Source-linked AI summary
Attentional Graph Convolutional Networks for Knowledge Concept Recommendation in MOOCs in a Heterogeneous View
Shen Wang, Jibing Gong, Jinlong Wang, Wenzheng Feng, Hao Peng, Jie Tang, Philip S. Yu
TL;DR
MOOC course recommendation overlooks students’ interest in specific knowledge concepts, while sparse interactions and heterogeneous relationships make fine-grained recommendation difficult. ACKRec addresses this gap with an attention-based heterogeneous graph convolutional network that combines content and contextual information through meta-paths. On real XuetangX data, the proposed approach outperforms strong baselines.
Problem
Course-level MOOC recommendation overlooks students’ interests in specific knowledge concepts, and sparse student–knowledge concept relationships limit recommendation performance.
Method
ACKRec models MOOC entities and relationships as a heterogeneous information network and uses attention-based graph convolution with meta-path guidance to combine content and contextual information.
Results
ACKRec outperforms strong baselines in comprehensive experiments on real data collected from XuetangX.
Takeaways & Limitations
Knowledge concept recommendation provides a more fine-grained recommendation target than conventional MOOC course recommendation.
Abstract
from arXiv · showhide
Massive open online courses are becoming a modish way for education, which provides a large-scale and open-access learning opportunity for students to grasp the knowledge. To attract students' interest, the recommendation system is applied by MOOCs providers to recommend courses to students. However, as a course usually consists of a number of video lectures, with each one covering some specific knowledge concepts, directly recommending courses overlook students'interest to some specific knowledge concepts. To fill this gap, in this paper, we study the problem of knowledge concept recommendation. We propose an end-to-end graph neural network-based approach calledAttentionalHeterogeneous Graph Convolutional Deep Knowledge Recommender(ACKRec) for knowledge concept recommendation in MOOCs. Like other recommendation problems, it suffers from sparsity issues. To address this issue, we leverage both content information and context information to learn the representation of entities via graph convolution network. In addition to students and knowledge concepts, we consider other types of entities (e.g., courses, videos, teachers) and construct a heterogeneous information network to capture the corresponding fruitful semantic relationships among different types of entities and incorporate them into the representation learning process. Specifically, we use meta-path on the HIN to guide the propagation of students' preferences. With the help of these meta-paths, the students' preference distribution with respect to a candidate knowledge concept can be captured. Furthermore, we propose an attention mechanism to adaptively fuse the context information from different meta-paths, in order to capture the different interests of different students. The promising experiment results show that the proposedACKRecis able to effectively recommend knowledge concepts to students pursuing online learning in MOOCs.
1 INTRODUCTION
MOOCs need better ways to capture student interests because course-level recommendation overlooks interest in specific knowledge concepts, while sparse interactions and heterogeneous relationships complicate recommendation. The paper proposes ACKRec, an attention-based heterogeneous graph convolutional framework that incorporates content and contextual information.
- Motivation: 5% is the reported overall course completion rate, motivating better understanding and capture of student interests.
- Motivation: MOOC course recommendation can overlook students’ interests in specific knowledge concepts within video lectures.Knowledge concept recommendation addresses this gap at a more fine-grained level.
- Motivation: Collaborative filtering suffers from sparse student–knowledge concept interactions, limiting recommendation performance.Side information is used to address sparsity in recommendation.
- Motivation: Different entity types and relationships expose semantic connections that single interaction types may overlook.Examples include users, courses, videos, teachers, and relationships among knowledge concepts across courses or videos.
- Proposed approach: ACKRec models MOOC data as a heterogeneous information network and uses attention-based graph convolution to learn entity representations.The framework incorporates content and heterogeneous context, propagating preferences under meta-path guidance.
2 PROBLEM STATEMENT AND SYSTEM ARCHITECTURE
The system recommends a top-N list of knowledge concepts for each user from interactive MOOC data. ACKRec extracts content and context features, selects meta-paths, and learns low-dimensional heterogeneous entity representations with graph convolution and attention.
- Problem statement: Given a user’s interactive MOOC data, the system predicts interest scores and recommends a top-N list of knowledge concepts.
- System architecture: ACKRec is organized around feature extraction, meta-path selection, and representation learning of heterogeneous entities.
- System architecture: Feature extraction uses knowledge-concept names and relationships among entities to construct concept and user features.The relationships include concept–video and concept–course relations.
- System architecture: Selected meta-paths guide graph convolutional representation learning to capture structural correlations among heterogeneous entities.An attention mechanism is then applied within the representation-learning process.
3 PROPOSED METHOD
ACKRec models MOOC data as a heterogeneous information network and learns entity representations from content and contextual relationships. Meta-path-guided graph convolution and attention fuse information across heterogeneous relations to represent users and knowledge concepts.
- 3.1 Feature Extraction: ACKRec extracts word-embedding content features from knowledge-concept and user names, then models learning interactions as contextual features.The contextual relations include user clicks on knowledge concepts, course enrollment, video watching, and courses taught by teachers.
- 3.2 Meta-path Based Relationship: The constructed HIN contains users, courses, videos, teachers, and knowledge concepts linked by heterogeneous relations.Its network schema represents semantic and relation information and supports semantic paths between entities.
- 3.2 Meta-path Based Relationship: A meta-path is a schema-level composite relation that connects object types through a sequence of relation types.The model uses selected meta-paths to describe heterogeneous relationships and guide representation learning.
- 3.3 Attention-based Graph Convolutional Networks for HIN Representation Learning: The graph convolutional network propagates representations over meta-path-specific adjacency matrices using normalized self-connected graphs and shared trainable weights.The propagation starts from content features, applies ReLU transformations, and shares weights across entities to reduce parameters.
- 3.3 Attention-based Graph Convolutional Networks for HIN Representation Learning: Attention weights fuse the representations learned from different meta-paths so their relative importance contributes to the final entity representation.The attention mechanism uses correlations among meta-path representations and normalizes them with a softmax function.
4 EXPERIMENTS
Experiments evaluate ACKRec on real MOOC click sequences using ranking metrics, meta-path combinations, model-parameter settings, baselines, and a recommendation case study. The results favor combining heterogeneous context with content features, while showing sensitivity to meta-paths, representation dimensions, latent factors, and GCN depth.
- Experimental Setup: The XuetangX dataset uses October 2016–December 2017 click sequences for training and January–March 2018 sequences for testing.Each training or test instance represents a user's click history, with the last clicked concept used as the target.
- Experimental Setup: Evaluation uses HR@K, NDCG@K, MRR, and AUC, with top-K settings of 5, 10, and 20.HR@K measures whether the ground-truth item appears in the top K, while NDCG@K accounts for its predicted position.
- Meta-path Combinations: Combining all four user-related meta-paths achieves the best performance, and larger meta-path combinations generally outperform smaller ones.Among individual paths, the reported ranking is MP3 > MP1 > MP2 > MP4; combinations follow the same tendency, although gains are not always large.
- Model Parameters: 30 latent factors, 100 representation dimensions, and approximately 3 GCN layers produce the reported optimal parameter settings.The latent-factor sweep covers 10–40, the representation-dimension sweep covers 20–200, and the layer comparison tests 1–4 layers.
- Baseline Comparison: HIN-based methods outperform the other compared methods, while ACKRecs+r, combining content and context features, achieves the best performance.The authors attribute the advantage to graph convolutional representations, adaptive attention over meta-paths, and richer heterogeneous information.
- Case Study: For one student, recommendations differ between a single meta-path and the combined MP1–MP2–MP3–MP4 setting.The case study uses two top-10 recommendation lists to illustrate how different relationship conditions change the recommendations.
5 RELATED WORK
Prior work applies heterogeneous graph neural networks and meta-paths to represent complex relationships in recommendation and other domains. ACKRec is distinguished by combining attentional graph convolution with heterogeneous user–knowledge concept representations.
- Heterogeneous graph neural networks have been applied to program behavior analysis and malicious-account detection.
- Figure 7 compares recommendations based on four meta-paths with recommendations based on the single path MP2, alongside a student's learning history.
- ACKRec uses attentional graph convolutional networks to represent users and knowledge concepts in heterogeneous information networks.
- Heterogeneous recommendation models use information networks and meta-paths to represent connectivity between users and items.
6 CONCLUSION
The paper concludes that ACKRec addresses overlooked knowledge concept recommendation in MOOCs through attentional heterogeneous graph convolution. Experiments on real XuetangX data report effective interest aggregation and performance above strong baselines.
- ACKRec targets knowledge concept recommendation, which existing MOOCs recommendation systems often overlook.
- The model propagates context information along meta-paths and uses attention to learn representations of different entities.
- Experiments on real XuetangX data report that the proposed approaches outperform a strong baseline.