Predição de sucesso acadêmico de estudantes: uma análise sobre a demanda por uma abordagem baseada em transfer learning

Student interactions with Learning Management Systems (LMS) generate logs, which are usually stored, allowing torecover each student activity. Analysis of these data with data mining and/or learning analytics techniques have beenprovided a better understanding of student behavior and teaching-learning processes.  In this context, a number ofstudies have been reporting promising results in the task of predicting student performance, which allows proactiveactions  to  avoid  academic  failures.   Usually,  data  mining  techniques  estimate  predictive  models  by  using  (past)historical data, assuming the premise that the estimated predictor will make predictions in future contexts that aresimilar to the (past) contexts which were used in its design.  Although it is reasonable to assume that the diversity ofexisting educational contexts is reflected in the data, few studies discuss the impact of the aforementioned premisein  the  area  of  Educational  Data  Mining  (EDM),  resulting  in  models  that  may  perform  poorly  when  used  underunforeseen  educational  conditions.   This  paper  proposes  an  empirical  analysis  to  verify  evidences  of  differencesbetween data from different educational contexts in the task of predicting students’ academic failure.  Logs of morethan 3,000 distance higher education students are used, and the adopted methodology is based on the supervisedclassification  approach,  commonly  used  in  prediction  tasks.   Specifically,  we  aim  to  verify  if  distinct  educationalcontexts are in fact separable in terms of the data they generate. Although data scenarios involve activities commonto students in the same subject, the experiments indicate an accuracy of up to 83% in the separation of data fromdifferent academic terms.  Although empirical, our results indicate a similar direction to that pointed out by otherstudies, contributing about the need of using transfer learning and/or domain adaptation techniques in the design ofpredictive models that aim to support proactive actions to prevent student failures.