LearnProML

Equipping teachers with an app running a machine learning built model that predicts students' learning proficiency assessment outcomes.

The problem

In a blog written by Silvia Montoya, Director of the UNESCO Institute for Statistics (UIS), and Karen Mundy, Chief Technical Officer at the Global Partnership for Education (GPE) titled “New Data Reveal a Learning Crisis that Threatens Development Around the World”, the authors state: “data show that 617 million children and adolescents worldwide are not reaching minimum proficiency levels in reading and mathematics”. Moreover, Part II of the latest World Development Report published by the World Bank is titled “The Learning Crisis” where the authors describe “many faces of the learning crisis” backed with strong evidence, and add “to tackle the crisis, it is necessary—though not enough—to measure learning.”

Many countries already conduct assessments of learning (cross-national, regional, or international) and there are several datasets of large-scale assessment available to the public. However, there is a major challenge: “assessment data is not used to inform policy” and “And there is even less evidence that large-scale assessment data has an impact on teaching and learning policies at the school and classroom level, as a recent study from the Asia-Pacific region shows.” These are quotes from a blog titled “Ensuring learning data matters” by Talia de Chaisemartin and Ursula Schwantner and published by GPE.In fact, large-scale assessment data are usually analyzed far away from the schools from where they were originally collected. But, disseminating reports and translating analysis results to truly impact schools are challenging. “Strategic reporting and dissemination” are steps to a solution being considered (blog by Talia de Chaisemartin and Ursula Schwantner)

My proposed solution

Equip teachers anywhere in the world with an easy-to-use standalone app, called LearnProML, running a machine learning built model capable of predicting the outcome of a learning proficiency assessment in reading and math.

Large-scale assessment data are already available to the public for several countries. Assessment data have several “inputs” or “features” and “outputs” such as reading and math assessment scores. I took PASEC 2015 data for Madagascar and built a predictive model using the xgboost algorithm. While the original data has over 600 features, I was able to cut this down to 15 yet maintain the same predictive accuracy using all features. The model's predictive accuracy is 80.35%. Using all features it was 80.5%.

Then, I embed model into a smartphone, iOS for now. Nowadays, a few tools (e.g. CoreML on iOS) are readily available that would ease this task. The app is designed to be “standalone”, i.e. no internet connection is needed to use it. Hence, it can be used by anyone with a smartphone or a tablet, anywhere. Internet connection is needed only for its installation and updates. Help from its “users” will be needed for its final version.

Changing the world

Our app LearnProML can help bring assessment data to teachers, school directors, and policy makers anywhere in the country, in any country, and can help them improve the level of proficiency in reading and math for hundreds of millions of students worldwide.

Currently, large-scale assessment data are analyzed using traditional statistical techniques. Furthermore, these analysis reports are mainly targeting policy makers and researchers. Hence, school directors and teachers that participated in the surveys would most likely not have even see the result of their school assessment.

To our knowledge, LearnProML will be the first app using a predictive model built with the help of machine learning technologies that can potentially be  available to teachers, school directors, policy makers and educators worldwide to potentially improve students' learning proficiency.

Our solution, once it is fully developed, in a standalone app called LearnProML. Its current version, developed for iOS,  is showing just its main functionality. At its core is a predictive model built using machine learning technologies. The model was built using the XGBOOST algorithm. Then the model was inserted within an app thanks to CoreML technology. The original data has over 600 features which was brought down to just fifteen (15), yet giving almost the same accuracy. Feature selection was the most challenging part of the model building. The model's predictive accuracy is 80.35%.

First, have a version of LearProML available to teachers, school directors, and policy makers in Madagascar. Learn how it's being used and how it is shaping policies and helping teachers to increase their students' proficiency and math and readings.

Then, build models for countries where PASEC assessment is being conducted.

A few other international assessment tests are being administered to several  countries in various regions around the world. A version of LearProML in many of these countries can help millions of teachers, school directors, and policy makers help their children, especially the least privileged ones, to be more proficient in math and languages.

The app will be made available for download through the usual app marketplaces namely Apple App Store and Google Play Store.

Our solution is still  being developed, hence it is not serving any one yet.

Our solution is intended to be used by at least teachers, school directors and policy makers. If there is an average of 5,000 of those per country, the total would be 25,000 for the first 12 months. 

I believe that the demand for a similar app will increase significantly. Hence in 3 years, it could easily reach 200,000.

I built the app from scratch with no help from anyone else. I'm proficient in math, statistics, programming (JAVA, Swift, Matlab, SAS), machine learning, and app building (iOS and Android). 

I've also been asked to lead a few troubled organizations and turn them around. I did so through skillful leadership practices.

I can also assemble PCs from scratch and, install operating systems, build a LAN network, and troubleshoot computers and networks. 

LearnProML is intended to be used by teachers and educators. I do not intend to sell it to them directly. Perhaps an international organization, NGO or governments working with developing countries would be interested in sponsoring its development or just buying it for them. I admit that at this point I don't have a clear idea of how this will progress.

First and foremost, I wanted to propose a major piece of a challenging puzzle.

Second, I believe that an initial grant from SOLVE would allow me to devote more time to the app's development. But more importantly, SOLVE will definitely help give the app a better visibility to potential partners.

To build the app, I need data, time, and a larger team. I believe that SOLVE can help with all of them but especially the last two. As this is so far a personal endeavor, I need help to gather the necessary resources so that more countries can benefits in a shorter period of time.