Learning-Oriented Question Recommendation Using Bloom’s Learning Taxonomy and Variable Length Hidden Markov Models - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 41

P12

P11

P10

P9

P8

P7

P6

P5

P4

P3

P2

P1

P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12

Fig. 6. Correlation matrices of user question orderings for the earth sciences domain

that recommends questions randomly. The SR corresponds to the approach pro-
posedin[ 10 ]. Table 4 shows the results obtained using a 10-fold cross validation
with input parameters: number of recommendations N = 5, maximum order of
the VLMCs, maxOrder =10and ˃ = the lowest prediction value among the
questions within the sequences used for training. For testing, the leave-one-out
technique was employed.

Table 4. Results

Data set Model acc avg-ll cov div lu

SR 0.65 93.84 0.29 0.47 0.45

Earth sciences LoR 0.31 137.69 0.60 0.44 0.69

RR 0.01 164.9 1 0.60 0

SR 0.52 112.34 0.30 0.50 0.41

Nutrition LoR 0.15 156.23 0.67 0.51 0.49

RR 0 165.38 1 0.40 0

SR 0.50 107.87 0.44 0.36 0.39

Homeschooling LoR 0.16 145.92 0.84 0.47 0.41

RR 0.06 150.06 1 0.42 0

Although the LoR did not achieve an accuracy ( acc ) and average log-loss
( avg ll ) as high as the SR, compared to the RR, it still had a good prediction

performance (see Table 4 ). However, the coverage ( cov ) and learning utility
( lu ) values of the LoR were much higher, whereas the diversity did not show
significant discrepancies among the three models. The coverage values show that
the LoR, compared to the SR, can recommend a larger percentage of questions.
The increased learning utility of the LoR shows that the prediction performance
of this recommendation model is more dependent on receiving as input learning
sequences, like the ones collected during our experiments. This means that the

−