Machine Learning made simple with Ruby

The post describes a Ruby-based way to automate category assignment for announcements without external prediction services. It uses the classifier-reborn gem, settles on Latent Semantic Indexing instead of a Bayesian classifier, and notes that LSI can also find texts similar to a query.

Reading notes

#

• It starts from a project where users add announcements by link, metadata is fetched from OpenGraph tags, and an administrator still needs to assign a category before publishing.
• The goal is to suggest an appropriate category automatically for newly inserted content so administration does not become harder.
• The author first considers a Bayesian classifier, then switches to classifier-reborn after reading the gem’s documentation.
• LSI is presented as a better fit because it relies on the idea that words used in a given context tend to share meaning.
• The gem is also shown supporting reverse use, which lets the software search for related texts.
• Training is required, and in this case it can be done from already manually classified publications.
• The training step is moved to an asynchronous task because it is time-consuming and only needs to run periodically.
• One command object receives the advertisements to classify and the training data, then feeds each training title and category id into the classifier.
• Another command object handles the actual classification, classifies each advertisement title, updates the category, and marks the advertisement as classified.
• If the classifier raises VectorZeroVectorError because it was not trained, the code logs the error and does not update the record.
• The example concludes with a successful classification result and mentions an unresolved issue: retraining currently uses all published announcements every time, and storing the trained structure for later updates could be a better approach.