Twitter-Sentiment-Analysis--BigData

Introduction:

• Sentiment analysis is used to analyse text data inorder understand the underlying sentiment (positive or negavtive).
• Sentiment analysis uses Natural language processing(NLP) and machine learning to determine emotional intent behind a communication.
• Twittershutdown, Riptwitter was trending when Elon Musk took charge and hundreds of twitter employees send in their resignations.
• This project will perform Sentiment anlaysis on tweets collected for hastags #Twittershutdown, #Riptwitter, #Elon Musk and so on.

Problem Statement:

• Collect tweets using Twitter Api by launching an AWS EC2 instance, stream the tweets using Kinesis firehose and store the data in AWS S3 bucket.
• Create a binary classification model to classify sentiment of each tweet (positive or negative), label= sentiment(0>negative, 1>positive) .
• Create a Quicksight dashboard for the data collected and also predictions from the classification model.

Tools used:

• AWS, Twitter Api, Amazon Kinesis firehose, Pyspark, Amazon Quicksight, Databricks

Data

Data Collection:

• 399333 tweets were collected using Twitter Api and stored in AWS S3
• Using Databricks environment connect to S3 bucket and mount the data by creating a spark session.

Data preprocessing:

• Created a pyspark dataframe object twitter data.
• Checked for null values and drop rows with Null values.
• Converted create_at to datetime column.
• Used regular expression to clean the tweet, location columns. .
• Textblob which is a library in python for text analysis can be used to assign sentiment for each tweet.
• Created a column Sentiment which will have values 0 if a tweet has nagative sentiment and 1 for positive sentiment.
• After cleaning we have 135,083 tweets out of which 45,760 tweets were with positive sentiment and 89,323 were tweets with negative sentiment.

Model:

Feature Engineering:

• Using library Tokenizer convert tweet column to lowercase and split it by white spaces, outputColumn="tokens"
• Remove stopwords from tokens using library StopWordsRemover,outputColumn="filtered" .
• Convert filtered tweets into matrix of token counts using CountVectorizer library,outputColumn="cv" .
• Inverse document frequency (IDF) library will check for relevant words in the tweet and remove sparse words, outputcolumn = "1gram_idf".
• Ngram (n=2) library is feature transformer that converts the input array of strings into an array of n-grams, outputcolumn= "2gram".
• HashingTF will map a sequence of terms to their term frequencies using the hashing trick, numFeatures=20000,outputcolumn= "2gram_tf".
• Again perform IDf to remove sparse terms, outputColumn="2gram_idf"
• VectorAssembler will merges "1gram_idf", "2gram_tf" columns into a vector column="rawFeatures"
• ChiSqSelector will select categorical features from rawFeatures, outputCol="features" and reduce the number of features to 16000

Model Development and Evaluation:

• Data was split into 90% train and 10% test data.
• Sentiment column is the label. 0 > negative sentiment, 1> positive sentiment
• We tried RandomforestClassifier and Logisticregression models to classify if the tweet in the test data is positive or negative
• With RandomForestClassifer we acheived 66% accuracy and 72.87% Roc-Auc score
• Classification report for RandomForestClassifer as follows:
• LogisticRegression gave us an accuracy score of 90.425 and Roc-Auc score of 92.83
• Classification report for LogisticRegression as follows:

LogisticRegresssion model gave us better accuracy, the predictions are saved back to AWS S3 bucket

QuickSight Dashboard

Tweets post data preprocessing:

• 66% of the tweets were with negative sentiment
• Top 10 location in terms of number of tweet, location as a feature doesnot seem to be a contributor in tweet sentiment as they almost have equal percentage of both negative and positive tweets

Predictions:

• from the 8.9K negative tweets, the model was able to correctly predict 8.19K tweets as tweets with negative sentiment.