help

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

978-1-4673-9802-2/16/$31.00© 2016 IEEE

Sampling Techniques for Streaming Dataset using Sentiment Analysis

D. Yazhini Priyanka1, Dr. Radha Senthilkumar2

Department of Information Technology1,2 Anna university-MIT campus

Chennai,India [email protected] , [email protected]

Abstract— Streaming data is very huge in size and it arrives continuously in a rapid manner. It may vary from time to time. Scalability, Continuous Availability, Workload Diversity, Data Security and Manageability are the challenges in Big data. Due to the huge size of data, it is difficult to analyze and it takes more time to complete. Rather than analyzing the entire streaming dataset, sampling provides an alternate solution to analyze in an efficient manner and thereby minimizing the computation time. Sampling should represent the properties of the entire dataset. Many sampling techniques such as Reservoir sampling(RS) is used to extract the sample. In our proposed work, Twitter dataset is extracted via Twitter Application Programming Interface(API) and analyze the dataset using Sentiment Analysis (SA) technique. SA technique is used to find out the polarity of the tweets and is applied for the sample dataset which are extracted from the complete dataset using Reservoir Sampling techniques. Further, the obtained results are analyzed and perceived that sampling technique will be precise for twitter dataset.

Keywords—Data analytics, Data preprocessing, Sampling, Sentiment analysis

I. INTRODUCTION Twitter is one of the second popular networking sites and

it act as a micro-blogging service. Twitter is a massive social networking site adapted towards fast communication. Every day more than 140 million active users tweet over 400 million "Tweets". Due to the limitation of the tweet characters length (140 length), it provides a good position for analysis of social network. Most of the researchers use twitter for analyzing social network sites such as user interaction, Tweet statistics etc.

SA is one of the most frequently used technique for

analyzing twitter. The purpose of the SA is to identify the emotion, and classify the polarity of the text. . SA is also known as opinion mining technique, to find out the sentiment of the tweet such as positive, negative and neutral. There are three classification levels of SA techniques. They are document-level, sentence level, and aspect level SA [21]. In Document-level SA is to classify the sentiment of the document, but the theme of the

document is uniform until the end. In sentence level SA is to classify the sentiment of the sentence. In aspect-level SA is to classify the sentiment of the aspects and entity.

In modern applications, data does not take the constant

form, but rather arrives in continuous, fast, large amount, and data may vary from time to time. The applications for streaming scenario include twitter, manufacturing, sensor data, and others. The streaming data model requires the stream data to scale down or the data mining algorithms to scale up . One of the feasible approaches for dealing with data streams is to select a sample and do data mining on the sample. Sampling is the process to select the sample from the large or streaming dataset. Sampling techniques are classified as Uniform Sampling (US) and Biased Sampling (BS) [22]. US is the process in which each and every element in the large dataset have the same probability of being selected as a sample. RS and Bernoulli’s sampling are under the US techniques. In BS, the element in large dataset has the different probability of being selected as samples. Weighted Sampling (WS) techniques are under the category of BS [22]. Meanwhile, fairly accurate answers are sufficient for several data mining applications. Sampling is the most widely used statistical approach for streaming dataset.

Every tweet has a limited size, The twitter dataset is used

for our experimental purpose. In our proposed work, SA is carried out for the tweet data. Then, RS technique is applied to fetch a sample from the complete dataset. Finally, SA is applied to the sample dataset. The accuracy of RS technique is evaluated.

II. RELATED WORK A. Data Collection

Twitter based Interface is used for programmatically accessing tweets by query term. Yazhe Wang et al [23] collected the tweets via API. The data collection and types of collecting tweets are briefly discussed in [15]. There are three types of data collection, such as Geo-tagged tweets, Tweets about a topic, Tweets from a group of users.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

Collecting Twitter data begins with identifying the topic

of interest using a keyword or hashtag [1], and requires the use of APIs. This API method is used for obtaining 1% of publicly available Twitter data. Twitter Firehoses act as a twitter providers, which can be used to extract 100% of Twitter data based on criteria. This is an optimal method for data collection. Pulkit Goyal et al [15] collecting twitter dataset via API and proposed a clustering analysis for twitter.

In our proposed work, the Twitter dataset can be

extracted via the Twitter API by creating a Twitter App and a keyword search method is also used to collect tweets. Twitter App needs to be authorized via ROuth.

B. Sentiment Analysis Walaa Medhat et al [22] discussed the sentiment analysis

types and its application. Algorithms and their originating references of various SA techniques are categorized and briefly explained. Sahana et al [20] proposed some feature selection techniques, which are used to choose the features from the high dimensionality feature set. This classification is performed using SVM provided by weka tool and also investigates that which is the best feature to extract sentiments from the reviews.

Bruno Ohana et al [7] proposed the use of

SentiWordNet. SentiWordNet is an outlook dictionary obtained from the WordNet database where each term is related to the score. The proposed approach determined sentiment orientation based on the positive and negative scores, and SentiWordNet is used for constructing a dataset of relevant features and applied to a machine learning classifier. The results showed that the SentiWordNet could be used as a significant resource for classifying the sentiment.

The beneficial of linguistic features for identifying the

sentiment of tweets and calculate the value of existing lexical resources are investigated in [9]. Features that confine information on the subject of the casual languages are used in microblogging. Efthymios Kouloumpis et al proposed a methodology to the problem, but leverage hashtags in the Twitter data for constructing training data. Using hashtags to collect training data from which data is collected based on sentimental emotions [5] [9].

Bahrainia et al [4] introduced and compared various

polarity detection algorithms and automatic aspect detection algorithm and built a sentiment summarization system in order to summarize opinionated text in the domain of consumer-product. Masahiro Ohmura et al [13] analyzed the social network sentiment through a tweet.In this paper, Improves sensitivity by reducing the adjectives.

Bongsug Chae proposed three methodologies combination for twitter analysis. They are Integrated text analysis with SA, Content Analysis(CA), Descriptive Analysis(DA) [5]. Ram Chatterjee et at [8] described the opinion mining deals with the SA and also analyzed various methods for extracting tweets. Windows 10 as a keyword search for extracting tweets in this paper. Bhuta et al [6] reviews the number of techniques for sentiment analysis in both lexicon based method and also learning based methods and explain the issues and challenges for analyzing twitter.

Apoorv Agarwal et al proposed the SA on twitter data. The main objective of the paper [3] are: (1) Introduce POS- that precise past polarity features. (2) The need for feature engineering is to search the use of tree kernel. The proposed work concluded that SA of Twitter data is not that diverse from SA for other data. Fajri Koto et al [12] introduced a feature called pos sequence in order to analyze the tweet sentiment analysis. Three forms of the pos tag are proposed such as a sequence of 2-tags, 3 tags and 5 tags. The proposed results obtained that the sentiment is classified as positive, negative, subjective, and objective tweets. Sarlen et al [16] proposed Twitter sentiment analysis to analyze customers outlook coming up from the critical to success in the marketplace. This paper suggested the use of machine-based learning approach and also natural language processing (NLP) which provides more accurate results for analyzing a sentiment of tweets. Alec Go et al [2] proposed an Approach for categorizing the sentiment of Twitter messages automatically. The twitter messages are categories as positive, negative as for query term. The proposed work presented the consequences of machine learning algorithms for categories the sentiment of Twitter messages with distant supervision. Hemalatha et al [10] proposed machine learning algorithm can achieve high accuracy for classifying the sentiment of the tweet [2][20]. In this paper, SA technique is proposed for twitter analysis. Sentence level SA and pos method are used to analyze the tweets.

C. Sampling for Streaming Data Yazhe Wang et al [23] proposed the comparison of samples and complete twitter dataset. A large number of Twitter users and tweets make it impractical to collect and maintain a complete record of activity. Therefore, most of the organizations rely on samples. In this sampling technique, sample size plays a major role. This paper proposed a comparative analysis of samples obtained from two of Twitter’s streaming APIs with the complete Twitter dataset. Wenyu Hu et al [22] surveyed the sampling of streaming data and explained the classification of sampling algorithms. At the same time, conversation and similitude of

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

sampling algorithms were performed. Owing to the obstruction of uniform sampling in a few applications, the significance of using biased sampling methods was entirely deserted. Consequently, this paper overlooked the applications of the sampling techniques, particularly those conventional sampling techniques in the data stream model. In our proposed method, the three existing sampling techniques are used such RS, Bernoulli’s A/R sampling and WS. Jeffrey Scott Vitter [11] proposed a fast algorithm for reservoir sampling technique. The proposed algorithm provides selected n random replacement from the collection of N records. A reservoir sampling algorithm provided the solution for the problem that processes the file consecutively. In our work, Statistical Sampling techniques such as RS is used to select the sample from the twitter dataset. The sample dataset and the whole dataset are compared and prove that a good sampling technique should represent the properties of the complete dataset.

III. SAMPLING TECHNIQUES AND SENTIMENT ANALYSIS

A. Problem definition Twitter dataset is very huge in size. In order to process

the whole dataset, it is a difficult task to analyze and also it will take more time for processing. To overcome this problem, Sampling Techniques are used.

B. Proposed Work In our proposed work, the first step is the extraction of

twitter dataset via API and preprocess the dataset to eliminate noise. Further, the RS Sampling technique is applied to choose the sample from the whole data set. The SA sentence level technique is applied to find out the sentiment of the tweet for the whole data set as well as the sample data set. Finally, this work is to find the accuracy of the RS sampling technique, Comparative analysis is carried out for the sample and whole dataset.

In Fig.I, RS technique is proposed. Let D be the twitter dataset whose size is not known in advance. To evaluate the twitter dataset, SA is performed. Rather than considering the complete data, Sampling gives the approximate result of analyzing the twitter data. RS provides the solution for anatomizing the dataset..

C. Data preprocessing The twitter data consist of noise such as stop words, non

English words, and punctuation marks. These types of noises are highly available in the tweet. Because, most of the people are using their own slang for posting tweet to

share their feelings. In order to remove the noise, preprocessing steps need to be carried out. After preprocessing step, the tweet is categorized as parts of speech (POS) Tag.

FIG.I FRAMEWORK OF PROPOSED WORK

D. Sentiment Analysis In SA phase, Sentence level SA is proposed in this work.

Here, Tweets are considered as sentences. To observe the polarity of the tweets, SA is used. SA is the opinion mining activity concerned with determining the sentiment orientation of the tweets [7]. The sentiment orientation can be classified as positive, negative and neutral.

The score of the tweet can be calculated by using the Eqn(1).

Score = Sum(positive words) – Sum(Negative words) (1)

Procedure 3.1: Sentiment analysis on tweets Input: Tweet whole dataset and sample dataset Output: classify the tweet as positive, negative and calculate 1 For each tweet 2 Evaluate the parts of speech for each sentence 3 Examine the word is either positive or negative 4 Calculate the score for each tweet 5 Score can be positive and Negative values. Score = positive values - negative values 6 Figuring the positive and Negative Score to

determine the Sentiment Orientation. 7 Calculate the Overall score for each twitter user.

Overall Score=Sum(Positive values)-Sum(Negative values)

EXTRACTION OF TWITTER DATASET

SAMPLING TECHNIQUE

RESERVOIR SAMPLING

SENTIMENT ANALYSIS

COMPARATIVE ANALYSIS AND FIND OUT THE

ACCURACY OF SAMPLE

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

In Procedure 3.1, describes the algorithm for SA. First

step is to analyze the word in the tweet that is subjective or objective. If it is subjective, then analyze the word is either positive or negative. Based on the analysis, Score can be calculated. If the score is positive, then it is conceded as a positive tweet. If the score is negative, then it is conceded as a negative tweet, under other condition it is a neutral tweet. The overall score can be calculated for the user, and detect the sentiment of the user.

E. Reservoir Sampling

RS is the Sampling techniques for streaming dataset. Streaming data size is not known in advance. Instead of storing and analyzing the whole data, RS is used to produce the sample which represents the whole data.

In procedure 3.2, an RS sampling algorithm for twitter

dataset is proposed for selecting sample twitter dataset. First k tweets insert into the reservoir array of size k. If k+1th tweet enters, it will generate a random number. If the random number is less than k, then it will randomly replace some other tweet in the reservoir array. It will continue until the streams are exhausted.

Procedure 3.2 Reservoir Sampling Input: Stream of tweets D not known in advance Output: Sample tweets of array size k 1 Initialize array Reservoir[k] 2 for each tweet i in 1: k do 3 Reservoir[i]:=sample[i] 4 done 5 for each tweet i in k+1 to D do 6 j:=random(1,i) 7 if(j<=k) then 8 Reservoir[i]:=sample[i] 9 continue until i=D

IV. EXPERIMENTS AND COMPARATIVE ANALYSIS

A. Data Collection Streaming dataset such as twitter dataset can be

collected by using the twitter API. First, create the twitter application on the twitter page. And then the consumer key, consumer secret, the access key and access secret token is necessary to authorize the Twitter application.

ROAuth is required to authorize the application. After the authorization process is complete, the tweet can be extracted by using search command. Using these steps, the dataset from twitter was extracted via twitter API. The dataset contains 5000 tweets and the size of the dataset is 1362 KB.

B. Performance Analysis In Fig.II describe the number of neutral, positive and

negative tweets. The extracted tweets can be analyzed by using sentiment analysis. Approximately, out of 5000 tweets extracted from the Twitter, 1800 tweets are neutral tweets, 2510 tweets are positive tweets, and 690 tweets are negative tweets.

FIG.II SENTIMENT ANALYSIS OF TWEETS

C. Comparative Analysis Compare the Sample twitter dataset and whole dataset

using SA. The user who tweets more than the threshold value will be considered as a regular user. Based on the regular user, Performance is evaluated for the RS technique.

FIG.III COMPARATIVE ANALYSIS BETWEEN WHOLE

DATASET AND SAMPLE DATASET The accuracy of the sample dataset can be

calculated by using the Eqn(2) S = Number of successful predictions of the tweets N = Total number of tweets

Accuracy of the RS = S/N (2)

-10

-5

0

5

10

15

20

A B C D E F G H I J K L M

Sc or

e

Name of the regular user

Comparative Analysis SAMPLE

WHOLE

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

Fig. III describes the comparative analysis of both the data sets. The X-axis represents the name of the regular user. The Y-axis represents the overall score of tweets for the user. It proves that the RS technique provides more than 69% accuracy (2) for the twitter dataset. In Fig.III, user A is considered as the regular user, whose score of tweets for the user is 16 for the whole dataset that means it is a positive value. So, the user A is considered as a positive person. The sample dataset also represents that user A is a positive person whose score of tweets for the user is 6. Finally, this analysis concluded that the sample dataset provides the approximate result when compared with the whole dataset.

FIG.IV ACCURACY OF THE RS

FIG.IV represents the accuracy of the RS technique. The X-axis represents the size of the sample dataset and Y-axis represents the accuracy of the RS sampling technique. This graph concluded that sample size is directly proportional to the accuracy. When the sample size is increased, RS technique will provide the finest result.

V. CONCLUSION AND FUTURE WORK

The twitter dataset has been extracted via twitter API. Sentiment Analysis has been performed on the twitter dataset to analyze the sentiment of the tweets. The score for tweets has been calculated. After that, performance analysis is done. Results obtained from the data set is analyzed and polarity is identified. Futher, RS technique is applied to select the sample dataset from the whole dataset. Similarly, SA is done for the sample dataset and Comparative analysis is done between the sample dataset and the whole dataset. The prospective work is to apply more statistical sampling techniques for the twitter dataset and find out the preciseness for the techniques.

REFERENCES [1] Abdelghani Bellaachia and Mohammed Al-Dhelaan

“Learning from Twitter Hashtags: Leveraging Proximate Tags to Enhance Graph-Based

Keyphrase Extraction“,IEEE Conference on Green computing and communication, pp. 348 – 357, Nov 2012.

[2] Alec Go, Richa Bhayani, Lei Huang,” Twitter Sentiment Classification Using Distant Supervision”,Processing, pp.1-6, 2009.

[3] Apoorv Agarwal Boyi Xie Ilia Vovsha Owen Rambow Rebecca Passonneau,”Sentiment Analysis Of Twitter Data”,ACM Digital Library,LSM’11 Proceedings Of The Workshop On Languages In Social Media, pp. 30-38, 2011.

[4] Bahrainia S.A. ,Dengel A.,” Sentiment Analysis and Summarization of Twitter Data”,IEEE on Computational Science and Engineering, pp. 227-234, 2013.

[5] Bongsug (Kevin) Chae, “Insights from hashtag #supplychain and Twitter Analytics: Considering Twitter and Twitter data for supply chain practice and research”,International Journal Of Production Economics, Volume 165, pp.247-259, July 2015.

[6] Bhuta, S.; Doshi, A.; Doshi, U.; Narvekar, M.,”A review of techniques for sentiment analysis Of Twitter data”, Issues and Challenges in Intelligent Computing Techniques, pp. 583-591, 2014.

[7] Bruno Ohana , Brendan Tierney,” Sentiment Classification Of Reviews Using Sentiwordnet” 9th. IT&T Conference, Dublin Institute Of Technology, Dublin, Ireland, pp. 22-23, October 2009.

[8] Chatterjee, R.; Goyal, M. ,”Tactics of twitter data extraction for opinion mining”, International Conference on computing for Sustainable Global Development, pp. 761-766, 2015.

[9] Efthymios Kouloumpis, Theresa Wilson, Johanna Moore,” Twitter Sentiment Analysis: The Good The Bad And The OMG!”, Fifth International AAAI Conference On Weblogs And Social Media, 2011.

[10] Hemalatha I, Saradhi Varma, Govardhan,” Sentiment Analysis Tool using Machine Learning Algorithms”, International Journal of Emerging Trends & Technology in Computer Science, Volume 2, Issue 2, pp. 105-109, April 2013.

[11] Jeffrey Scott Vitter,”Random Sampling With Reservior”,ACM Transactions On Mathematical Software, volume 11, Issue 1, pp. 37-57, March 1985.

[12] Koto, F.; Adriani, M.,“The Use of POS Sequence for Analyzing Sentence Pattern in TwitterSentiment Analysis”, IEEE 29th International Conference on Advanced Information Networking and Applications Workshops (WAINA), pp. 547-551, 2015.

[13] Othmura M, Kakusho K, Okadome T, “Social Mood Extraction from Twitter Posts with

60 65 70 75 80 85

30 40 50

A C

C U

R A

C Y

(I N

P

E R

C E

N T

A G

E )

SAMPLE SIZE(IN PERCENTAGE)

ACCURACY OF RS

ACCURACY

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.

2016 FIFTH INTERNATIONAL CONFERENCE ON RECENT TRENDS IN INFORMATION TECHNOLOGY

Document Topic Model”, International Conference on Information Science and Applications (ICISA), pp. 1-4, 2014.

[14] Pavlos S. Efraimidis, “Weighted Random Sampling over Data Streams”,December 2010.

[15] Pulkit Goyal, Sapan Diwakar, “Data Mining And Analysis On Twitter”, January 14, 2011.

[16] Sarlan A, Nadam C, Basri S, “Twitter Sentiment Analysis”, International Conference on Information Technology and Multimedia”, pp. 212-216,2014.

[17] S. D. Lee, David W. Cheung, Ben Kao, “Is Sampling Useful in Data Mining? A Case in the Maintenance of Discovered Association Rules”, ACM transaction on Data Mining and Knowledge Discovery, Volume 1, Issue 3, pp. 233-262, Sep 1998.

[18] Septianto, G.R.; Mukti, F.F.; Nasrun, M.; Gozali, A.A.,”Jakarta congestion mapping and classification from twitter data extraction using tokenization and naive bayes classifier”, Asia Pacific Conference on Multimedia and Broadcasting (APMediaCast), pp. 1-6, 2015.

[19] Shahheidari S, Hai Dong, “Twitter Sentiment Mining:Domain Analysis”, Seventh International Conference on Complex, Intelligent, and Software Intensive Systems(CISIS), pp.144-149, 2013.

[20] Shahana P.H,Bini Omman,“Evaluation of Features on Sentimental Analysis”, International Conference on Information and Communication Technologies,Volume 46, pp. 1585-1592, 2015.

[21] Walaa Medhat, Ahmed Hassan, Hoda Korashy," Sentiment Analysis Algorithms And Applications:A Survey", Ain Shams Engineering Journal ,Volume 5,Issue 4,pp. 1093-1113, December 2014.

[22] Wenyu Hu, Baili Zhang, “Study Of Sampling Techniques And Algorithms In Data Stream Environments”, 9th International Conference On Fuzzy Systems And Knowledge Discovery , pp. 1028 – 1034, May 2012.

[23] Yazhe Wang, Jamie Callan, Baihua Zheng, “Should We Use the Sample? Analyzing Datasets from Twitter’s Stream API”, ACM Transactions on the Web(TWEB), volume 9 Issue3,June 2015.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on October 17,2021 at 14:55:15 UTC from IEEE Xplore. Restrictions apply.