Recommender Systems (RSs) are largely used nowadays in many areas to generate items of interest to users. Recently, they are applied in the Technology Enhanced Learning (TEL) field to let recommending relevant learning resources to support teachers or learners' need. In this paper we propose a novel recommendation technique that combines a fuzzy collaborative filtering algorithm with content based one to make better recommendation, using learners' preferences and importance of knowledge to recommend items with different context in order to alleviate the Stability vs. Plasticity problem of TEL Recommender Systems. Empirical evaluations show that the proposed technique is feasible and effective.
Web development has created a need for new techniques to help users find what
they want and also to know that information exists, these techniques are called
Recommender Systems (RSs). These systems are built generally based on two different
types of methods that are Content Based Filtering (CBF) and Collaborative Filtering
(CF). RSs suffer from several problems defined in [
TEL can be differentiated into formal and non-formal learning settings. In nonformal learning, the learners are acting more self-directed and they are responsible for their own learning. The learning process is not designed by an institution or responsible teachers like in formal learning, but it depends on individual learners’ preferences or choices, which is similar to consumers looking for products on the internet. So, lifelong learners are need to have an overview of the available learning activities and materials to decide which of them match better their personal needs, preferences, prior knowledge and current situation. Where the need to use Personalized Recommender Systems (PRS) that use efficiently the available resources in the network and propose learning resources and activities depending on individual needs, learning goals, context, and increase collaboration between learners. But the learner’s need and preferences may change over period of time, also in the same time where he wants to learn from resources with different context. This creates the need of designing Adaptive RSs (ARS) capable of generating recommendations with different tastes depending on the learner’s profile evolution. ARSs design is a great challenge because of the Stability vs. Plasticity problem of these systems.
Whereas recommendations in TEL field depend not only to learner’s preferences
but on the context as demonstrated in [
The paper is organized as follows. Section two, presents the RSs field and the third section, contains some works deployed RSs in the TEL field. Then, we outline our proposed fuzzy hybrid technique to recommend learning resources with different tastes, in section four. Empirical results are presented in the fifth section. Finally, we give some conclusions and lines of future work. 2
RSs provide adequate information to people in need using a representation of the
user called "User Profile". This profile is compared with different profiles available to
determine those to which they correspond [
RSs are built generally based on two different types of methods that are Content
Based Filtering (CBF) and Collaborative Filtering (CF). The CBF approach
generates content recommendations based on the characteristics of users or items, while the
CF method just use the evaluations made by users on the items to predict
the unknown ratings of new user-item pair. Typical CF algorithms can be categorized
into two classes: neighborhood methods and factorization methods. Generally
factorbased algorithms are considered more effective than those based on neighborhood.
But they are often complementary and the best performance is often obtained by
blending them [
One of the major problems of RSs is the Stability problem of these systems
compared to the dynamic profile of the user (Stability vs. Plasticity Problem) [
In order to offer all needs of the active learner that fit their different tastes, we
propose a fuzzy based clustering algorithm to regroup learners including the active
learner, and that guarantees a multi-affectation of learners to nearest clusters allowing them
to receive partial recommendations generated in each cluster according to their
membership degrees. Due to the two major challenges for the CF based systems, which are
the Scalability and Sparcity Problems, the application of traditional FCM algorithm
can confront some difficulties. From this point, our goal was to design an efficient CF
algorithm that guarantee a multiple assignment of a user to different clusters, by
modifying the FCM objective function to a Matrix Factorization (MF) one[
Many RSs have been deployed in TEL, as surveyed in Manouselis and al.[
Generally RSs in e-learning deal with information about the learners and learning activities and would be able to track the evolution of the learner profile (behavior) during his different learning levels. For this aim, we propose a new hybrid technique that combines CF (using MF) with CBF to better predict the learner’s need. The proposed technique allows generating learning resources recommendations to lifelong learners that correspond to their different interests, tracking their profiles evolution. 4
To improve the recommendation quality, we are conducted toward hybridization between CF and a CBF to enhance the CF accuracy in TEL Recommender Systems in order to deal with the sparcity and scalability problems.
Our proposed approach can be divided into two main phases; the first one contains the description of the fuzzy-based CF algorithm and the CBF one, with their messing scores predictions. Then, it presents the hybrid scheme that blends the two predictions in order to obtain a full learner-course matrix. The second phase contains the recommendation algorithm adapted to TEL field by incorporating the learner’s performances in order to generate effective recommendations. 4.1
( ,
The universe of discourse considered in our system is based on pair-wise relationships between two types of entities u and t, which we call “user” and “item”, or “learner” and “course”, respectively. We envision a world with: ─ A set of learners U = { , , …… }; - A set of courses C = { , , …, }. ─ Each item is described by a set of descriptors D(t)= {d1,d2, ..,dn} such that |D (t)|≥1.
A taxonomic descriptor d is an ordered sequence of topics p denoted by d = {p0, p1,…pq} where d ⊆ D(c), c ⊆ C. The topics within a descriptor are sequenced so that the former topics are super topics of the latter topics, when the super topic covers the general term of the domain and sub-topic covers a more specific term. ─ , The evaluation of course c made by learner u. All evaluations made by the learner u form a vector , that represents his profile. The evaluation matrix is R. ─ , The probability that learner u belongs to cluster k; Z= ( , is the probability matrix U × K, where U, K are number of learners and clusters, respectively. ─ , The average of evaluations made by members of cluster k to item t, and C = is the matrix of centroïds K × T, where T is the number of items. 4.2
As mentioned above, this part contains our novel CF algorithm description. From
the literature survey on the CF algorithms, we have the main steps of our algorithm:
─ First, the automatic construction of groups in the system from the evaluation
matrix using the Non-Negative Matrix Factorization (NNMF) method. The reason
behind this choice and use of this method, is the reduction of the scalability
problem that occurs when adding a new user or a new item
─ In addition, the resulting probability matrix can be used to process data to
solve large-scale problems of CF more efficiently.
─ Then, for the neighborhood selection, we propose to consider just the K-nearest
neighbors belonging to the C-nearest clusters following the principle of [
In this step we will factorize the evaluation matrix R into two matrices Z and C. where Z is the probability matrix and C is the matrix of cluster centers.
We will use a modified version of FCM into NNMF following the same principle
of WU and LI [
To resolve this problem, we have used the ACLS algorithm (Alternating
Constrained Least Squares Algorithm) proposed in [
An important step in the CF algorithm is the search for neighbors of the current
learner. Traditional methods generally need to search the entire database, which
definitely suffer from the scalability problem. We proposed an adjusted version of the
fuzzy neighborhood algorithm following the same principle as in [
We proposed to use the difference between membership degrees to the same cluster as a similarity measure between the active learner and members of FCNP. Where, the similarity between two learners increases when the difference between their degrees of belonging tends to 0.
Similar to the idea presented in [
, Where , and , are user-based and item-based predictions.
After the application of CF-based prediction methods, values in the cells of learner-course matrix are recalculated and updated. So, the sparseness of the matrix is therefore reduced. However, there may still be some empty cells due to the inadequate number of nearest neighbors for that learner. For this reason, it is necessary to use content information to make prediction for each learner-course pair. Then, merging the two predictions types to make full evaluation matrix. (2) 4.3
To predict messing values based on content, we must have a set of features to describe the items’ content in order to correlate similar items. In our system, items are courses and features are topics (information used to describe the courses’ content).
We propose to calculate the occurrence frequency of each topic in all evaluated items by the active learner ua. Then, we will give a score to each topic to promote courses according to the topics’ appearance and evaluations made by learner ua to each course. The reason is that two topics and belong to two courses and , respectively, can have the same occurrence frequency in the set of items evaluated by ua, but the course had a better evaluation against the course . Hence, the learner’s preferences should promote ∈ over ∈ through their scores in the preferences’ vector. So, the score assigned to the topic in the preferences’ vector of the learner ua is computed as follows
∑ , . , V || (3)
Such that |c(ua)| is the number of items rated by . rating(ua,c) is the evaluation made by ua to the course c containing topic , and represents the occurrence frequency of the topic in the set of items evaluated by ua.
After have given a score for each topic, we calculate the similarity between the test course and the set of courses assessed by the active learner to select the Tnearest courses to the test course. We propose to use the cosine similarity measure to calculate the similarity between two course vectors.
Finally, we make the content-based prediction of the messing values. The rating prediction for an unseen course is formulated as follows _, ∑ ∑ ∈, , . , (4) Where ratingu , m represents the evaluation made by the learner ua to course ∈ and sim(c,m) is the similarity calculated in the previous section.
This type of prediction use topics to predict messing ratings. So, it needs predictive features to achieve a good prediction which limit the effectiveness use of this prediction lonely. To address limitations of the CF-based and Content-based predictions, we are conducted toward hybridization between them. 4.4
In this section, we will present our hybrid prediction scheme that combines between the CF-based prediction and the Content-based prediction in order to obtain a full user-item matrix. Our proposed hybrid prediction scheme is defined as follows _ =α× _ ( , ) + (1−α)× _ ( , ) (5) Where α is used to control the weight between the two predictions.
After applying the hybrid algorithm cited above, we obtain predictions of the unviewed items by the active learner. Then, we apply the procedure for generating the recommendation. The first step is to calculate the scores of items based on clusters’ preferences and the learner preferences’ prediction, to select the Top-N items in each group. Then, generating a list of Top-K items selected from the Top-N items.
The preferred items (courses) will be determined by the number of nearest learners who evaluated the course (popularity) and their mean explicit evaluations by: _ ( , c)= ( , c) + (1− ) ( ,c) (6)
This formula is based only on the explicit evaluations. To apply this formula in the
TEL field, we introduce the importance of knowledge proposed in [
Moodle1 is a free source e-learning software platform. Due to the lack of no data sets have been made publicly available for formal and non-formal learning, we used a database very known in RSs, BX-Book-Rating2 and we consider that each book is a learning resource or a course. We restricted our validation to a subset of this base by selecting just 21 learners, 20 courses and we have added information about the knowledge level of the learner, which are his test scores. And we integrated it with our technique in the Moodle platform. As showed in Fig.1.
|| ∑ ,
1
We choose the Mean Absolute Error (MAE ) as the evaluation metric to calculate the performance of our CF scheme.
∑ , | | ,| , | (13) N is the number of test evaluations. More MAE is lower, the performance is better.
As we are in the TEL field, we will apply the novel MAE metric proposed by [
To evaluate the performance of Top-K recommendation, we used the metric,
Where P and R are the precision and recall respectively. They are calculated as ,
N: The total number of items; number of relevant items : Number of relevant items found and : Total 5.4
As the data sample on which we applied our algorithm is smaller the used by
[
(14) (15) (16)
From Fig.2. We observe that the MAE has an inverse relationship with the clusters size K, and the different values of α (0.3+0.8), ie. Most K and α are large; most the value of MAE is small. We can notice that the new MAE in almost all cases is smaller than usual MAE, which is due to the subtraction of both products of the values on the y-axis and we know that the levels of RS accuracy are better when the new metric is applied, this is due to the favorable weighting of the users knowledge.
The figure below shows the evolution of the F1metric with number of recommended courses. We observe from Fig.3 that the F1 metric increase until 15 courses evaluated. The F1 values are varied depending on the number of relevant items. It can be seen also that the recommendation performance of the system is good.
Recommender Systems are widely used recently in Technology Enhanced Learning which creates the need to adapt these systems to e-learning. For this and we proposed, in this paper, a novel approach which uses an adapted RS to TEL field. Especially when recommending learning objects that belong to different contexts. Experimental results show that the proposed approach can improve the recommendation accuracy. In the future work, we will elaborate this technique to generate multicontext recommendations taking in the account the temporal dynamics effect.