Knowledge Tracing Transformer with Masked Attention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current digital tutoring systems either rely solely on human-labeled item to skill associations, which are error-prone and do not tolerate multiple mappings, or machine-learned associations that discard human expert knowledge and require frequent student interactions, failing to effectively address learner forgetting behavior and combine human expertise with statistical patterns.
Innovation Solution
A digital learning system incorporating a machine learning-based knowledge tracing engine with a transformation architecture and masked attention layer that initializes an interaction-skill mapping matrix using expert-assigned skill tags, adjusts attention weights based on time gaps between interactions, and combines human expertise with machine-learned patterns to predict learner interactions and mastery levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If human-labeled item to skill associations are used, then expert knowledge is incorporated, but errors and multiple mappings cannot be tolerated
Solution Approach 1:
The patent combines human-labeled item to skill associations with machine-learned associations by initializing the interaction-skill mapping matrix W using expert skill tags while allowing machine learning to adjust weights. This merging approach preserves reliable expert knowledge while accommodating errors and multiple mappings through probabilistic weighting, where Wi represents the weight associated with all latent skills for learner interaction xi.
2Productivity
If machine-learned item to skill association is used, then statistical patterns are learned, but human expert knowledge is discarded and frequent student interactions are required
Solution Approach 1:
The patent performs preliminary action by initializing the interaction-skill mapping matrix W with expert skill tags before machine learning begins. This preliminary incorporation of expert knowledge reduces the amount of student interaction data needed for effective learning, as the system starts with a knowledgeable baseline rather than learning from scratch.
3Measurement precision
If current knowledge tracing models are used, then learner understanding is tracked, but forgetting behavior is not addressed
Solution Approach 1:
The patent applies dynamics by making the attention weights time-dependent through the time bias function b(Δtj−i). The attention mechanism dynamically adjusts weights based on time gaps between interactions, allowing the model to capture forgetting behavior where older interactions have diminishing influence on current knowledge state estimates.
4Ease of operation
If traditional attention mechanisms are used, then all interactions are weighted equally, but time gap effects are ignored
Solution Approach 1:
The patent changes parameters by introducing time bias b(Δtj−i) as an additional parameter in the attention weight calculation. This modifies the attention mechanism to account for temporal decay, where the time gap Δt between interactions influences the weight assigned to each interaction, improving reliability without significantly complicating the mechanism.
Data Source
AI summary
Digital learning or tutoring systems as described herein embed, by a trained machine learning knowledge tracing engine, an array for learner interactions X into a static representation ej corresponding to a prior learner interaction xj and determine a contextualized interaction representation hj based on this. Digital tutoring systems described herein calculate, by a masked attention layer of the trained machine learning knowledge tracing engine, an attention weight Aij based on a time gap between two learner interactions with the system, and can calculate a contextualized interaction representation hj, wherein the contextualized interaction representation hj is proportional to the attention weight Aij. The systems can provide for display at the GUI a second question item based on the contextualized interaction representation hj, the second question item corresponding to a recommended learner recommendation.