Sprint 3 (MVP)

Backend:

• Investigated SVMs and logistic regression classifier much deeper
• Finalised our statistical model, at 74.5% accuracy, using reviewer features and SVC with grid search
• Investigated three forms of neural network architectures, FFNN, CNN and RNN, in a POC fashion
• Cross compared the performance of the these three architectures using BOW and word2vec
• Created custom word embeddings over our datasets using Google's word2vec (attached) and Facebook's fastText
• Did an experiment investigating FFNN architectures with BOW and word2vec
• Created and hosted our first neural network model (attached), a FFNN running alongside our SVM returning feature weights
• Remodelled and revamped the wiki for documenting
• Toyed around and read up on Grove, DCU's GPU instances that we will use next semester to train models
• Researched deep learning and neural networks extensively and documented our research in the wiki

Frontend:

• Got rejected by the Yelp API, however..
• Integrated with Google Places API, and used an ensemble of Yelp Fusion and Google Places to return Google reviews
• Set up a NoSQL OO database on our Yelp dataset to make our data queryable, allowing us pseudo-Yelp access as a backup
• Did extensive research on data visualization and color theory, documented in the wiki
• Implemented a word cloud indicating the most important words to a particular classification
• Grouped best and worst classified reviews to make the result easier to read

Sprint 2

Added term weight visualisations to the webapp, along with integrated Yelp search.
Did a whole bunch more experiments on the data with multiple statistical classifiers, and cross-compared accuracies.
Did a bunch of research on the most significant papers in the opinion spam detection field and proposed a novel hypothesis for improving the cutting edge.

Sprint 1

Niall & Stefan

• Conda environment for managing Python dependencies and versions
• Docker environment for replication on any machine
• Jupyter Notebook integration detailing experiments and classifier results and metrics
• Dataset conversion into Protobuffers
• Cross comparison and metrics of 4 classifiers: Naive Bayes, Logistic Regression, k-NN and Linear SVM
• Feature extraction of data to replicate Stanford paper: POS, structural, sentiment
• Understanding of Naive Bayes
• Unit testing with Pytest and linting with Pylint
• Python API serving a /classify endpoint using a pickled classification model to serve webapp results

Kirill:

• Continuous Integration / Continuous Delivery pipeline via CircleCI
• Node backend server hosted on Redbrick machine
• ReactJS web application using Webpack, Bulma.io and Redux
• Unit testing and integration testing using Jest and Chai
• End-to-end communication of frontend and backend API classification model and displaying of results