International Journal of Innovative Research in Computer and Communication Engineering

International Journal of Innovative Research in Computer and Communication Engineering

Menu

ISSN ONLINE(2320-9801) PRINT (2320-9798)

All submissions of the EM system will be redirected to Online Manuscript Submission System. Authors are requested to submit articles directly to Online Manuscript Submission System of respective journal.

Prognosis of Failed Back Surgery Syndrome Based On Feature Extraction Method

Alda Kika1, Ridvan Alimehmeti2
  1. Department of Informatics, Faculty of Natural Science, University of Tirana, Tirana, Albania
  2. Neurosurgical Department at University of Medicine, Tirana, Albania
Related article at Pubmed, Scholar Google

Visit for more related articles at International Journal of Innovative Research in Computer and Communication Engineering

International Journal of Innovative Research in Computer and Communication Engineering

Abstract

Failed Back Surgery Syndrome (or FBSS) refers to patients with persistent or new pain after spinal surgery for back or leg pain. Multiple factors can contribute to its onset. We studied different techniques of data mining to determine the prognosis of patients with FBSS. Different machine learning algorithms are tested to find the best algorithm that predicts the factors that influence FBSS from the set of 305 patients operated for lumbar disc herniation. Since the data is unbalanced other criteria rather than accuracy is used as the evaluation criteria. The tools that are developed using WEKA API and the approach of feature selection test different machine learning algorithms to evaluate the best algorithm that maximize AUC or F-Measure, but on the same time maintaining false negative low. The results of the experiments are discussed and the factors that mostly influence on this syndrome are identified.

Keywords
aspect identification; aspect Ranking; Consumer review; Product aspects; Sentiment classification.

INTRODUCTION
In the recent years the use of e-commerce is grown very rapidly. Most retail Websites promotes consumers to write their feedbacks about products to express their opinions on various aspects of the products. An aspect, which can also be called as feature, refers to a component or an attribute of a certain product. A sample review “The sound quality of Sony Experia is amazing." reveals positive opinion on the aspect “sound quality" of product Sony Experia. Many forum Websites also provide a platform for consumers to post reviews on number of products. For example, CNet.com involves more than seven million product reviews; These numerous consumer reviews contain rich and valuable knowledge, which is becoming an important resource for both consumers and firms [1]. Before purchasing a product, consumers commonly seek quality information from online reviews and firms can use these reviews as feedbacks for better product development, consumer relationship management and marketing.
Generally, a product may have number of aspects. For example, a Smart Phone has hundreds of aspects such as “screen size," “camera," “memory size," ”sound quality.” one may say that some aspects are more important than the others, and have strong influence on the consumers’ decision making as well as firms’ product development strategies. For example, some aspects of Smart Phone e.g., “camera" and “memory size," are considered important by most of the consumers, and are more important than the others such as “color" and “buttons." Hence, the identification of important product aspects plays an essential role in improving the usability of reviews which is beneficial to both consumers and firms. Consumers can easily make purchasing decision by paying attention to the important aspects, while firms can focus on the improvement of product quality so that product reputation is enhanced. However, manual identification of important aspects is impractical. Therefore, an approach to automatically identify the important aspects is highly demanded. Motivated by the above observations, we made a survey on different techniques used to find important product aspects automatically from online consumer reviews.
In this paper we present the methodology in section II and techniques used for the product aspect identification and product aspect classification in the section no. III and section no. IV respectively and section V illustrates the product aspect ranking.

METHODOLOGY
The process of product aspect ranking consisting of three main Steps: (a) aspect identification; (b) sentiment classification on aspects (c) Product aspect ranking. Given the consumer reviews of a product, first identify the aspects in the reviews and then analyze these reviews to find consumer opinions on the aspects via a sentiment classifier and finally rank the product based on importance of aspect by taking into account aspect frequency and consumers’ opinions given to each aspect over their overall opinions.
Reviews can be posted on the webs in three different types:
Type (1) - Pros and Cons: The reviewer is asked to describe Pros and Cons separately.
Type (2) - Pros, Cons and detailed review: The reviewer is asked to describe Pros and Cons separately and also write a detailed review.
Type (3) - free format: The reviewer can write freely, i.e., no separation of Pros and Cons.
Different types of reviews may need different techniques to perform the tasks such as product aspect identification, product sentiment classification and product aspect ranking as mentioned in fig. 1
For type (1) and (2), opinion orientations are known because Pros and Cons are separated and thus there is no need to identify them. Only product features need to be identified from the comments of customers. For type (3), we need to identify both product features and opinion orientations.
In order to obtain identification of aspects, the Pros and Cons reviews are used as supporting knowledge to assist the identification of aspects in the free text reviews. In particular, first split the free text reviews into sentences, and parse each sentence using parser. After that the frequent noun phrases are extracted from the sentence parsing trees as candidate aspects. Since these candidate aspects may contain noises, further the Pros and Cons reviews are used to assist them in identification of aspects from the candidates. Then all the frequent noun terms extracted from the Pros and Cons reviews are collected to form a vocabulary. Each aspect in the Pros and Cons reviews is represented into a unigram feature, and all the aspects are then used to learn a one-class Support Vector Machine (SVM) classifier [2]. The resultant classifier is used to identify aspects in the candidates extracted from the free text reviews.
This task of analyzing the sentiments expressed on aspects is called aspect-level sentiment classification [3]. Many techniques are used for sentiment classification which includes the supervised learning approaches and unsupervised approaches such as the lexicon-based approaches. The lexicon-based method uses a sentiment lexicon which contains a list of sentiment words, phrases and idioms, to determine the sentiment orientation on each aspect [4]. On the other hand, the supervised learning methods train a sentiment classifier by using training dataset. The classifier is then used to predict the sentiment on each aspect. In the subsequent sub sections we will discuss the various methods for aspect sentiment classification. Finally a probabilistic aspect ranking algorithm is used to identify the important product aspects from reviews.

ASPECT IDENTIFICATION TECHNIQUES

Supervised learning
Supervised learning technique use the collection of labeled reviews to learn an extraction model. This extraction model called as extractor is then used for the identification of aspects in ne reviews. Most of the supervised learning techniques are based on the sequential learning. Various literatures show the different technique for the learning of extractor.
Wong and Lam [5] uses the HMM model and conditional random field to learn the extractor.
Li et al [18] uses skip CRF and tree CRF i.e. to integrated CRF variation to learn the extractor. The main disadvantage of this method is that it require Labeled sample for training. These methods are very time consuming to label the samples.

Unsupervised learning
In this method the aspects are considered noun or noun phases and occurrence frequency of noun and noun phrases is calculated. The frequent noun or noun phrases are considered as aspects. Hu and Liu [3] use this unsupervised technique for aspect identification. Main disadvantage of this method is that identified aspects candidates may contain noise.
Wu et al [21] uses a phrase dependency parsing. Phrase dependency parsing takes the sentence as input and segments it into phrases. Then these segments are linked with directed arc. Phrase dependency parsing focuses on phrases and not on single word inside phrase. To make sure that identified aspects candidate is to be aspect language model which is based on product reviews is used to predict score of candidates. Model filter out low score candidates. Such model may be biased to frequent terms in the review and cannot sense precisely the related score of aspect terms as a result cannot filter out noise efficiently.
Popesu and Etrioni [19] developed their own systems for aspect identification. They developed OPINE system which is based on KnowItAll web information extraction system which extract aspects from reviews. Su et al [20] design a reinforcement strategy. This strategy cluster together product aspect and opinion words by iteratively using both content and sentiments

ASPECT SENTIMENT CLASSIFICATION TECHNIQUES

Lexicon Based Approach
Opinion words are used in many sentiment classification tasks. Desired states are express with Positive opinion words while, undesired states are expressed with negative opinion words. All opinion words, opinion phrases and idioms are together called as Lexicon.
Hu and Liu [3] use this method. They utilize synonym /antonym relations defined in WordNet to bootstrap the seed word set and finally obtain a sentiment lexicon. Fig 2 shows example of Bipolar adjective structure where synonym relationship is represented by arrow (→) and antonym relationship is represented by dash arrow (--->).
Three approaches used to collect this opinion word list are manual, dictionary based and corpus based approaches. Manual approach is very time consuming and it is not used alone. It is usually combined with the other two automated approaches to avoid mistakes that resulted from automated methods. The two automated approaches are presented in the following subsections.

Dictionary-based approach:
[7,8] presented the main strategy of the dictionary-based approach. A small set of opinion words is collected manually with known orientations. Then, synonyms and antonyms of these words is added to this set which is grown by searching the words in the well known corpora WordNet [6] or thesaurus [9]. The newly found words are added to the seed list then the next iteration starts. This iterative process stops when no new words are found. After completion of process list is checked manually to remove or correct errors.
This method is unable to find opinion words with domain and context specific orientation which is the main drawback of this method.

Corpus-based approach:
The drawback of dictionary based approach is overcome in Corpus- based approach which helps to solve the problem of finding opinion words with context specific orientations. Its methods depend on syntactic patterns.

Holistic Lexicon Based Approach:
Ding et al [22] presented a Holistic lexicon based method Holistic lexicon-based approach improves the lexiconbased method in [23] by addressing two issues that the opinion of sentiment words would be content sensitive and conflict in the review.
This method does not look at the current sentence alone rather it uses the external information and evidences in other sentence and other reviews. Some linguistic conventions in natural language expression are used to find the orientation of opinion word. This method required prior domain knowledge or user inputs are needed. This approach is highly effective when sentence contain multiple contradictory opinion words.

Supervised Learning Techniques

Naïve Bayes Classifier (NB):
Naïve Bayesian networks are composed of acyclic graph with only one parent and several children. There is a very strong assumption of independence with child nodes in the context of their parents. Independence model can be represented with :
image
When these two probabilities are compared, larger probability is more likely to be the actual class label. Advantage of naive bayes classifier is its short computational time for learning the dataset. Bayes classifiers are usually less accurate than that of other learning algorithms.

Maximum Entropy Classifier (ME):
Another classifier is Maximum Entropy classifier. The name Maximum Entropy comes from the fact that the classifier finds the probabilistic model which is the simplest and least constrained. Yet it has some specific constraints. The idea behind maximum entropy is that one should prefer the most uniform models that also satisfy any given constraints. This Classifier is used to converts labeled feature sets to vectors using encoding. This encoded vector is then used to compute weights for each feature that can then be combined to determine the most likely label for a feature set. This classifier is controlled by a set of X{weights}, which is used to combine the joint features that are produced from a feature-set by an X{encoding}. Each C{(featureset, label)} pair is mapped to vector with encoding scheme. The probability of each label is calculated using the following equation:
image
ME classifier was used by Kaufmann [25] to detect parallel sentences between any language pairs with small amount of training data. Other tools were developed to automatically extract parallel data from non-parallel corpora use language specific techniques or require large amounts of training data. Their results showed that ME classifiers can generate useful results for almost any language pair. This can allow the formation of parallel corpora for many new languages.

Support Vector Machines Classifiers (SVM):
Support Vector Machines (SVMs) are the newest supervised machine learning technique. SVM uses the notion of a “margin”- a hyperplane that divide two data classes.. An upper bound on the expected generalization error can be reduced by maximizing the margin and thereby largest possible distance between separating hyperplane instances on either side of it. In Fig. 3, X, O are 2 classes and A, B and C are the three hyperplanes. Hyperplane A provides the best separation between the classes, because the normal distance of any of the data points is the largest, so it represents the maximum margin of separation. In the case of linearly separable data, once the optimum separating hyperplane is found, data points that lie on its margin are known as support vector points whose solution is represented as a linear combination of only these points . Other data points are ignored. Therefore, the model complexity of an SVM is unaffected by the number of features encountered in the training data. For this reason, SVMs are well suited for learning tasks where the number of features is large with respect to the number of training instances.

PRODUCT ASPECT RANKING
After the aspect sentiment classification, classified aspects are then ranked. Zheng et al [26] described the probabilistic aspect ranking algorithm in their literature.

CONCLUSION
This survey paper presented an overview on the product aspect ranking techniques to identify important aspects of products. Product aspect ranking process contains three main steps i.e. product aspect identification, aspect sentiment classification and aspect ranking. We have conducted a survey which illustrates various methods for aspect identification and sentiment classification.
 

Figures at a glance
Figure 1 Figure 2 Figure 3
Figure 1 Figure 2 Figure 3
 

References