The comparison of articles and methods that perform the cascade of SVMs in parallel. Here, we highlight the articles in the survey that have interesting findings using cascades of SVMs for training the samples. The description of each column is as follows,

"Algorithm" indicates the corresponding algorithm's name or approach.

"Problem Type" shows whether the algorithm solves binary, multi-class, linear or/and non-linear classifications.

"Parallelism Type" highlights whether the algorithm uses the shared memory parallelism, distributed big data architectures, distributed HPC architecture, or/and GPU parallelism.

"Main Idea" simply shows the main contribution of the author(s).

"Pros" explains the remarks and findings of the algorithm. Besides, it highlights the advantages of the algorithm compared to the similar approaches, if possible.

"Cons" describes the shortcomings or the place for further improvements of the corresponding algorithm. Besides, it shows the weak points of the algorithm compared to the similar approaches, if possible.

Note the table can be sorted by clicking on the title of each column.
Algorithm	Problem Type	Parallelism Type	Main Idea	Pros	Cons	Reference
Algorithm	Problem Type	Parallelism Type	Main Idea	Pros	Cons	Reference
Alternating feedback cascade SVM	Non-linear and binary classifications	Shared memory parallelism	Obtains SVs using samples' distance mean and alternating feedbacks	Data reduction Reduces the number of SVs Improves the classification accuracy	Concerns regarding efficient decomposability of the problem Limited shared memory	Zhang et al. [2005]
Parallel SVM in .NET	Unclear	Shared memory parallelism using .NET	Implements a cascade SVM in parallel using .Net framework	Achieves superlinear speedup Good scalability in terms of the number of cores and threads Faster than the standard cascade	Unclear number of passes through the cascade to achieve the desired accuracy Unclear scalability in terms of the number of training data Benefits not fully clear compared to the similar approach by Zhang et al. [2005]	Hu and Hao [2010]
Improved cascade with crossed feedbacks	Non-linear and binary classifications	Distributed HPC architectures	SVs are obtained by an alternating feedback in a crossed way similar to the work proposed by Zhang et al. [2005]	Improves the classification accuracy Improves the performance in terms of training time compared to the standard cascade	Benefits not fully clear for large-scale problems Benefits not fully clear compared to the similar work proposed by Zhang et al. [2005] Unclear number of passes through the cascade to achieve the desired accuracy Unclear scalability in terms of the number of cores or threads	Yang et al. [2006]
A simple parallel cascade	Non-linear and binary classifications	Distributed HPC architectures	Implements a bagging-like strategy to aggregate SVs in the last layer of the cascade and improves the parallelizability	Improves the classification accuracy Reduces communication overhead Better parallelizability than the standard cascade An adaptive termination criterion Almost doubled speedup	Unclear benefits in some cases Non-optimal results for some of evaluated datasets Unclear scalability Lack of parameter tuning	Meyer et al. [2014]
PSMR	Non-linear and multi-class classifications	Distributed HPC architectures	Employs ontology-based strategy to improve the classification accuracy	Reduces the number of SVs Higher speedup compared to the standard cascade Studies impact of the number of partitions on the performance Higher speedup for the increasing number of partitions	The optimal number of partitions is unclear, i.e., the number of partitions that training data is divided into without deterioration of accuracy Limited explanation of the ontology-based semantic used for improving classification accuracy	Xu et al. [2014]
Incremental cascade	Non-linear and binary classifications	Distributed HPC architectures	Implements one layer cascade with data fusion in which the cascade is combined with incremental learning for intrusion detection	Heavily reduces the number of SVs Reduces the number of training data Avoids re-constructing of SVMs for whole data to train a new coming data Higher speedup compared to the standard cascade Filters redundant data	Unclear scalability in terms of the number of training samples Limited explainability of the conducted benchmarks Unclear scalability in terms of the number of samples and cores	Du et al. [2009]
CA-SVM	Non-linear and binary classifications	Distributed HPC architectures	Combines the cascade with a divide-and-conquer like method and implements a communication avoiding classifier using initial clustering of data	Good scalability in terms of the number of processors (up to 1536 processors) Heavily reduces the communication between distributed nodes Higher speedup compared to the similar approaches Improved isoefficiency	Modest loss of classification accuracy	You et al. [2015]
P2P SVM	Non-linear and binary classifications	Distributed HPC architectures	Reduces the number of SVs through a peer to peer network	Improves the classification accuracy for a large number of SVs Robust to imbalance classes Quadratic time and memory complexity with respect to the number of SVs Scales well with the size of peers Exists an upper bound for communication overhead	Not robust if the number of SVs is small Unstable for a small number of SVs Deterioration of the classification accuracy	Ang et al. [2008]
RASMO	Non-linear and binary classifications	Distributed big data architecture using map-reduce	Conducts load balancing using a genetic algorithm for image classifications	Load balancing Scales well with the number of mappers Good scalability in terms of the number of cores Improves the performance in terms of computing time while increasing number of mappers Supports computation in a heterogeneous environment	Complicated load balancing strategy Suffers from map-reduce startup overheads for small-scaled problems Benefits not fully clear for a small number of training samples Load balancing may take longer time than the time spent on training sub-problems	Guo et al. [2015]
Parallel cascade SVM for IDS	Linear and non-linear classifications	GPU parallelism	Parallelizes the standard cascade SVM using GPU-based parallelism	Improves the classification accuracy Higher speedup compared to the similar approaches Finds the suitable number of partitions to split a large problem into smaller problems to get maximum performance Scales well with the number of training samples	Speedup is saturated for more than 64 processors Suffers from communication overhead	Tarapore et al. [2016]
Parallel cascade SVM for pedestrian recognition and tracking	Two-layer linear and non-linear classifications	GPU parallelism	Implements a real-time classifier with selective feature extraction for video-based classifications	Higher speedup compared to the similar approaches in some cases Demonstrates a real-time behavior Good scalability in terms of the number of training samples Reduces features	Complex implementation Limited explanability of results Benefits not clear in some cases Not significant speedup in some cases	Weimer et al. [2011]