Youtube Analysis

• The first iteration of this project was created for the UCSB Data Science Club with Matthew Peterschmidt, Shon Inouye, and Conor O'Brien (see original repo here)

Goals:

1. Perform sentiment analysis on video comments to better convey audience reaction to content
2. Create a functional dashboard to provide more usable and visual results

Technologies

Packages: Pandas, NumPy, NLTK, Plotly's Dash, and Scikit-learn
Models: Naïve Bayes, Logistic Regression, Support Vector Machine, K-NN, Neural Network, and Random Forest
Languages: Python, Javascript, HTML, possibly some SQL
APIs: Google Development's Youtube API

Dashboard (In Progress):

Table of Contents:

1. Youtube API Call
   • pulls comments for a specified video from the API
2. User Prompt to Select Video
   • prompts the user to enter one of three videos (OK Go's "Obsession" music video, President Trump's 2017 inauguration speech, and highlights from the 2014 FIFA World Cup between Brazil and Germany). These three videos were selected because their topics were somewhat controversial, allotting us a variety of sentiments in the comment section.
3. Machine Learning Jupyter Notebook
   • conducts natural language processing to parse comments
   • performs sentiment analysis through machine learning algorithms to classify data as positive, negative, or neutral
4. Dashboard Script
   • provides an interactive platform for visually analyzing results
   • written in Plotly's Dash, which is built on the Plotly.js, React, and Flask platforms

Preliminary Instructions:

1. Set up an Google Dev account and get a personal API key for accessing the Youtube API
2. Download the repo and open it with a text editor
3. Create a file named config.js in the main directory of the repo. Edit the script and type in your key with the following format (replacing the text within the single quotes with your key):
   • The secret key should be about 39 characters in length and consists of upper and lower case letters as well as numbers

var config = {
	SECRET_KEY : 'hfhfhfhfh12345fhfhfhfhfhfhf' 
}

3. In your terminal, cd into the master file and run the user_prompt.py script by typing the following:

   Python3 user_prompt.py

   • The script will ask you for some user input, follow the instructions provided in the terminal

4. Check that the comments were correctly outputted in the data/data.csv file

5. Run the machine learning and sentiment analysis code in a notebook by typing jupyter notebook YoutubeAnalysis.ipynb into your terminal.

   • Click shift + enter to run each box
   • Make sure to change the os.chdir() path to match your current directory
   • change the pd.read_csv('...') to pd.read_csv('data.csv') to work with the comments from the video you selected
   • NOTE: still in progress of cleaning this part up

6. For the Dashboard capabilites, type python -W ignore dash/Dashboard.py >/dev/null to run the plotly Dash application

   • for Windows, the slashes are reversed: Python3 dash\Dashboard.py
   • the -W ignore component of that command is to ignore the warnings to be printed in bash and the >/dev/null supresses any other output and downloading messages. This is generally not good practice, but for now it helps with readability of prior terminal commands, so I left it in place. 6.b. Once you have the dashboard running locally through your terminal, open up the link http://127.0.0.1:8050/ in your browser.
   • Hitting ^C followed by y in your command line will kill the dash.
   • Note: Right now the dashboard is running off of the data produced from Variables.py script within the dash folder. This is simply a python-only version of the jupyter notebook Youtube_Analysis.ipynb in the main file. Eventually, the flow of the different scripts and components will be smoothed out so there is less "jumping around" between files.

TO DO

• Create a more integrated/ cohesive workflow between scripts
• Run the ApiCall with user inputted video
• Finish Dashboard
  • add table of comments
• Improve API call output format (fix encoding issues)
• [N] Add cross validation, neural nets, and grid search to Machine Learning notebook
• Add gradient descent, fix class imbalance
• Add n-grams
• [N] Use AWS + grid search to improve model accuracy

part-of-speech (POS) tagger, the named entity recognizer (NER), the parser, the coreference resolution system, and the sentiment analysis tools Stanford CoreNLP, CoreNLP, LingPipe and SentiWordNet

recursive deep learning sentiment treebank

base forms of words, their parts of speech, whether they are names of companies, people, etc., normalize dates, times, and numeric quantities, mark up the structure of sentences in terms of phrases and syntactic dependencies, indicate which noun phrases refer to the same entities, indicate sentiment, extract particular or open-class relations between entity mentions, get the quotes people said, etc.

the part-of-speech (POS) tagger, the named entity recognizer (NER), the parser, the coreference resolution system, sentiment analysis, bootstrapped pattern learning, and the open information extraction tools

https://stanfordnlp.github.io/CoreNLP/

sentiment analysis: polarity analysis, categorisation, scale on emotion

Affective Norms for English Words, wordnet, RapidMiner or ViralHeat

topic modeling: latent dirichlet allocation, where words are automatically clustered into topics, with a mixture of topics in each document probabilistic latent semantic indexing, which models co-occurrence data using probability.

Dan Sullivan: “The idea there is you train a bunch of classifiers and then essentially they vote on it and you take the majority answer,” he said. “Heterogeneous ensemble algorithms use a combination of different kinds. So you might take a SVM [support vector machine], a perceptron, a nearest centroid, a naïve bayes [cluster classifier] and combine those results.

what accuracy score to aim for in terms of precision and recall is 0.8 or above

TF–IDF can be a “crude” analytics practice because it throws away a lot of information on syntax and words that are semantically similar, such as hotel and motel.

conditional random fields for NER event extraction: Alan Ritter and Sam Clark