Learning goals recommendation system

The Case:

Early childhood programs aim to promote the well-being of young children and to assist them in their mental and physical development. In the West, these programs provide comprehensive guidelines on how teaching curriculums should be designed to suit every child’s needs.

The Challenge:

Due to the complexity involved in early childhood development, these guidelines are often hundreds of pages long and vary from state to state. With some states having more than one set of guidelines with different goals based on the age of the child.


This makes it difficult for childcare personnel to keep track of what the learning goals for each child should be since goals are based on the physical health, cognition, language, and social and emotional development of the child.


For an automated system to recommend the correct goal it needs to understand the meaning behind the teacher’s written assessment of the child’s progress as well as the meaning behind the goals that it can recommend taking into account the needs of the child.


This makes rules-based systems unfeasible due to the large number of scenarios and parameters that need to be taken into account. Furthermore, a lack of data as to what effect each recommendation has had in the past, as well as lack of feedback, makes this a challenging problem to solve with statistics and traditional machine learning.

The Solution:

To solve this problem we built a multi-stage recommendation system. In the first stage, the algorithm would try and match the meaning of the teacher’s assessment of the meaning of the goal.


This simple approach proved to be quite successful in the cases where there was no data for the given goal framework.


The 2nd stage used a few shot learning algorithm to quickly train a highly accurate model using only a few hundred samples. This improved the accuracy of the model from the previous stage and allowed the system to learn from teacher feedback.

The Outcome:

The final system was able to quickly integrate new learning frameworks and make suggests with 90% accuracy from as few as 500 samples of data and is now deployed to preschools in several English speaking countries.