How Learning Pool (formerly HT2 Labs) is Becoming a Machine Learning Trailblazer

• Home
• Blog
• Learning & Development

Chris Dunne, Digital Marketing Manager

Diagram of the Community of Inquiry Framework

Earlier this year Learning Pool (formerly HT2 Labs) introduced the award-winning Social Intelligence Dashboard, SID, as part of its growing army of tools to help organizations all over the world make data-driven decisions about their learning practices.

SID has been inspired by the Community of Inquiry (CoI) framework to reveal insight about a learners progression through analyzing the text they contribute in a social learning environment. This work has become a key differentiator for Stream LXP (formerly Curatr), our Learning Experience Platform.

SID is predicated on Machine Learning techniques that attempt to label comments according to a marking schema that we invented (based on CoI). The idea is to give a percentage certainty that a new comment has been marked as being of high quality, in that it demonstrates an individual’s stage in the learning process. But whilst we were developing our algorithms we found ourselves constantly tweaking to eek out every last drop of accuracy possible. As you iterate towards the best solution and fine tune hyperparameters (human-defined settings, as opposed to the ‘black box’ machine learning stuff), you find yourself in a long and arduous process. And it was becoming clear from early customer feedback that a single generic classification algorithm would not suffice – we would need to build multiple models, deploying different marking schemas. Not everybody would want to assess ‘quality’ in the same way we do.

Enter xAI – our Model Manager

Our xAI Model Manager allows us to both edit existing models and new models quickly and reproducibly using a simple graphical interface, allowing us to load up a high performing model and quickly make changes that help adapt it to new data or to a new use case.

Screenshot of the xAI Model Manager

We can visualize exactly how a model was trained, and on what data, as well as earmark favourite models for production or sort models based on characteristics such as language or any other number of data or model specific attributes.

Using a model pipeline it is possible to feed filtered comments from one model into another or return multiple scores from multiple models reflecting a variety of criteria demonstrated in comments.

Using tools like Live inference we can visually see how the new model reacts to specific sentences and words, look at its training summary as a visualization and assess its accuracy and precision on unseen data with a comprehensive breakdown demonstrating its decisions against the human classification for each line of input text.