Time Series Forecasting of Political Bias in Twitter and GAB.

This is the git repository of our paper

Introduction.

In this study, an innovative research venture is embarked upon, with the aim of predicting political bias in online social media platforms, such as Twitter and Gab. This is framed as a time series forecasting challenge, where a time series of politically biased data is input, and the forecasting model is used to predict future developments or the next steps in the series. By accurately anticipating how political sentiments evolve over time, insights into the formation of ideological groups and the dissemination of biased information within social media platforms are sought. This knowledge can be deemed invaluable in devising effective strategies to combat misinformation and echo chamber effects. It is worth noting that previous research in forecasting political bias data is limited, rendering our exploration groundbreaking in this field. Existing time series forecasting models are utilized to assess their suitability for this task. Through an analysis of how well these models forecast the time series of political leanings, an attempt is made to uncover their strengths and weaknesses in capturing the temporal dynamics of political bias. As each social media platform has its unique user engagement, popularity, and political ideologies, experiments are conducted using two social media datasets, Twitter and Gab, collected during the same timeframe. In summary, two main contributions are offered by our paper:

Contribution-1: A novel challenge of predicting political bias in online social media posts is introduced, which is pivotal for understanding the political stance of social media on various topics or events.
Contribution-2: Diverse time series forecasting models are experimented with to quantify trends in different political biases within a social media platform. In the analysis, two social media platforms with distinct political biases, Twitter and Gab, are examined.

Gist of used Timeseries forecasting models.

Model	Definition	No.of lookbacks(Previous days count)	No.of next days predicted
SARIMA	A statistical method used	0	1
	to forecast time series
	based on the average of lags and
	seasonality trends of
	previous time steps of data.
LSTM-I	A deep learning method which is	1	1
	an extension of RNN that is
	designed to learn a sequential
	data by preserving them.
LSTM-II	Same as LSTM-I but only	14	1
	difference is in LSTM-I we use
	Only one previous day for
	predicting next day but here
	We can use as many lookbacks
	as we want for the next day.
MULTISTEP FORECASTING	A LSTM model,combination of	14	7
	Above two models in which we can
	use as many previous days'
	patterns to predict next
	sequence of days.
GRU	A simplified version of LSTM in	14	1
	which it has only 2 gates
	reset and update whereas
	LSTM has 3 gates(input, output
	and forget gates.

NOTE: Multiple lookbacks are done vector concatenation and given as input for the models which have 14 days of lookback. And for predicting next 7 days in Multistep time-series forecasting a method called teacher enforcing (Teacher forcing is a method for quickly and efficiently training recurrent neural network models that use the ground truth from a prior time step as input) is established internally which uses previous day's ground truth as input for today and this process continues for vector of 14 days.

Datasets :

In this study, we made use of publicly accessible datasets obtained from Twitter in 2018 and Gab in 2018. The Twitter dataset comprises tweets that contain links to news articles related to political topics from selected news sources. This dataset covers the period from January 2018 to October 2018 and includes a total of 722,685 tweets. On the other hand, our Gab dataset from 2018 is extensive, containing 40 million posts, including replies, reposts, and quotes, all featuring URLs and hashtags. The data spans from 2016 to 2018. To ensure a fair analysis, we only considered a subset of Gab posts that shared news article URLs during the period from January 2018 to October 2018, which corresponds to the time frame of the Twitter data. It's worth noting that unlike Twitter, Gab data isn't specifically concentrated on any particular subject, which accounts for its larger number of posts.

Implementation of Methodology:

• 1. Initial data preprocessing:

  • At first, the raw data in the form of JSON is converted to a data frame.
  • Next, expanded URLs and tweet creation timestamps are extracted and stored in another data frame.
  • An additional Excel sheet is provided which contains media channel names with their sub urls and respective leanings.
  • Now from these datasets, we need to regroup all left-leaning media suburbs into left- left-leaning, all right media suburbs into right, all centre-leaning media sub URLs to center, all left media to left, and all right-leaning media to right-leaning. For that, HashMap has been implemented in which keys are leaning and values are the list of sub URLs that are associated with their respective leaning.
    
  • After iterating over the whole additional Excel sheet now hashmap has just 5 keys of left, right, center, left-leaning, and right-leaning with a list of values of sub URLs associated with them.
  • With this hashmap of 5 keys,5 different data frames are generated in which we get left-leaning media sub URLs to one data frame, right-leaning media sub urls to another data frame, right media sub urls to another data frame, left media sub urls to another data frame, and center media URLs to another data frame .
    -Later, the main task here is to split timestamps from the whole initial dataset to left, right, center,left-leaning, and right-leaning. For that to happen we use the presence of sub urls in main urls and categorize them according to 5 dataframes that were created before.
    
  • We will convert the objects type of sub urls and expanded URLs to string types in all data frames. -As the initial data frame is very huge it requires a huge amount of time to iterate. So to make it faster we will first convert to vectorized array. -Now, by using re .find() we will check if the respective sub URL is present or not in the main expanded urls. if present then we add timestamps to a list. -The above step is repeated for 5 times to get timestamps in 5 different lists of left, right, center, left-leaning, and right-leaning.
  • At present, we will have 5 lists of timestamps that are left, right, center, left-leaning, and right-leaning. (These lists have duplicates and that is required). -Getting how many duplicates for each timestamp in each list gives you the frequency of respective leaning tweets per day. So to get the frequency of each leaning tweets per day, we use hashmaps where in iteration itself if one timestamp is not in the key of hashmap then we put in it with value 1. If existed then we increment the respective timestamp value by 1.
    
  • At this point in time we will be having 5 hashmaps with timestamps as keys and frequencies as values. These 5 hashmaps are left, right, center, left-leaning, and right-leaning timestamps with frequencies.
    
  • From these hashmaps, we will be converting to 5 data frames .
    

• 2. Next step of data preprocessing:

  • As we have left, right, center, left-leaning, and right-leaning timestamps with frequencies we take a range of required data, here timestamps between May to October 2018 have taken and converted to time series because of outliers and fewer postings from January to May.
    

• 3. Application of grid search for SARIMA parameters and SARIMA MODEL FITTING.

  • Grid search is an algorithm with which we can select parameters for the model. It automatically tries all the combinations of SARIMA parameters for the dataset and gives the best combination which showed less RMSE. With these parameters, we fit 5 time series to the SARIMA model and we will be getting 5 different RMSEs for left, right, center, left-leaning, and right-leaning.

• 4. Keras tunning and application of LSTM-1, LSTM-2, GRU, and Multistep forecasting.

  • The time series have been split in a window of 60-80, 20-40 percentages of training and testing datasets, and applied Keras tuner which gives the best combination of LSTM parameters with low rmse of all.

Outputs :-

The outputs folder contains 4 images of multi-line plots conveying which average tweet frequency or Gab posts, and total Tweet likes or Gab post likes per month, 2 images of stacked plots in which 1st image contains the Average Compound score per day of tweet content, and another image contains the Average Compound score per day of Gab post content. These are the results that were obtained after applying VADER sentiment analysis. And 2 Excel sheets in which one is RMSES of Twitter posts, likes and another one is RMSES of Gab posts and their likes which were recorded after forecasting with above models.

• 5. Conclusion.

  • All these models are compared based on RMSE'S (Root Mean Squared Errors) of all left, right, center, left-leaning, and right-leaning time series after fitting to the above models.

Note:

• 1. The same methodology Implementation has been applied to the GAB dataset as well but the way of preprocessing will be different because of the different JSON format.
  

• 2. 2 tasks have been implemented with this methodology and the second task is a more specified version of task-1. Task-1 is predicting polarized tweets or Gab posts next few days, whereas task-2 is adaption rate which is predicting Tweet likes or Gab post likes.
  

• 3. The only difference between task-1 and task-2 is that for task-2 in addition to expanded urls, we extract likes count.