Feedbank

• 

  Dashboard without tweets

• 

  Tweet

• 

  Sign-in

• 

  Dashboard with tweets

Inspiration

We decided to pursue this project as it was a specific sponsor challenge at the hackathon. Additionally, members of this team have a common interest in building solutions that leverage the power of Machine Learning, and all have a passion for FinTech.

What it does

Feedbank uses the twitter api to retrieve tweets based on a query. Once the tweets are retrieved, we send the text data to the Python script that is running the machine learning model. The model ran inference on each tweet that was sent, and the result was sent back to the node script. Based on the result from the inference, Feedbank adds tags to each tweet, and displays them so the business can see all of the feedback.

How we built it

Dataset populator: Created a 3 step process that converts twitter data into a dataset that can be used by three machine learning models that learn off of the three corresponding parameters. Positivity vs negativity, technical vs non-technical, and the relevancy of the data. This was accomplished using 3 Python scripts. The first converted twitter API json files to CSVs. The second was a command line tool we wrote to allow us to fill in the necessary data for each tweet. The last one sorted and formatted the data according to the input specifications of the machine learning model.

Machine Learning model: Once the dataset populator was complete, we ran some data normalization techniques to make the dataset uniform. The techniques used to normalize the data were removing stop words (in, of, at, etc), lemmatization, and n-grams. The model architecture is a Support Vector Machine (SVM), which is very common in classification problems such as this one.

Website: The web application was created using the MEAN stack. MongoDB was used to store user profiles which contain information such as; email addresses, and saved handles. A REST API was created using express.js to serve tweet information to the frontend. This REST API was also useful when creating CSVs as it gave easy access to a large amount of tweets. Firebase was integrated with Angular to allow for secure email/password authentication.

Challenges we ran into

The first challenge occurred on making sure that raw json files do not contain string that will change formatting to something that’s not convertible (like “\n”). We needed to find all the possible key strings and erase them before they were loaded as a dictionary.

Tensorflow was not able to install (as we were using an older version and was needed to update to a newer version) , instead, scikit-learn was used to create and train the machine-learning model Our team had different versions of Python, which caused some issues interfacing the Javascript and Python code using PythonShell

There was no established datasets that we were able to use to train the model. Instead, we had to create a dataset populator ourselves and manually enter a boolean parameter for each tweet that would then be used to train our model.

Accomplishments that we're proud of

Fluid method of converting raw Json objects of tweets to data that can be fed to the model

We were able to achieve 80% accuracy on the current model that is for RBC tweets. Additionally, we were able to achieve up to 90%+ accuracy on established datasets for similar tasks (IMDB movie classification)

Very modular class for the machine learning model will allow for scalability in the future

Simple, intuitive, and easy to use UI

What we learned

As a group we learned a new machine learning architecture that will potentially help in future projects.

Learned CSV input, output and specific CSV file manipulation using Python

Some of our team members learned Git

What's next for Feedbank

Expand target companies to other companies

Refactor the codebase.

Sorting by tag to allow for easier consumption of tweet data.

Built With

• css
• html
• javascript
• python
• typescript