Related works on collaborative filtering based on learner’s relationship

Collaborative filtering based on learner’s relationship is proposed as to provide asymmetric inter personal influence with implicit method (section 5.3). In this sub section, we review collaborative filtering, inter-personal influence in collaborative filtering and multidimensional recommendation model.

1) Collaborative filtering

Collaborative filtering was proposed to automate the process of “word-of-mouth” [Shardanand and Maes, 1995] by leveraging like-minded users’ opinions. It is an information filtering technique that depends on human beings’ evaluations of items. It infers the interests/preferences of an individual based on the interests/preferences of others with similar tastes. According to [Sarwar et al., 2000], algorithms for collaborative filtering can be grouped into two kinds: collaborative filtering based on user (CF-U) and collaborative filtering based on item (CF-I). Collaborative filtering based on user [Resnick et al., 1994] [Sarwar et al., 2000] [Shardanand and Maes, 1995] is the most successful recommending technique to date, and is extensively used in many commercial recommender systems.

Chapter 2: Related Works 33

These schemes rely on the fact that each person belongs to a larger group of similarly behaving individuals. Consequently, items (i.e. products) frequently purchased by the various members of the group can be used to form the basis of the recommended items. To address the scalability concerns of CF-U algorithms and provide better explaining for recommendation to users, collaborative filtering based on item (CF-I) techniques have been developed [Billsus and Pazzani, 1998] [Breese et al., 1998] [Wolf et al., 1999] [Sarwar et al., 2000]. These approaches analyze the user-item matrix to identify relations between the different items, and then use these relations to compute the list of top-N recommendations. The key motivation behind these schemes is that a user will be more likely purchase items that are similar or related to the items that he/she has already purchased. Since these schemes do not need to identify the neighborhood when a recommendation is requested, they lead to much faster recommendation engines. A number of different schemes have been proposed to compute the relations between the different items based on either probabilistic approaches or more traditional item-to-item correlations.

In contrast to content-based filtering, collaborative filtering is applicable to any of content [Guido and Leuven, 2005], while it can also capture concepts that are hard to represent, such as quality and taste [Herlocker et al., 2002]. Additionally, collaborative filtering does not restrict the spectrum of recommendations to items similar to the ones that the user has previously evaluated. In education, collaborative filtering holds promise not only for the purposes of helping learners and educators find useful resources, but as a means of bringing together people with similar interests and beliefs, and possibly as an aid to the learning process itself [Recker et al., 2003]. Collaborative filtering has become the preferred technology for personal recommendation [Burke, 2002, 2007]. Many approaches and systems, such as Amazon.com [Linden et al., 2003], GroupLens [Resnick, et al., 1994] and etc., adopt this technique.

[Breese et al., 1998] gave classifications of collaborative recommendations, and divided them into two general classes: memory-based and model-based. Memory-based systems are more efficient, in that they generate their recommendations without a need for any preprocessing. Nevertheless, they suffer from serious scalability problems, user-based collaborative filtering [Resnick et al., 1994], and

34 Chapter 2: Related Works

content-based filtering both belong to this category of filtering algorithms. A different approach is taken by model-based systems [Breese et al., 1998]. These algorithms, which often approach the problem from a probabilistic perspective [Chen and George, 1999], produce their predictions by first developing a model of user ratings. The construction of that model requires time but once created, the generation of the recommendations can be really fast. The model building process is performed by different machine learning algorithms such as Bayesian networks [Breese et al., 1998], singular value decomposition with neural network classification [Billsus and Pazzani, 1998.], and induction rule learning [Basu et al., 1998] and clustering [Hsu, 2008] and so on.

2) Symmetric and asymmetric inter-personal influence

In collaborative filtering, similarity – which is a symmetric relationship between users – plays a central role. However, it is in fact the inter-personal influence – an asymmetric relationship – that most directly and effectively supports the automation of the word-of-mouth process. For instance, asymmetric relationships such as employer-to-employee, teacher-to-student, and physician-to- patient have a much stronger influence on decisions than their reverse relationships [Song et al., 2006]. And not all groups exert the same amount of influence on an individual. Researchers have proposed trust-based collaborative filtering [Donovan and Smyth, 2005] [Riggs and Wilensky, 2001]

[Massa and Avesani, 2004]. Such methods derive the neighbors’ trust explicitly or implicitly and use it as a supplementary criterion of similarity to select more credible neighbors.

In education, students of social relations have stressed two relational categories: (1) communion (solidarity) relations (e.g., liking) marked by expected symmetry, and (2) power relations (e.g., influencing) marked by expected asymmetry (if A influences B, then B is expected not to influence A) [Brown and Gilman, 1960]. And symmetric relations would convey communion, asymmetric relations would convey power [Guido and Leuven, 2005]. Consequently, the analysis of social interactions among learners is important. When choosing learning resources, some learners’ decision is strongly influenced by learner with professional expertise. Therefore, recommender systems of e-learning are the most paradigmatic in the asymmetric situation. It is in fact the inter-personal influence, an asymmetric relationship such as advanced learner- to- beginner learner, which most

Chapter 2: Related Works 35

directly and effectively supports the learning resource recommendation. Enhancing traditional recommender system by incorporating social factors can gain accuracy and earn greater user acceptance.

[Song et al., 2006] shows that symmetric similarity is not as direct and effective as asymmetric inter-personal influence between people for recommendation. They also recently proposed a novel asymmetric recommendation algorithm based on early adoption based information flow network using such asymmetric interpersonal relationships [Song et al., 2007]. This algorithm is based on the diffuse of innovation theory [Rogers, 1995][Manzoni and Angehrn, 1997/1998][Angehrn, 1995], and pair-wise comparisons to describe how likely each user make adoptions earlier than the other user and it model information flow behavior of a node by using ergodic Markov Chains. Kawamae et al.

adopt this model to their research to search relative innovator [Kawamae et al., 2007].

Based on the previous mentioned reasons, recommendation of learning resources is based on asymmetric relationship. The method which provides asymmetric relationship based on the diffuse of innovation theory proposed by [Song et al., 2006, 2007] opens profoundly opportunities for the new approach for collaborative filtering, but we found some weaknesses using it in e-learning domain. For example, a learner didn’t understand the learning resources, but he goes on learning, the approach could infer he is an advanced learner based on only his learning process. Therefore, it is necessary to filter out the learning resources that learner didn’t understand from the learning process.

Second, recommendation approach reply upon implicitly or explicitly acquired behavioral data denoting learners’ interests and the factors that characterized the learners themselves such as innovators, but ignored concerned with factors such as learners proficiency level and knowledge level. To address these weaknesses, in our research, we focused our efforts on considering learners’

learning activities, and automatically extract some interaction indicators based on learners learning activities. Then, we use these interaction indicators to generate some comparison indicators. These indicators are as symbols in order to describe a situation and relative degree which knowledge and understanding are socially distributed among group learners. Thirdly, we use machine learning approach for acquiring relationships of learners according to the indicators. That is, we identify the

36 Chapter 2: Related Works

relationships of learners based on the learners’ learning processes and interaction indicators by using the Markov Chain Model implicitly (Chapter 4).

3) Multidimensional recommendation model (MD model)

A multidimensional recommendation model (MD model) makes recommendations based on multiple dimensions and, therefore, extends the classical two-dimensional (2D) Users × Items paradigm [Adomavicius, 2005]. For example, the time, place and movie-viewing companion can be considered as categories with which to augment a recommender like MovieLens with contextual dimensions [Luca and McLoughlin, 2002]. This method uses the reduction-based approach that uses only the ratings that pertain to the context of the user-specified criteria in which a recommendation is made.

Moreover, this method combines some of the multi-strategies and local machine learning methods with On-Line Analytical Processing (OLAP) and marketing segmentation methods to predict unknown ratings. A multidimensional approach to recommender systems that can provide recommendations based on additional contextual information besides the typical information on users and items used in most of the current recommender system has been proposed [Adomavicius, 2005]. And a multidimensional rating estimation method capable of selecting two-dimensional segments of ratings pertinent to the recommendation context and applying standard collaborative filtering or other traditional two-dimensional rating estimation techniques to these segments is also presented [Adomavicius, 2005].

In our approach, we use three-dimensional (3D) User(learner) × Item(learning resource)

×User(relationship) paradigm to make recommendations. In this research, a recommendation approach, which was developed based on the multidimensional collaborative filtering, is presented to show how learning resources can be recommended to each learner based on asymmetric personal influence. Moreover, a method of combining learning step impact coefficient with an individual’s recommendation is proposed.

Chapter 2: Related Works 37