Find the Datastory Webpage here
The outcome of the 2016 US election were a surprise, few experts predicted a Trump win. Since then, it has been revealed that the Russian might have had an influence on the election through their action on social media. To test this claim, we will compare the IRA dataset and the "2016 Elections Poll". Moreover, the scope of the tweets dataset does no stop to the election. The dataset was last updated on March 2018, so we can also draw insights from the current presidency. Is there a correlation between the russian tweets and the approval raitings ?
First, we will classify and understand the dataset.
- We want to start by defining the classes: what are the words associated with "RightTroll" or "LeftTroll". What is a typical right/left troll?
- Quantify each variables. # of tweets by subject/target.
Correlation: we want to understand the effect of the tweets.
- How does a tweet activity correlates with approval ratings? Is it significant?
- Is the tweet effect more important during an election campaign or a president's tenure?
- Are tweets more influential against Trump or Clinton?
We use the IRA dataset given in the course. We use two additionals datasets from FiveThirtyEight: "2016 Election Poll" (https://www.kaggle.com/fivethirtyeight/datasets) and "Trump Approval Rates" (https://projects.fivethirtyeight.com/trump-approval-ratings/?ex_cid=rrpromo) We use Wiki Portal Curent Event (https://en.wikipedia.org/wiki/Portal:Current_events) for scrapping information on certain dates.
- Nicolas Gandar : Datastory Writting / WebScrapping / Correlations / Backgroung historical research ...
- Jean-Baptiste Prost : Datastory Writting / Word2Vec / WebScrapping / Bokeh plotting ...
- Antoine Spahr : Correlations / D3 Plots / WebPage construction / Datastory Writting ...
Everyone has help/supervised/developped the work of each other
Out work is gathered in the Final.ipynb files that is composed of the following notebooks (each notebook can be found it the Notebooks folder).
PreProcess_Word2vec_TopWords.ipynb
In this notebook we focused on the treatment of the tweet’s content.
- First, the tweets are cleaned and tokenized.
- Then a Word2vec model is used to represent each word in a 200 dimensions vector according to it linguistic context. We were hopping that the model would enable to observe cluster of similar words. To do so, we reduced to output 9 component with a PCA. We 2d-plotted each combinaison of those nine components but none of them gave a clear clustering. We could assed that the model was effective by simply looking at similar words by their cosine similarity.
- The most popular words, hashtags and mentioning were computed in this notebook. Moreover, we have assigned a continuous score characterizing each expression by it troll orientation (right or left). this provided additional information on the vocabulary used by the two trolls category.
- Finally, we displayed the top hashtag per day.
WebScrap.ipynb
In this notebook, we defined the (25) topics of the tweets.
- A first short keyword list for each topic was first written. Then, the same Word2vec model was used to extend the list. The model enables to find similar words in our vocabulary. As a result, the lists were enriched by ten folds.
- As we noticed that each tweets topic over time had spiking behavior, we wanted to understand what were the cause of those peaks. Hence, we implemented an event detector function. We defined a adapting threshold to each topic : mean ± 2.25 * standard deviation . When a spikes crosses the threshold, the date is retrieved and the corresponding page on Wiki Portal CurrentEvents is scrapped. Then, we look for potential word matching between the content of the topic list and the content of each paragraph of the web page text. If a sufficient match occurs, an event is detected.
The function could still to be optimized but some very interesting output are provided
Poll_Descriptive_Analysis.ipynb and TrumpPresidency.ipynb
Those notebook integrates data from several pollsters during the election campaign (autumn 2017) and the Trump’s presidency respectively. They follow the same structure:
- The polls data is extracted and the average over all pollster of a given day is taken
- Then the data is smoothend with a rolling mean of 7 days
- The number of Right/Left tweet per day is obtained from the IRA dataset
- Then the polls data and the tweet data are merged based on the date. Then the correlation between the Trump/Clinton (Approval/Disapproval) Polls and the Right/Left activity is assessed via scatter plots and spearman coefficients calculation.
- The correlation is also studied similarly with the number of people reached by the tweets instead of the number of tweet. The correltion between the polls variation and the tweet activity is also studied (the polls variation is obtained by the derivative of the polls with regards to time).
- Finally the Correlation is studied between the number of tweets for all the categories obtained above and the polls values.
In all cases no significant correlations were found.