Prévia do material em texto
1 MACHINE LEARNING EM SISTEMAS DE RECOMENDAÇÃO EDUCACIONAIS: QUESTÕES ÉTICAS A SEREM ENFRENTADAS Fábio Josende Paz 1,2 , Francielle Marques do Nascimento 3, Raphael Leite Campos 1, 4, Priscila Piccolo 3 1 Pós Graduação em Informática na Educação - Universidade Federal do Rio Grande do Sul (UFRGS) 2 Universidade da Região da Campanha (URCAMP) 3 Pós Graduação em Ciência da Computação - Universidade Federal do Rio Grande do Sul (UFRGS) 4 Universidade do Vale do Rio dos Sinos (UNISINOS) fabiopaz@urcamp.edu.br, francielle.nascimento@inf.ufrgs.br, prof.raphaellc@gmail.com, priscila.piccolo@ufrgs.br Abstract: Educational Recommendation Systems (ERS) involves a degree of complexity when it comes to ethical and social aspects, in this sense the objective of this study is to analyze and present some Ethical Issues involved in machine learning in educational recommendation systems. This study is characterized as a bibliographical review of the theme. The specific contribution deals mainly with critical issues derived from the bibliography found on the integration of subjects, such as reflections on the ethical and technical implications of artificial intelligence such as Machine Learning (ML) applied in these systems. The present study intends to arouse interest and help to guide the first steps in ethical issues to be analyzed in initiatives in the field of Recommendation Systems focused on education. 1. Introduction There is a growing demand to deal with an ever-increasing volume of data in order to promote its processing to make it useful as information. Therefore, the information goes on to subsidize decision-making systems, increasing the potential of computational tools that offer different possible solutions, with an increasing impact on people's lives and consequently on education. Computational Intelligence (CI) is a field whose main purpose is to study adaptive mechanisms so that intelligent behavior can solve complex problems that traditional approaches do not have as much effectiveness (ENGELBRECHT, 2007). Among the techniques of CI, Machine Learning (ML) uses concepts of human learning to develop software skills, from experience. According to Brownlee (2014), ML is the training of a model from data that generalizes a decision in relation to a measure of performance. Models can be generated from supervised and unsupervised approaches. The supervised approach requires data sets with defined classes, where training is done based on examples. While in the unsupervised approach, pattern recognition comes from groupings performed on a data 2 set. Recommendation Systems use this approach in an adaptive process and classifications from large databases. Recommendation Systems are considered to be very useful tools, especially for users of online systems, because they can handle properly with large databases. (RICCI et al., 2011), but the Educational Recommendation Systems (ERS) aims to produce predictions and recommendations for its users, usually students and teachers (LIU et al., 2013; IMIRAN et al., 2014; FULANTELLI et al. al., 2015). They can still be used to provide teaching materials and useful information that take students' preferences into account and help them to achieve the learning objectives of the course (LIU et al., 2013, COSTA et al., 2013). However, Xing et al. (2015) and Siemens & Baker (2012) show that many do not present explanations or justifications for these recommendations, leading users to have no confidence in them. Johnson (2014) affirms that the benefits of such techniques are significant, however they present a number of ethical and social challenges. Concerns about ethical issues are increasingly evident in Artificial Intelligence (AI) (BOSTROM and YUDKOWSKY, 2011), Machine Learning and Recommendation Systems are subareas of AI. In this context, some ethical issues arise regarding application of Machine Learning in ERS, but little has been studied on this subject which demonstrates the bibliometric study that is presented in this article. Therefore, the purpose of this study is to analyze and present the ethical issues involved in learning machines in educational recommendation systems. This paper is structured in six sections, including the introduction and the theoretical studies for research background. In section 3, the materials and methods are presented, and in section 4, the analysis of results and discussion. Finally, the final considerations and the references of this study are presented. 2. Theoretical reference In this section we will briefly present the Machine Learning issues in Educational Recommendations Systems, the Moral Dilemmas in Educational Recommendation Systems, in addition to present some Educational Recommendations Systems. 2.1 Machine Learning in Educational Recommendations Systems The recommendation systems are intended to use algorithms to provide recommendations according to the user's interest. Recommendations are made to help the user to find new items based on user information or the recommended item. Techniques such as Machine Learning, Deep Learning, and algorithms in general seek trends in user access, either through (Collaborative Methods), or (Content-Based Methods), or a combination of the two strategies (Hybrid Methods) (PORTUGAL, ALENCAR, COWAN, 2017). These three methods use strategies to make the recommendation to the user in different perspectives of interests. Content-Based uses similar items / products previously accessed by the user. In Collaborative Methods the items are presented to the user based on the preferences and similar tastes of other users. And the Hybrid method can present several models with different combinations, as an example, they produce lists of 3 recommendation of each approach and join the results, or use weights for the filtrations (enhance items that had more accesses) (ADOMAVICIUS, TUZHILIN, 2005). In this context, some Machine Learning techniques that are recurrently applied in Recommendation Systems are Decision Trees, Naive Bayes Classification, Support Vector Machine, Clustering Algorithms, Neural Networks, Deep Learning, Neighbor- based, among others. Most of these techniques are able to classify user profiles, collecting data related to user preferences and behavior, such as income, age, gender, access history, and information that may be useful in the discovery of RS knowledge and application. Likewise, the Educational Recommendation Systems seek through these technological devices to identify and suggest ideal contents for students, based on the characteristics of each group or individual. In this perspective, the teaching and learning process is the central object of the system, since, from a modeling, the system can filter educational resources, make suggestions, personalize content, with the objective of providing the student a better experience of learning process. So, the educational process becomes more interactive, dynamic and innovative. Based on the information and data collected, the ERSs allow a more in-depth assessment of the methodology applied and the resources used in the learning systems. In addition, the technique allows students to focus their studies based on their difficulties, either through reinforcement learning (reinforcing subjects already studied) or to acquire new knowledge (COSTA, AGUIAR, MAGALHÃES, 2013). 2.2 Moral Dilemmas in Educational Recommendation Systems Information and data collected for an ERS application may violate privacy, as they collect characteristics of an individual, either explicitly or implicitly (BERDICHEVSKY, NEUENSCHWANDER, 1999). ERSs use this data to infer a group's profile or an individual's profile, which may present or incorrectly generalize the interests of a class. Although, the purpose of the educational system is to utilize this knowledge to provide differentiated learning and experiencefor the student, the recommendations made by the system can persuade the student to follow a line of reasoning, rather than providing resources for student to form their own opinions. Knowing this, in the development of ERSs, must take into account the information security, privacy, and the impacts that will be generated in the formation and development of the student. In addition, educational systems available on the Internet must have “clear and complete information about the collection, use, storage, processing and protection of its personal data” (BRASIL, 2014), the use of this information is only allowed when specified in the services or terms of use and privacy policy (BRASIL, 2014). In relation to moral dilemmas some Educational Recommendation Systems, among them MELOD, presents a dashboard with indicators that should be used by teachers to analyze students' behavior in relation to their interactions during the learning experience and thus to carry out educational interventions (FULATELLI et al., 2015). 4 Imiran et al. (2014, 2016) in their studies present an ERS tool, the PLORS that connects with Moodle to obtain information from undergraduate students, the focus of the tool is to present which learning objects within the course are most useful to them, considering the learning object they are visiting, as well as the learning objects visited by other students with the same profile. Another ESR found is the ELAT (Learning Analytics Toolkit), which allows teachers to explore and relate the use of learning objects, user behavior (student), as well as evaluation results based on dashboards. I t is intended to use technology to enhance teaching and learning scenarios and to identify opportunities for interventions and improvement (DYCKHOFF et al., 2012). 3. Materials and methods This study is characterized as a bibliographical review on the ethical theme in the application of Machine learning in Educational Recommendation Systems, for which a quantitative bibliometric study was carried out in the Google Scholar database, where articles were searched with the keywords “Machine Learning”; “Educational Recommendation Systems”, “Machine Learning and Ethics”, and “Educational Recommendation Systems and Ethics”. The period for the search for results was between 2012 and March 2018, using articles in the English language published in this period. In general terms, the methodological procedures used were the following: In the bibliometric research the following keywords were used: "Machine Learning"; "Educational Recommendation Systems", "Machine Learning and Ethics" and "Educational Recommendation Systems and Ethics" in the Google Scholar database; It was carried out the crossing of keywords searched in the database; An analysis was carried out by the number of articles published; Additionally, to complement the study, international articles about ethical issues involved in ML and ERS were analyzed. The following are the results and analyzes of what could be observed during this research. 4. Results and discussions The purpose of the quantitative bibliometric review was to analyze the number of publications on the topic of this research, as well as to search for a publication perspective in the stipulated period. Table 1: Quantitative analysis of terms 5 Tabl e 1 sho ws the resul ts of the bibli ome tric study, which shows the importance of Ethics in Machine Learning with 16700 studies in recent years, however, the low number of studies of the union of the keywords “Educational Recommendation Systems” and “Ethics” demonstrates a field which still to be explored, which was chosen as the focus of this study. We make it clear that this article is not intended to bring answers, but rather to bring up some ethical discussions about the application of Machine Learning in Educational Recommendation Systems. Following are some discussions raised in the readings done. Johnson (2014) raises many questions that should be well evaluated in the implementation of Educational Recommendation Systems, including issues related to students' privacy and individuality. The privacy may be violated if personal information is used in an incompatible manner with the context, i.e. information that would not allow some people comfortable with disclosure. Willis et al. (2013) raises an interesting question “Is it worth exchanging privacy for other benefits within the new social economy?” Another risk can be electronic reputation by presenting student performance to teachers and staff through electronic dashboards (FULANTELLI et al., 2015). In the studies by Imiran et al. (2014, 2016) groupings of students are carried out through their profiles. It is noted that the grouping of similar students can be a risk by not taking into account the individuality of the student. In ERS the tendency to treat the individual as a collection of attributes rather than an entire individual. The autonomy of the student can also be withdrawn by implanting suggestions to the students of both contents, and even in the choice of subjects to be taken, always leading them to the ways defined by the system, that is, trying to make the ideal student for the institution and not let him follow his own ways. In the study by Willis et al. (2013) it is presented a list of questions that educational institutions should make when implementing Educational Recommendation Systems, they are: Does college administration allow students to know that their academic behaviors are being monitored? What and how much information should be provided to the student? How should the teacher react to the data? Should the teacher contact the student? In what way? Should the data influence student perceptions and classification tasks? What amount of resources should the institution invest in students who are unlikely to succeed in a course? What is the student's obligation to ask for help? They are perceived as interesting ethical questions and should be analyzed with some care. 6 How can the use of artificial intelligence, based on machine learning from patterns found in historical data series, impact the learning process of human beings? Standardization may come with cultural prejudices and political-ideological biases. In addition to not contemplating the emergence of the learning process. (FERREIRA, 2017). In addition, we must also consider a potential risk that is to use great students to carry out training in the computational models and send these models to not so good students, generating frustration and consequently dropping out of school. In the same line as the AI impacts questionnaire presented by the above authors, the Ethics and Artificial Intelligence Committee of the IEEE (IEEE 2018) adds that moral values are included in intelligent systems, but are included by programmers, engineers who design systems which reflect reality, or the context and should be evaluated. “[...] Computers and robots already reflect values in their choices and actions, but these values are programmed or designed in by the engineers that build the systems. Many of the existing experimental approaches to building moral machines are top-down, in the sense that norms, rules, principles, or procedures are used by the system to evaluate the acceptability of differing courses of action, or as moral standards or goals to be realized.” (IEEE, 2018, p. 43). According to IEEE (2018) the inclusion of rules, principles, norms and ethical procedures in the systems to evaluate the different courses of action are made top-down. However, AI techniques already exist that allow the construction of bottom-up standards, in other words, these systems would learn, just as a child learns to recognize ethical and moral standards in a society. Also in (Ferreira, 2017) the situation is presented in which these systems identify courses of action thatwere not predicted. It is evident how this issue dialogues with the non- consideration of the emergence in the data and updating of the members of the groups. In addition, imperfect datasets impact AI learning for decision making, as punctuated in: “However, biases may still emerge from imperfections in the norm identification process itself, from unrepresentative training sets for machine learning systems, and from programmers’ and designers’ unconscious Embedding Values into Autonomous Intelligent Systems assumptions.” (IEEE 2018). The question of how to assess the impact of decision-making on such cases is pertinent. Considering imperfect data, machine learning techniques such as classification, grouping and association will generate imperfect models. Consequently, supporting Manouselis et al. (2011) affirms that in the case of systems of recommendations aimed at the teaching and learning process, the theme is quite complex, because there are several dimensions of evaluation, such as the pedagogical dimensions that should be taken into account in its construction. 7 5. Final Considerations There is no doubt that Educational Recommendation Systems are useful in teaching, many students, teachers and institutions enjoy their benefits, however ethical issues cannot be left aside when creating and using ERS. We believe that this work will assist institutional researchers who must understand the risks in this deployment and ensure the contextual integrity of the information flow to protect the actors involved. From this study, the authors hope to arouse interest and help guide the first steps in ethical issues to be analyzed in initiatives in the field of Recommendation Systems focused on education. The limitations of this work are in the bibliographical research carried out which some important articles may have been excluded from this study and for future work it is desired to deepen the studies on the ethical issues raised in this article as for example the privacy and individuality of the students, besides the search of the maintenance of its autonomy. References BOSTROM, N.; YUDKOWSKY, E. THE ETHICS OF ARTIFICIAL INTELLIGENCE. Draft for Cambridge Handbook of Artificial Intelligence. P. 1 - 20, 2011 Available in https://nickbostrom.com/ethics/artificial-intelligence.pdf. Acesso 22 abr, 2018. BERDICHEVSKY, Daniel; NEUENSCHWANDER, Erik. Toward an ethics of persuasive technology. COMMUNICATIONS OF THE ACM. Vol. 42, No. 5, May, 2009. Available in Access on: April 25th, 2018 BRASIL. Lei Nº 12.965, de 23 de Abril de 2014. Estabelece Princípios, Garantias, Direitos E Deveres Para O Uso Da Internet No Brasil. Brasília, DF. Abril de 2014. Available in :http://www.planalto.gov.br/ccivil_03/_ato2011-2014/2014/lei/l12965.htm . Access April 26th, 2018. COSTA, E., AGUIAR, J., & MAGALHÃES, J. Sistemas de Recomendação de Recursos Educacionais: conceitos, técnicas e aplicações. In Jornada de Atualização em Informática na Educação (JAIE), Campinas, SP, 2013. DYCKHOFF, A. L., ZIELKE, D., BÜLTMANN, M., CHATTI, M. A., & SCHROEDER, U. Design and Implementation of a Learning Analytics Toolkit for Teachers. Educational Technology & Society, 15 (3), 58–76, 2012. ENGELBRECHT, A. P. Computational Intelligence: An Introduction, 2nd edn, Wiley Publishing. 2007 FERREIRA, GISELLE M. Dos S.; et. al. Educação e Tecnologia: abordagens críticas. Rio de Janeiro: SESES, 2017. Available in . Accessed on: 8 FULANTELLI, G.; TAIBI, D.; ARRIGO, M. A framework to support educational decision making in mobile learning. Computers in Human Behavior, N. 47, p. 50–59, 2015. IEEE IEEE Standards Association. Ethically Aligned Design A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems. 2018. Available in https://standards.ieee.org/develop/indconn/ec/autonomous_systems.html Accessed on: IMRAN, H.; ZADEH, M. B.; CHANG, T.W.; KINSHUK; GRAF, S. A Framework to Provide Personalization in Learning Management Systems through a Recommender System Approach. ACIIDS 2014, p. 271-280, 2014. IMRAN, H.; ZADEH, M. B.; CHANG, T.W.; KINSHUK; GRAF, S. PLORS: a personalized learning object recommender system. Vietnam Journal Of Computer Science, N. 3, p. 3-13, 2016. JONHSON, J. A. The Ethics of Big Data in Higher Education. International Review of Information Ethics. Vol. 7. 2014. LIU, C.; CHANG, C. TSENG, J. The effect of recommendation systems on Internet- based learning for different learners: A data mining analysis. British Journal of Educational Technology, Vol. 44, N. 5, p. 758–773, 2013. SIEMENS, G., & BAKER, R. S. Learning analytics and educational data mining: towards communication and collaboration. In Proceedings of the 2nd international conference on learning analytics and knowledge (pp. 252–254), ACM, 2012. RICCI, F.; Rokach, L.; and Shapira. B. Introduction to recommender systems handbook. Kantor, editors, Recommender Systems Handbook, pages 1–35. Springer, 2011. XING, W.; GUO, R.; PETAKOVIC, E.; GOGGINS, S. Participation-based student final performance prediction model through interpretable Genetic Programming: Integrating learning analytics, educational data mining and theory. Computers in Human Behavior, N. 47, p. 168–181, 2015. MITCHELL. T. M. Machine Learning. McGraw-Hill, Inc., New York, NY, USA, 1 edition, 1997. BROWNLEE, Jason. What is Machine Learning?, 2013. Available in: https://machinelearningmastery.com/what-is-machine-learning/. Accessed on April 22nd, 2018. PORTUGAL, Ivens. ALENCAR, Paulo. COWAN, Donald. The use of machine learning algorithms in recommender systems: A systematic review. 2017. Available in: https://www.sciencedirect.com/science/article/pii/S0957417417308333. Accessed on April 23rd 2018. ADOMAVICIUS, Gediminas. TUZHILIN, Alexander.Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions. IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 17, NO. 6, JUNE, 2005. 9 MANOUSELIS, N., DRACHLER, H., VUORIKARI, R. & HUMMEL, H.. Recommender Systems in Technology Enhanced Learning. (F. Ricci, L. Rokach, B. Shapira, & P. B. Kantor, Eds.) Learning, Springer US, p. 387-415, 2011. Available in: http://www.springerlink.com/index/10.1007/978-0-387-85820-3 WILLIS, J. E.; John P. Campbell, J. P.; Pistilli. M. D. Ethics, Big Data, and Analytics: A Model for Application. EDUCASE REVIEW online. Available in: https://er.educause.edu/articles/2013/5/ethics-big-data-and-analytics-a-model-for- application. Accessed on April 23rd 2018,2013. 1. Introduction