Elsevier

Information Sciences

Volume 479, April 2019, Pages 231-249
Information Sciences

Learning peer recommendation using attention-driven CNN with interaction tripartite graph

https://doi.org/10.1016/j.ins.2018.12.003Get rights and content

Abstract

Learning peer recommendation (LPR) is one of the effective solutions to overcome the information load of learners. This paper presents a multi-objective LPR framework for online learning. Using a dynamic interaction tripartite graph (DITG), we characterize and model the complex relationships among learners, learning content, and interaction behaviours, followed by capturing the dynamic interactions among learners with an attention-driven convolution neural network (CNN). The proposed attention-driven CNN is leveraged to tune the weights of interaction behaviours according to the features of the learning content. A multi-objective function composed of three conflicting metrics, interaction intensity, diversity and novelty, is optimized to achieve simultaneous multiple recommendations for a group of learners. Compared to the state-of-the-art approaches, the proposed LPR framework and algorithms perform favourably.

Introduction

With the rapid development of online learning and social networks, online learners not only interact with each other at any time and from anywhere, but also make full use of learning from their peers’ wisdom [17], [29]. Confronted by a large body of information, sometimes learners can feel lonely and overwhelmed [6]. However, a learner’s needs can be identified by analyzing a large amount of available information, such as grades, learning content and interactions among learners [35]. Therefore, personalized recommendation is becoming increasingly important to alleviate information overload [46]. Unlike traditional recommendation systems for commercial products such as movies, different kinds of interactions among learners will significantly affect the performance of a recommendation system [40].

Traditional approaches for personalized learning recommendation mainly consider the static factors of learners. Static factors refer to learner profile attributes that remain unchanged over a given period, such as historical tags, scores and other information on learning content or learners [23]. As an example, we can use a tripartite graph to model such factors. As we know, interactions among learning peers take place frequently and usually keep changing [43]. From educational perspectives, it is insufficient to use a tripartite graph to model static factors (tags or scores) [12], [32], [44]. Therefore, in this paper, we propose to use a dynamic interaction tripartite graph (DITG) to model peer interactions.

To extract deep interaction features from the collected data, conventional neural networks such as multilayer perceptions (MLPs) with the backpropagation algorithm [10], [42] perform poorly for learning peer recommendation. In the context of big data, advanced learning models, such as CNN [11] and stochastic configuration networks [37], are better alternatives for capturing complex and dynamic relations in learning analysis [11], [13]. It has been observed that highly accurate recommendations can be easily achieved for popular learners [38], but the recommendation diversity is undoubtedly lost. Thus, it is useful and challenging to design a learning peer recommendation system which is well balanced between multiple objectives, such as accuracy, diversity and novelty. In this paper, we propose a multi-objective framework for learning peer recommendation based on DITG and an attention-driven CNN (LPRACNN). Specifically, we first construct a DITG with manually assigned weights that reflect the complex relationships between learners and content from learning objective perspectives. Then, we devise two novel layers of a scaler layer and an attention-driven CNN to tune the initial weights of the DITG. After obtaining different interaction intensities among learners, we optimize the proposed system in terms of diversity, novelty and interaction intensity. From the experiment results, it can be seen that the proposed LPRACNN achieves significant improvements in recommendation performance. The following summarizes our contributions:

(1) The development of an attention-driven CNN-based framework for recommending learning peers: multi-objective learning peer recommendation to balance the interaction intensity, novelty and diversity of recommendation lists.

(2) The use of a dynamic interaction tripartite graph (DITG) for peer interactions: considering the dynamic nature of the online learning process, we use DITG to integrate dynamic interaction behaviours into a tripartite graph with dynamic weights in a real-time manner.

(3) Learning the parameters in our model by both manual settings in DITG and automatic adjustment in an attention-driven network: first, the initial weights in DITG are set manually according to the education design principle. Second, we design an attention-driven network to adjust these weights through a normal training process. This differs from the common initial random weights in that it saves training time and takes pedagogies into account.

The remainder of this paper is organized as follows. Section 2 briefly reviews the related work. Section 3 formulates the problem of learning peer recommendation and briefly introduces our framework. Section 4 describes the proposed attention-driven CNN model for multi-objective learning peer recommendation. Section 5 reports our experiment results with comparisons, and Section 6 concludes this paper with suggestions for future studies.

Section snippets

Related work

This section briefly reviews some related work, including learning recommendation systems, graph-based recommendation, and deep learning in recommendation.

Problem formulation

To match the optimal learning peers with target learners Lt, we present an LPR approach that utilizes an attention-driven CNN model based on constructing DITG. Before describing our learner recommendation model, we introduce the notations used in this paper.

Assumption 1 (Learning objective). In terms of target learners’ learning objectives LOLt running through the whole learning process, Lt should be tightly objective-centric. Therefore, there is a hypothesis that the learning objectives of Lt

Attention-driven CNN based learning peer recommendation model

In this section, we present a deep learning based learning peer recommendation model with DITG. As shown in Fig. 1, the proposed DITG can represent the interaction behaviours of learners. Therefore, we employ deep learning to develop a learner recommendation model using the DITG records as training data. We detail the model in the following section.

Performance evaluation

This section reports the experiment results to demonstrate the performance of our proposed LPRACNN framework. Our experiments consist of three parts. The first is to validate the proposed multi-objective optimization. The second is for assessing attention-driven dynamic performance, while the last one is to comprehensively compare our framework with other relevant ones.

Conclusion

As online learning environments are popular nowadays, learning peer recommendation is becoming important. In this paper, we presented a novel framework for recommending learning peers using a tripartite graph and CNN. Our framework takes into account dynamic interaction behaviours and an attention mechanism. From the perspective of educational design, we integrated five interaction behaviours with corresponding weights into ternary relations in a dynamic interaction tripartite graph. A deep CNN

Acknowledgments

This work was supported by the Major Project of the National Social Science Fund of China (No. 18ZDA334), the National Natural Science Foundation of China (No. 61877020 and 61370178), and the S&T Project of Guangdong Province (No. 2015A030401087).

References (46)

  • D. Wang et al.

    Stochastic configuration networks: fundamentals and algorithms

    IEEE Trans. Cybern.

    (2017)
  • X. Wang et al.

    Dynamic attention deep model for article recommendation by learning human editors’ demonstration

    Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    (2017)
  • J. Wu et al.

    Trust-aware media recommendation in heterogeneous social networks

    World Wide Web

    (2015)
  • S. Ye et al.

    Measuring message propagation and social influence on twitter. com

    Int. J. Commun. Netw. Distrib. Syst.

    (2013)
  • T. Alashkar et al.

    Examples-rules guided deep neural network for makeup recommendation

    Proceedings of the 31th AAAI Conference on Artificial Intelligence

    (2017)
  • S. Bhagwani et al.

    Semantic textual similarity using maximal weighted bipartite graph matching

    Proceedings of the First Joint Conference on Lexical and Computational Semantics

    (2012)
  • C. Castro et al.

    Novel cost-sensitive approach to improve the multilayer perceptron performance on imbalanced data

    IEEE Trans. Neural Netw. Learn. Syst.

    (2013)
  • S. Cavallari et al.

    Learning community embedding with community detection and node embedding on graphs

    Proceedings of the 2017 ACM on Conference on Information and Knowledge Management

    (2017)
  • J. Chen et al.

    Micro tells macro: predicting the popularity of micro-videos via a transductive model

    Proceedings of the 2016 ACM on Multimedia Conference

    (2016)
  • J. Chen et al.

    Attentive collaborative filtering: Multimedia recommendation with item-and component-level attention

    Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval

    (2017)
  • X. Gao et al.

    Attention driven multi-modal similarity learning

    Inf. Sci.

    (2017)
  • Y. Gong et al.

    Hashtag recommendation using attention-based convolutional neural network.

    Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI)

    (2016)
  • Z. Huang et al.

    Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering

    ACM Trans. Inf. Syst.

    (2004)
  • Cited by (18)

    • Personalized Recommendation via Multi-dimensional Meta-paths Temporal Graph Probabilistic Spreading

      2022, Information Processing and Management
      Citation Excerpt :

      Najafabadi, Mohamed and Onn (2019) considered the time impact and modeled user profiles using the graph-based model. Hu, Han, Lin, Huang, and Zhang (2019) used an attention-driven CNN to acquire learners' interactive behaviors in a tripartite graph for peer recommendation. Sánchez and Bellogín (2020) proposed a soften time decay function to measure the similarity between two users for recommendation.

    • Cross-modal recipe retrieval via parallel- and cross-attention networks learning

      2020, Knowledge-Based Systems
      Citation Excerpt :

      Its success is mainly due to the reasonable assumption that human recognition only focuses on selective parts of the whole perception space. By utilizing the merit of dynamic attention weights learning, the attention network has been incorporated into various research fields (e.g., emotion analysis [24–26], sentiment classification [27–29], and item recommendation [30–32]) and achieves great performance improvement. Furthermore, as an extension of the conventional plain attention network, the hierarchical attention network is finely designed to manipulate the hierarchical structure-based tasks, such as multimedia recommendation [33], rock descriptions generation [34], and text summarization [35].

    • Adaptive resource prefetching with spatial–temporal and topic information for educational cloud storage systems

      2019, Knowledge-Based Systems
      Citation Excerpt :

      We report the experiment results of the proposed techniques in the next section. This section reports the results from both simulation experiments and a real-world case study on an educational application named WorldUC [47]. In our experiments, we use several metrics to evaluate the performance of the proposed techniques.

    View all citing articles on Scopus
    View full text