Stress Detection on Social Media (presentation)

Social media is a major platform where people express their worries and stresses across the world. Learn2Relax was built in order to analyze content and identify stress from Reddit dataset by deploying NLP techniques. Word embeddings were pre-trained on unlabeled data and deployed by both discrete and neural supervised models.

• Clone the GitHub repository

  git clone https://github.com/Evaaaaaaa/Learn2Relax.git

• Change working directory

  cd Learn2Relax

• The model is tested on Python 3.7 with dependencies listed in requirements.txt. To install these Python dependencies, please run

  pip install -r requirements.txt

  Or if you prefer to use Conda

  conda install --file requirements.txt

Notice: If NLTK package cannot be installed via above command, make sure you have Xcode installed if you use MacOS and you are not using Conda. To install the “minimum version” of Xcode, simply download the Command Line Tools DMG file from here and follow the installation instructions. If you are using Windows/Linux and the installation of NLTK does not work, try

sudo apt-get install python3 python3-pip ipython3 build-essential python-dev python3-dev

then install NLTK package via

pip install nltk

• Install tensorflow for GPU to run BERT model on GPU

  pip install tensorflow-gpu==1.15

• Download and install Docker to create a containerized application for the inference demo. If you are new to Docker, here’s a quickstart guide.
• The models use preprocessed data files in the data/preprocessed repository. However, if you want to reproduce the tokenization steps from scratch using raw datafiles in data/raw, you need to install NLTK datasets/models

  python configs/config.py

The live demo is deployed and scaled up online using Google Kubernetes Engine(GKE) and Google Cloud Platforms(GCP). If you want to run the app on your own, below are three options:

• To run the Streamlit web app locally, make sure dependencies listed in requirements.txt are installed then run

  streamlit run app.py

  If no browser window pops up, point your browser to the External URL and you will be able to see the app as below

• To create a containerized application locally

  • in the project directory Learn2Relax, run

    docker build -t learn2relax-streamlit:1.0 .

  • you can run this image as a container via

    docker run -p 80:80 learn2relax-streamlit:1.0

  Point your internet browser to localhost:80 to see the app.

• To deploy the app via GKE with your own GCP account:

Prerequisites

1. Google Cloud SDK
2. Kubenetes SDK, run the following command to install
   

   gcloud components install kubectl

3. Docker (installation guide in above Additional Setup section)
4. GCP account with your GCP project ID {PROJECT_ID}, which you can find in the GCP console
5. Domain name

Workflow

1. Dockerize the app
   

   docker build -t gcr.io/{$PROJECT_ID}/learn2relax-streamlit:v3 .

2. Test the container locally and point your internet browser to localhost:80 to see the app
   

   docker run -p 80:80 gcr.io/{$PROJECT_ID}/learn2relax-streamlit:v3

3. Push image to Google Container Registry (GCR)
   

   gcloud auth configure-docker
docker push gcr.io/{$PROJECT_ID}/learn2relax-streamlit:v3

4. Create a container cluster (in this example with 2 nodes). If you have already created a cluster with the gcloud container clusters, only the last step is necessary.

   gcloud config set project {$PROJECT_ID}
gcloud config set compute/zone {$COMPUTE_ZONE}
gcloud container clusters create {$CLUSTER_NAME}
gcloud container clusters get-credentials {$CLUSTER_NAME} --zone {$COMPUTE_ZONE} --project {$PROJECT_ID}

5. Use Google-managed SSL certificates for your domain name: Edit the certificate.yaml from the cloned repo and change the host name to your respective choice. You will need to create an A record for your chosen host name with the IP address reserved above. Once you edit the certificate.yaml, run
   

   kubectl apply -f certificate.yaml

6. Deploy app to GKE: Update the deployment.yaml and replace PROJECT_ID with your project ID. Once you update the yaml file, execute

   kubectl apply -f deployment.yaml

7. Apply the Ingress Configuration

   kubectl apply -f ingress.yaml

You can now head over to the Google Cloud console and under Kubernetes Engine -> Services & Ingress you can see the Ingress being created. Wait for the Ingress to be created before you continue. Once completed, you can now visit your deployed application. For my demo you can visit: https://www.datatranslator.space/

• Features for the labeled dataset were generated by five different feature extraction methods: unigram TF-IDF, bigram TF-IDF, Word2Vec with TF-IDF as weights and BERT embeddings.
  Word2Vec embeddings were also trained with 190k unlabeled Reddit posts.

• After feature extraction, 9 classification models were trained: Logistic Regression, Naive Bayes, SVM, AdaBoost, Gradient Boosting, Decision Tree, Random Forest, XGBoost and BERT.

Best model for each featurization technique and their performances:

• Recall is the most important metric here because we want to best prevent misclassification of stress posts as non-stressful which helps us better understand the stressful contents in social media.

• BERT is the most robust model with all four metrics the highest.

• All models are able to provide a confidence score in addition to prediction.

Picture below shows word similarities in the raw dataset

Top 20 frequent words in stressed posts and non-stressed posts are

The labeled data is retrieved from Elsbeth Turcan & Kathleen McKeown.

Predicting Movie Reviews with BERT on TF Hub
Run Streamlit.io on Google Cloud Kubernetes
Step-by-Step Streamlit App Deployment Via GKE