Upgrade to Pro — share decks privately, control downloads, hide ads and more …

[ICCE2021] Prior Knowledge on the Dynamics of S...

Qiushi Pan
November 24, 2021

[ICCE2021] Prior Knowledge on the Dynamics of Skill Acquisition Improves Deep Knowledge Tracing

Proceedings p.265 https://icce2021.apsce.net/proceedings/volume1/
GitHub: https://github.com/qqhann/KnowledgeTracing

Abstract: Knowledge tracing (KT) is the task of modeling how students' academic skills change over time. Given a sequence of a student's learning history, one goal of KT is to predict how well he/she will perform in the next interaction. Unlike in BKT (Bayesian knowledge tracing), the models in DKT (Deep knowledge tracing) cannot be improved simply by introducing elaborate prior knowledge about the task domain. Instead, we need to observe how trained models behave and identify their shortcomings. In this paper, we examine a problem in existing models that have not been discussed previously: the inverted prediction problem, in which the model occasionally gives predictions that are opposite to a student's actual performance development. Specifically, given an input sequence where a student has solved several problems correctly in a row, the model will occasionally estimate his/her skills to be lower than when he/she could not solve them. To tackle this problem, we propose pre-training regularization, which incorporates prior knowledge by supplying synthetic sequences to the neural network before training it with real data. We provide regular, simplistic synthetic data to a sequence- processing neural network as a specific implementation of pre-training regularization. This method solves the inverted prediction problem and improves the performance of the model in terms of AUC. We observed its effect qualitatively and introduced a quantitative measure to assess the improvement also. For ASSISTments 2009, ASSISTments 2015, and Statics 2011, improvements in AUC scores were 0.2 ~ 0.7 %, which are significant considering the scores are already high (around 70~80%). We developed an open-source framework for DKT with pre- training regularization. It also contains user-friendly hyperparameter optimization functionality.

Qiushi Pan

November 24, 2021
Tweet

More Decks by Qiushi Pan

Other Decks in Research

Transcript

  1. % %BUBTFU"44*45NFOUTDPSSFDUFE  • &BDIRVFTUJPOJTMBCFMFECZBTLJMM ,$ LOPXMFEHF DPODFQU UIBUJTSFRVJSFEUPTPMWFJU •

    %BUBTFUTDPOUBJOTRVFTUJPOBOTXFSJOHMPHT 4LJMM*% $POUFOU q = 13 a = q = 15 a = q = 41 a = … q = 13 a = q = 38 a = $POUFOUGSPN "44*45NFOUTDPSSFDUFE %BUBTFU 4FRVFODF 4FRVFODF 4FRVFODF
  2. *OWFSUFEQSFEJDUJPOQSPCMFN   • 8FJOWFTUJHBUFEUSBJOFE%,5NPEFMTXJUIEVNNZEBUBTFRVFODF 
  JG BOE PUIFSXJTF

    sq k = ((q, a1), (q, a2), …, (q, aT)) at = 1 t > T − k at = 0  1SFEJDUFE1FSGPSNBODF 0.0 s0 s20 s10 s0 s20 s10 s0 s20 s10 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1SFEJDUJPODVSWF B  q = KC30 1SFEJDUJPODVSWF C  q = KC83 1SFEJDUJPODVSWF D  q = KC98
  3. 2VBOUJ fi DBUJPOPGUIFJOWFSUFEQSFEJDUJPOQSPCMFN   • $MFBSBOESPVHIJOEJDBUPS r1 = 1

    Q ∑ rq 1 = 1 Q ∑Q q=1 1 { ̂ y(sq T , q)> ̂ y(sq 0 , q)} rq=KC98 1 = 0 rq=KC83 1 = 1 rq=KC30 1 = 1  1SFEJDUFE1FSGPSNBODF 0.0 s0 s20 s10 s0 s20 s10 s0 s20 s10 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1SFEJDUJPODVSWF B  q = KC30 1SFEJDUJPODVSWF C  q = KC83 1SFEJDUJPODVSWF D  q = KC98
  4. 2VBOUJ fi DBUJPOPGUIFJOWFSUFEQSFEJDUJPOQSPCMFN   • &YQSFTTEFHSFF r2 = 1

    Q ∑ rq 2 = 1 Q ∑Q q=1 NDCG ( ̂ y (sq 0 , q), …, ̂ y (sq T , q), (0,…, T)) rq=KC98 2 = 0.51 rq=KC83 2 = 0.74 rq=KC30 2 = 0.83  1SFEJDUFE1FSGPSNBODF 0.0 s0 s20 s10 s0 s20 s10 s0 s20 s10 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1SFEJDUJPODVSWF B  q = KC30 1SFEJDUJPODVSWF C  q = KC83 1SFEJDUJPODVSWF D  q = KC98
  5. 1SFUSBJOJOHSFHVMBSJ[BUJPO  • "MMXSPOHBOEBMMDPSSFDUTZOUIFUJDEBUBGPSFBDITLJMMT • 1SJPSLOPXMFEHFUIBU   X′ 

    wro = {(qi ,0), …, (qi ,0)}, y′  wro = (qi ,0) X′  cor = {(qi ,1), …, (qi ,1)}, y′  cor = (qi ,1) qi a = qi a = qi a = qi a = qi a = qi a = qi a = qi a = … …
  6. 1SFUSBJOJOHSFHVMBSJ[BUJPO  qi a = qi a = qi a

    = qi a = qi a = qi a = qi a = qi a = q1 a = q1 a = q2 a = q2 a = qt a = qt+1 a = qt+1 a = qt a = 4ZOUIFUJDEBUB 3FBMEBUB … … … … 1SFUSBJOJOH 5SBJOJOH
  7. /%$(EJTUSJCVUJPO  • /FBSPQUJNBMTLJMMT /%$(TDPSF JODSFBTFEJOBMMEBUBTFUT rq 2 ≈ 1

    'SFRVFODZ /%$(TDPSF /%$(TDPSF /%$(TDPSF /%$(TDPSF "TTJTUNFOU "TTJTUNFOU 4ZOUIFUJD 4UBUJDT QSFUSBJO QSFUSBJO
  8. -FBOJOHDVSWF  • )JHIFSJOJUJBM"6$BGUFSQSFUSBJOJOH • 'FXFSJUFSBUJPOTMFBEUPFBSMZTUPQQJOHBUIJHIFS fi OBM"6$ "6$ &QPDI

    &QPDI &QPDI &QPDI "TTJTUNFOU "TTJTUNFOU 4ZOUIFUJD 4UBUJDT QSFUSBJO QSFUSBJO
  9. 3FGFSFODFT • <>$PSCFUU "5BOE"OEFSTPO +3,OPXMFEHFUSBDJOH.PEFMJOHUIFBDRVJTJUJPO PGQSPDFEVSBMLOPXMFEHF 6TFS.PEFMJOHBOE6TFSBEBQUFE*OUFSBDUJPO 7PM /P 

    QQr   • <>1JFDI $ #BTTFO + )VBOH + (BOHVMJ 4 4BIBNJ . (VJCBT -+BOE4PIM %JDLTUFJO +%FFQLOPXMFEHFUSBDJOH "EWBODFTJO/FVSBM*OGPSNBUJPO1SPDFTTJOH 4ZTUFNT QQr   • <>:FVOH $,BOE:FVOH %:"EESFTTJOHUXPQSPCMFNTJOEFFQLOPXMFEHF USBDJOHWJBQSFEJDUJPODPOTJTUFOUSFHVMBSJ[BUJPO BS9JWQSFQSJOUBS9JW