learning algorithms and
evaluation measures different from those used for classification. Vapnik’s principle’’ (Vapnik 1998),
says to solve the problem directly without any intermediate step. Based on this we tested sentiment
quantification on text, showing that it substantially outperforms several state-of-the-art baselines.
1.1 Outline
This paper is structured as follows. Section 2 is methodology which is categorized into two sections 2.1
Sentiment Classification and 2.2 Text Quantification. In 2.2 Quantification specific methods for
adjusting the errors caused by classifiers and also probabilistic and iterative approaches for estimating
the prevalence of classes by post-processing and by calibrating the priori is experimented. Various
evaluation measures are used to compare the performance of quantification. Section 3 displays the
observations and interpretation of the experiment . Section 4 discusses the results of our experiments,
while Section .5 concludes and summarizes.

2 Methodology

This work, presents a methodology similar to density estimation ie. to find the prevalence of each
component in a mixture, like an Urn containing a mixture of three classes of grains. The obvious
solution is to ‘classify and count’ but classifiers provide suboptimal results and therefore should be
solved directly. The methodology is in two parts, Classification and Quantification for which we need
to build classifiers and Quantifiers .
2.1 Sentiment Classification

The following modules assisted in building an accurate classifier .


Data Acquisition : Following are the sites we used for making our datasets.

ISSN: 2005-4238 IJAST 462


Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

Data Source
Kindle Reviews jmcauley.ucsd.edu/data/amazon/
Kindle Reviews https://www.kaggle.com/datafiniti/consumer-reviews-of-mazon-products

Preparing the Training Dataset : The various characteristics of training data could also mean
requiring less data for training, or fewer computational resources.
Preparing the Test Dataset : The following are the Test Datasets prepared which are different
versions of Kindle Reader :

Datasets No._reviews Datasets No._reviews


1. Kindle_fire1 2331 6. Kindle_Am2 2850
2. Kindle_fire2 2290 7. Kindle _Echo 3369
3. Kindle_pw1 1634 8. Kindle _Kids 1685
4. Kindle_pw2 2271 9. Kindle_Voyage 799
5. Kindle_Am1 2000 .

Data Preprocessing System : Preprocessing of review included correcting spellings using spell
Check and Removed all punctuations, symbols, numbers and expanded words like can’t, wouldn’t
etc ,removed the stopwords and potter stemmer performs stemming.
Vectorization : Vectorization defines a good numerical measure for texts. Unstructured reviews are
converted in structured ones to extract sentiment words from it. Generally adjectives and adverbs are
sentiment words ,and it can be positive, negative or neutral. A Polarity list (like WordNet) is referred
to assign a score to these words depending upon its intensity and 4 categories of words that is very
Negative, Negative, Positive ,very Positive having numerical values as -2,-1,+1,+2. This Scoring
module determines the score of sentiments during the analysis of data and converts it into numerical
format.
Handling Negation in Reviews : Negation identification and detecting its scope within a sentence (text)
are necessary in finding out the sentiments from a piece of text. Sentiment polarity identification and
calculation is a nested process.
Word Sense Disambiguation using WordNet and Lesk Algorithm : Classifiers interpreted words
which have more than one meaning differently thus changing the polarity of the sentence. Lesk
algorithm computes the semantic relatedness of texts to the given context with a topic from a
knowledge source.
Training the Model - Learning Curves : Learning curve measures the performance of the model
based on variations in the training data supply. The learning curve plotted below helps to identify the
correct spot on your curve where bias and variance are minimized.

Figure 1 Learning Curves


The rate of learning becomes stable at 400 training data for an Support Vector Machine(SVM) classifier
with RBF kernel.
Categorize Aspects : In this study we extracted five most frequent nouns for categorization like
(screen, battery , price, weight, library) using statistical function TF-IDF. These mined review
categories were stored in separate text files.

ISSN: 2005-4238 IJAST 463


Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

Figure 3 Important word features of dataset extracted using TF-IDF


Appearance of Words in each
Sno Category
Category Words
Touch, eye, glare, backlit,
break uneven, freeze,
1 Screen
lighting strain,
navigation
Charge,life,
replacement ,time, port,
2 Battery
time, connecting , hours,
weeks, recharge, weak
Cost, afford, sale,
3 Price purchase, expensive, buy,
money, Cheap
Carry, easy, hand,
handbag ,light , weight,
4 Weight
convenient, portability,
heavy , travel
Books, storage, download,
5 Library read , carry, store , cheap,
font, travel

Evaluation of SVM classifier on test dataset : Confusion matrix tabulates the performance of any
classifier and shows the relation between correctly and wrongly predicted reviews. Here TP (True
Positive) ,FP (False positive). TN (True Negative) FN (False Negative) is number of negative reviews
predicted as positive by the classifier. The difference between Actual and Predicted no. of reviews in
each class is called the Distribution Drift. The performance of sentiment classification can be
evaluated .
Accuracy = (TP+TN+TT)/(TP+TN+TT+FP+FN+FT) = 0.88
All the test datasets is classified using SVM and the results in the confusion matrix is used further in
Quantification Process.

2.2 Sentiment Quantification Methods

Quantification methods belong to two classes ; Aggregative : Require classification of each items and
post processing their prevalence and Non-aggregative : Quantification without performing
classification of each item. We used the following methods for our experimental Analysis (Esuli and
Sabastiani 2014)
Classify and Count Approach : (CC) for quantification consisting of training a classifier from Tr via
a standard learning algorithm (SVM) , classifying the objects in Te, and estimating its prevalence pTe by
simply counting the fraction of objects in Te that are predicted to belong to the class. If by we
denote the event “class c has been assigned by the classifier", so that pTe( ) represents the fraction of
test documents that have been assigned as c by the classifier, this corresponds to computing
= pTe( ) =
Probabilistic Classify and Count : (PCC) estimates pTe by simply counting the expected fraction of
documents predicted positive. If by p(c|x) we indicate the posterior probability, i.e., the probability of
membership in c of test object x as estimated by the classifier, and by E[x] we indicate the expected
value of x, this corresponds to computing
= E[pTe( =

ISSN: 2005-4238 IJAST 464


Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

Calibrated probabilities of prevalence ps(c) of a class c in a set s is obtained by the logistic function
p(c|x) = ;
Adjusted Classify and Count : (ACC) method is composed of two phases: a prediction phase by the
classifier and then an adjustment phase. Let the prevalence of a class , being the estimate of
CC is adjusted to prevalence . It is based on correcting the estimate provided by the CC approach,
depending on the characteristics of the underlying classifier and the law of total probability. It holds
that
pTe( ) = (1)
Here, represents the fraction of test documents belonging to ci that have been falsely
assigned cj by the classifier. Once the classifier has been trained and applied to Te, the quantity
pTe( ) can be observed, and the quantity can be estimated from Tr via k- fold cross-
validation; the quantity is adjusted.
Probabilistic Adjusted Classify and Count (PACC) :The PACC method is called (‘Scaled
Probability Average’) is a probabilistic variant of ACC. Its underlying idea is to replace, in Eq.(1)
and with their expected values. Equation (1) is thus transformed into
E[pTe( )] =
Where E[pTe( )] =
E =
and indicates the set of objects in Te whose true class is c i. Like for ACC, once the classifier has
been trained and applied to Te, the quantity can be estimated from Tr via k fold cross-
validation
Expectation Maximization for Quantification : (EMQ) Saerens (2002), is an instance of Expectation
Maximization is an iterative algorithm for finding maximum likelihood estimates of the class
prevalences for models that depend on unobserved class labels. EMQ incrementally updates the
posterior probabilities by using the class prevalences computed in each iteration, , in a recursive
fashion.

2.2.2 Evaluation measures for quantification

By Ʌ ( , p, S, C) we will indicate a quantification loss, i.e., a measure of the error made in estimating
a distribution p defined on set S and classes C by another distribution ; we can simply write Ʌ ( , p)
when S and C are clear from the context. They are as follows.
Absolute Error : (AE) is the simplest measure for quantification, which corresponds to the average
(across the classes in C) absolute difference between the predicted class prevalence and the true class
prevalence i.e.,

Relative Absolute Error : (RAE) addresses this Problem by relativizing the value |ˆp(cj)−p(cj)| in
Equation 1 to the true class prevalence, i.e.,

Normalized Absolute Error : A normalized version of AE(NAE) that always ranges between 0 (best)
and 1 (worst) can thus be obtained as

Kullback-Leibler Divergence : (KLD) is a member of the family of f -divergences is normalized


cross-entropy, better known as Kullback-Leibler Divergence defined as

ISSN: 2005-4238 IJAST 465


Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

KLD ranges between 0 (perfect coincidence of p and ˆp) and +/- 1

3. Observations and Analysis

Table 5 lists our 9 datasets in increasing order of their distribution drift, i.e., of how
differently the classes are distributed across the three classes of interest returned by KLD .
The 9 datasets into broken into three groups, based on the distribution drift:
Table 1. Datasets used in this work and their Characteristics .
Dataset No. of Training No of Test Total no of Distribution drift
reviews reviews reviews Absolute Error
LD Kindle_fire2 500 500 2800 0.002475
LD Kindle_pw2 500 500 2271 0.003028
LD Kindle_fire1 500 500 3331 0.005148
MD Kindle _Echo 500 500 4369 0.005343
MD Kindle_Am2 500 500 2850 0.005766
MD Kindle _Kids 500 500 2685 0.006924
HD Kindle_Am1 500 500 2000 0.008333
HD Kindle_Voyage 500 500 1799 0.008761
HD Kindle_pw1 500 500 1634 0.008976

1. Kindle_fire2, Kindle_pw2, Kindle_fire1 are the low-drift (LD) datasets, characterized by values
KLD( pTe, pTr)<0.0052;the resuls are reported in Table 5.
2. Kindle_Echo, Kindle_Am2, Kindle_Kids are the medium-drift (MD) datasets, characterized by
values 0.0053<KLD( pTe,pTr)< 0.0070; reported in Table 5.
3. Kindle_Am1, Kindle_Voyage, Kindle_Pw1 are the high-drift (HD) datasets, characterized by
values KLD( pTe, pTr) > 0.0080 ; results reported in Table 5.
In this experiments we compare the Quantification-specific learning algorithms against a baseline (CC)
which is a classification-specific learning algorithm .
Table 2 : Quantification accuracy on three low-drift datasets;
Dataset System Pos Neu Neg AE NAE RAE KLD NKLD
/ Drift
Kindle_Fire1 [Training] 0.822 0.092 0.086 – – – – –
LD [Test] 0.810 0.096 0.094 – – – – –
CC 0.821 0.089 0.090 0.003 0.004 0.028 0.0001 0.0000
PCC 0.819 0.093 0.087 0.001 0.002 0.005 0.0009 0.0004
ACC 0.831 0.089 0.091 0.005 0.009 0.033 0.0046 0.0023
PACC 0.826 0.091 0.088 0.004 0.005 0.016 0.0023 0.0012
EMQ 0.821 0.091 0.088 0.003 0.003 0.013 0.0001 0.0000
Kindle_pw2 [Training] 0.750 0.152 0.098 – – – –
LD [Test] 0.746 0.148 0.106 – – – – –
CC 0.746 0.144 0.110 0.008 0.014 0.060 0.00072 0.00036
PCC 0.749 0.148 0.104 0.003 0.006 0.028 0.00035 0.00017
ACC 0.749 0.144 0.108 0.006 0.011 0.053 0.00039 0.00019
PACC 0.749 0.148 0.103 0.003 0.005 0.027 0.00027 0.00014
EMQ) 0.750 0.143 0.123 0.011 0.019 0.103 0.00556 0.00278
Kindle_Fire2 [Training] 0.748 0.148 0.104 – – – – –
LD [Test] 0.746 0.148 0.106 – – – – –
CC 0.748 0.148 0.103 0.000 0.001 0.002 0.00041 0.00021
PCC 0.748 0.148 0.104 0.000 0.000 0.001 0.00020 0.00010
ACC 0.748 0.148 0.104 0.000 0.000 0.002 0.00024 0.00012
PACC 0.748 0.148 0.104 0.000 0.000 0.001 0.00012 0.00006
ISSN: 2005-4238 IJAST 466
Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

EMQ 0.747 0.148 0.120 0.006 0.009 0.051 0.00618 0.00309


Average LD CC 0.004 0.006 0.030 0.00040 0.00020
PCC 0.002 0.003 0.011 0.00047 0.00024
ACC 0.004 0.007 0.029 0.00176 0.00088
PACC 0.002 0.003 0.015 0.00091 0.00046
EMQ 0.006 0.010 0.056 0.00395 0.00197

For PCC, ACC, PACC we use the confusion matrix generated by the SVM-based classifier and
calculated the priori and posterior probabilities generated by the classifier as described in Section
2.

Table 3. Same as Table 6, but obtained on Three MD Medium -drift datasets


Dataset System Pos Neu Neg AE NAE RAE KLD NKLD
/Drift
[Traini
Kindle_Echo
ng] 0.788 0.140 0.072 – – – –
MD [Test] 0.792 0.132 0.076 – – – – –
CC 0.785 0.132 0.083 0.0070 0.0114 0.0686 0.0004 0.0002
PCC 0.788 0.135 0.077 0.0035 0.0058 0.0352 0.0001 0.0001
ACC 0.787 0.131 0.082 0.0051 0.0104 0.0662 0.0003 0.0002
PACC 0.787 0.131 0.082 0.0037 0.0056 0.0291 0.0000 0.0000
EMQ 0.788 0.132 0.080 0.0025 0.0038 0.0183 0.0000 0.0000
[Traini
Kindle_Am2
ng] 0.576 0.132 0.292 – – – –
MD [Test] 0.570 0.126 0.288 – – – – –
CC 0.574 0.129 0.297 0.0035 0.0061 0.0146 0.0000 0.0000
PCC 0.575 0.131 0.295 0.0018 0.0030 0.0073 0.0000 0.0000
ACC 0.575 0.129 0.296 0.0026 0.0045 0.0124 0.0002 0.0001
PACC 0.575 0.131 0.294 0.0013 0.0022 0.0062 0.0001 0.0000
EMQ 0.572 0.128 0.190 0.0340 0.0584 0.1171 0.0504 0.0105
[Traini
Kindle_Kids
ng] 0.762 0.148 0.090 – – – –
MD [Test] 0.758 0.144 0.098 – – – – –
CC 0.758 0.141 0.101 0.0073 0.0121 0.0584 0.0000 0.0000
PCC 0.760 0.144 0.095 0.0036 0.0060 0.0292 0.0002 0.0001
ACC 0.761 0.142 0.100 0.0057 0.0094 0.0509 0.0009 0.0004
PACC 0.762 0.144 0.095 0.0032 0.0053 0.0281 0.0000 0.0000
EMQ 0.758 0.134 0.116 0.0147 0.0249 0.1301 0.0023 0.0012
Average MD CC 0.0059 0.0099 0.0472 0.0002 0.0001
PCC 0.0030 0.0049 0.0239 0.0001 0.0001
ACC 0.0044 0.0081 0.0432 0.0005 0.0002
PACC 0.0027 0.0044 0.0211 0.0001 0.0000
EMQ 0.0170 0.0290 0.0885 0.0176 0.0039

Table 4 Quantification obtained on three HD(High drift) datasets

Dataset System Pos Neu Neg AE NAE RAE KLD NKLD

Kindle_Am1 [Training] 0.642 0.106 0.252 – – – – –


HD [Test] 0.634 0.102 0.244 – – – – –
CC 0.644 0.102 0.254 0.00267 0.00445 0.01626 0.00004 0.00002
PCC 0.640 0.103 0.250 0.00222 0.00372 0.01207 0.00291 0.00146
ACC 0.649 0.102 0.252 0.00624 0.00600 0.01505 0.00149 0.00074
PACC 0.646 0.104 0.252 0.00180 0.00270 0.00777 0.00079 0.00039
ISSN: 2005-4238 IJAST 467
Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

EMQ 0.640 0.102 0.250 0.00381 0.00574 0.01084 0.00493 0.00246


Kindle_Voyage [Training] 0.466 0.086 0.448 – – – –
HD [Test] 0.450 0.086 0.438 – – – – –
CC 0.459 0.086 0.454 0.00445 0.00731 0.01086 0.00004 0.00002
PCC 0.463 0.086 0.451 0.00223 0.00365 0.00543 0.00001 0.00001
ACC 0.462 0.085 0.451 0.00266 0.00436 0.00873 0.00049 0.00024
PACC 0.464 0.086 0.450 0.00133 0.00218 0.00437 0.00024 0.00012
EMQ 0.455 0.086 0.138 0.10194 0.16733 0.22719 0.21772 0.01050
Kindle_pw1 [Training] 0.688 0.196 0.116 – – – – –
HD [Test] 0.678 0.190 0.112 – – – – –
CC 0.682 0.188 0.130 0.00808 0.01393 0.06019 0.00809 0.00404
PCC 0.674 0.197 0.120 0.00609 0.01038 0.01863 0.00399 0.00200
ACC 0.686 0.189 0.128 0.00702 0.01208 0.04805 0.00137 0.00069
PACC 0.687 0.191 0.122 0.00392 0.00670 0.02610 0.00022 0.00011
EMQ 0.680 0.123 0.149 0.03514 0.31762 0.21700 0.02091 0.01045
Average HD CC 0.00507 0.00857 0.02911 0.00272 0.00136
PCC 0.00351 0.00592 0.01204 0.00231 0.00115
ACC 0.00531 0.00748 0.02394 0.00112 0.00056
PACC 0.00235 0.00386 0.01275 0.00042 0.00021
EMQ 0.04696 0.16356 0.15167 0.08119 0.00781

In Tables 2, 3 , 4. shows results obtained by each method and five evaluation measures (columns 6–
10), we report the class distribution computed for each dataset by each method (columns 3–5), which
can be easily compared with the actual class distributions in the training set and in the test set.
Table 5 - Quantification accuracy averaged over all 9 datasets,
SYSTEM AE +% NAE +% RAE +% KLD +% NKLD +%
PACC 0.002 0.004 0.016 0.0005 0.0002
PCC 0.003 10 0.005 18 0.016 0 0.0010 98 0.0005 98
ACC 0.005 84 0.007 94 0.032 102 0.0011 128 0.0006 128
CC 0.005 99 0.008 114 0.035 123 0.0011 129 0.0006 129
EMQ 0.023 851 0.068 1662 0.099 523 0.0342 6926 0.0046 1770

The overall results of experiments are reported in Table 5 ; for each quantification method we present
(a) the score (averaged across the 9 datasets) obtained by the method according to each of the five
evaluation measures we consider, and (b) the relative deterioration in accuracy brought about by each
method with respect to the best-performing system in table 5 .

4. Result and Discussion

The observations gives us an idea about which of the five measures we use are is better. If we want
to compare the learning methods with each other, Table 5 clearly says that PACC is the best method,
according to all 5 measures. Since the key to PACC’s performance is in adjusting the CC measures and
then calculating the posterior probabilities it receives as input.
In the results on the individual datasets (see Tables 2, 3, 4), the superiority of PACC seems even more
clear . Out of 45 combinations of 9 datasets and 5 evaluation measures, PACC is the best performer in
5, while ACC is in 0, PCC is in 3, and EMQ is in 1.
Out of 45 combinations of 9 datasets, 5 evaluation measures, PACC is the best performer in 28, while
PCC is in 8, and SVM(KLD) is in 7 and ACC is in 2. If we look at the average results on low-drift and
mid-drift datasets, PACC is always the best performer; on high-drift datasets
Wilcoxon Statistical Significance Test : It checks if the differences in performance among the five
quantification methods, are statistically significant. we have performed apair wise two tailed Wilcoxon
signed-ranks (Wilcoxon 1945) on the 9 datasets, based on the KLD results. As a non-parametric
alternative to the paired t test, the Wilcoxon test ranks the absolute differences in the performance
ISSN: 2005-4238 IJAST 468
Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

measure (KLD, in our case) of two algorithms for each dataset, and compares the ranks for the positive
and the negative differences . When the computed p value is lower than the significance level α = 0:05,
we reject the null hypothesis and claim that the difference of the two algorithms is statistically
significant.
Table 6 Results of Wilcoxon tests for our 8 quantification methods,
CC PCC ACC PACC EMQ
CC - < < >
PCC > - > < >
ACC > < - < >
PACC > < > - >
EMQ < < < < -
Symbol ‘‘>’’ indicates that the method on the row is better, in a statistically significant way, than the
method on the column; symbol ‘‘<’’ means the opposite; symbol ‘‘ ’’ means that there is no
statistically significant difference among the two; symbol ”_’’ indicates “not applicable “

The results of our Wilcoxon test shows the superiority of PACC over most other algorithms is
statistically significant,
Results using SVM Classifier and PACC Quantification. : To detect the sentiment issues of reviews
related to various aspects of the product Kindle for the Kindle_fire dataset we categorized five most
prominent aspects into five different files .These were quantified and the positives for each of the
categories using the four approaches for quantification is as shown in figure 4.. We can find that the
Battery aspect of the Kindle is showing less positive sentiments . So an indepth analysis was done on
this category to study the reason behind why many people are opposing it. On Quantifying this
particular category again in the same process we get the results in the graph in Figure 5 and Figure 6 .

Figure 4 - Positives for each of the categories


Here we come to the conclusion that reviewers are referring to the battery replacement and battery
getting heated during their usage. Similarly Sentiment Analysis can be done for all other categories
also.

ISSN: 2005-4238 IJAST 469


Copyright ⓒ 2019 SERSC
International Journal of Advanced Science and Technology
Vol. 28, No. 11, (2019), pp. 462-470

Figure. 5 Graphical Analysis of the Figure. 6 Analysis of the Quantified


Sentiments of each Category Sentiments of the Problem Category

The mean error using KLD measures = 0.0036

5. Conclusion

Sentimental Analysis is an easier and cost effective way to understand how the people are feeling about
a particular aspect of a product using online reviews. Applying techniques like Vectorisation ,
Negation Module and Word Sense Disambiquation Module improved the classification accuracy to
a commendable height . Quantification is applied to find the prevalence of various sentiment classes in
each aspect of the product Kindle from the dataset. We see from the results obtained that CC, the naive
classification based method to estimate class prevalences, is inferior, in a statistically significant sense,
to methods (PACC and PCC) which is explicitly designed for quantification. It estimates prevalences
directly, without using the classification of individual items as an intermediate step, but by using
calibarated and adjusted posterior probabilities as estimated by the classifiers.
Our model detects correctly five most important aspects of the product Kindle fire about which the
customers have referred in their text reviews and quantified its sentiments using PACC, thus mining
the important information contained in it. It depicts the actual problem faced by the customers very
correctly.
We have experimentally shown, on a multiple datasets, that more accurate prevalence estimates may be
obtained by considering quantification as a task in its own right, i.e., by using quantification-specific
learning algorithms which directly attempt to solve the prevalence estimation problem, rather than
viewing quantification as a by product of classification.

References

1. A. Esuli, F. Sebastiani F (2014) Explicit loss minimization in quantification applications


(preliminary draft). In: Presented at the 8 th international workshop on information filtering and
retrieval (DART 2014), Pisa, IT.
2. G,Forman (2005) Counting positives accurately despite inaccurate classification. In: Proceedings
of the 16th European Conference on machine learning (ECML 2005), pp 564–575, Porto, PT
3. W,Gao , F,Sebastiani (2016) ,” Tweet Sentiment: From Classification to Quantification’,
Springer-Verlag Wien
4. M.Saerens , P.Latinne , C. Decaestecker (2002) Adjusting the outputs ofa classifier to new a priori
probabilities: a simple procedure. Neural Comput 14(1):21–41
5. Vapnik V (1998) Statistical learning theory. Wiley, New York.

ISSN: 2005-4238 IJAST 470


Copyright ⓒ 2019 SERSC

Potrebbero piacerti anche