uow csit940 research methodology structure of online social networks mirrors

Questions:
The answers to the six questions follow a reference to each paper.
Dunbar, R. I. M., Arnaboldi, V., Conti, M., & Passarella, A. (2015). The structure of online social networks mirrors those in the onine world. Social Networks, 43 , 39-47.
What problem (or research question) is being solved (addressed) and why is it important?
The amount of time expended on friends and the level of closeness have been found to exhibit a layering e ect in natural social networks. This has been explained by the constraints on the amount of time an individual can spend on people in their social networks and the cognitive constraints brought about by the social brain hypothesis. This paper tests the hypothesis that the seeming lack of time constraints in Internet-based communication might remove the layering e ect observed in natural social networks. Results of this study may deepen our understanding of how humans operate within social networks in both online and o ine environments. It also has implications for the design and promotion of online social environments.
What have others done about the problem?
There has been several studies that examined the size of social networks main-tained by individuals in online environments for extroverts and introverts. These studies used Twitter and Facebook datasets to reach conclusions about the average size of these social networks.
What solution is being proposed by the authors?
Based on Twitter and Facebook datasets, the authors used statistical cluster-ing techniques (k-means and partitioning clustering) to discover the possible presence of a layered structure in the communications (social networks). They also determine the optimum number of the layers. The authors use the clus-tering techniques to discover the number of clusters of the ego networks of each dataset.
What result was obtained?
The study con rmed the existence of a layered structure in the ego networks. On average 4 clusters were found in the Facebook datasets, implying that there are 4 layers corresponding to the levels of closeness maintained by the individuals. The Twitter dataset indicated that the average number of clusters is about 6. These results are similar for both clustering techniques.
How does the solution/result compare/contrast with previous results?
The analysis con rmed that layered structure found in o ine face-to-face so-cial networks exist in online social networks. The scaling ratio (i.e. the relative sizes of the layers) are extremely close to those observed in o ine networks.
What further work is proposed?
The authors plan to explore the sociological similarities between online and o ine communities based on the results of present study.
Wu, J., Yu, Y., Huang, C., & Yu, K. (2015). Deep multiple instance learning for image classi ca-tion and auto-annotation. In Proceedings, IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (p. 3460-3469).
What problem (or research question) is being solved (addressed) and why is it important?
Multiple instance learning is a weakly supervised learning paradigm that holds promise for several real-world problems. Deep learning has recently been de-veloped for fully supervised learning problems and has shown signi cantly good results. This paper addresses the problem of designing and building a deep learning architecture for weakly supervised learning problems. Speci – cally, the multiple instance learning was studied. The architecture was tested on image classi cation and automatic annotation problems.
What have others done about the problem?
There has been prior work on the application of neural networks to weakly supervised problems including multiple instance learning. However, these are shallow networks without the feature revealing properties associated with deep networks. The few cases of the use of deep learning in weakly supervised settings have not been directed at the integrated problem of classi cation and annotation.
What solution is being proposed by the authors?
The authors use the well known deep convolutional neural networks as a basis of their architecture and redesign the last hidden layer for multiple instance learning.
What result was obtained?
Mean average precision and mean accuracy have been used as evaluation cri-teria. Based on publicly available datasets – PASCAL VOC 2007 the results obtained were on average better than three others used for comparison. They also created a new dataset speci cally for the evaluation of annotation. They evaluated the performance of their architecture on image-level and patch-level annotation.
How does the solution/result compare/contrast with previous results?
Their results on average were superior to those of the algorithms in the com-parative evaluation.
What further work is proposed?
This is an exploratory paper and one of the motivations was to inspire more research in the application of deep learning to weakly supervised learning.
Zhang, X., Pham, D.-S., Venkatesh, S., Liu, W., & Phung, D. (2015). Mixed-norm sparse representa-tion for multi view face recognition. Pattern Recognition, 48 (9), 2935-2946.
What problem (or research question) is being solved (addressed) and why is it important?
Conventional methods of recognizing faces from images are based on a single query image. This paper addresses the problem of using multiple query face images of the same person for recognition. This is important because there are occasions when a single face image is insu cient to perform the recognition task. It is also possible to acquire multiple face images of the same person as would happen when a video of the person is available under di erent pose conditions.
What have others done about the problem?
Most face recognition algorithms are based on the nearest neighbour classi ca-tion paradigm. Methods based on the nearest subspace formulation have also been proposed. In the sparse representation approach a dictionary designed and minimum number of atoms is selected from within-class and across-class sections of the dictionary. The joint sparse representation approach assumes that the query images share the same sparsity pattern at the atom level. This was extended in the joint dynamic sparse representation method where as-sumption of atom level sparsity pattern similarity was replaced by class-level pattern similarity. In all the sparse representation methods the l
1
norm was used in the optimization.
What solution is being proposed by the authors?
The authors propose to use a mixture of l
1
and l
2
norms in the sparse rep-resentation formulation of the problem. This proposed method exploited the correlation structure in the query images and also allow exibility in the se-lection of atoms for representation.
What result was obtained?
Evaluation of the proposed method was conducted on publicly available face image datasets – CMU-PIE, Yale B and Multi-PIE. Results of experiments with di erent number of views and dimensions were provided as well as com-putaional e ciency. The proposed method was superior (higher recognition rate) to the other sparse representation methods used in the comparative evaluation. Results of experiments with unseen pose also suggest that the proposed method performed much better than the other sparse representa-tion methods as well as the basic methods of principal component analysis and linear discriminant analysis.
How does the solution/result compare/contrast with previous results?
Based on the datasets used in the evaluation and the methods chosen for comparison, the mixed-norm sparse representation method of multiple view face recognition is e ective.
What further work is proposed?
The authors did not suggest further work.

Read less