"1. INTRODUCTION. In the last years, distance learning has become very popular, especially, its modality of online education. In this regard, online students have different characteristics, needs and preferences to the ones who attend face-to-face classes. For instance, online learners are often over 30 years old, which means having learners with different professional and educational backgrounds in the same classroom. This heterogeneity entails helping students with different needs at the same time. Likewise, because most learners are adult, they have a lot of responsibilities, such as: working, taking care of children, etc. Consequently, it is hardly surprising that they prefer studying at their convenience, i.e. anytime, anywhere and at their own pace. As a result of the aforementioned particularities, online education should avoid using a teacher-centered paradigm, since it treats all learners as if they were the same and forces them to do the same things in the same amount of time [32]. Instead, online education, as any kind of distance learning, involves a student-centered approach in which the instructor is a facilitator and students engage in peer learning [19, 25]. Thus, students are required to take primary responsibility for their learning process [4]. To do this, learners need to have some specific skills, e.g. self-regulation. However, some studies [23, 26] provide evidence that a lot of students need some help to learn these skills, since most of them are not able to achieve these abilities on their own. Consequently, the role of teachers shifts from a masterful figure to a facilitator who guides learners throughout the course and arranges meaningful learner-centered experiences [29]. This role is essential in online education, since learners may become disoriented, frustrated or confused if they do not receive the support, feedback or guidance needed to be successful [8]. In order to carry out a good instruction in an online environment, instructors need appropriate means to set up a monitoring process so that they can be aware of the students’ learning process and provide learners with just-intime assistance. In addition, monitoring allows teachers to forecast potential problems (e.g. dropouts) and avoid them in time. In this regard, instructors can use tracking information (e.g. logs) and monitoring tools which are available in learning management systems (LMSs). However, a comparative study on tracking functionalities of various LMSs [10] concluded that none of them offer much tracking ability. One of the reasons of this devastating conclusion may be the fact that these platforms often provide tracking data in a tabular format which is commonly poorly structured, incomprehensible and difficult to understand [21]. To avoid the aforesaid drawbacks, this paper presents a novel graphical educational monitoring tool based on faceted browsing and data portraits. On the one hand, faceted browsing, as an exploratory search technique, helps instructors to narrow the class roster down until finding those learners who meet teachers’ requirements. On the other hand, data portraits depict summarized information about an individual student in a single image. The rest of the paper is structured as follows: a categorization of educational monitoring tools can be found in the next section. In section 3, an educational monitoring tool based on faceted browsing and data portraits is proposed. Finally, conclusions and future work are in section 4. 2. CLASSIFICATION OF EDUCATIONAL MONITORING TOOLS. This paper classifies educational monitoring tools into two categories: (1) according to the data processing techniques that they use, and (2) the element which they monitor. They both are described below. 2.1 Data Processing Techniques. The main goal of any monitoring tool, whether this has an educational purpose or not, is to give users insights on the data at which they are looking. To this end, two different data processing techniques can mainly be used. On the one hand, information visualization (IV) techniques and, on the other, data mining (DM) algorithms. Next, they both are explained in sections 2.1.1 and 2.1.2, respectively. At the same time, some relevant educational examples related to each method are described briefly. Finally, a brief discussion about the main differences between IV and DM is presented in the section 2.1.3. 2.1.1 Information Visualization. One of the most well-known definitions of information visualization (IV) was proposed by Card et al. [3]: “the use of computer-supported, interactive, visual representations of abstract data to amplify cognition”. Thus, IV encompasses a set of techniques that transform data into effective graphical representations by taking properties of human visual perception into consideration. Thereby, these visual representations reveal facts and trends which allow users to infer some unknown information by combining the visual inputs with their knowledge of data. To stress that the data processing which is performed in IV is mainly based on simple mathematical and statistical functions, such as sum, calculation of percentages, mean, median, mode, and so on. In the educational context, some proposals based on IV techniques have been suggested. Two of the most relevant tools are CourseVis [21] and, its successor, GISMO [22]. The former is a stand-alone visualization tool that obtains tracking data from WebCT, transforms them into a form convenient for processing and generates graphical representations that can be explored and manipulated by instructors. Thereby, teachers can examine social (it uses a 3D visualization of discussion boards), cognitive (it uses a matrix in which each cell is the grade attained by a specific student in a particular quiz), and behavioral (it shows information about access and participation in a timeline) aspects of students. CourseVis mainly uses 2D visualization techniques, but it also uses color and shape as a third dimension. As far as GISMO is concerned, this also uses LMS tracking data, but in this case from Moodle, to display graphical representations (e.g. bar charts, matrices, etc.) about overall classroom accesses and detailed information of a specific student. This was developed as a Moodle block, but its interface is detached from Moodle’s one. Similarly, Zhang et al. [35] designed Moodog, a visual student tracking data plug-in for Moodle which also sends automatic reminder e-mails to students. This displays information about the course, students, resources and access time. Unlike GISMO, the information provided by Moodog is integrated into Moodle’s interface, keeping the original Moodle’s layout as much as possible. There are many other works that have also proposed different representations based on IV techniques. For instance, Hardy et al. [9] constructed, as a graph, the route taken by one student through the course material during a single work session; Hijon-Neira and Velazquez-Iturbide [11] used an interactive graph (for students’ grades) and a data mountain (for access) by using Prefuse API; Juan et al. [14], in turn, proposed scatter plots and quadrants as well as an evolution graph in order to represent students’ performance; and Bakharia et al. [2] proposed SNAPP, a social network analysis tool that displays, as a graph, the evolution of participants’ relationships within an LMS’s discussion forums. 2.1.2 Data Mining. According to Romero and Ventura [28], data mining (DM) techniques can be classified into four categories: (1) clustering, classification and outlier detection; (2) association rule mining and sequential pattern mining; (3) text mining; and (4) statistics and visualization. However, the latter is not universally seen as a category of DM [1], since the data processing that is carried out is minimum compared to the rest of categories, which are based on artificial intelligence algorithms. In fact, the fourth category proposed by Romero and Ventura accords with the definition of information visualization (IV) presented in the section 2.1.1. Hence, in this paper, only the first three categories defined by Romero and Ventura are considered as belonging to data mining. DM techniques are able to infer underlying patterns from a large database, generating some new type of valuable information such as student models or predictions. These new data can be delivered in writing or visually. For some time now, more and more educational monitoring tools are based on DM. For example, Kosba et al. [17] developed TADV, a system that builds student, group and classroom models by using fuzzy logic. From these models, TADV gives instructors advice, e.g. to advise a learner to review a concept. Thereby, teachers have extra information to make appropriate decisions during the course. On the other hand, Hung and Zhang [12] used statistical models and machine learning algorithms to analyze patterns of online learning behaviors and, at the same time, to make predictions on learning outcomes. Zorrilla et al. [37], in turn, proposed a decision support system that utilized different techniques. First, two clustering algorithms, Expectation-Maximization (EM) and KMeans, were used to characterize, on the one hand, students and, on the other, sessions. The output of EM, a probability distribution, allowed to determine the number of clusters with which K-Means would be executed. Depending on the input data, the clusters obtained from K-Means described either student behavior models or session patterns. In addition to the information provided by the clusters, the proposal of Zorrilla et al. also indicated the resources that were commonly used together. To obtain these data, Apriori, an algorithm for finding association rules, was employed. More proposals of educational monitoring tools based on DM techniques can be found in [1, 28]. 2.1.3 Information Visualization vs. Data Mining. One of the main differences between information visualization (IV) and data mining (DM) is how new information is inferred. Due to the complexity of the algorithms used in DM, these are able to suggest latent information with an explanation based on text, rules, clusters, etc., e.g. “student S is about to drop out because she has not acceded to the classroom for two weeks”. By contrast, in the case of tools based on IV techniques, new information is not inferred by an algorithm, but by users through observing graphics and taking advantage of their knowledge of the domain. As a result of the manner of inferring new information, two issues arise: (1) reliability, and (2) computational cost. With regard to reliability, the users of tools based on DM must rely on the accuracy of the information suggested by the algorithm, whereas the users of tools based on IV infer reliable information based on their own expertise. As far as computational cost is concerned, IV techniques calculates statistical data that require minimum processing, while tools based on DM executes complex artificial intelligence algorithms which are usually time-consuming. Another important aspect is user-friendliness. Merceron and Yacef [24] claim that it is essential to use techniques and measurements which are fairly intuitive and easy to interpret. In this regard, the explanation provided by the tools based on DM often requires users to master the algorithm so as to understand it. Unfortunately, a lot of users neither have this knowledge nor can make time for acquiring it. On the contrary, tools based on IV seem to meet these requirements better, since they use simple statistical data and focus on displaying information in a visual and effective way. Despite the differences between information visualization and data mining, they can work together in an application. For instance, a system can first infer some latent information from a large database by using a DM algorithm and then display it by using a visual representation based on IV techniques. Thereby, users can benefit from the advantages of each technique: on the one hand, the possibility of inferring new data automatically and, on the other, the capability of visual representations to convey information effectively. This combination of DM algorithms with IV techniques is called visual data mining [15]. There is every indication that this new discipline will become a promising research area. 2.2 Monitored Element. Although the most popular name is student monitoring tools, it would be more correct to call them educational monitoring tools. The reason is that they can monitor other items in addition to learners. In this regard, this section describes the elements that are commonly monitored in an educational context. Finally, to stress that both IV and DM techniques can be used to track any of the following items. 2.2.1 Classroom. The most common item is the whole classroom, i.e. all students are considered a unique entity. This entity is characterized by the overall information that comes from combining all students’ tracking data (e.g. class’s average grade in each assignment). Thereby, instructors are provided with an overview of their learners’ performance. Thanks to this, teachers can make decisions that affect the whole class. Most educational monitoring tools show data of the classroom. 2.2.2 Student Group. A particular case of the previous monitoring is the one which focuses on groups. This is useful when group activities are proposed, e.g. tasks that belong to a project-based learning. The supervision of groups allows to obtain information about how students interact each other, who are the most and least participative members of a group, which group has the best and the worst performance, etc. [2, 14, 27] are examples of group monitoring tools. 2.2.3 Individual Student. Teachers often need to take a closer look at a particular learner or make comparisons between students. Consequently, they need tools that provide detailed information about an individual learner (e.g. how many times a student has accessed the LMS). These data may help instructors to gain understanding of the reasons why a specific learner has a particular behavior. Thereby, teachers can offer each student a better support and a tailored learning experience. Regarding this, CourseVis [21] and Moodog [35] are two tools that display information of a particular student. 2.2.4 Resources. The term resource encompasses a wide spectrum of elements, from learning materials (e.g. documents, videos, quizzes, etc.) to educational tools (e.g. forum, chat, LMS’s pages, etc.). Regarding this, there are multiple types of data related to resources that can be tracked, e.g. how many times a document has been read, how many messages a forum has, the path that a learner followed while she was navigating through an LMS, etc. Hence, monitoring resources may provide teachers with valuable information about the instructional design (e.g. to detect a bad design of content pages) and students’ performance (e.g. to detect whether learners are engaging in the course thanks to forums participation). An example of this kind of monitoring is [37]. Likewise, a graphic can simultaneously provide information about various items depending on how this is read. For instance, Mazza and Dimitrova [21] use a matrix to show students’ performance on quizzes. Thereby, if teachers paid attention to a specific column, they would observe the performance of a particular student in all course quizzes. By contrast, if a row were observed, then teachers would analyze the performance of the whole class in a specific quiz. 3. PROPOSAL. From the observations done in the previous section, this paper presents a graphical interactive educational monitoring tool which uses information visualization (IV) techniques. This allows instructors to monitor the class and, at the same time, look details of a particular student. The first part, monitoring of the classroom, is based on faceted browsing, a type of exploratory search. As far as the second part is concerned, a novel technique called data portrait is used to depict the information of a specific learner. A detailed explanation of both parts of the tool is exposed below. Nevertheless, before explaining the proposal of this paper, a brief list of the most common characteristics of educational monitoring tools based on IV is given. This will help to better understand the work presented in this article. 3.1 Features of Educational Monitoring Tools Based on Information Visualization 3.1.1 Stand-alone vs. Built-in. There are tools which have their own interface and collect data from an LMS [21, 22], whereas there are others that, due to the success of LMSs (e.g. Moodle), are integrated into their framework as a plug-in [35]. 3.1.2 Representation 2D vs. 3D. There are two kinds of representation according to the number of axes: 2D and 3D. In this regard, 2D graphics are firmly established. They consist of two dimensions (or axes) in which each of them represents a variable (or attribute) of information (e.g. students, grades, etc.). To a lesser extent, 3D graphics has also been proposed, e.g. the scatter plot for discussion boards suggested in [21]. In general, 2D representations are usually more intuitive for instructors than 3D ones. For that reason, most of educational monitoring tools display 2D graphics. 3.1.3 Multivariate Data. Much more variables than the number of axes of the graphic (i.e. 2D or 3D) can be displayed thanks to the use of different techniques. Next, some of them are described briefly: • Single-axis composition [18]: this is a method whereby an axis represents a large set of variables (e.g. content topics, number of accesses, etc.). This is used in [21]. • Visual components: elements such as color, shape and size, are often used as a data dimension. • Quadrants: these allow to organize data into four different groups. These are employed in [14]. • Mouse events: different mouse actions, such as click and rollover, are also used to show information (e.g. relationships between elements [11]). • Multivariate representations: unlike most of the proposals that depict data as points, bars, lines and so on, there are tools that use a single representation to display multivariate observations with an arbitrary number of variables. This is the case of star plots [13], which show each observation (e.g. a learner) as a star-shaped figure with one ray for each variable (e.g. grades, participation, etc.). 3.1.4 Manipulation. Zoom, rotation (in 3D) and filtering are typical actions that educational monitoring tools allow instructors to do. With regard to filtering, this refers to hide some visual elements so as to emphasize others. Thereby, some values in the graphic act as layers that can be shown and hidden. Take the example of a bar chart that shows the number of posts written by the students of a class. In this case, the graphic has the variables “student” (X-axis), whose values are {S1 ,S2 ,S3 ,S4 ,S5 }, and “number of posts” (Y-axis) with values {4,6,7,5,8}. If the graphic focuses on the students, then this has 5 layers (i.e. S1 -S5 ). Hence, a teacher can indicate that the graphic only shows the bars that belong to the students S2 and S5 . However, she cannot set a general constraint that asks to show those learners who have written between 3 and 9 messages, since the layers 3 and 9 do not exist. Similarly, she cannot filter the information based on other variables that are not layers (e.g. number of accesses). A lot of educational monitoring tools with filtering options are based on static layers, therefore this limits the exploratory search that an instructor can carry out. 3.2 Monitoring of the Classroom. 3.2.1 What Is Faceted Browsing?. Faceted browsing is becoming a popular method to allow users to interactively search and navigate through complex information spaces [16]. This is widely used on a lot of ecommerce websites such as eBay or Wal-Mart. A faceted browser provides users with facet-value pairs that are used for query refinement. Faceted browsing is made up of three stages [34]: opening, middle game and end game. In the opening, the interface shows the whole collection and all facets that can be used. The middle game, in turn, allows users to iteratively narrow down the result set by defining constraints on the values of one or several facets, which refines the search query. Finally, the end game occurs when the user finishes the search by selecting an individual item from the result set and its information is detailed. 3.2.2 Why Is Faceted Browsing Suitable for Educational Monitoring in an Online Environment? As said in section 1, teachers in online education should be guides. To carry out this role, they need to observe students’ behavior by analyzing any feature associated with learners. Thereby, teachers may detect any kind of problem in time. Unlike the other proposals, which show graphics with predefined attributes or queries, faceted browsing allows teachers to perform an exploratory search strategy by using a wide range of orthogonal variables (i.e. facets). Thereby, instructors gain an insight into the behavior of their students by iteratively submitting tentative queries based on more than one facet at the same time. This iterative process (i.e. the middle game) finishes when teachers find relevant information that they did not know or when the result set meets a specific set of requirements that teachers wanted. In that moment, instructors can either use the information found to make decisions (e.g. to send an e-mail or start a new search), or click on a student to see detailed information about her. As seen, faceted browsing seems to be good at monitoring a classroom, since it allows instructors to define any tailored query and find relevant information that they did not know while they explore/refine the result set. Figure 1: The opening stage of the faceted browser proposal. Students remain anonymous. 3.2.3 Proposal of Faceted Browser. This paper presents a built-in faceted browser (see Figure 1) that collects data from an ad hoc learning management system [7]. This uses information visualization techniques in order to effectively display data, leaving knowledge inference in teachers’ hands. Therefore, data mining techniques are not used to obtain underlying information about students. Regarding its interface, 2D graphics are used. Its design is divided into two areas. The main one shows the class roster as a set of cards in which each one includes student’s name and photo. The use of a photo is better than identifying students by looking at points, squares, bars, etc. Actually, Zhao et al. [36] states that there is evidence for the existence of a dedicated face processing system in human’s brain. On the left, there is a menu that has different facets whereby teachers can narrow down or sort out the class. Table 1 gathers the facets that have been included along with their possible values. The chosen facets encompass the three aspects studied by Mazza [20]: sociality (i.e. forum participation), cognition (i.e. assignments grade) and behavior (i.e. studying pace). To stress that facets related to resources, such as the number of times that an item has been accessed, were not included because the platform, in which the faceted browser is placed, does not provide these data. As seen in Table 1, facets can have nominal or numeric values. Those facets that allow gradation are represented with sliders. To indicate the number of students in each facet value, numbers in parentheses and histograms are used. Their values are updated depending on the data set that is shown in the main area every time. This helps teachers to set more meaningful queries during the middle game. As said, teachers can sort out the result set by student’s name, studying pace or the average grade of assignments, in ascendant or descendant order. Finally, different visual elements are used to represent more variables in the main area. The background color of each card indicates the learner’s studying pace. Thereby, the more orange the background is, the more advanced the student is. Besides the background color, the card can have a red border that means that the average grade of assignments is C- or less and, hence, that student would not pass the course if the term finished at that moment. Moreover, the border of the photo can be drawn with black dashes. This means that the student is repeating the subject. Likewise and, because the result set can be ordered, the position of the cards is another visual element that gives extra information. Thereby, if a teacher sorted learners by the average grade of assignments in ascendant order, then students with lower grades would be on the top positions. 3.2.4 An Example of Using the Faceted Browser. From the opening stage shown in Figure 1, a teacher may define any query, e.g. “to retrieve, ordered by studying pace in a descendant way, those students who, regardless of their participation in forums, are not repeating the course and, at the same time, have achieved 24% of activities as well as they have an average grade of assignments equal or greater than C+”. If the previous query were executed, then the result set would be that shown in Figure 2b. In this regard, Figure 2a shows the transition between the opening (see Figure 1) and the middle game (see Figure 2b). As seen, instructors can see how students change their positions or even they dissapear. Thanks to the possibility of seeing the transition, teachers can gain extra understanding of their learners. Finally, it is worth emphasizing that the result of the Figure 2b may be either an intermediate step of the middle game or the last one. This depends on whether the teacher is satisfied with the result and selects one student, or prefers to narrow down the result set by changing some facets. Table 1: Set of facets with their values. Figure 2: Execution of a query in the proposed faceted browser. 3.3 Monitoring of an Individual Student. So far, only information about the whole classroom has been represented. However, instructors need to know details of a specific learner quite often. In this regard, the great majority of monitoring tools usually show the same kind of graphic for both the classroom and the learner. Instead, the present paper suggests using a novel technique, called data portraits, to display an individual student’s information. 3.3.1 What Is a Data Portrait?. According to Donath [5], “data portraits depict their subjects’ accumulated data rather than their faces. They can be visualizations of discussion contributions, browsing histories, social networks, travel patterns, etc. (...). Data portraits depict a person through their digital archive”. In short, the idea behind data portraits is to compactly convey a large amount of information from an individual in a single graphic. An example of data portrait used for discussion forums is PeopleGarden [33]. This is a metaphor in which each user is a flower and, hence, the forum is a garden. Each petal symbolizes a message written by the user. Thereby, the number of petals indicates the user’s posting frequency. Thus, the more petals a flower has, the more active the user is. Likewise, the petal’s color represents if the message is an initial post (in magenta) or a reply (in blue). Moreover, pistil-like circles are used on top of the petals to show how many responses each message has received. To display how old a message is, petals, like in the real life, fade over time. Another similar proposal is daisy maps [13]. A daisy map is a star-shaped glyph that displays the scores received on different parts of an assignment (e.g. reading, writing, etc.). Thereby, each student is a daisy map and each ray has a color depending on the score attained. Lexigraphs [6] is, in turn, a group of data portraits in which each user is represented as a face-like outline. Each one is drawn by the words written in the user’s Twitter account. Thus, silhouettes are updated with each new tweet. Finally, Authorlines [31] is an horizontal timeline with vertical monthly dividers that represents the user’s yearly posting behavior in a set of newsgroups. Each month is divided into weeks, and each week is shown as a vertical lineup of circles. Each circle represents a conversation and its size indicates the number of author’s messages in that thread. Authorlines places threads that were initiated by the author above the timeline, whereas the rest of threads to which she contributed are placed underneath the timeline. 3.3.2 Proposal of Student Data Portrait. The proposed data portrait creates a snapshot of an individual student from the data of her learning process. Each portrait appears when the instructor moves the mouse over a card of the proposed faceted browser. The items shown in the data portrait are the facets of Table 1, except gender. The data portrait (see Figure 3) is a bar that is divided into five squares: number of initial posts (in dark blue), number of replies (light blue), number of read messages (pink), the average grade of the messages (green) and the number of highlighted messages (orange). Therefore, these five squares represent the student’s forums participation. Each one of them change its opacity in order to indicate the level of achievement. The lighter the color of the square is, the lower student’s performance in that item is, and vice versa. Figure 3: Different examples of student data portraits. Likewise, the border of the bar can be red or black. Red means that the student would not pass if the course finished at that moment. Moreover, this can be drawn with dashes. This indicates that the learner is repeating the course. Finally, there are two red markers above and underneath the bar. The former represents the studying pace (in percentage), i.e. how many activities the learner has already finished. Thereby, the top of the bar works as a continuous axis that goes from 0% to 100%. As far as the second marker is concerned, this indicates the average grade of the assignments. In this case, the bottom of the bar works as a discrete axis whose values are: NA/NP, D, C-, C+, B and A. Each value coincides with the ends of each square. 3.3.3 Examples of Using the Student Data Portrait. Figure 3 shows different students’ behaviors by using the proposed data portrait. For example, the learner in Figure 3a participates actively in forums, since she writes (opaque dark blue) and replies (opaque light blue) a lot and, at the same time, she reads most of messages (opaque pink). On the other hand, the student in Figure 3b initiates a lot of conversations (opaque dark blue), but she does not reply to other messages (transparent light blue). Therefore, she only writes when she initiates the conversation. By contrast, the learner in Figure 3c helps classmates by participating in threads initiated by others (opaque light blue). However, she does not initiate any conversation. Moreover, the teacher has highlighted some of her messages (opaque orange). As seen, the students in Figure 3b and Figure 3c participate in discussion forums, but in an opposite way. According to Taylor [30], the learner in Figure 3d would be a lurker. A lurker can be defined as a user that writes occasionally or not at all (transparent dark and light blues), but she regularly participates as a reader (opaque pink). Likewise, lurker students usually obtain similar grades to more active learners. For instance, in Figure 3d, this student has a B as the average grade of her assignments. The last three data portraits show: Figure 3e) a student who is repeating the course (dashed border) and, at that moment, she would pass the course (she has a C+); Figure 3f) a lurker who does activities, but she would fail (red border and bottom marker in the beginning); and Figure 3g) a student who has dropped out (all squares are transparent, her studying pace is low and she has not handed in any assignment) and, moreover, is repeating (dashed border). As seen, thanks to the proposed data portrait, it is possible to compact a student’s learning process in a single image. Thereby, teachers can get an overall idea of a student at a glance and compare learners with each other easily. 4. CONCLUSIONS AND FUTURE WORK. Online education involves a student-centered approach in which the instructor is a facilitator. Consequently, teachers should be guides who help their students to achieve learning goals successfully. To carry out this role p"