Power-of-Choice/sentiment_analysis

Directory Structure

├── train_dnn_text_MLflow.py                   # sentiment analysis jupyter notebook. Use MLflow to tracking training metric, hyperparameters and model.
├── plot_res.ipynb                             # jupyter notebook to produce figures
├── train_dnn_text_V3.py                       # python script for sentiment analysis
├── params.py                                  # args_parser for train_dnn_text_V3.py
├── util_text.py                               # utils functions for sentiment analysis
├── models.py                                  # models
├── data_preprocessing.py                      # text data cleaning for sentiment analysis
├── Sent140                                    # raw data for sentiment analysis
├── comm_helpers.py
├── README.md

Dataset Sent140 is download from "http://help.sentiment140.com/for-students"

Glove from "https://nlp.stanford.edu/projects/glove/"

• Pre-trained 200D GloVe embedding.

parameters applied for training

• batch size: b = 32
• learning rate: η = 0.05, without learning rate decay.
• number of clients: K = 314, each tweeter account is a client.
• number of selected clients for each communication round: m = 8
• local trainig update for each communication round: τ = 100
• The output labels 0 as positive, and 1 as negative.