iMath Project - Optlearn Algorithm (University of Genova)
The strategy for question selection is twofold: for the initial question, a random selection is made since there is no available information on past test answers to inform the predictive model. Subsequently, from the second to the last question, a Random Forest model comes into play. This model predicts the probability of the user correctly answering each potential question, and the question with the highest predicted probability becomes the next in line. The underlying goal of this approach is to bolster student motivation by providing questions deemed more manageable, fostering a positive and encouraging learning experience.
The aim of this code is to generate 5 questions-math tests personalized for each student (questions are chosen accordingly to student's background and past results).
The approach is structured as follows:
• First question is chosen randomly since there is no information yet on past answers of the current test which can be used in the prediction model.
• From second to last question: using a Random Forest model to predict the probability that the user will correctly answer the next question, for all the possible available questions; then choosing as next question the one which maximizes that probability.
The algorithm has the following main features:
• The input features are: student’s background, past tests’ results and data on past answers of the current test, statistical quantities of the questions (ex. average time spent on the question by all the users).
• Using the training data, the most important features are selected and only these data are given in input to the RF in order to improve its accuracy.
• Input data are scaled: categorical features have been encoded using one-hot encoding, numerical features have been scaled between (0,1).
• For each question of the test, a RF model is built. Each RF model differs only for the size of the past data’s window we used as input to the model.
Each RF model has the following settings:
• Number of trees in the forest: 900.
• Criterion to choose the best split: entropy.
• Number of features to consider when looking for the best split: square root of the total number of features.
The aim of this code is to generate 5 questions-math tests personalized for each student (questions are chosen accordingly to student's background and past results).
The approach is structured as follows:
• First question is chosen randomly since there is no information yet on past answers of the current test which can be used in the prediction model.
• From second to last question: using a Random Forest model to predict the probability that the user will correctly answer the next question, for all the possible available questions; then choosing as next question the one which maximizes that probability.
The algorithm has the following main features:
• The input features are: student’s background, past tests’ results and data on past answers of the current test, statistical quantities of the questions (ex. average time spent on the question by all the users).
• Using the training data, the most important features are selected and only these data are given in input to the RF in order to improve its accuracy.
• Input data are scaled: categorical features have been encoded using one-hot encoding, numerical features have been scaled between (0,1).
• For each question of the test, a RF model is built. Each RF model differs only for the size of the past data’s window we used as input to the model.
Each RF model has the following settings:
• Number of trees in the forest: 900.
• Criterion to choose the best split: entropy.
• Number of features to consider when looking for the best split: square root of the total number of features.
Scientific Area:
Python
Language/Environments:
Learning
Target Group:
Advanced
Cite as:
Cademartori, G. and Oneto, L. "iMath Project - Optlearn Algorithm (UNIGE)", 2024. https://github.com/aipereira/iMath_Public/tree/main/Algorithms_and_reports/UNIGE/unige_before_ipb_integration, available at January, 2024.
Author of the review:
Gabriel Leite
IPB
You have to login to leave a comment. If you are not registered click here