EXPLORING ML-BASED CLASSIFICATION SYSTEM FOR DIGITAL LEARNING PLATFORMS: A MAPPING TECHNIQUE OF MASSIVE OPEN ONLINE COURSES

The object of research is massive open online courses. One of the most problematic areas in online learning is how to improve the quality assurance of digital learning systems. Analysis and classification of massive open online courses is a difficult task, given the variability of massive open online courses structures, contents, designs, platforms, providers, and learner profiles. To overcome this challenge, this study aims to propose an automatic and large-scale machine learning based classification system for massive open online courses according to their learning objectives by making use of the six cognitive levels of Bloom’s taxonomy. During the course of the research, it is shown that analyzing learning objectives associated with modules and programs can further enhance the quality of digital learning system. As a result of the research, a representation and a detailed analysis of the dataset for experimentation with the different models are provided. Further research can focus on the privacy implications of the current control on developments of artificial intelligence taking into account creativity, and innovation which can hardly be performed by machines.


Introduction
Given the cognitive level of learning objectives (LOs), a few scholars proposed an evidence-based massive online open course (MOOC) ontology, that served as a standard to unify the representation of MOOCS and facilitate interoperability between MOOCs platforms [1][2][3]. The aim of the study is to enrich this ontology with metadata about learning objectives classified according to Bloom's taxonomy. The ontology serves as a basis for the design and implementation of a linked data repository. The aim is to automatically extract semantically rich descriptive metadata from different MOOC providers and integrate this metadata into a repository accessible through a simple protocol and RDF query language (SPARQL) endpoint.
Various theoretical frameworks have been adopted for the evaluation and classification of MOOCs. Author of [1] associated good learning with quality learning. It was critical, in her opinion, to meet the characteristics of good learning in order to accomplish effective learning. Author of [2] based the 12-dimensional assessment framework, as well as the 7Cs for learning design framework on this principle [3].
On the other hand, according to author of [3], the first principles of instruction he proposed constituted the foundation of all present pedagogical models and theories.
Author of [3] suggested five guidelines for the development of learning activities.
Author of [4] used theoretical frameworks to drive their research into the development of evaluation frameworks focused on open-ended questions and necessitating the assistance of an expert. Author of [4] used a Web-based online instruction approach to drive their evaluation of MOOCs based on three global design dimensions: information, instruction, and learning.
The ultimate aim of this research is to evaluate and classify the user interface dimension of MOOCs based on their learning objectives. The theoretical foundation of Bloom's taxonomy is most appropriate for our context since it co vers the different levels of cognitive learning and allowed for classifying learning objectives according to six hierarchical levels. The Bloom's taxonomy includes a two-dimensional framework consisting of knowledge and cognitive processes [5]: -the first dimension includes the subcategories of the first level of the original taxonomy; -the second dimension renamed the six levels as verbsremembering, understanding, applying, analyzing, evaluating, and creating. No research has been done on the automatic classification of LOs. Machine learning has been used most often, followed by the rule-based approach. The deep learning approach has been used less often; only the ANN architecture ISSN 2664-9969 has been tested in this context. BERT was used in a single study for cognitive classification purposes [5]. There has been some research comparing BERT and other machine learning or deep learning models [6].
The BERT (Bidirectional Encoder Representations from Transformers) model is based on two stages: pre-training and fine-tuning [4,5,7]. During pre-training, the model is trained on a large unlabeled corpus. The model is then finetuned, starting with the pre-trained parameters and refining all parameters with task-specific labeled data [6,8]. A simple transformer consists of an encoder that reads text input and a decoder that generates a task prediction. BERT requires only the encoder depicted in Fig. 1 because its objective is to develop a model of the language representation [9].
BERT is based on the attention mechanism [10] that was invented to allow a model to comprehend and remember the contextual relationships between features and text.
BERT represents a single sentence or a pair of sentences as a sequence of tokens with the following characteristics [11]: -the first token in the sequence is CLS; -when there is a pair of sentences in the sequence, they are separated by the token SEP; -for a given token, its input representation is constructed by summing the corresponding token, position, and segment embeddings (Fig. 2).  BERT architecture: C -at the end; T 1 , T 2 , …, T N -the learner, will be able, to analyze; E CLS -the module; E 1 , E 2 , …, E N -aims to equip the learner; CLS -the module's objective'; tok 1 , tok 2 , …, tok N -to enhance analytical thinking skills BERT is a leading model for a variety of Natural Language Processing (NLP) tasks, demonstrating its efficiency and potential.

Thus, the object of research is massive open online courses.
The aim of research is to propose an automatic and largescale classification system for MOOCs according to their learning objectives by making use of the six cognitive levels of Bloom's taxonomy. The study concludes with a mapping MOOC learning objectives and Bloom's taxonomy levels based on a cognitive classification of MOOCs.

Research methodology
2.1. A BERT-based cognitive approach for classifying MOOCs. The existing research has addressed one of the following [8,9]: -frameworks developed for quality assurance that are generalist and lack means to operationalize the review of MOOCs' quality; -case studies that detailed the design of individual MOOCs to highlight best practices and pedagogical models which are not based on a well-defined evaluation framework; -descriptive frameworks that were intended for designing MOOCs from scratch; -evaluation frameworks that dealt with several dimensions including the pedagogical. This study aims to classify MOOCs on a large scale. Author of [10,11] used machine learning for a large analysis of MOOCs. However, the number of MOOCs they examined remained limited, and their data collection methods were manual [12]. These researchers [12] used machine learning for the analysis of about 20 MOOCs. Nevertheless, the result of their clustering cannot be generalized given the limited number of MOOCs they analyzed.
BERT is the state-of-the-art technique in NLP [13,14] and it has demonstrated its performance on small datasets. The contributions of this study are: -the automatic classification of MOOCs according to their pedagogical approaches based on the cognitive levels of Bloom's taxonomy; -exploring the impact of choosing different classifiers, from a simple softmax classifier to a more complex classifier like dense layers, LSTM, and Bi-LSTM.
2.2. Fine-tuning strategies. ML models can efficiently scale across accelerators while model weights are duplicated across accelerators for partitioning and distributing the training data.
BERT fine-tuning involves training a classifier with different layers on top of the pre-trained BERT transformer to minimize task-specific parameters. Fine-tuning for a specific task can be done using several approaches, either by fine-tuning the architecture or by fine-tuning different hyper-parameters such as the learning rate or the choice of the best optimization algorithm [15,16].
As the aim is the cognitive classification of MOOCs according to their learning objectives, this study recommends to adopt the basic architecture of BERT and then to add an output layer for the classification. The output layer can be either a simple classifier like softmax or a more complicated network like the bidirectional Bi-LSTM. ; z i values are the elements of the input vector to the softmax function; K is the number of classes in the multi-class classifier. The output node with the highest probability is then chosen as the predicted label for the input.
For preprocessing, one can simply clean the text of nonalphabetic characters and converted it to lower case (Fig. 3). Fig. 3. BERT-based fine-tuning architecture: C -at the end; T 1 , T 2 , …, T N -the learner, will be able, to analyze; E CLS -the module; E 1 , E 2 , …, E N -aims to equip the learner; CLS -the module's objective'; tok 1 , tok 2 , …, tok N -to enhance analytical thinking skills 2.2.2. Combine with fully connected layers. The fully connected layer took the output of BERT's 12 layers and transformed it into the final output of six classes that represented the six cognitive levels of Bloom's taxonomy (Fig. 4). This layer consists of three dense layers. Fig. 4. BERT with fully connected layers architecture: C -at the end; T 1 , T 2 , …, T N -the learner, will be able, to analyze; E CLS -the module; E 1 , E 2 , …, E N -aims to equip the learner; CLS -the module's objective'; tok 1 , tok 2 , …, tok N -to enhance analytical thinking skills

Combine with deep network layers.
In previous architectures, the CLS output was the only one used as input for the classifier. In this architecture, one can use all the outputs of the last transformer encoder as inputs to an LSTM or Bi-LSTM recurrent neural network as shown in Fig. 5. The next section provides a representation and a detailed analysis of the dataset for experimentation with the different models.

Research results and discussion
3.1. Data analysis. Researchers can start by collecting LOs (Table 1) from the MOOCs providers, Coursera, and edX, and then manually annotate them based on Bloom's taxonomy action verbs list. However, this could lead to ambiguity about the actual meaning of the required cognition [17][18][19]. Create a Docker container in which you will implement a Web application by using a flask in a Linux environment ISSN 2664-9969 3.2. Evaluation metrics. Several considerations, including class balance and expected outcomes, can guide the selection of the best measures to evaluate the performance of a given classifier on a certain dataset. Given a dataset with an approximately balanced number of samples from all classes, one can use the accuracy measure to evaluate the performance of a model and compare it with other models [20].
Accuracy is the sum of true positive (TP) and true negative (TN) items divided by the sum of all other possibilities: where TP -True Positives; TN -True Negatives; FN -False Negatives; FP -False Positives.
3.3. Environment setup. One can use the Google Colab and Tensorflow environment as well as Keras Tensorflow to build the BERT models. Keras TensorFlow is an opensource mathematical software library used for machine learning applications. It has tools to run on graphic processing units, which can significantly reduce training and inference times on some models. Keras is a high-level APi for TensorFlow and has a modular and easily extensible architecture, and it allows users to create sequential models or a graph of modules that can be easily combined.
The library contains many different types of neural layers, cost, and activation functions. Researchers can implement different fine-tuning strategies of BERT on Tensorflow Hub (TFHub). TFHub provides a way to try, test, and reuse machine learning models.

Implementation details.
For the implementation of the models adopted, one can use the Keras Layer function of Tensorflow Hub to build our BERT layer. Then, one can tokenize text based on the variables of this layer. This would allow having the first input of BERT model, (e.g: input_word_ids). Next, one can add the embeddings of position input_mask and segments segment_ids.
3.5. Discussion. The rise of AI makes it impossible to ignore a serious debate about its future role in our lives. Given complex algorithms designed by programmers that can transmit their own biases, an in-depth discussion is critical to promote, and develop knowledge and wisdom.
One of the limitations of this study is that collection of LO s from several MOOCs can cause an interference with their privacy policies. Nevertheless, given the aim of the study about the improvement of MOOCs, these MOOC providers can still show a willingness to contribute to the study. Another limitation is the accessibility and availability of Google Colab or Tensorflow environments as not every researcher might be equipped with these tools. These constraints should be taken into account when trying to apply in practice.
There is a need for further research on the privacy implications of the current control on developments of AI taking into account creativity, and innovation which can hardly be performed by machines.

Conclusions
This study proposes an automatic and large-scale ML-based classification system for MOOCs according to their learning objectives by making use of the six cognitive levels of Bloom's taxonomy. During the course of the research, it is shown that analyzing learning objectives (LOs) associated with modules and programs can further enhance the quality of digital learning system. As a result of the research, a representation and a detailed analysis of the dataset for experimentation with the different models are provided. In alignment with the research objective, these results show that its distinctive algorithmic features make it possible to solve the quality issue of MOOCs which can provide certain advantages over other existing frameworks and models.
The results imply that the use of ANN architecture for automatic classification of LO s can be used in MOOCs to improve the quality of digital learning environments.

Conflict of interest
The author declares that she has no conflict of interest in relation to this research, whether financial, personal, authorship or otherwise, that could affect the research and its results presented in this paper.

Financing
Presentation of research in the form of publication through financial support in the form of a grant from SUES (Support to Ukrainian Editorial Staff).

Data availability
Data cannot be made available for reasons disclosed in the data availability statement.