This project aims at exploring the validity of a simple hypothesis that the presidential tweets influence major national polls significantly. In order to perform this analysis, data mining from text strings and their correlations with numerical data from polling agencies are needed. We decided to use one of the most powerful techniques in the field, namely ‘Sentimental Analysis (SA)’. Sentiment analysis is used to graphically represent selected human sentiment using major keywords that associate with similar sounding words in any given public internet comment/tweet/feedback. Over the past few decades, corporate giants like Amazon and Google, along with many influential politicians, have adopted the notion of feedback from their customers, clients and fellow citizens. Twitter data have been successfully used to determine the role of fake media stories in recent presidential elections [ 9 ] [ 10 ]. There have been several reviews of tools for analytical mapping of opinion mining and sentiment analysis research that are used for this literature survey [ 11 ].

Sentiment analysis is a fast growing research area in computer science, making it hard to perform a comprehensive literature survey that would encompass all developments and findings in the area. In recent times, using the Web 2.0, consumers express and share their opinions regarding day-to-day activities and global issues on social media. and this provides a transparent platform to share views across the world. Methods have been developed to measure mass sentiments using these electronic Word of Mouth (eWOM) [ 12 ]. Guided by the method described in Mantlys, Graziotin and Kuutila [ 13 ], we have used Google Scholar and taxonomy of research topics to select publications related to the use of SA in social media texts from twitter and Facebook micro-blog texts. Success of quantitative and qualitative analysis of sentiments and latent factors using online review of doctors based on a regressive analysis show strong correlations of state-level measures of quality healthcare with patient likelihood of visiting their primary care physician within 14 days of discharge [ 5 ]. Political organizations around the world have been increasingly using Twitter micro-blogging tools that give platform for divergence of political discourse compared to traditional media [ 2 ]. The SA tools are used to study the consumer satisfaction in various fields such as medical visits, airline reservations and hotel industry [ 5 ], [ 14 ]. While these are confined to a section of the general population, it is not clear from these studies whether the SA tools can quantify the opinion of a wider population and relate them between two independently measured quantities namely, sentiments derived from Twitter micro-blogs and national polls. This project aims at examining the relationship between them using the SA tools. The intellectual merit of the project is to understand the mass following in a society by the leader who can effectively communicate using modern social media. The broader perspective is to determine the specific SA tools that can yield useful results in the field of capturing mass opinion that can be related to independently collected national polls.

• The first set of data consists of time series of Twitter micro-blogs posted daily by President Trump. The second data set comprises of daily presidential approval ratings.

TABLE I Tweet Count Distribution Sr.No. Source Tweets Retweets 1 All Devices 9438 10645 2 iPhone 7298 8456 3 Android 1478 1479 • Twitter micro-blog data is collected from an archive website which organizes daily postings by date, time and device. The website “https://www.trumptwitterarchive.com" allows users to export data in CSV or JSON format. It can display retweets, permits word search and retweet count. The Table 1 gives the tweet count distribution over the period of this study. • It is clear from Table 2 that over the past three years, President Trump micro-blogging post have significantly transitioned from android system to iPhone. Hence, the source data used in this research is taken from iPhone devices only. The two types of data for this study is collected from public sources between January 20,2016 and January 1, 2019.

The Real Clear Politics team was kind enough to provide with average approval score in percentage data file for Trump going back to the first date. The approval ratings cross checked with other data sources collected from the following websites: (1) https://www.realclearpolitics.com, (2) https://www.projects.fivethirtyeight.com/trumpapproval-ratings

The data analysis was carried out, following the protocol described below, using Python development language. Here, we describe the role of each step in the analysis process. Extraction of relevant data is possible even without sufficient know-how of Natural Language Processing(NLP). The complexity of code involved in this project is benign and execution is performed using Python kernel in Jupyter notebook. Pandas library was first imported for reading and writing purposes. The source CSV file which uses comma delimiter is read using the read_cv command.

Fig. 2. President Trump’s approval rate time series. This is divided into three periods. The time before the dotted line represents his approval rating prepresidency whereas the solid line represents post-presidency approval rating. The time between is the interim period where is yet to take an oath for Presidency but has been elected as President of the United States Of America. • Number of words : The word length of each tweet is calculated using the split function in Python programming language. • Number of characters : The total number of characters are calculated based on the previous logic of sentiment score. Spaces are also included. • Average word length : The sum of the length of all the words when divided by the total length of the tweet. • Number of stopwords : ‘Stopwords’ is calculated in order to understand the extent information which will be lost after processing the data. Natural Language Toolkit(NLTK) is a basic NLP library in python, which is used for this calculation. • Number of special characters : Special characters like hashtags represent extra information in a tweet and are undesirable during data analytics. ‘Starts with’ function is invoked since hashtags (or mention) generally appear at the beginning of a word. • Number of numeric : The method involved for the total count of numeric is similar to that of word count method. It is calculated using digit() function which lists all numeric present in a tweet. • Number of uppercase words : Word such as ‘hate’ which expresses anger and rage, is often commented in uppercase letters. Thus, converting uppercase words to lowercase is pertinent before advancing to text processing. of the same word. ‘Text’ and ‘text’ is the same word but during the word count step, they are considered as two different individual words. • Punctuation removal : Punctuations do not supply any extra information during the data analytics process. Hence, cleaning up such instances allows to reduce the size of the training data. • Stopwords removal : Commonly occuring words or stop words are removed from the source text as it creates undesirable redundancy in the training data. NLTK predefined library ‘stopwords’ is imported which consists of all the stopwords present in English language. • Frequent words removal : After detecting commonly occurring words in a general sense, it is important to remove these words from the text data. The top 15 most frequently occurring words in the training data is found, these words are removed. • Rare words removal : This method is similar to the common word removal process where rarely occurring words are located and removed. Since these words are very scarce, noise generally dominates the association between them and other words. Rare words with a more general form will have higher count and are calculated using value_count() function.[ 25 ] • Spelling correction : Twitter users often hastily post online which leads to profusion of spelling mistakes. This also creates superfluous copies of similar words. For example, “Analytics" and “analytics" is treated as two different words even though they are used in a similar tone. TextBlob natural language processing library is referred in this step to achieve spelling correction. • Tokenization : It is defined as the process of chopping up a piece of text into a sequence of individual words or sentences. TextBlob NLP library is highly efficient for performing this activity with a large dataset.[ 25 ] • Stemming : This method involves in suffice removal like ‘ing’,‘ly’ etc. using a simple rule-based approach. PorterStemmer is imported from the NLTK library and used with stem() and split() functions. • Lemmatization : The process of lemmatization is a finer method when compared to stemming. In this scenario, a word is converted into its root word instead of stripping the suffices. It accounts for vocabulary and executes morphological analysis to procure the root word. ‘Word’ sub-library is imported from TextBlob and lemmatize() function is used along with the lambda() function to perform this task.

Data cleansing protocol is crucial for accurate text and feature extraction. It allows in reduction of vocabulary clutter which will play an important role during advanced text processing. This is achieved by following the basic pre-processing steps on the training data.

• Lower case : All the tweets are transformed into lower case letters in order to avoid storage of multiple copies

All the pre-processing steps involved in data modification and cleansing are now complete. The final step involves extracting features from the refined text using NLP techniques. • N-grams : N-grams refers to combination of multiple words which are used together. For N=1, n-grams is called ‘Unigrams’ whereas ‘Bigrams’ and ‘Trigrams’ represent N=2 and N=3 respectively. Unlike Bigrams and Trigrams, Fig. 3. Word cloud after successful data cleaning. Nonessential words which represent noise are removed. This shifts the focus on the frequently occurring words which are pertinent towards sentiment analysis.

Unigrams do not consist of significant information. The n-grams principle is based on capturing the language structure i.e. what letter or word is supposed to follow a given one. The quantity and quality of context depends upon the length of the n-grams. TextBlob provides a ngrams() function which is deployed for data extraction from the twitter data module.[ 25 ] • Term Frequency : Term frequency is the ratio of word count present in a sentence to the length of the sentence. value _counts() function is called upon to perform this task. TF = (Number of times term T appears in the particular row) / (number of terms in that row) • Inverse Document Frequency : IDF is used to calculate the uniqueness of each word. It is the log of the ratio of the total number of rows to the number of rows in which the word is present. IDF = log(N/n), where N is the total number of rows and n equals to the number of rows in which the word exists.[ 25 ] • Term Frequency-Inverse Document Frequency (TF-IDF) : TF-IDF is the multiplication of the TF and IDF values. In order to achieve the TF-IDF value without calculating TF and IDF values, sklearn function is used. This function can also perform simple preprocessing tasks like lowercasing and removal of stopwords. • Bag of Words : BoW model is a simplified representation where a text (a sentence or a document) is represented in a multiset (bag) of words. Similar text fields should contain similar kinds of words and hence, will have a similar bag of words. Sklearn provides a separate function for this task called ‘CountVectorizer’. • Sentiment Analysis : Our original problem is to calculate the sentiment score for each tweet published by the @RealDonaldTrump twitter handle. TextBlob is used to calculate the polarity and subjectivity of each twitter feed The analytical methodology will provide the values for polarity and subjectivity. The sentiment score of each tweet and the corresponding published date is exported into a CSV file. The flat file is then imported as a table shown in Query 1 in Microsoft SQL Server named ‘SentimentScore’. A select query is fired to calculate the sum of sentiment scores for each day (Query 2). Approval rate is plotted on the Y-axis against sentiment score in the X-axis to achieve the final graph using Microsoft Excel charts [dbo].[SentimentScore] ( [SL NO] varchar(50), [TEXT] varchar(50), [CREATED_AT] varchar(50), [WORD_COUNT] varchar(50), [CHAR_COUNT] varchar(50), [AVG_WORD] varchar(50), [STOPWORDS] varchar(50), [HASHTAGS] varchar(50), [NUMERICS] varchar(50), [UPPER] varchar(50), [SENTIMENT] varchar(50) )

Query 1: Schema of SentimentScore table SELECT SUM(SENTIMENT) AS ’SA’ , CAST (CREATED_AT AS DATE) DateAdded FROM DATABASE.TABLE_NAME GROUP BY CAST(CREATED_AT AS DATE) ORDER BY CAST(CREATED_AT AS DATE)

Query 2: SQL code used for daily sentiment summation calculation

The absolute value of the approval rating, A(t), and the twitter sentiment score, S(t), can give insights into how news events can influence both metrics. To analyze the relationship between the two, we computed the cross correlation between the two time series at multiple lags l using the standard expression: S · A(l) = hS(t), A(t − l)i, where hx, yi represents the correlation coefficient between the data x and y. • Approval rate time series graph is created in Fig. 2 using the data collected from Real Clear Politics. • A Twitter sentiment time series is plotted in Fig. 4 where sentiment score of President Trump‘s microblog feed is scaled from 2016 to 2019 • Approval rate is mapped against sentiment score in Fig.

5. • The cross-correlation plot where in the X-axis lag is accounted for i.e. it denotes a delay in approval rating change using a number of days count. The Y-axis represents the normalized value. (Fig. 6) • Epoch plots are created for President Trump‘s approval rating and twitter activity sentiment score before the 2016 presidential election, during the interim period where he was elected but yet to take office and after he made the transition to the Oval office. Sentiment analysis is classified into two categories : • Polarity is a float value within the range [-1.0 to 1.0] where 0 indicates neutral, +1 indicates a very positive sentiment and -1 represents a very negative sentiment. • Subjectivity, on the other hand, is a float value within the range [0.0 to 1.0] where 0.0 is very objective and 1.0 is very subjective. Subjective sentence expresses some personal feelings, views, beliefs, opinions, allegations, desires, beliefs, suspicions, and speculations where as objective sentences are factual.

The primary purpose of this study is to understand the impact of President Trump‘s daily twitter feed on the Presidential approval rating. The data collected only consists of the tweets posted by iPhone device since his presidential term began. Retweets and other sources have not been taken into consideration because the intention is to understand the relationship between President‘s published twitter posts using the device, he uses most commonly and the impact it has on his popularity.

The graphs developed from the sentiment analysis has the following implications : • Sentiment score for Donald Trump‘s twitter activity before he took office is averaged at 0.25986. After he was elected President of the United States of America, the average sentiment score increased by 60 percent to 0.556785. This increment clearly states that President Trump‘s tweets have more influential words since he took over the Oval office. The epoch plot and data used here in Fig 7(b) was used to determine this result. • The plot in Fig 5 is unevenly distributed and leads to the conclusion that there is no direct correlation between the Presidentâs twitter activity and his approval rating. • However, after analyzing the results of cross-correlation plot, it is evident that approval rating of President Donald Trump is subjected to change based on his twitter activity. It takes five days on an average for the approval rating to be affected by any significant amount of twitter activity. The cross-correlation plot is hits the peak around the five day limit.(Fig. 6)

One possible explanation for this result revolves around the fact that news takes time to spread. This change is not instantaneous but rather gradual as the country will first be exposed to this news, then the people will digest and discuss amongst themselves which ends up shaping a common opinion within different groups of voters. This study has used data mining and analytical technique to verify the direct impact of social media platform on poll data. For future discussion, the increase of sentiment score in the post presidency period may be verified if such analysis is carried for the next one and a half years.