Hybrid Recommender System for Personalized Pedagogical Resource Recommendations in E-Learning Platforms

Recommender Systems (RS) are essential software operated within a framework to make sure the management of specific date types online. These RS supply users with overwhelming data and furnish a user a high quality navigation experience. They are widely used in several fields, such as movies, documents, music, and news [1]. Like other platforms, e.g. social services networks and e-commerce, E-learning platforms also try to benefit from recommender systems in order to facilitate and improve learning. These E-learning platforms supply learners with appropriate educational tools and help them construct a better pedagogical environment, including resources sharing and collaboration [2]. Online pedagogical resources are digital entities that involve educational design characteristics, such as online books, articles, and courses [3]. Generally, they are included in E-learning platforms to build an intelligent educational platform.

Recommender systems approaches are arranged into content-based, popularity-based, collaborative filtering-based, and hybrid-based.

Popularity-based recommender systems (PR) are considered a basic form of collaborative filtering (CF), as we will see later, where items shared by most users are considered popular [4]. This system works with trends. Essentially, use currently popular articles. For example, suppose every new user commonly purchases a product. In that case, it can potentially recommend that product to users who have just logged in. Recommending the preferences of most users is somewhat impersonal because it recommends the majority of users' preferences [5].

In Content-based (CB) recommendation systems, the system utilizes the content of items such as categories, descriptions, etc., to construct a profile that includes the extracted information about these items [6]. Typically, the profile consists of all the items liked previously by the user and is automatically updated based on their comments. It uses information search and filtering techniques to match certain candidate items with items that users have previously rated, aiming to recommend the best items to users [7]. 

In Collaborative Filtering (CF), any given user has mostly similar tastes and preferences to others [8]. The underlying assumption of the collaborative filtering method is that it is possible to make reasonable predictions about the preferences of active users for items they have not yet rated based on the opinions of other users. These methods assume that users with the same preference for one set of items may have the same preference for another that they have not yet rated [9]. CF usually has two main sub-families: model-based and memory-based methods.

Memory-based takes into account all user ratings available when computing recommendations. This approach uses a similarity measure to predict recommendations based on user-wise or item-wise. Where, user-based recommends to users items referring to other like-minded users' opinions on the item. In contrast, item-based collaborative filtering provides a user with an item as a recommendation based on other items with high relevance by calculating their similarity [10].

The fundamental concept behind model-based recommender systems is to construct a "model" using ratings from a dataset. Once user patterns are identified, personalized predictions are automatically generated based on their profiles [10]. The matrix factorization (MF) is the most known model-based approach [11].

To compensate for the weaknesses of each method and provide the best recommendation, a hybrid recommender system (HRS) combines at least two methods. Combining multiple techniques to address the shortcomings of each and exploit their advantages individually has attracted researchers’ interest [12]. In general, hybrid filtering uses some strategy to combine the recommendation set to create the final recommendation for the user. Burke identified seven mixing strategies: weighting, cascading, switching, meta-level, feature combination, mixed, and feature augmentation [13]. Among hybrid recommender frameworks having been explored in the context of e-learning, we cite the next online learning platforms:

• LinkedIn Learning which assembles content-based techniques to collaborative filtering to suggest courses. Collaborative filtering determines courses referring to similar users’ behavior. Content-based filtering is used to recommend courses that match skills and interests identified from the user’s LinkedIn profile and activities [14].
• Coursera which is a massive open online course (MOOC) platform. It uses a hybrid approach to combine collaborative with content-based filtering. Collaborative filtering is used to recommend courses based on very similar students and their behaviors. Content-based methods are used to identify courses that match the user’s education, career goals, and implicit interests [15].
• edX which is a MOOC platform. It integrates collaborative filtering with personalized learning paths. Collaborative filtering is used to recommend courses based on the behavior of similar learners. Personalized learning paths are designed using a content-based approach, taking into account the learner's ability level, referred learning style and previous course completion [15].
• Smart Sparrow, which is an adaptive e-learning platform. It combines collaborative filtering with learning analytics. Collaborative filtering recommends courses and activities based on similar learners’ preferences. Learning analytics, integrating real-time data on learner interactions and performance, are integrated to dynamically adapt recommendations based on learners’ current needs and progress.

In addition to enhancing user navigation and handling large datasets, recommender systems also encounter challenges such as the issue of cold-start, which arises where new users (students in e-learning systems) have no prior information available in the system [16]. Another problem with recommender systems is data sparsity [14], where the rated items number is tiny compared to unrated items. Furthermore, scalability is a major challenge when recommender systems require abundant training data [17].

To enhance the usability and performance of RS, it is important to overcome the issue of cold start. There are many manners to deal with the issue of cold start [18]. Among them:

• Content-Based Recommendations: Those analyze items content and use this information to provide initial recommendations to users with little or no interaction history. This method is efficacious for new items or users because it depend on the items intrinsic properties;
• The use of demographic and contextual information like age, gender, location, etc., to provide initial recommendations to users with limited interaction history;
• Hybrid Models that combine multiple recommendation techniques to take advantage of different approaches. They generally integrate content-based recommendations, demographic information, or other knowledge-based methods with collaborative filtering or matrix factorization. Hybrid models can provide more meaningful recommendations, especially during the cold start phase;
• And finally, the use of popular items to recommend items that are generally popular or highly rated by the entire user base. This helps to provide a starting point for new users.

Our research proposed to assemble collaborative filtering and popularity-based approaches for pedagogical resources by using some hybridization strategies. The popularity-based approach is used with new learners to tackle the issue of cold start by suggesting popular items to new users referring to the user’s overall preferences. We use two kinds of CF approaches: Singular Value Decomposition (SVD) and Low-Rank Matrix Factorization (LRMF) enhanced with deep neural networks (DNN) in the CF approach with standard learners.

The SVD is a statistical technique that decomposes one dimension matrix to three matrices, taking the latent factors and the underlying structure of the original matrix. In the recommender systems context, the matrix user-item interaction, with users as rows, items as columns, and entries are user-item interactions (e.g., ratings), is decomposed using SVD. This decomposition helps to identify hidden patterns or hidden features in the data, such as user preferences or item characteristics. So, SVD technique is based on dimensionality reduction and is one of the techniques to generate higher-quality recommendations for users. While, matrix factorization (MF) is a way of decomposing a matrix to a product of two matrices or more, highlighting the latent factors that contribute to the observed data Matrix factorization is used to model user-item interactions. The matrix user-item interaction is decomposed to matrices containing latent features for users and objects. These latent features capture latent patterns, preferences, or characteristics that affect user choice. LRMF enhanced with DNN is a popular technique in recommender systems. It outperforms the traditional MF model and improves the accuracy.

This work proposes our recommender system architecture. It comprises two main modules. The first one represents the separate popularity-based approach, while the second module encapsulates the collaborative filtering approach. Our built system switches between cases based on the learner's criteria. This suggestion addresses the issue of cold start and provides the learner with a list of relevant propositions.

This article is structured as follows. Section 2 presents related work about recommender systems in e-learning. Section 3 presents our proposed architecture. Section 4 discusses data collection, preprocessing, and experiments. Section 5 contains obtained results discussion. Finally, we give a conclusion of our article in section 6.

The proposed recommender system suggests relevant pedagogical resources that match students' interests. A hybrid approach is developed. It comprised two principal techniques, starting with popularity and then applying SVD-based collaborative filtering. In addition, we have used two hybrid strategies: switching and cascading. In Figure 1, the proposed architecture comprised two modules. The first contains a popularity-based component, while the second module includes the SVD-based CF method.

1.png

Figure 1. General architecture of our recommender system

Referring to the criteria of the state of learners, this system switches between two cases. The first one is used when there is a newly registered learner to collect information about him. In this case, the final recommended list is based on two phases. The first phase recommends the most popular items and saves the user preferences using the cascading strategy. Then, in the second phase, and to better enhance the recommended list of the first phase, the system uses the SVD-based collaborative filtering to generate other non-popular items based on user preferences. As a result, new learners get both popular and non-popular recommendation items. However, the system switches to the second case when there is a standard learner. In this case, the system already has information about the learner's preferences. Here, the system uses only the second module that contains the SVD-based collaborative filtering approach to provide the recommendation list. 

Our architecture is explained as follows:

3.1 Popularity-based recommender system

In this module, we are interested in supplying users with the pedagogical resources which are most popular and having not seen before. We can present the steps of the popularity algorithm as:

Step 1: For every user/item pair, we aggregate the implicit interactions’ weights W regarding this item. Because in our system, we have five implicit user interactions (View, Like, Comment, Follow, and Bookmark). To each interaction type is assigned an appropriate weight depending on the level of interaction. For example, Wu, view = 1, Wu, comment = 2.5. In this step, we aggregate all the interactions of the user u regarding a given item i by taking into the count the level of interaction w_u,i. where, i represent one of the implicit interactions. W_u,i represents the user's interest level regarding the item i. It is calculated as follow:

$W_{u, i}=\sum_{k=1}^K \log \left(1+w_{u, i}\right)$              (1)

where, $k \in \mathrm{K}$ representing the number of the possible interactions in the system.

Step 2: Collect all the interaction weights regarding each item to find the popular items in the system.

$\operatorname{tot}\left(W_i\right)=\sum_{u=1}^U W_{u, i}$                    (2)

where, $u \in \mathrm{U}$ represents the number of users interacting with item i. tot(W_i) represents the interest level regarding a given item i of all the users that have interacted with this item. Thus, the items with the higher interaction weight value are considered popular.

Step 3: Generate a candidate list of popular items L1 sorted decently by weight strength.

Step 4: Create new users' primary profiles that contain the generated list L1 and store it in the system.

3.2 Collaborative filtering

In this module, we have adapted two different types of collaborative filtering: SVD collaborative filtering and Low-Rank matrix factorization technique enhanced with DNN architecture.

3.2.1 SVD-based collaborative filtering

We have adapted the SVD matrix factorization technique to address the scalability and data sparsity issues and improve recommendations accuracy. In the recommender system context, SVD decomposes the user/item utility matrix (a sparse matrix) A, defined in the step2 bellow, into three matrices (factors) using equation 3.  SVD is used to reduce the utility matrix dimensionality by extracting its latent factors. Each user and item is mapped to an r-dimension latent space. If A initially contains users as rows and resources as columns, then the k-dimensional rows of US contain latent factors for users and those for resources are contained in the columns of SV^T also k-dimensional. After the decomposition of the original matrix A, the dimensionality reduction, and the reconstruction of a new matrix which represent an optimal approximation to the original matrix A by using equation 3., we obtain this optimal approximation without sparsity and scalability.

We can define the SVD steps as follow:

Step 1: For every user/item pair, we aggregate the implicit interactions’ weight W regarding this item.

Step 2: The utility matrix A of user/item is built. It contains all U = {u1, u2, …, um} representing users and arranged as rows, and all items I = {i1, i2, …, in} placed as columns. The cells contain the different interactions' weights where 0 is assigned to non-interacted items see Table 1.

Table 1. Example of a user/item utility matrix A

Step 3: Break the utility matrix A into its factors with rank(A) =r [25].

$A=U . S . V^T$             (3)

where, $U_{m, r}$, $V_{r, n}$ are orthogonal matrices whose columns represent the eigenvectors of A. $A^T$ and $A^T . A$, respectively.

$S_{r, r}$ is the diagonal matrix with r non-zero elements. Where: (∂1, ∂2, …, ∂r) are the r initial diagonal elements of S with ∂1 ≥ ∂2 ≥ ... ≥ ∂k > 0.

Step 4: We perform the matrix factorization of A by providing the optimal approximation to the matrix A, and by applying the multiplication of the three smaller matrices [25].

$U_k, S_k, V_k^T=S V D_{-}$function $(A, r)$               (4)

In the matrix S, we keep the largest k singular values, where first k rows and first k columns of S are set equal to $S_k$ of (k × k matrix). In addition, we keep the largest k right singular vectors, where first k rows and first k columns of $V^T$ are set equal to $V_k^T$ of (k × n matrix).   Finally, first k rows and first k columns of U are set equal to $U_k$ of (m× k dimension) to keep the k left singular victor [25].

Step 5: Get and sort the user's predictions.

Step 6: Recommend the pedagogical resources with the highest predicted ratings that the user has not yet seen.

3.2.2 LRMF enhanced with DNN architecture

This module employs a LRMF technique enhanced with DNN architecture to ameliorate recommendation accuracy. We have built the model by combining the MF linearity and the non-linearity of the multi-layer perceptron in order to model the latent structure of user elements. The algorithm and architecture are explained below:

Step 1: For every user/item pair, we aggregate implicit interactions’ weights W regarding this item.

Step 2: Create the utility matrix $y_{u i}$ of user-item, which comprises the users U = {u1, u2 …um} placed as rows and the items I = {i1, i2 …in} placed as columns. This obtained matrix is very sparse, with non-interactive elements assigned as 0.

Step 3: Divide the utility matrix into two latent factors which are low-rank, the factor $p_u$ of the user, and the factor $q_i$ of the item [26]:

$\widehat{y_{u l}}=f\left\{u, i \mid p_u, q_i\right\}=\sum_{k=1}^K\left(p_{u k} \times q_{k i}\right)$            (5)

1a.png

Figure 2. Neural low-rank MF architecture

Step 4: Construct the DNN architecture presented in Figure 2 [27]:

• The first network layer is the input layer, with two input vectors $v_u$, and $v_i$, representing a user u and an item i, respectively. These are sparse real-value vectors of user-item pair W_ui interaction weights.
• Apply dropout operation to the input layer to simulate the low-ranking of low-rank matrix factorization by randomly deleting some Historical removal interaction.
• Build an embedding layer. This completely connected layer (with some dropout) mainly projects sparse representations into dense vectors. As a result, we have user-item latent vectors.

where, $\mathrm{P} \in R^{m \times k}$ and $\mathrm{Q} \in R^{k \times n}$ represent users and items compressed latent-factor matrix.

$\theta_f$ represents the hyperparameters of the interaction function model. f is a multi-layer neural network that is formulated as follows:

$\begin{gathered}f\left\{P^T \cdot v_u, Q^T \cdot v_i \mid P, Q, \theta_f\right\} =\emptyset_{\text {out }}\left(\emptyset_X\left(\ldots \emptyset_2\left(\emptyset_1\left(P^T \cdot v_u, Q^T \cdot v_i\right)\right) \ldots\right)\right)\end{gathered}$             (7)

• To predict scores, we feed the embedding layers into a neural architecture of multi-layer with varied number of neurons and hidden layers to find new latent structures from the interactions user-item. We combine the user and item paths to construct a proficient deep learning-based recommender system. However, capturing interactions between latent features of items and users while the concatenation of a simple vector is not enough. So we tackled this issue by adding hidden layers into the combined vector and used multi-layer perceptron (MLP) to acquire how latent vectors of the user and item interact.
• Construct the final layer that gives the predicted interaction score $\widehat{y_u}$ (implement Relu activation function).
• Train the network model, where the objective is minimizing the pointwise loss of $\left(\widehat{y_{u l}}, y_{u i}\right)$ representing the predicted  and the  actual value respectively.

Step 5: For every user, we generate a ranked candidate list of items according to the predictions given by the network.

5.1 Evaluation of performance

This section evaluated our recommendation list's effectiveness for each approach. To do that, we have used the strategy of the cross-validation holdout, in which our data is randomly splited into 80% for the train set and 20% for the test set. Otherwise, we have used the Top-N sample strategy to rank the items for each user and reduce the time consumed to rank all items in the dataset. In our case, we have taken 80 items.

1b.png

Figure 3. Recall measures for the two cases

1c.png

Figure 4. Precision measure for the two cases

To evaluate our system's quality performance, it is interesting to emphasize that the metrics used (accuracy, f1_score, precision, and recall) are modified according to the selected cases. In the first one, which is the case of a new student, we have adapted the cascading hybridization strategy between the SVD CF and the popularity-based. The recall measured 0.42 which means (42%) of the interacted articles of the test set are ranked by the model in top-10 items (Figure 3). It scored a precision of 0.116 (Figure 4.) which means (11.6%) of recommended resources are relevant in top-10 items. The f1_score is 0.495 (Figure 5.), and the accuracy is 0.856 (Figure 6.). Otherwise, in the second case, we have a standard student with information about his preferences in the database. We have provided the SVD-based collaborative filtering to generate the recommendation list. The recall measured 0.519 (Figure 3). In addition, the precision measured 0.20 (Figure 4), the f1_score measured 0.54 (Figure 5), and the accuracy measured 0.857 (Figure 6). All the results of the two cases are presented in the figures below.

1d.png

Figure 5. F1_score measure for the two cases

Table 2. Users with a few interactions

To answer our questions about:

Does the proposed system provide a recommendation for new students and prevent user cold start problems?

We can take an example of our experiment with users with few (less than 5) interactions in the system (see Table 2).

When we consider the user with the ID "-9196668942822132778," who has only one interaction in the system, we can observe that our system provides a recommendation to him (Table 3). According to this example, we can conclude that our system that uses a hybrid approach can prevent the issue of cold-start and supply new learners with recommendations because our system adopts as baseline a popularity-based method to supply the most trending articles to new users.

According to our experiment, the system recorded encouraging results in the above figures. In addition, we have provided a LRMF technique enhanced with DNN architecture in the second case and compared it with SVD. The obtained results are presented in the Figure 7.

In Figure 7, the F1_score and the recall of SVD outperform those of the DNN. The precision of the DNN surpasses the SVD precision. However, the two models are very close in terms of accuracy.

5.2 Comparison and discussion 

Many research studies have been conducted in the RS field in order to provide recommendations to users. A comparison between our proposed model and some other previous propositions is illustrated in Table 4. The recommender system presented in Mediani [24] proposed three models, including a content-based model using the TFIDF technique (CB_TFIDF), a collaborative filtering using SVD technique (CF_SVD), and a hybrid model combining CB_TFIDF and CB_TFIDF. While the work prsented in [29] uses a popularity approach, a collaborative filtering using SVD approach and a Hybrid model.  This Table presents global recall and precision scores, ranking among the top 10 articles in the field of recommender systems for pedagogical resources.

Table 3. Some recommended resources for a given user with a few interactions

1.png

Figure 7. The difference between metrics of DNN and SVD in the case2

Table 4. Comparison of the methods of the recommender system

As shown in the table above, our SVD and DNN methods recorded the best results according to the other results. We all used the same dataset. However, we have used some data engineering explained in the section above to better train the model. In addition, we have proposed a new architecture adapted to the E-learning platform allow us to tackle the cold start problem.