Indian Election Trend Prediction Using Improved Competitive Vector Regression Model

Transcription

1 Indian Election Trend Prediction Using Improved Competitive Vector Regression Model Navya S 1 1 Department of Computer Science and Engineering, University, India Abstract Election result forecasting has become an important requirement in many emerging markets. Industries, stock markets etc are strongly depend on election forecasting results. Most accurate election forecasting is needed and this necessitates accurate models to forecast. In this paper, we propose a social network based election forecasting model, which uses tweets collected week wise to build the model and use the model to predict for next weeks. Current models for forecasting are not adaptive and highly dependent on sampling time during which the public opinion was collected, but deviating from this we design a continuous trend prediction model. We test this model for Indian elections. I. INTRODUCTION Political elections, especially presidential election and House race, are extremely influential events in INDIA. Many consulting agencies, interests groups, think-tanks, journalists, and political analysts have devoted tremendous effort to predict or forecast the outcome of the election. These predictions are influential in two aspects on the one hand, the voting population and interest groups are interested in the predictions. On the other hand, the candidates and parties may adjust their campaign strategies according to the predictions. Until today, the highest authority resource for predicting the Indian election is opinion polls, which randomly choose small samples of the voting population and aggregate their opinions. Due to the nature of the Indian elections, two kinds of relevant opinion polls, nationwide and state wide, are relevant to election prediction. The sampling size of each poll is usually about one thousand or some comparable number. The candidates for the polling investigation are either registered voters or likely voters. Due to the importance of the presidential election, there were many different polls administered by different agencies. However, polls based forecasting has shown limitations. It has been criticized for survey design, the sampling bias and so on. Indeed, the low percentage of sampling rate inevitably introduces sampling bias and causes prediction errors. Recent research has shown trend of increasing skepticism on polling predictions of election outcomes. A recently emerging trend in data analytics is to rely on social media data from the web for prediction and forecasting, where the actions of individual Internet users can be properly pooled to indicate macro trend. But this kind of social media based election forecasting is new to INDIA and this paper we propose a improved competitive vector regression to predict the election trend based on data collected from social networks. We evaluate out approach for a sample data set from twitter and prove the accuracy. II. RELATED WORK In this section we survey the current solutions for election forecasting. DOI: /IJRTER MAJ7L 1

2 In [1], authors employed LIWC text analysis software to investigate whether Twitter is used as a forum for political deliberation in German federal election. Their results show that Twitter is indeed used extensively for political deliberation, where the mere number of messages mentioning a party reflects the election results. Moreover, joint mentions of two parties are in line with real world political ties and coalitions. In [2] authors forecasted the vote share of the 2010 UK General Election with the motivation of wisdom of the crowds, employing the ARIMA model that accurately predicted the votes for the Labour party, Liberal-Democrats and Conservatives. However, there are counter argumentsthat question whether prediction using Twitter data is reliable enough. They find that electoral predictions using the published research methods on Twitter data are not better than chance In [3] authors found no correlation between existing analysis techniques using Twitter messages and election outcomes. In [4] authors illustrated non-significant correction between Likes of candidates in Facebook and their share of votes in the election. Their analysis indicated that this was due to the bias in the social network itself; and may be solved by developing robust models and accumulating abundant data. Note that it is easy to make fake Twitter follower In [5] authors model successfully predicted the winning party in all the 50 states of the 2012 U.S. presidential election. However, the details of his model have not been published. In his interview, he talked about the importance of big data, and distinguishing signal and noise as the key to improving the accuracy of prediction. In [6] discussed how to refine noisy tags and geo tags using image features and ecological phenomena including ground snow cover, snow fall and vegetation density. In [7] author used more than ten million Flickr images for semantic understanding and topic modeling. In [8] tried to better understand and evaluate image data by utilizing social curation data. In [9] authors tried to enhancing the annotation of Flickr images via Logistic Canonical Correlation Regression. Motivated by these studies, as well as the attempt author in [110] to use Flickr images to forecast the 2008 presidential election and the sales of popular IT products and achieved promising All Rights Reserved 389

3 III. PROPOSED SOLUTION The architecture of the proposed solution Figure 1: Architecture diagram Twitter data is extracted week wise and all politics related tweet is alone extracted from the tweets. The related tweets are found by defining a key word dictionary for Indian election with involved parties, personalities and issues. Each tweet is tokenized and the tokens are matched to dictionary to find if the tweet is related to politics or not. The related tweet is then analyzed for sentiment and based on sentiment analysis results week wise popularity is calculated. In our model, popularity is not only count of positive sentiment of a party, but also the negative sentiment of other parties. Once week wise popularity is calculated for each week for each party, ARIMA model is build on the popularity. With ARIMA model, the popularity for next and subsequent weeks can be predicted. Popularity distribution for week wise for each party can be seen in our system. IV. RESULTS We implemented the proposed system in Java for three different classifier ARIMA, CVAR (competitive vector auto regression) and Improved CVAR for different weeks of twitter political comments. Figure 2: Creating Data Set All Rights Reserved 390

4 In the system tweeter data set folder is given and the how many weeks of data for which model must be used. Figure 3: Analyzing tweets For each tweet which is a political tweet, sentiment is extracted from the tweet and the popularity is calculated. Figure 4: Forecasting Once forecasting is completed, we can check the result of forecasting for next week. Figure 5: Popularity results The popularity values for parties over week wise is predicted using the improved competitive auto vector regression model and the result is All Rights Reserved 391

5 Figure 6: Forecast Accuracy The accuracy values for week wise distribution are compared and from the graph we see that the proposed Improved competitive vector auto regression system is able to achieve higher accuracy than other prediction models. V. CONCLUSION We implemented the proposed improved competitive vector based regression to predict the election trend. From the results we showed that our system was able to predict the trend in line with actual results every week wise. Right now we implemented only for national level elections and we will extend the solution for state wise election also in our future work. REFERENCES [1] A. Tumasjan, T. O. Sprenger, P. G. Sandner, and I. M. Welpe, Predicting elections with twitter: What 140 characters reveal about political sentiment, in ICWSM 10, 2010 [2] F. Franch, Wisdom of the crowds: 2010 UK election prediction with social media, JITP, vol. 9, no. 4, 2012 [3] D. Gayo-Avello, P. T. Metaxas, and E. Mustafaraj, Limits of electoral predictions using twitter, in ICWSM 11, [4] F. Giggliest, If likes were votes: An empirical study on the 2011 Italian administrative elections, in ICWSM 12, [5] N. Silver, The Signal and the Noise: Why So Many Predictions Fail-but Some Don t. Penguin Press, 2012 [6] H. Zhang, M. Korayem, D. J. Crandall, and G. LeBuhn, Mining photo-sharing websites to study ecological phenomena, in WWW 12, [7] G. Kim, L. Fei-Fei, and E. Xing, Web image prediction using multivariate point processes, in SIGKDD 12, [8] K. Ishiguro, A. Kimura, and K. Takeuchi, Towards automatic image understanding and mining via social curation, in ICDM 12, [9] L. Cao, J. Yu, J. Luo, and T. S. Huang, Enhancing semantic and geographic annotation of web images via logistic canonical correlation regression, in ACM MM 09, [10] X. Jin, A. Gallagher, L. Cao, J. Luo, and J. Han, The wisdom of social multimedia: using flickr for prediction and forecast, in ACM MM 10, All Rights Reserved 392