Train a classifier to recommend relevant tweets.

• character-level neural network with bi-GRU architecture based on Tweet2Vec implementation

• Short test: THEANO_FLAGS='floatX=float32' python train.py

• Train: THEANO_FLAGS='floatX=float32' python run.py

• pymongo

• tweets with the Computer Science conference hashtags
• ArchiveTeam JSON Download of Twitter Stream 2017-02

It is important to provide input samples equally balanced between all the classes otherwised the results are skewed towards the most frequent classes.

1. CS topics

Dataset: 5 classes * 2,000 tweets each = 10,000 tweets in total Random guess: 0.2 uniform probability distribution Model: 0.234375

2. ML vs NLP ?

Dataset: 2 classes * 5,000 tweets each = 10,000 tweets in total Random guess: 0.5 uniform probability distribution Model:

3. CS vs random tweets

Dataset: 2 classes * 30,000 tweets each = 60,000 tweets in total Random guess: 0.5 uniform probability distribution

• Run 1

split: 0.8 x 0.1 x 0.1

learning rate: 0.3

Epoch 12 Training Cost 0.00590342712402 Validation Precision 0.776462743791 Regularization Cost 3.82598352432 Max Precision 0.859375

Test: 0.6875

• Run 2

split: 0.6 x 0.2 x 0.2

learning rate: 0.3

Epoch 11 Training Cost 0.00385822719998 Validation Precision 0.742083333333 Regularization Cost 3.74113941193 Max Precision 0.8125

Test: 0.8125

• Run 3

split: 0.6 x 0.2 x 0.2

learning rate: 0.1

• Tweet2Vec