ACML-2018

Motivation and Objectives

As AlphaGo defeated the world's best Go player in 2016, AI is brought into the classroom to individualize learning in the form of adaptive learning. It analyzes the students and note their weaknesses and strengths, then changes the course around so that students can polish up areas which they may be struggling with. It also responds to the students' needs and personalize the course to best fit their talents. We take this chance to discuss the most recent development of machine learning technology used in education and to provide a forum for communication of researchers active in machine learning used in education.

Topics of Interest

Interested topics include, but are not limited to:

• Personalized learning paths: Students learn knowledge points at their individual speeds. Machine learning can help incorporate adaptive learning in classrooms through algorithms that let the students move ahead based on their speed of mastering knowledge points. Teachers can thus assess the understanding of an individual student or a class as a whole. This insight allows teachers to adjust their pace and delivery according to student progress, and be able to help each student individually, if required.

• Content Analytics: Content analytics refers to machine learning platforms that optimize modules for students and teachers. Through machine learning, content taught to students can be analyzed for maximum effect and optimized to take care of student needs. It enables educators and content providers to not just create and manage their e-Learning content, but also gain important insights into learner progress and understanding through a powerful set of analytics.

• Scoring: Machine learning also helps teachers in scoring assessments in less time and with greater accuracy. Current scoring systems primarily rely on humans, but with machine learning assessments can be scored in an automated manner. As an example, software that detects plagiarism in essays is already used by educators worldwide.

• Automating repetitive tasks: Teachers in traditional classroom setups spend a lot of time on repetitive tasks such as attendance, collecting assignments, etc. With technology, these tasks can be automated, allowing teachers more bandwidth to spend time with students tackling modules, concepts, and discussing higher-order thinking.

• Learning analytics: Machine learning also has a pivotal role to play in tracking student learning and progress. Learning analytics is more than just providing data to teachers. Machine learning algorithms create value for the system through designing predictive learning paths for students. As students progress through a course with adaptive learning software, machine learning algorithms decide if reinforcement is required through additional content or if they have adequately mastered concepts and can move ahead. Learning analytics, thus, focuses on tracking student knowledge and enhancing their learning environment.

Schedule

08:30 - 08:40 Opening

08:40 - 09:30 Keynote talk by Tingshao Zhu, Chinese Academy of Sciences

Title: Identifying Personality by Content Analysis

Abstract: Because of its richness and availability, social media has become an ideal platform for conducting psychological research. We proposed to predict active users' personality traits through micro-blogging behaviors. 547 Chinese active users of micro-blogging participated in this study. Their personality traits were measured by the Big Five Inventory, and digital records of micro-blogging behaviors were collected via web crawlers. After extracting 845 micro-blogging behavioral features, we first trained classification models utilizing Support Vector Machine (SVM), differentiating participants with high and low scores on each dimension of the Big Five Inventory. The classification accuracy ranged from 84% to 92%. We also built regression models utilizing PaceRegression methods, predicting participants’ scores on each dimension of the Big Five Inventory. The Pearson correlation coefficients between predicted scores and actual scores ranged from 0.48 to 0.54. By using the predicting model, we can conduct some noval research to investigate the personality of individuals in humanities.

09:30 - 10:10 Invited talk by Jing Zhang, Renmin University of China

Title: User Alignment and Personalized Course Recommendation in MOOCs

Abstract: This talk will cover two topics: how to align the same users across different social networks, and based on the integrated networks, how to conduct effective personalized recommendation. For the first topic, we target at inferring whether two users in heterogeneous networks are the same person in the physical world or not. We focus on dealing with the challenge of heterogeneous profile attributes and diverse neighbors in different networks. I will introduce the proposed model that can effectively leverage the user profiles and the network structures together to integrate two networks. For the second topic, we take the course recommendation in MOOCs as an example and study how to effectively distinguish the effects of different historical courses when recommending different courses. I will introduce the proposed reinforcement learning method to deal with the challenge.

10:10 - 10:30 Coffee Break

10:30 - 11:10 Invited talk by Yupei Zhang, Northwest Polytechnical University

Title: Machine Learning for Personalized Teaching and Learning

Abstract: Educational machine learning (EML) is a promising tool for personalized teaching and learning. Different from traditional educational studies, EML aims to learning the educational patterns from data. EML is a young field and has a lot of challenging problems. In this talk, I mainly introduce the work progress of EML group in NWPU, including cognitive diagnose, next-term grade prediction, and student early warning. To probe the status of knowledge mastery, we propose the robust cognitive diagnose method from 1-bit data by learning a knowledge dictionary with a sparse guess-and-error matrix. And we integrate two graphs created by side information into robust matrix factorization to carry out the next-term grade prediction. Finally, we extract the features from multi-sources multi-mode data and then classify the students for early warning, using the deep learning model. In addition, we show the conclusions and discussions on the educational data set of NWPU.

11:10 - 11:50 Talk by Zhen Xue, Yixue Squirrel AI

Title: Adaptive Learning System and Its Promise On Improving Student Learning Performance

Abstract: In online education products, adaptive learning, by definition, adjusts the content and guidance offered to individual learners. Systems that offer adaptive learning grew out of computer science research that aimed to replicate the dynamic interactions between human tutors and learners. Studies have shown that these systems can be effective learning tools. This talk introduces an adaptive learning system, “Yixue,” that was developed and deployed in China. It diagnostically assesses students’ mastery of fine-grained skills and presents them with instructional content that fits their characteristics and abilities. The Yixue system, first developed in 2016, has been used by over 10,000 students in 17 cities in China for learning 12 subjects in middle school in 2017. This talk describes major features of the Yixue system and its implementation model in class. We evaluated the efficacy of the Yixue compared with whole class instruction by expert human teachers and the other learning systems. The result suggests that students learned more from using Yixue adaptive learning system than whole classroom instruction by teachers and the other learning systems.

11:50 - 13:30 Lunch

13:30 - 14:10 Invited talk by Jie Tang, Tsinghua University

Title: Learning Intervention with Implicit Feedback

Abstract: Massive open online courses (MOOCs) boomed in recent years and have attracted millions of users worldwide. For example XuetangX.com, a platform similar to Coursear and edX, is offering thousands of courses to more than 10,000,000 registered users. However the completion ratio is always very low. I will introduce how we leverage users’ implicit feedback, e.g., user clicks, to help improve learning effectiveness.

14:10 - 14:50 Invited talk by Yu Lu, Beijing Normal University

Title: Data-Driven Learning Analytics and its Applications

Abstract: Learning analytics emphasizes on understanding and optimizing the learning content, process and environment, where the data-driven approach is specifically adopted to conduct the analytics on domain knowledge, learners and their contexts. This talk introduces the data-driven learning analytics research at Beijing Normal University, which includes the auto-generated knowledge graph for education, prerequisite-driven knowledge tracing, learner status recognition and task-driven question-answering solutions. The talk also showcases the implemented system that integrates the above research efforts, and how it can be applied in the real-world scenarios for education.

14:50 - 15:30 Invited talk by Dan Bindman, Yixue Squirrel AI

Title: Knowledge Assessment and AI Adaptive Learning: Now and the Future

What are the key components for a strong adaptive learning system? (1) Very strong content. This means the questions and lessons must cover everything that needs to be covered in the course, with great quality questions, explanations, and lessons. (2) An automated system (AI) that uses each student’s recent history in the course to precisely map the student’s current knowledge at a very high resolution—determining exactly which questions, lessons, or topics the student has fully mastered, not mastered, or partially mastered. (3) An automated system (AI) that uses this high resolution map of the student’s knowledge, to custom choose learning material that is ideal for that particular student to learn—neither too easy nor too hard, but at just the right level of difficulty given the student’s current knowledge. In this talk, we will focus on these three points: strong contents as mentioned in (1) and a new model that can help with (2) and (3). We measure student knowledge with the Probabilistic Knowledge State (PKS). It is assumed that a student’s PKS accurately and completely reflects the underlying knowledge of the student. That is, the PKS gives the actual probabilities correct for the student at any given time t; they would not be closer to 0 or 1 even if we were “all knowing” and knew everything about the student. So in this model, students can have “partial knowledge” or “partial mastery” of a given question such that the probability correct for this question is substantially higher than the probability of a lucky guess, but still far from the maximum probability correct that the students could attain if they had completely mastered the question. This is a very important aspect of the model.

Organizers

Dr. Wei Cui, Shanghai Yixue Education Technology Ltd., China, cuiwei@songshuai.com，http://www.squirrelai.us/cui_wei.html

Prof. Xiangen Hu, University of Memphis, xhu@memphis.edu，http://www.memphis.edu/psychology/people/faculty/hu.php

Dr. Sam Wang, SRI international (Stanford Research Institute), USA, sam.wang@sri.com，https://www.sri.com/about/people/sam-wang

Dr. Zhen Xue, Shanghai Yixue Education Technology Ltd., China, xuezhen@songshuai.com, https://jim-zhenxue.github.io/Zhen-Xue/

ACML-2018 Homepage