11-5-17 Report

Amal

• beautified notebook and streamlined/standardized ML pipeline
• added more classifiers
• added some comments to explain some esoteric stuff I forgot why I did
• also guys for future reference please BRANCH instead of FORK from the repo - its much easier to manage IMO
• converted apiTest nb to python tweet crawler and made run script and sample tweet json

TODO

• used word2vec for this pass -- but will also try w/ LDA/LSI/docv2vec as well
• @Fred the above featurization schemes should answer your question as to the different ways to featurize the tweets - I will tentatively assign comparison to you
• Note also there is no preprocessing - which should improve model performance
• We also need to tune hyperparameters

Comments

• Any of the above are good things for you guys to do
• when in doubt copy the notebook when you edit to avoid merge conflicts
• guys for future reference please BRANCH instead of FORK from the repo - its much easier to manage IMO

Midterm report TODO

Amal

• @Fred - it is very easy to use a mlp for classification - in general it will achieve similar performance as random forest but will take longer to train, making iterating other aspects of our model pipeline more difficult
• I will add a mlp test in the notebook and include the time to demonstrate my point only if we get a larger dataset will a richer model like a NN be usefull
• also I am seeing a submission link for a crawler code - I'll de notebookify what I already did and make a sample tweet json
• I'll add two new classifiers (SVC, MLP)
• denotebookify crawler and create run script as well as sample tweet json file

Jonny

• also @Jonny Hurwitz it is important you include the ROC and PR/RC plots and associated AUC statistics
• also pls add f1 score - it is a good way to combine recall and precision
• I'll add two new classifiers (SVC, MLP) and we can include the plots for those guys too maybe have a quick discussion of relative merits
• I would also say a discussion of other features not currently being used is warranted (tweet date, user metadata, images, etc...)
• when you get a chance pls resolve merge conflicts and submit a pull request

11-2-17 report

Amal

• took a first pass at sentiment analysis
• Note the general workflow in the notebook
  • generate vectors to represent words
  • aggregate word vectors for each tweet into a single vector for each tweet
  • train a classifier on tweet vectors
• was going to annotate w/ comments, but think its easier to explain in person if you don't understand

TODO

• used word2vec for this pass -- but will also try w/ LDA/LSI/docv2vec as well
• Also --- will test other classifier models
• Note also there is no preprocessing - which should improve model performance
• We also need to tune hyperparameters

Comments

• Any of the above are good things for you guys to do
• when in doubt copy the notebook when you edit to avoid merge conflicts

TODO

• tweet preprocessing - Amal
• tweet featurization - Amal
• tweet clasification - Amal
• comparing word featurization - Jonny
• comparing word composition - Fred
• comparing tweet featurization - Fred

IDEAS

• extending a featurization mtd w/ twitter metadata
• including external link metadata
• [ ]

MVP Problem: mood detection

-Need to featurize tweets -First solve the problem of classifying tweets as positive or negative

Tasks -Analyze images -Pagerank for external links to determine reputable sources and then parse the page -Make amount of shares/likes per tweet a feature -Bag of words -word2vec -LSI -LDA -doc2vec

-Best way for featurizing words -Best way for aggregating word vectors into document vectors -Best hyperparameter sweep for training the model

Amal -MVP with mood detection

Jerry

Jonny

Fred