Text stream mining for Massive Open Online Courses: review and perspectives

Figures & data

Table 1.  Advantages and limitation of MOOCs.

Advantages	Limitations
More effective than a professor monologuing to a large class	Inability of educators to assess student learning
It offers quizzes for retrieval practice which is an established method to improve learning	No accreditation
Open opportunities for millions of people who cannot access universities	Validation and plagiarism
Provides global access to education and can be scheduled to work with family and personal commitments	Lack of in-depth evaluation models to evaluate projects and assignments
Can be used as support materials for face to face courses	Lack of effective communication and feedback

Figure 1. Dendogram of text documents.

Figure 2. Agglomerated document clustering. (a) Single-linkage document clustering and (b) complete-linkage document clustering.

Figure 3. Partitional document clustering. (a) Clusters generated by k-means and (b) cluster-generating process.

Table 2.  Text-clustering algorithms summary.

Algorithm	Category	Description	Limitation and computational efficiency
Single linkage	Hierarchical	Merge data points based on the least interconnect dissimilarity	Chaining phenomenon, computationally efficient. Time complexity O(N log N). Space complexity O(N)
Group-average linkage	Hierarchical	Merges data points based on the average similarity of all possible combinations of documents	Overcome chaining phenomenon. Computationally expensive O(N^3). Space O(N^2)
Complete linkage	Hierarchical	Merges data points based on the worst similarity	Overcome chaining phenomenon. Computationally expensive O(N^2). Space O(N)
k-means	Partitioning	Starts by a set of kernels documents not necessarily from the original corpus and build the clusters around these documents using closest similarity. The centroid of the cluster is used in the next iteration. Bisecting k-means is a variation of k-means	Computationally efficient O(N log N). Requires few iterations to converge. Outperforms k-medoids
k-medoids	Partitioning	Starts by a set of kernels documents from the original corpus and build the clusters around these documents using average similarity. The quality of the clusters is improved using objective function	Robust clusters generated
Bipartite graph	Graph based	Documents and words are represented as bipartite graph. Cut function is used to cluster documents	NP complete problem. Heuristic function is used for optimal solution
Buckshot and fractionation	Hybrid	k-means-based clustering method. With improvement in the kernels set selection using buckshot and fractionation	Multiple runs of this algorithm generates different clusters
FTC and HFTC	Frequent item	Reduce the dimensionality of documents by representing document using its frequent terms set. FTC is bottom-up clustering. HFTC is top-down clustering	FTC generates clusters without overlapping. HFTC generates overlapped clusters. Both FTC and HFTC outperform k-means and k-medoids
Star-satellites	Graph based	Data points are represented as similarity graph. A cosine similarity function is used. A star is the centre of the cluster. Documents that are above user-defined threshold similarity are satellites	Similarity between satellites is not guaranteed. Theorem exists in this method failed to prove satellites similarities.Time complexity O(N log^2 N)
Windowing-based	Partitioning	It divides the document into k-gram sub-strings. Then, a subset of these sub strings represents the document fingerprints. Documents share two or more fingerprints are considered similar and clustered	Outperforms k-means and FTC

Table 3.  Topic detection techniques summary.

Approach	Category	Description
LSA	Vector space model	Using SVD to project documents and words into lower dimensional space. Fold-in and semantic space updating are two methods to enhance computational efficiency
PLSA	Probabilistic	Statistically model co-occurrences information using aspect model. It applies probabilistic framework to discover latent semantic structure
LDA	Probabilistic	Generative probabilistic modelling aims to capture multiple topics exist in a document
Hierarchical generative model (Bigram)	Probabilistic	Extends LDA where word position and co-occurrence are considered. Bigram model based on marginal and conditional counts is used. Space complexity O(V^2K), V: vocabulary, K: topics
Discriminative model	Probabilistic	Time factor is considered. Documents are represented by words and times vectors. The temporal discriminative probabilistic model is used for online and online documents
Experimental model	Cluster-based	Clustering documents based on the topics exist in these documents. Experimentally many techniques were implemented. Average link clustering outperforms other used clustering techniques

Figure 4. MOOCs feedback management system architecture.