Team Members
	- Sanjay Roberts 
	- Jeff Coady
	- Chris Haddad

Project Submission:
	- EDA.ipynb
	- model.ipynb
	- toxic_cnn.ipynb
	- Presentation.ipynb

Justification	
	- Our problem involves NLP and predictive modeling.  Jeff Coady has a background in linguistics, 
	- Chris Haddad has experience with predictive modeling, and Sanjay Roberts has some experience with both NLP and modeling.
	- Further, Sanjay Roberts has experience with Kaggle competitions and the rest of the team is anxious to learn about the process of competing.


Problems/Motivation
	- Detect toxic comments and minimize unintended model bias
	- Our goal is to build a model that recognizes toxicity in comments and minimizes unintended bias with respect to mentions of identities. 
	- https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/overview




Libraries and Tools (What you already know and what else you need to evaluate)
	- Known Packages:
		+ matplotlib
		+ pandas
		+ tensorflow
		+ scikit-learn
		+ scipy
		+ imbalanced-learn
		+ keras
		+ ipython-autotime
		+ psutil
		+ nltk
		+ gensim
	- Potential Packages:
		+ elmo
		+ bird
		+ spacy
		+ textblob
	
Data Collection
	- We will be using a Kaggle dataset.  The dataset is labeled for identity mentions and optimizing a metric designed to measure unintended bias. 
	- The dataset can be found here:
		+ https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data
	
Features:
	+ male
	+ female
	+ transgender
	+ other_gender
	+ heterosexual
	+ homosexual_gay_or_lesbian
	+ bisexual
	+ other_sexual_orientation
	+ christian
	+ jewish
	+ muslim
	+ hindu
	+ buddhist
	+ atheist
	+ other_religion
	+ black
	+ white
	+ asian
	+ latino
	+ other_race_or_ethnicity
	+ physical_disability
	+ intellectual_or_learning_disability
	+ psychiatric_or_mental_illness
	+ other_disability
	
Any Literature review
	- Our intention is to begin by reviewing the leading kernels here:
		+ https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/kernels
	- And the following papers:
		+ https://www.hergertarian.com/detecting-toxic-comments-with-multitask-deep-learning
			- ompl
		+ https://becominghuman.ai/my-solution-to-achieve-top-1-in-a-novel-data-science-nlp-competition-db8db2ee356a
			- kaggle and NLP
		+ https://arxiv.org/pdf/1802.09957.pdf
			- ensemble that outperforms individual deep or shallow models
		+ http://demo.clab.cs.cmu.edu/ethical_nlp/
			- ethics 
		+ https://medium.com/@nehabhangale/toxic-comment-classification-models-comparison-and-selection-6c02add9d39f
			- LSTM v CNN, tokenization, embedding
		
Required work detail before build model
	- data cleansing
	- Tokenize and Pad
	- FastText Embedding

What is the predictive task and model detail.
	- multinomial classification of toxicity classes
	- seeking to use ensemble methods
	- grid search 
	- k-folds
	- imbalance-learn
	
Model evaluation and selection strategy.
	- k-folds for cross validation
	- Matthews Correlation Coefficient for accuracy
	- start with LSTM and CNN, move to ensemble
	- use kaggle submissions to test against other competitors
		+ retrain, resubmit to kaggle, evaluate rank, retrain, ...

How a user is going to test the final model. is there any webpage/command line interface.
	- TBD

Tentative time line of activities.
	- Week of April 22nd
		+ read literature
	- Week of April 29th
		+ make ML ready dataset
	- Week of May 6th
		+ test LSTM using grid search for hyper parameter tuning, k-folds for bias analysis
	- Week of May 13th 
		+ test CNN, grid search, k-folds
	- Week of May 20th
		+ begin testing ensemble learning method
	- Week of May 27th
		+ continue testing ensemble learning method
	- Week of June 3rd
		+ final submission to kaggle with best results
	- Week of June 10th
		+ summarize results