Research On Learning Engagement Assessment Algorithm In Online Learning Scenarios

Online learning has gradually become a popular and common way of education with the increasing development of technology and electronic devices.Despite learners can learn anywhere and anytime due to online learning’s diverse resources and ease of use,it lacks realtime feedback compared with traditional education.Assessing learning engagement in real-time and accurately in the online learning process is of great significance to the development of online education.Moreover,it can give learners enough supervision and feedback to maintain their learning state and give teachers timely feedback to improve their teaching quality.In order to simplify the way of getting data and meet the applicability of the algorithm,we focus on the online learning scenarios and capture videos of learners’ learning process with the cameras commonly used on electronic devices.Besides,we mainly research algorithms of learning engagement assessment based on video data.Previous methods can only assess the video-level learning engagement for the short-length videos.However,the entire process of online learning usually lasts for dozens of minutes,making it challenging to map the low-level frame features to the top-level labels for long videos.Also,the previous methods need to be trained on the data with fine-grained annotation to achieve real-time learning engagement assessment,which means the data needs to be annotated every few seconds or several frames.In addition,the requirement of providing feedback in real-time also raises the problem of balancing the accuracy and efficiency.Focusing on the problems mentioned before,the main work has three aspects:(a)We propose a temporal multiple instance learning attention module based on weakly supervised learning to avoid fine-grained videos.This module can be placed after the feature extraction module as a flexible plug-in and can assess fine-grained learning engagement for a clip with 30 frames in our paper under the supervision of coarse-grained annotated videos only with video-level labels.(b)We construct a lightweight hierarchical neural network model to obtain the relationship between the low-level features and top-level labels.We combine two layers of feature aggregation modules consisting of a feature extraction module,a multiple instance module and a feature fusion module.The bottom layer can capture the mapping relationship between framelevel and clip-level features and the top one can capture the relationship between clip-level features and video-level features.(c)We collect a new data set with long videos that can better reflect the real online learning scenarios to verify the effectiveness of the model proposed.In conclusion,we innovatively propose a real-time learning engagement assessment model based on weakly supervised learning to meet the needs of practical applications,which can obtain state-of-the-art results on both the video-level and clip-level learning engagement assessment tasks.