Design a personalized e-learning system based on item response theory and artificial neural network approach Ahmad Baylari, Gh.A. Montazer IT Engineering Department, School of Engineering, Tarbiat Modares University, Tehran, Iran Expert Systems with Applications, Volume 36, Issue 4, May 2009, Pages 8013-8021 Reporter:Yu Chih Lin Outline Introduction Item response theory Test construction process System architecture System evaluation Conclusion Introduction In web-based educational systems the structure of learning domain and content are presented in the static way Without taking into account the learners’ goals experiences existing knowledge ability interactivity Introduction Personalization and interactivity will increase the quality of learning This paper proposes a personalized multi-agent e-learning system Item response theory (IRT) • Presents adaptive tests Artificial neural network (ANN) • Personalized recommendations Item response theory Item response theory (IRT) was first introduced to provide a formal approach to adaptive testing Three common models for ICC One parameter logistic mode(1PL) Two parameter logistic model(2PL) Three parameter logistic model (3PL) Item response theory 1PL models i : only one parameter bi :item difficulty D:1.7 𝜃: ability scale 𝑃𝑖 ( 𝜃) = 1 1+e(−D(𝜃−𝑏𝑖 )) Item response theory 2PL model • 𝑎𝑖 : discrimination degree, added into the item characteristic function • bi : item difficulty • D : 1.7 • 𝜃 : ability scale 𝑃𝑖 ( 𝜃) = 1 1 + 𝑒 (−𝑎𝑖 𝐷(𝜃−𝑏𝑖 )) Item response theory 3PL model 𝑐𝑖 : guess degree to the 2PL model Potential guess behavior of examinees 𝑃𝑖 ( 𝜃) = 𝑐𝑖 + frac1 − 𝑐𝑖 1 + e(−𝑎𝑖D(𝜃−𝑏𝑖 )) Item response theory An estimation method called maximum likelihood estimator (MLE) • Effectively estimate item parameters and examinee’s abilities 𝑃𝑖(𝜃) : learner can answer the i th item correctly Q𝑖(𝜃) : learner cannot answer the ith item correctly(1 - 𝑃𝑖(𝜃) ) ui : 1 (correct answer) or 0 (incorrect answer) 𝑛 𝐿 𝜃|𝑢1 , 𝑢2 , … . . 𝑢𝑛 𝑛 𝑃𝑖 (𝜃)𝑢𝑖 𝑄𝑖 (𝜃)1−𝑢𝑖 = 𝑖=1 Item response theory Item information function (IIF) is the subject of the amount of information Effectively distinguish between subjects potential ability to reduce the estimation error Test information function (TIF) • Sum of the amount of information the test results for each subject Item response theory IIF&TIF: (IIF) (TIF) P’(θ) is the first derivative of Pi(θ) and Qi(θ) = 1 - Pi(θ) I (θ) : amount of information for item,1~N Test construction process Three types of tests • pre-test • post-test • review tests(延後測) All of these tests have 10 items Use IRT-3PL model to test construction • appropriate post-test selection for learners Test construction process For posttest construction System architecture Proposed a personalized multi-agent e-learning system Middle layer contains four agents Activity agent : records e-learning activities Planning agent : agent plans the learning process Test agent : based on the requests of planning agent , presents appropriate test type to the learner Remediation agent : analyzes the results of review tests, and diagnoses learner’s learning problems System architecture System architecture System design and development Collect stutdents’s responses and tests • • Diagnose their learning problems Recommend them appropriate learning materials Maximum number of recommended LOs were five Large number of responding states to a test • 210 states for each test System design and development Experiment the remediation agent Essentials of information technology management course • Divided into several Los • A few codes were allocated for all Los System design and development I1 to I10 columns are item codes R1 to R10 are corresponding responses which code 1 : correct response 0 : incorrect response System design and development Use a back-propagation network(BPNN) • Learning Data Use 20 input nodes Use 5 output neurons System design and development Use items responses data as input data the neural network Output neurons are recommended LOs System design and development Normalization of data within a uniform range 0–1 Prevent larger numbers from overriding smaller ones Prevent premature saturation of hidden nodes No one standard procedure Input Output System design and development Scale input and output variables (xi) in interval 𝑋𝑛 : normalized value of 𝑋𝑖 , 𝑋𝑖𝑚𝑖𝑛 , 𝑋𝑖𝑚𝑎𝑥 𝜆1 , 𝜆2 Normalized the input and output data , range 0.1 ~ 0.9 𝑋𝑛 = 𝜆1 + 𝜆2 − 𝜆1 𝑋𝑖 − 𝑋𝑖𝑚𝑖𝑛 𝑋𝑖𝑚𝑎𝑥 − 𝑋𝑖𝑚𝑖𝑛 System design and development ANN requires partitioning of the parent database into three subsets Training Test Validation Training used 60% of all data Validation 10% for data Remaining data for testing the network System design and development Use one or two hidden layer Trained with various neurons in each layer in MATLAB software In hidden layer • Use sigmoid function as activation function For output neurons First use linear activation function Second use sigmoid activation function System design and development Two different criteria used to stop training • Training error(MSE) :10−4 • Maximum epoch: 1000 epochs Training back-propagation network • • With a random set of connection weights (weight initialization) Train with different architectures System design and development For example in training a network 15 neurons in one hidden layer • Output layer neurons • With sigmoid activation function With linear activation function MSE error 7.13 • System design and development For example in training a network 15 neurons in one hidden layer and 10 neurons in second hidden layer • Output layer neurons • With sigmoid activation function With linear activation function MSE error 0.0158 System design and development Summarizes the results of trained networks with different architectures Network configuration 20-20-5 (network No. 11) System evaluation Recommended LOs from the network compared with recommended LOs from a human instructor Output was exactly the same as the target output, 25 of 30 tests (83.3%) Conclusion Proposed a personalized multi-agent e-learning system Estimate learner’s ability using item response theory Diagnose learner’s learning problems Recommend appropriate learning materials to the learner Neural network approach to learning material recommendation