• Journal of Electronic Science and Technology
  • Vol. 22, Issue 2, 100262 (2024)
Wei-Wei Gao1, Hui-Fang Ma1,*, Yan Zhao1, Jing Wang1, and Quan-Hong Tian2
Author Affiliations
  • 1College of Computer Science and Engineering, Northwest Normal University, Lanzhou, 730070, China
  • 2Computer Center of Gansu Province, Lanzhou, 730070, China
  • show less
    DOI: 10.1016/j.jnlest.2024.100262 Cite this Article
    Wei-Wei Gao, Hui-Fang Ma, Yan Zhao, Jing Wang, Quan-Hong Tian. Enhancing personalized exercise recommendation with student and exercise portraits[J]. Journal of Electronic Science and Technology, 2024, 22(2): 100262 Copy Citation Text show less
    Exercise recommendation: (a) students’ mastery of knowledge with past response records and (b) difficulty of exercises pertinent to each knowledge concept and the exercises-knowledge concepts indication.
    Fig. 1. Exercise recommendation: (a) students’ mastery of knowledge with past response records and (b) difficulty of exercises pertinent to each knowledge concept and the exercises-knowledge concepts indication.
    Model framework of the presented PER, which consists of three main tasks: (a) finer-grained portrait of the student and the exercise construction of CSEG, (b) importance ranking of the exercises through a joint random walk, and (c) final list of exercise recommendations with multi-objective optimization.
    Fig. 2. Model framework of the presented PER, which consists of three main tasks: (a) finer-grained portrait of the student and the exercise construction of CSEG, (b) importance ranking of the exercises through a joint random walk, and (c) final list of exercise recommendations with multi-objective optimization.
    Influence of similar students (exercises).
    Fig. 3. Influence of similar students (exercises).
    Influence of portraits of students (exercises).
    Fig. 4. Influence of portraits of students (exercises).
    Impacts of the candidate, i.e., top-P (recommendation), i.e., top-L (exercise) number on performance.
    Fig. 5. Impacts of the candidate, i.e., top-P (recommendation), i.e., top-L (exercise) number on performance.
    Performance comparison with the change in the student (exercise) similarity threshold .
    Fig. 6. Performance comparison with the change in the student (exercise) similarity threshold .
    NotationDescription
    S, E, KThe set of students/exercises/knowledge concepts
    RStudent exercise response matrix
    QExercise and knowledge concept incidence matrix
    msThe degree of student mastery of knowledge concept
    csThe knowledge concept coverage of student response
    deThe exercise difficulty
    qeThe knowledge association
    Ws, WeThe student/exercise similarity matrix
    JThe probability transition matrix
    D0The set of candidate exercises
    DThe final list of recommended exercises
    Table 1. Several important mathematical notations.
    Algorithm 1: PERP algorithm
    Input: R, Q
    Output: Final recommendation list D
    1.  A, B$ \leftarrow $training NeuralCD model;
    2.  for i = 0 to N do
    3.   ${{\bf{m}}_{{s_i}}} = {{\mathrm{softmax}}} \left( {{\bf{x}}_{{s_i}}^{\mathrm{T}}{\bf{A}}} \right)$
    4.   ${{\bf{c}}_{{s_i}}} = {{\mathrm{softmax}}} \left( {{{\bf{x}}_{{s_i}}}{\bf{RQ}}} \right)$
    5.  end for
    6.  Ws$ \leftarrow $student similarity matrix;
    7.  for j = 0 to M do
    8.   ${{\bf{d}}_{{e_j}}} = {{\mathrm{sigmoid}}} \left( {{\bf{x}}_{{e_j}}^{\rm{T}}{\bf{B}}} \right)$
    9.   ${{\bf{q}}_{{e_j}}} = {\bf{x}}_{{e_j}}^{\mathrm{T}}{\bf{Q}}$
    10.   ${{\bf{e}}_j} = { {\bf{d} }_{ {e_j} } } \odot { {\bf{q} }_{ {e_j} } }$
    11.  end for
    12.  We$ \leftarrow $exercise similarity matrix;
    13.  J$ \leftarrow $transition probability matrix J by (7);
    14.  for j = 0 to t do
    15.   ${\bf{v}}_s^{t + 1} = \beta {\bf{Jv}}_s^t + (1 - \beta ){\bf{v}}_s^0$
    16.  end for
    17.  D0$ \leftarrow $choose exercises with the top-P largest scores;
    18.  while termination criterion is not satisfied do
    19.   D1$ \leftarrow $the top-L exercises of D0;
    20.   D2$ \leftarrow $replace some of the exercises in D1;
    21.   $ {\bf{D}}_1' $$ \leftarrow $distance matrix of the exercises in D1;
    22.   ${\bf{D}}_2' $$ \leftarrow $distance matrix of the exercises in D2;
    23.   if ${\bf{D}}_1' $ > ${\bf{D}}_2' $
    24.    D1$ \leftarrow $D2
    25.   else
    26.  $\gamma = \exp \left( {\frac{ { - \left( { {{\rm{mean}}} \left( {{\bf{D}}_2^\prime } \right)} \right) - \left( { {{\rm{mean}}} \left( {{\bf{D}}_1^\prime } \right)} \right)} }{ { {k_B}T} } } \right)$
    27.    r$ \leftarrow $random(0, 1)
    28.    if r >$\gamma $
    29.     D1$ \leftarrow $D2
    30.    else
    31.     D1$ \leftarrow $D1
    32.    end if
    33.   end if
    34.  end while
    35.  D$ \leftarrow $largest distance
    Table 2. Description of PERP algorithm.
    DatasetASSISTments 2009-2010Algebra 2006-2007
    Students41631338
    Exercises1774691771
    Knowledge concepts123491
    Records278607222314
    Table 3. Real dataset statistics.
    ModelASSISTments 2009-2010Algebra 2006-2007
    NoveltyAccuracyDiversityNoveltyAccuracyDiversity
    KNN0.9340.8880.2540.7830.7470.407
    KGEB-CF0.9120.8790.5240.6760.6310.674
    MBHT0.9460.8930.3430.7980.8590.468
    DKT0.6020.8800.4660.6210.8550.583
    NeuralCD0.5830.8940.4950.6450.8590.668
    DTransformer0.7130.8920.4520.6610.8610.516
    HB-DeepCF0.9140.8230.7580.7390.6950.619
    KCP-ER0.9570.8950.7650.8180.8630.743
    PERP0.9590.8970.7810.8210.8650.758
    Table 4. Performance of all methods on all datasets.
    ASSISTments 2009-2010PERP+PERP–
    Novelty0.9590.961
    Accuracy0.8970.887
    Diversity0.7810.779
    Table 5. Right and wrong answer recommendations on the ASSISTments 2009-2010 dataset.
    Exercise numberKnowledge concepts
    Actual answer records7,33,960, 962,1098, 1831,3090, 3102,3145, 3151,7398, 7465,7516, 8724,8729, 10274,12339, 12358,12811, 17125,17162Equation solving, Histogram as table or graph,Number line,Line plot,Stem and leaf plot,Table,Mode,Addition and subtraction fractions,Ordering fractions,Conversion of fraction decimals percents,Finding percents,Scale factor,Unit rate,Pattern finding
    KNN8318,8306,8304,8252,14864,8254Multiplication and division integers,Addition whole numbers,Division fractions
    NeuralCD10224,52,70,8286,37,7380,8307Equation solving,Stem and leaf plot,Addition and subtraction fractions,Circle graph,Finding percents
    PERP196,246,202,235,184,200Box and whisker,Congruence,Ordering integers,Square root,Equivalent fractions
    Table 6. Student with ID.219 answered and recommended the situation.
    Wei-Wei Gao, Hui-Fang Ma, Yan Zhao, Jing Wang, Quan-Hong Tian. Enhancing personalized exercise recommendation with student and exercise portraits[J]. Journal of Electronic Science and Technology, 2024, 22(2): 100262
    Download Citation