ITS 2018 SARLR: Self-adaptive Recommendation of Learning Resources Authors: Liping Liu, Wenjun Wu and Jiankun Huang Institution: State Key Lab of Software Development Environment Department of Computer Science, Beihang University
01 Introduction Self-Adaptive 02 Content Recommendation 03 Experiments 04 Conclusions
01 Introduction
Introduction s
1. Introduction Rule-based Data-driven Recommendation Recommendation Require domain experts to evaluate Compare similarity among students learning scenarios and learning objects h Define extensive recommendation Be more scalable and general rules Fail to consider the impact of Only be applied in specific learning difficulty of learning objects and domains dynamic change
1. Introduction Contributions SARLR, a novel learning recommendation algorithm T-BMIRT, a temporal, multidimensional IRT-based model, incorporates the parameter of video learning An evaluation strategy for recommendation algorithms in terms of rationality and effectiveness
Self-Adaptive 02 Recommendation
2. Self-Adaptive Recommendation The Overall architecture of the SARLR algorithm
2. Self-Adaptive Recommendation IRT T-IRT • 𝛽 : question discrimination 1 The Temporal IRT extend IRT model by • 𝛾 : question difficulty 𝑞 𝑡𝑟 = modeling the student’s knowledge state over 1 + 𝑓𝑦𝑞[−(𝛽 𝑟 𝜄 𝑡 − 𝛾 𝑟 )] • 𝜄 : student’s ability time as a Wiener process 1.2 𝑄 𝜄 𝑢+τ 𝜄 𝑢 = 𝜚 𝜄 𝑢 ,𝜑 2 𝜐 𝜄 𝑢+τ 𝜄 𝑢+𝜐 − 𝜄 𝑢 ~𝑂(𝜄 𝑢 , 𝑤 2 𝜐) Probability of corrent response 1 0.8 0.6 𝑄 𝜄 𝑢+τ 𝜄 𝑢 = 𝜚 𝜄 𝑢 ,𝜑 2 𝜐 𝜄 𝑢+τ 0.4 0.2 0 -6 -4 -2 0 2 4 6 Student ability Item Characteristic Curve(ICC)
2. Self-Adaptive Recommendation T-BMIRT 𝑄 𝜄 𝑡,𝑢+τ 𝜄 𝑡,𝑢 , 𝑚 𝑡,𝑢 ,𝜑 2 𝜐 𝑚 𝑡,𝑢 = 𝜚 𝜄 𝑡,𝑢 + 𝜄 𝑡,𝑢+τ 1.2 𝑒 𝑡 𝑢 1 𝑚 𝑡,𝑢 = ∙ 𝑢 ∙ 1 𝑒 𝑢 𝜄 𝑡,𝑢 𝜄 𝑡,𝑢 ∙ ℎ 𝑢 1 + 𝑓𝑦𝑞 − − ℎ 𝑢 0.8 ℎ 𝑢 Skill 2 𝑄 𝜄 𝑢+τ 𝜄 𝑢 = 𝜚 𝜄 𝑢 ,𝜑 2 𝜐 𝜄 𝑢+τ 0.6 𝑚 𝑡,𝑢 : the knowledge that student 𝑡 gains from the video 𝑢 0.4 𝜄 𝑡,𝑢 ∙ ℎ 𝑢 − ℎ 𝑢 𝑢 : the knowledge of the video 𝑢 ℎ 𝑢 0.2 ℎ 𝑢 ℎ 𝑢 : is the prerequisites of video 𝑢 0 0 0.2 0.4 0.6 0.8 1 1.2 𝑒 𝑡 𝑢 is the duration in which student 𝑡 watches video 𝑢 Skill 1 𝑒 𝑢 is the total length of the video 𝑢 We use vector projection method to get the value that student’s ability exceed the video requirements.
2. Self-Adaptive Recommendation Search and Extraction SARLR Phase 1: Search and Extraction INPUT : … • Set of students 𝑇 = {𝑡 1 , 𝑡 2 , … , 𝑡 𝑜 } , target student 𝑡 𝑌 ∈ 𝑇 Video n Assessment n Video 1 Assessment 1 Video 2 • Matrix of abilities 𝐵 = [𝜄 𝑡,𝑢 ] , where 𝜄 𝑡,𝑢 is the ability value of student s at time t • Set of learning resources 𝐹 = {𝑓 1 , 𝑓 2 , … , 𝑓 𝑛 } 10 OUTPUT : learning path 𝑞 9 1: search for similar students MS, where 𝑡 𝑙 ∈ 𝑁𝑇 8 and 𝜄 𝑡 𝑙 ,𝑢 0 is similar to 𝜄 𝑡 𝑌 ,𝑢 0 7 2: for each 𝑡 𝑗 ∈ 𝑁𝑇 do 6 Skill 2 5 find 𝑡 𝑐 = 𝑏𝑠𝑛𝑏𝑦(𝑒𝑗𝑡𝑢𝑏𝑜𝑑𝑓(𝜄 𝑡 𝑗 ,𝑈 𝑡𝑗 − 𝜄 𝑡 𝑗 ,𝑢 0 )) , 3: where 𝑈 𝑡 𝑗 is the time of 𝑡 𝑗 completing learning 4 3 4: end for 2 5: extract the learning path 𝑞 = (𝑓 𝑗 1 , 𝑓 𝑗 2 , … 𝑓 𝑗 𝑈 ) of 𝑡 𝑐 1 6: return 𝑞 0 0 1 2 3 4 5 6 7 8 Skill 1
2. Self-Adaptive Recommendation Adaptive Adjustment INPUT : 1 𝑞 𝑡𝑟 = • SARLR Phase 2: Adaptive Re-planning Set of students 𝑇 = {𝑡 1 , 𝑡 2 , … , 𝑡 𝑜 } , target student 𝑡 𝑌 ∈ 𝑇 1 + 𝑓𝑦𝑞 − 𝜄 𝑡,𝑗 ∙ 𝛽 𝑟 − 𝑐 𝑟 • Matrix of abilities 𝐵 = [𝜄 𝑡,𝑢 ] , where 𝜄 𝑡,𝑢 is the ability value of student s at time t INPUT: 1 • Set of learning resources 𝐹 = {𝑓 1 , 𝑓 2 , … , 𝑓 𝑛 } 𝑞 𝑡𝑓 = • Target student 𝑡 𝑌 , recommended learning path 𝑞 = (𝑓 𝑗 1 , 𝑓 𝑗 2 , … 𝑓 𝑗 𝑈 ) 𝜄 𝑡,𝑗 ∙ ℎ 𝑓 OUTPUT : learning path 𝑞 • Result of 𝑡 𝑌 interacted with learning resources in 𝑞 1 + 𝑓𝑦𝑞 − − ℎ 𝑓 1: search for similar students MS, where 𝑡 𝑙 ∈ 𝑁𝑇 and 𝜄 𝑡 𝑙 ,𝑢 0 is similar to 𝜄 𝑡 𝑌 ,𝑢 0 ℎ 𝑓 OUTPUT: new learning path 2: for each 𝑡 𝑗 ∈ 𝑁𝑇 do 1 : for each 𝑓 ∈ 𝑞 do find 𝑡 𝑐 = 𝑏𝑠𝑛𝑏𝑦(𝑒𝑗𝑡𝑢𝑏𝑜𝑑𝑓(𝜄 𝑡 𝑗 ,𝑈 𝑡𝑗 − 𝜄 𝑡 𝑗 ,𝑢 0 )) , where 𝑈 𝑡 𝑗 is the time of 𝑡 𝑗 completing learning 3: 2 : if 𝑓 is a video and 𝑞 𝑡𝑓 < 𝐷 𝑡𝑓 do 4: end for 𝑞 𝑡𝑟 : the probability of student 𝑡 correctly answering return SARLR Phase 1 to re-plan path 𝑞 3 : 5: extract the learning path 𝑞 = (𝑓 𝑗 1 , 𝑓 𝑗 2 , … 𝑓 𝑗 𝑈 ) of 𝑡 𝑐 exercise 𝑟 4 : else if 𝑓 is an exercise and 𝑡 𝑌 failed it and 𝑞 𝑡𝑟 < 𝐷 𝑡𝑟 do 6: return 𝑞 𝑞 𝑡𝑓 : the degree of knowledge that student 𝑡 can 5 : return SARLR Phase 1 to re-plan path p 6 : end if acquire from the video 𝑓 7 : end for
03 Experiments
3. Experiments Datasets A publicly accessible data set • Assistments Math 2004-2005 • From Assistment online platform • Including 224,076 interactions, 860 students, 1,427 assessments and 106 skills A proprietary data set • blended learning data • From our blending learning analysis platform • Including 14,037,146 learning behavior data from 140 schools and 9 online educational companies
3. Experiments Experiments for T-BMIRT • Frequency method : predict the student Assistments Blended learning data correctly answer the assessment when his One-dimensional Multidimensional One-dimensional Multidimensional Models history correct rate is greater than 50%. ACC AUC ACC AUC ACC AUC ACC AUC • IRT : two-parameter ogive model. Frequency method 0.694 N/A 0.683 N/A 0.702 N/A 0.688 N/A • MIRT : multidimensional item response. IRT 0.716 0.779 0.701 0.758 0.721 0.784 0.706 0.752 • T-IRT : temporal IRT with 𝜑 = 0.5 , which were selected in exploratory experiments. MIRT 0.714 0.771 0.721 0.786 0.718 0.775 0.722 0.783 • T-BMIRT : temporal blended T-IRT 0.738 0.805 0.712 0.769 0.744 0.801 0.717 0.764 multidimensional IRT with 𝜑 = 0.15 and T-BMIRT 0.743 0.815 0.738 0.803 0.757 0.820 0.748 0.816 α = 10 −4 .
3. Experiments Rationality Evaluation • 𝑓 𝑗 ∈ 𝑞 : the learning resources in a recommended path, 𝑛 is the length of the path 𝑞 𝑡𝑗𝑛𝑗𝑚𝑏𝑠𝑗𝑢𝑧(ℎ 𝑓 𝑗 , 𝐿𝐷 𝑡 𝑦 ) 𝑓 𝑗 • 𝐿𝐷 𝑡 𝑦 : the knowledge components which 𝑡 𝑦 is learning in RC s x = 𝑛 the current chapter 𝑞 𝑡𝑗𝑛𝑗𝑚𝑏𝑠𝑗𝑢𝑧(ℎ 𝑓 𝑗 , 𝜄 𝑡 𝑦,𝑗 ) 𝑓 𝑗 DC s x = • similarity() : the adjusted cosine similarity of the two 𝑛 vectors in the parentheses. Model Relevance accuracy Difficulty accuracy UCF 0.86 0.77 ICF 0.71 0.83 LFM 0.87 0.84 SARLR 0.97 0.92
3. Experiments Effectiveness Evaluation 𝑇 ′ : the students whose learning paths are strictly • recommended • 𝐻 = 𝐹 𝑆 𝑇 ′ − 𝐹 𝑆 𝑇 𝑇 the students whose learning path are randomly selected • and 𝐹 𝑆 𝑇 : the students’ average score in the last 𝐹 𝑆 𝑇 ′ 𝐹 𝑆 𝑇 online assessment. Expected gain Model 1 2 3 4 5 6 UCF -0.04 -0.06 0.07 -0.03 0.08 0.01 ICF 0.05 0.04 -0.03 0.07 -0.02 0.05 LFM 0.04 0.12 0.09 0.10 0.03 -0.05 SARLR 0.11 0.27 0.24 0.23 0.17 0.06
04 Conclusions
4. Conclusions Establishes conditions to adaptively adjust recommendations towards the dynamic needs of the students Adaptively Evaluation T-BMIRT Strategy criteria Performs well on the prediction task of For personalized learning recommendation multi-dimensional skills assessments in terms of rationality and effectiveness
THANKS
Recommend
More recommend
