Educational Data Mining

Learning management systems record colossal sets of data. However, this data is used only for building superficial tables about what tasks are executed. Furthermore, these summaries, in most cases, are done only on individual level. In other words, we have a data with all elements to diagnose individual learners in detailed level but we skip it. We are approaching Web 3.0: the Web with intelligence. How about Learning 3.0?

To solve this, we have built Educational Data Mining tool, called EduMiner. The solution is based on teachable AI (same as is in AnimalClass's AI) that can learn conceptual structures in terms of conceptual learning.

In AnimalClass the learner can teach conceptual structures about mathematics, sciences, languages and arts to virtual pets (teachable agents). A data model behind AnimalClass is an applied semantic network. The data model is formed during the game play. A key term is teaching phase that should be defined in details in order to understand the results of this study. A teaching phase consists of a question creation and evaluation of the answer. Each teaching phase adds new relations into the conceptual structure. Furthermore, if the concept is not taught before, the new concept is also added into the conceptual structure during the teaching phase. The following example (figure 1) briefly describes the development of conceptual structures in the agent’s mind during teaching phases. The understanding of how an agent’s conceptual structure develops during playing is important in order to be able to interpret the results of the study.

According to research results (e.g. Ketamo & Suominen 2010; Ketamo 2009; Ketamo, Alajääski & Kiili 2009), the conceptual structure taught for the AI correlates highly to users’ real knowledge measured by tarditional paper tests. In fact, there are more variance between paper tests that between average paper tests and conceptual structure constructed while playing AnimalClass or using Mathematics Navigator.

Currently, we are building a open interface to import all SCORM tagged data into our system. In fact, we can easily import SCORM data, but the difficulties and challenges are related on how the meta data is constructed. In AnimalClass and Mathematics Navigator, all definitions are done by authors who have instructed to use meta tags in proper way. Though SCORM is well defined, the information inside XML is not standardized. For example keywords, difficulty levels and object outcomes are always more or less subjective.

More: www.gameminer.fi

Teaching a virtual pet

Playing educational games, as well as learning in general, should be seen as being an active process that is meaningful from the subject's point of view. The importance of motivational themes should also be considered. The positive relationship between cognitive and motivational themes in sciences and mathematics learning has been widely studied (e.g. Rao, Moely & Sasch 2000; Lapointe, Legault & Batiste 2005; Mason & Scrivani 2004). While a positive relationship between learning and motivation exists, there is no absolute understanding that increased motivation automatically increases learning outcomes. Because of this, motivation and educational outcome should be studied as unconnected phenomena.

In the Animal Class game series (e.g. Ketamo & Suominen 2008; Kiili & Ketamo 2007) motivation to learn is based on a need to teach a virtual pet and on the challenge of competing with this virtual pet against the schooled pets of others. One of the most important features of the Animal Class game series is the freedom of teaching. A pupil can freely teach whatever s/he wants in a given subject area and the game mechanics do not restrict the process. The community, including other pupils, teacher or parents, is the directing force in learning. Animal Class is a game series meant for pupils aged 6-12 years. While playing pupils are asked to teach mathematics, natural sciences, languages and arts to their virtual pets. To date 22 different Animal Class games have been published. When allowing learners to use their knowledge and letting them see how their knowledge fits into the framework of the game, the game retains the idea of 'Learning by Doing' (Dewey 1938). With this approach, the game is no longer the teacher: the game only ensures that the learner has the skills and knowledge required to pass the game.

One of the key elements during the development process of Animal Class was related to the artificial intelligence underlying the game. The traditional goal of AI is to make machines perform cognitive tasks that humans can do, or try to do. In the game industry, the definition of AI is extended so that the most important task of a game's AI is to entertain. It is allowable for game AI to cheat or be 'stupid' in order to achieve the illusion of intelligent behavior (Scott 2002). The balancing issue is also challenging within the domains of game development and AI research: It is easy to create a poor or perfect opponent; the challenge is building a reliable and entertaining opponent. (Liden 2003; Scott 2002). In Animal Class the intelligence of opponents and game balancing is constructed by game players themselves: the virtual pets are teachable intelligent agents and game mechanics are based upon their behavior.

Technically, the game AI in Animal Class is based on a dynamically extensible Bayesian (belief) Network (e.g. Reye 2004). The extensible user model is relatively close to Semantic Bayesian Networks (e.g. Kim, Hong & Sho 2007) and the Qualitative Probabilistic Network (e.g. Lucas 2005). Parallel methods, behavior recording (e.g. Bowling, Furnkranz, Graepel & Musick 2006, Houlette 2003) and behavior mining (e.g. Mukkamala, Xu & Sung 2006, Kuo, Huang, Chen & Jeng 2005), have been studied and used in the game industry for some time. Behavior recording refers to game AI studies and behavior mining usually refers to intrusion detection in networks.

More: www.gameminer.fi