Their model outperformed every model on each of these three dataset:



San Diego State University

1. Experimental Setting
2. Results
3. Discussion

Experiments (DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills)

As it was described in the paper,  they split student population in 5 disjoint groups and cross-validation is made on this basis. Every feature weight and embedding components follow a normal prior distribution. They implemented their models in python. DAS3H was compared DASH, IRT, PFA, and AFM for 0, 5 and 20 embedding dimensions. Three datasets were used in their experiments: ASSISTments (assist12), Bridge to Algebra (bridge06) and Algebra I (algebra05).

Experimental Setting (DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills)

Results (DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills)

As  they found it wasn't clear what kind of impact embedding dimension can have on the performance of DAS3H. Even though on every dataset model had lower performance when embedding dimension was set to 20 still, when the dimension was 5 sometimes it had better performance than when it was 0.

It was observed that using time window features(instead of regular
skill wins and attempts counts) and different parameters for distinctive skills in $h_{\theta}$ increased AUC of the model. "Including both item-skill relationships and forgetting effect improves over models that consider one or the other."

Additionally, DAS3H can be well-suited for the adaptive scheduling. It can be useful for adaptively optimizing personalized skill practice schedule.


Learn Before

Related