Correct evaluation of Machine Learning based sequencers require large data availability, large scale experiments and consideration of different evaluation measures. Such constraints make the construction of ad-hoc Intelligent Tutoring Systems (ITS) unfeasible and impose early integration in already existing ITS, which possesses a large amount of tasks to be sequenced. However, such systems were not designed to be combined with Machine Learning methods and require several adjustments. As a consequence more than a half of the components based on recommender technology are never evaluated with an online experiment. In this paper we show how we adapted a Matrix Factorization based performance predictor and a score based policy for task sequencing to be integrated in a commercial ITS with over 2000 tasks on 20 topics. We evaluated the experiment under different perspectives in comparison with the ITS sequencer designed by experts over the years. As a result we achieve same post-test results and outperform the current sequencer in the perceived experience questionnaire with almost no curriculum authoring effort. We also showed that the sequencer possess a better user modeling, better adapting to the knowledge acquisition rate of the students.

Currently, a lot of research in the field of intelligent tutoring systems is concerned with recognising student’s emotions and affects. The recognition is done by extracting features from information sources like speech, typing and mouse clicking behaviour or physiological sensors. According to the state-of-the-art support vector machines are the best performing classification models for those kinds of features. However, single classification models often do not deliver the best possible performance. Hence, we propose an approach for further improving the affect recognition performance, which is based on ideas from ensemble approaches and feature selection methods. The approach is proven by experiments on low-level speech features extracted from data which was collected in a study with German students solving mathematical tasks. In these experiments the proposed approach reached on average an affect recognition performance improvement of about 59 % in comparison to a single SVM.

Usually, in intelligent tutoring systems the task sequencing is done by means of expert and domain knowledge. In a former work we presented a new efficient task sequencer without using the expensive expert and domain knowledge. This task sequencer only uses former performances and decides about the next task according to Vygotsky's Zone of Proximal Development, that is to neither bore nor frustrate the student. We aim to support this task sequencer by a further automatically to gain information, namely students affect recognized from his speech input. However, the collection of the data from children needed for training an affect recognizer in this field is challenging as it is costly and complex and one has to consider privacy issues carefully. These problems lead to small data sets and limited performances of classification methods. Hence, in this work we propose an approach for improving the affect recognition in intelligent tutoring systems, which uses a special structure of several support vector machines with different input feature vectors. Furthermore, we propose a new kind of features for this problem. Different experiments with two real data sets show, that our approach is able to improve the classification performance on average by 49% in comparison to using a single classifier.

Most of the artificial intelligence and machine learning researches deal with big data today. However, there are still a lot of real world problems for which only small and noisy data sets exist. Hence, in this paper we focus on those small data sets of noisy images. Applying learning models to such data may not lead to the best possible results because of few and noisy training examples. We propose a hybrid neural network plait for improving the classification performance of state-of-the-art learning models applied to the images of such data sets. The improvement is reached by (1) using additionally to the images different further side information delivering different feature sets and requiring different learning models, (2) retraining all different learning models interactively within one common structure. The proposed hybrid neural network plait architecture reached in the experiments with 2 different data sets on average a classification performance improvement of 40% and 52% compared to a single convolutional neural network and 13% and 17% compared to a stacking ensemble method.