1 Introduction

Online learning is a revolutionary way to provide virtual education in modern life, thus benefiting every day more and more people. A recommender system is a piece of software that helps users to identify interesting and relevant learning information from a large amount of educational information. Recommender systems aim to provide students with search relevant results adapted to their needs by performing predictions on their preferences and delivering those educational contents that could be closer than expected [1]. In addition, recommender systems must use different information sources such as educational databases, Learning Object Repositories (LORs), federations of LORs, among others [2].

Educational recommender systems (ERS), can be classified into several kinds as follows [3]: Content-based ERS in which recommendations are performed only by using the already created student profile. Collaborative ERS wherein recommendations are based on the similarity degree among users by applying collaborative filtering algorithms. Knowledge-based ERS, use user’s browsing history to provide appropriate educational resources. Finally, hybrid ERS seek to integrate some of the recommendation techniques, in order to gather the best accurate features adapted to the user’s profile hence providing better recommendations.

It is important to highlight that ERS for LOs use students’ characteristics and needs in order to support their learning-teaching processes [4]. Generally, ERS are facing three important issues: sparsity, scalability, and cold-start. Using the hybrid ERS we seek to solve these problems by integrating the results of recommendation techniques.

The aim of this paper is to present a student-centered LO recommender system based on a hybrid recommendation technique that combines three following approaches: content-based, collaborative and knowledge-based. In addition, those LOs adapted to the student profile are retrieved from LO repositories using the stored descriptive metadata of these objects.

The rest of the paper is organized as follows: Sect. 2 presents the conceptual framework of this research. Section 3 reviews some related works analysis. Section 4 describes the proposed model. Section 5 explains the model validation and the results of the proposed model. Finally, the main conclusions and future research directions are shown in Sect. 6.

2 Basic Concepts

Following are the main concepts related to hybrid recommender systems, learning objects, and student profile.

2.1 Learning Objects, Repositories and Federations

According to the IEEE, a LO can be defined as a digital entity involving educational design characteristics. Each LO can be used, reused or referenced during computer-supported learning processes, aiming at generating knowledge and competences based on student’s needs [5, 6]. LOs have functional requirements such as accessibility, reuse, and interoperability. The concept of LO requires understanding of how people learn, since this issue directly affects the LO design in each of its three dimensions: pedagogical, didactic, and technological. In addition, LOs have metadata that describe and identify the educational resources involved and facilitate their searching and retrieval. LORs, composed of thousands of LOs, can be defined as specialized digital libraries storing several types of resources heterogeneous, are currently being used in various e-learning environments and belong mainly to educational institutions [7].

Federation of LORs serve to provide educational applications of uniform administration in order to search, retrieve and access specific LO contents available in whatever of LOR groups [8].

2.2 Recommender Systems

Recommender Systems (RS) aim to provide users with search results close to their needs, making predictions of their preferences and delivering those items that could be closer than expected [1, 9]. In the context of LO, Educational Recommender Systems (ERS) deliver educational materials according to the student’s characteristics, preferences and learning needs. In order to improve recommendations, ERS must perform feedback processes and implement mechanisms that enable them to obtain a large amount of information about users and how they use the LOs.

ERS can be classified into several kinds as follows [3, 10]:

• Content-based ERS. In this kind of systems, recommendations are performed based on the user’s profile and created from the content analysis of the LOs that the user has already assessed in the past. The content-based systems use “item-by-item” algorithms generated through the association of correlation rules among those items.

• Collaborative ERS. These systems hold promise in education not only for their purposes of helping learners and educators to find useful educational resources, but also as a means of bringing together people with similar interests and beliefs, and possibly as an aid to the learning process itself.

In this case, the recommendations are based on the similarity degree among users. To achieve a good collaborative recommendation system, i.e. that provides qualified recommendations, it is necessary to use good collaborative filtering algorithms aiming at suggesting new items or predicting the utility of a certain item for a particular user profile based on the choices of other similar user profiles.

• Knowledge-based ERS. The knowledge-based ERS attempt to suggest LOs based on inferences about a user’s needs and preferences. Knowledge-based approaches are distinguished in that they have functional knowledge: they have knowledge about how a particular item meets a particular user need, and can therefore reason about the relationship between a need and a possible recommendation. In addition, these systems are based on the user’s browsing history and his/her previous LO elections.

• Hybrid Recommender Systems. The hybrid approach seeks to combine several ERS techniques in order to complete their best features and thus make better recommendations. The proposed hybrid filtering approach transparently creates and maintains user’s preferences.

To make the hybridization of recommendation techniques – using at least two of them – Burke [11] describes the following different methods that could be applied.

Weighted: the score of different recommendation components are combined numerically.

Switching: the system chooses among recommendation components and applies the selected one.

Mixed: recommendations from different recommenders are presented together.

Cascade: recommenders are given in strict priority, with the lower priority ones breaking ties in the scoring of the higher ones.

Feature combination: features derived, from different knowledge sources, are combined together and given to a single recommendation algorithm.

Feature augmentation: one recommendation technique is used to compute a feature or set of features, which is then part of the input to the next technique.

Meta-level: one recommendation technique is applied and produces some sort of model, which is then the input used by the next technique.

2.3 Student Profile

For a SR deliver tailored results they need profiles that store the information and the preferences of each user [12]. The student profile stores information about the learner, its characteristics and preferences, which can be used to obtain search results according to its specificity. To handle a user profile can be used to support a student or a teacher in the LO selection according to its personal characteristics and preferences [13].

Some research works present the student’s learning style as the most important feature for the delivery of appropriate educational resources [14, 15]. In fact, learning styles are currently used to organize collections of new information and represent different ways through which students can learn. There are different models to represent a student’s learning style. For instance, Duque [16] presents a combination of VARK (Visual, Aural, Read/Write, Kinesthetic) and FSLSM (Felder and Silverman Learning Style Model) models with good results to characterize the student profiles and thus, provide students with learning materials tailored to their specific learning styles.

3 Related Works

Following some hybrid ERS research works will be described. Salehi et al., use genetic algorithms and perform two recommendation processes. The first uses explicit characteristics represented in a matrix of student’s preferences. The second recommendation process assigns implicit weights to educational resources that are considered as chromosomes in the genetic algorithm for optimizing them based on historical values [17].

The authors deliver educational materials adapted to the user profile, combining several types of filtering methods with the available information about objects and users. The first method preselects the LOs from repositories, using a search based on LO metadata, then those objects passed by other filtering processes to obtain a final list which will be the best that suits the user. It is important to highlight that this research work combines several filter criteria: content-based, collaborative activity, and demographics [18].

Vekariya and Kulkami [10] perform a review of some hybrid ERS concluding that the hybrid filter obtained by integrating collaborative and content-based filtering approaches improves predictions made by the recommender. Although this research works quite well in hybrid recommender systems, however, no recommendations tests have been made on educational materials recovered from LORs.

By contrast, the research presented by Sikka et al. [19] works well on learning materials and thus some recommending activities are provided within an e-learning environment by using web mining techniques and software agents. However, authors implement just a unique collaborative recommendation filter rather than using a hybrid approach.

4 Model Proposed

The adaptive recommender model proposed consists of six modules (see Fig. 1) according to a hybrid recommendation technique that combines three following approaches: content-based, collaborative and knowledge-based. In addition, LOs are retrieved from LORs, and federations of LORs, using the stored descriptive metadata of these objects. The student profiles are also available having their personal information, preferences and learning style. Thus, there are three recommendation modules one for each of the selected techniques. A fourth module that performs the hybridization (integration) process uses intermediate recommendations results (see Fig. 1), and finally, the last two modules handle information about student profiles and LO metadata respectively.

Fig. 1.
figure 1figure 1

Adaptive recommender model composing of six modules

Full size image

Among hybridization techniques previously described in Sect. 2.2 we selected and applied the Cascade method wherein LOs recommenders are given in strict priority, consequently recommenders with lower priority affects the scoring of the higher ones.

Following the three main recommendation techniques used in the hybrid model are described.

Content-based Recommendation: this recommendation technique is based on the student profile. The LOs metadata such as Learning-Resource-Type, Interactivity-Level, Intended-End-User-Role, Context, Description, and Language are matched with the student’s learning style.

This process is performed using production rules such as the following:

Collaborative Filtering Recommendation: the aim of this kind of recommendation technique is to deliver LOs that liked or interested to students with similar profiles. The similarity among students can be defined as the numeric representation of the coincidence degrees according to all the characteristics that define their profile. To do so, a similar profile is at first searched, and then, the cosine distance for this case was selected as similarity measure according to the study presented in [7]. The students’ characteristics used are the following: education level, learning style, level of education, language preference, choice of subject, and format preference. The result of the recommendation are those LOs that users with similar profiles have positively assessed.

To perform its calculation the cosine distance is usually used along with vectors whose elements are numeric values and thus mathematical operations on such elements can be performed. This application was extended to categorical data given by formula 1.

$$ {\text{Cosine distance}} = \sum\nolimits_{1}^{n} {\left( {Pi *Qi} \right)} /\sqrt {(\sum\nolimits_{1}^{n} {Pi^{2} } * \sum\nolimits_{1}^{n} {Qi^{2} } } ) $$
(1)

where:

Pi: frequency term i on vector 1

Qi: frequency term i on vector 2.

Knowledge-based Recommendation: This recommendation technique search for similar LOs that the user positively assessed in the past. First, the process starts searching similar LOs that have been assessed in the past, through the following metadata: title, description, and keywords. To calculate the semantic distance among LOs, we used the cosine distance, as appeared in formula 1, which measures the similarity between arrays. In this case, such arrays are the words contained in the title description and keywords of the metadata.

Hybridization process of intermediate results: As previously mentioned the selected recommendation technique is Cascade. This technique applies a filter to each result obtained by implemented recommendations. The process is applied by stages, starting with a recommendation technique to produce an initial rating for each of the candidate items and then, a second technique refines the recommendation among the set of candidates given by the first. In fact, each of the recommendations techniques refines the recommendations given by the others. Figure 2 presents the Cascade hybridization process.

Fig. 2.
figure 2figure 2

Cascade hybridization process

Full size image

Each recommendation technique executes its process at a given time and the final results appeared after applying the Cascade hybridization strategy. In this model, the first selected recommendation technique is the Content-based Recommendation, since this technique according to previous experiments delivers a greater number of LOs. Later, intermediate results are filtered using the Collaborative filtering technique, and finally, the Knowledge-based Recommendation technique is applied to filter again results. LOs results obtained after applying Cascade hybridization process are delivered to students.

Following the two useful external modules used by the hybrid model will be described.

User profile module: this module handles the student profile information using an ontological representation. The ontology proposed contains the student information given by the student profile such as schooling level, learning style, educational level, language, learning topics, and format preferences.

Learning Object Metadata Information module: Finally, this module handles LO metadata. This module has access to LORs and federations of LORs in order to extract the required metadata at each stage of the hybrid recommendation process.

5 Experiments and Results

The system delivers to the student a list of recommended LOs from similar profiles of students sharing learning style, historic behavior, and preferences. The recommendation process starts using a search criterion that can be expressed by keywords or educational skills wishing to be achieved.

The experiment was performed using the LOs stored in FEB (http://feb.ufrgs.br/feb/), the Brazilian Federation of LORs. Initial searches were performed with Portuguese words in order to select the LOs that would initially enter to the recommendation process. Each module executed the recommendations as follows: content-based recommendation module applies the inference rules among LO metadata and student’s learning style. The collaborative recommendation module seeks similar user profiles to deliver items that have been assessed by students with similar profiles and knowledge-based recommendation module searches some LOs similar to those that the student had previously assessed. The integration module performs the hybridization process to deliver the student with the most relevant and appropriate LOs.

In addition, students of Information System Management program at National University of Colombia branch Manizales were selected to (1) use the hybrid recommender system, (2) register his/her user profile, and (3) rank the relevance of the recommendation results.

Burke [11] establishes that precision measurement helps to evaluate the results according to the relevance value given by the student. Thus, the precision measurement analyzes the quality of recovered educational materials regarding to students [20]. A LO is relevant if it supports the student learning process by adapting to his/her preferences and needs. A group of students, who qualify the relevance, determines this relevance value. Formula 2 shows the way this measurement is calculated.

$$ Precision = \frac{Relevant \,LOs}{Relevant\, LOs + Retrieved\, LOs} $$
(2)

After executing the hybrid model proposed using the experimental group of students the results for each technique applied are the following, on average: the content-based technique recovered around 83 LOs for each student and, 42 of them were relevant. The collaborative filtering technique obtained 7 LOs since there are few students registered on the system, with similar profiles, and also, there were few LOs previously evaluated. Thus, just 5 LOs were relevant. The knowledge-based recommendation produced 19 LOs similar to those that the student evaluated previously, and relevant results are only 13 LOs.

The hybrid recommendation system proposed delivered 4 LOs, wherein 3 of them are relevant for students. Table 1 presents the results obtained by applying the precision measurement for each recommendation technique and also the hybridization technique by using the Cascade method.

Table 1. Precision measurement results
Full size table

We can conclude that using hybrid recommendation techniques on learning environments might be promissory because it improves the precision measurement. The problem using the Cascade hybridization technique is the number of resulting LOs. In some cases, the recommendation result was zero, since any technique did not recover LOs or in the process of hybridization, some LOs were lost. The collaborative recommendation technique has a high precision measurement and this result is due to the low volume of recovered LOs.

6 Conclusions and Future Work

An adaptive student-centered LO recommender model is proposed composing of six modules according to a hybrid recommendation technique that combines three following approaches: content-based, collaborative and knowledge-based. In addition, LOs are retrieved from LORs, and federations of LORs, using the stored descriptive metadata of these objects. The student profiles are also available having their personal information, preferences and learning style. Thus, there are three recommendation modules one for each of the selected techniques; a fourth module that performs the hybridization (integration) process with obtained recommendations, and finally, the last two modules handle information about student profiles and LO metadata respectively. In fact, the proposed model delivers educational materials adapted to students’ needs and cognitive characteristics according to different criteria based on recommendation approaches such as content-based, collaborative and knowledge-based. During the testing phase a precision measurement was used to assess the quality of relevance of the recovered LO. By applying this precision measurement, it can be concluded that the hybrid recommendation approach enhances the results of the recommendation in terms of the relevance of the educational material to assist and hence to improve the student’s learning process. The case study performed in order to validate the proposed hybrid recommender model demonstrated the effectiveness of using this kind of approaches in virtual learning environments.

As a future work, we envisage to explore and incorporate more hybridization techniques to the model and perform new case studies in order to compare their performance with previous results.