Ricardo Matsumura de Araújo

Hashtags Functions in the Protests Across Brazil

In this article, we discuss the communicative functions of hashtags during a period of major social protests in Brazil. Drawing from a theoretical background of the use of Twitter and hashtags in protests and the functions of language, we extracted a sample of 46,090 hashtags from 2,321,249 tweets related to Brazilian protests in June 2013. We analyzed the hashtags through content analysis, focusing on functions, and co-occurrences. We also qualitatively analyzed a group of 500 most retweeted tweets to understand the users’ tagging behavior. Our results show how users appropriate tags to accomplish different effects on the narrative of the protests.

Clustering Learning Objects for Improving Their Recommendation via Collaborative Filtering Algorithms

Collaborative Filtering can be used in the context of e-learning to recommend learning objects to students and teachers involved with the teaching and learning process. Although such technique presents a great potential for e-learning, studies related to this application in this field are still limited, mostly because the inexistence of available datasets for testing and evaluating. The present work evaluates a pre-processsing method through clustering for future use of collaborative filtering algorithms. For that we use a large data set collected from the MERLOT repository. The initial results point out that clustering learning objects before the use of collaborative filtering techniques can improve the recommendations performance.

Predição de estudantes com risco de evasão em cursos técnicos a distância

The present paper describes an approach for detecting possible dropout students in technical distance learning courses. The proposed method uses only the count of students interactions inside a Learning Management System (LMS), along with other attributes derived from the counts. Such strategy allows better generalization in different platforms and LMS, as it does not rely on the differences among the interaction types, or use other darta sources besides LMS logs. Predictive models were trained and tested with data from 4 technical distance learning courses in two different scenarios: 1) train and test with data from one course, and 2) train with data combined from 3 courses and test with data of the remaining course. Results point out it is possible to predict dropout students in the first weeks of the courses with average accuracy rates of 75% in most of scenarios, achieving 95% in the best case scenarios. Resumo. O presente trabalho apresenta uma abordagem para a detecção de alunos em risco de evasão em cursos técnicos a distância que utiliza apenas a contagem de interações dos estudantes dentro do AVA, além de atributos derivados dessas contagens. A premissa inicial é de que essa estratégia permite uma maior generalização em diferentes plataformas e AVA, uma vez que não utiliza diferenciações entre os tipos de interações, nem informações de outra ordem encontradas fora do AVA (dados demográficos, exames, questionários, etc). Os modelos de predição foram testados e treinados com dados de 4 diferentes cursos técnicos EAD em dois cenários diferentes: 1) treino e teste com dados de um mesmo curso, e 2) treino com dados de 3 cursos e teste com dados do curso restante. Os resultados apontam a possibilidade de predição de estudantes em risco de evasão já nas primeiras semanas dos cursos com taxas de desempenho próximas a 75% na maioria dos cenários, e chegando a 95% nos melhores casos.

A comparison among approaches for recommending learning objects through collaborative filtering algorithms

Purpose The purpose of this paper is to present the results of a comparison among three different approaches for recommending learning objects (LO) inside a repository. The comparison focuses not only on prediction errors but also on the coverage of each tested configuration. Design/methodology/approach The authors compared the offline evaluation by using pure collaborative filtering (CF) algorithms with two other different combinations of pre-processed data. The first approach for pre-processing data consisted of clustering users according to their disciplines resemblance, while the second approach consisted of clustering LO according to their textual similarity regarding title and description. The three methods were compared with respect to the mean average error between predicted values and real values. Moreover, we evaluated the impact of the number of clusters and neighborhood size on the user-coverage. Findings Clustering LO has improved the prediction error measure with a small loss on user-coverage when compared to the pure CF approach. On the other hand, the approach of clustering users failed in both the error and in user-space coverage. It also became clear that the neighborhood size is the most relevant parameter to determine how large the coverage will be. Research limitations The methods proposed here were not yet evaluated in a real-world scenario, with real users opinions about the recommendations and their respective learning goals. Future work is still required to evaluate users opinions. Originality/value This research provides evidence toward new recommendation methods directed toward LO repositories.

Generating models to predict at-risk students in technical e-learning courses

Different studies attempt to generate and evaluate models for predicting students success in distance courses. Those works normally rely on a number of different features that are context dependent, thus becoming difficult to generalize the models and the results to other scenarios. The present article describes an alternative approach for predicting at-risk students that only uses the counting of students interactions inside the virtual learning environments as a way to produce more generalizable models.