This project presents an intelligent hashtag recommendation tool for use with Twitter, that makes it easier for users to compose tweets with relevant hashtags, as well as aiding the search and navigation of Twitter as a whole. It integrates a Naive Bayes classifier with a novel stream processing framework to give functional, non-personalised hashtag suggestions and search query expansions.

A thorough and detailed explanation of this project can be read in the final report.

The work contained within this repository has been submitted as the 3^rd Year Project for the award of MEng Computer Science by Jamie Davies (jagd1g11@ecs.soton.ac.uk).

The system is based on a client-server architecture. The main classification system is encapsulated in the server, which is accessible through a RESTful API. Client interfaces (including the demonstration one included within this repository) can then use the features available through the classification server as they see fit.

The RESTful API for the server is as follows:

• /api/classify - POST
  Classify a tweet to provide a list of hashtag suggestions.

  • text - The text of the tweet to classify
  • results - The number of classifications (recommendations) to provide (optional)
    

• /api/status - GET
  Returns a JSON object with useful information and statistics about the server.

• /api/hashtags - GET
  Returns an ordered list of all hashtags and their counts.

  • num - The number of hashtags to return (optional)
    

• /api/tokens - GET
  Returns an ordered list of all tokens and their counts.

  • num - The number of tokens to return (optional)

• /api/hashtag/<string:hashtag> - GET
  Returns an ordered list of all tokens and their counts that have been seen with the given hashtag.
  

  • num - The number of tokens to return (optional)
    

• /api/token/<string:token> - GET
  Returns an ordered list of all hashtags and their counts that have been seen with the given token.

  • num - The number of hashtags to return (optional)
    

The project uses the pip tool to manage Python dependencies, and Bower to manage HTML/CSS/Javascript dependencies. Before the project can be used, the dependencies must first be installed:

$ git clone https://github.com/daviesjamie/3yp hashtag_recommendation
$ cd hashtag_recommendation
$ pip install -r requirements.txt
$ cd client
$ bower install

Next, you need to register the application with Twitter to get the OAuth tokens necessary to use data from the live Twitter stream. You can do this by going to https://apps.twitter.com/, and then entering the information you are given into oauth.json files in both the client and server directories.

The classification server can then be run through the server.py script inside the server directory. If no arguments are supplied to the script, Tornado is used to provide access to the WSGI application. If the dev argument is supplied, then Flask is used to serve the application directly, which provides much more verbose output.

By default, the server will serialise its state and write it out to a .pickle file once every hour. A development server can be started from any .pickle file by simply passing in the file as a second argument: server.py dev state.pickle.

The client is a simple Django application that is provided only for demonstration purposes. It can be run through the server.py script inside the client directory.

More screenshots of the client interface can be found in the screenshots folder.