Article Preview
Top1. Introduction
Recent years have acknowledged the related views of data in many areas including recognition of patterns, mining social network data, link-based document databases, multilingual documents and so on. Here, the document content is represented as one view, whereas link relations are considered as another view in case of link-based document databases. As far as multilingual document is concerned, each language is considered as a separate view.
In this paper, we aim to design a novel framework that considers the balance between the similarity and coverage due to avoidance of implicit trust across different views ensuring high quality clustering results and consistent across different views. A two-term global objective function is developed, which simultaneously maximizes the rating coverage while reduces the complexity analysis across different views.
A similarity-based clustering approach for multi-view clustering called Multi-View Affinity Propagation (MVAP) was designed by Wang et al. (2016) with two objectives being satisfied. The two-term global objective takes into account are the sum of similarities between each data point where each view was assigned with its exemplar. On the other hand, the second objective considers explicit clustering consistency across several views based on the item-wise exemplar consistency. This in turn ensured high quality clustering results.
However, similarity and correlation relationships between the data were not provided in MVAP. To address this issue, two different views that take into account are rating patterns and social trust relationships by Guo et al. (2015). A support vector regression model was used for users in two different clusters, whereas a probabilistic method was used for users which could not be clustered due to insufficient data. In this way both accuracy and coverage of recommendation were ensured.
With the development in the field of hardware technology, a large amount of multi-view data hold different representations that generate real-world applications in recent years. Weighted Multi-View Clustering with Feature Selection (WMCFS) by Xu et al. (2016) was not only focused on multi-view clustering, they also focused on the feature selection in a simultaneous manner enhancing the clustering performance. A Pairwise sparse subspace representation model was designed by Yin et al. (2015) to maximize the correlation and the correlation between different views. However, the missing value prediction was not guaranteed.
To address the missing value prediction gray values was used in Huang and Lee (2004). A bilingual document clustering in the form of multi-view was constructed by Ye et al. (2015) using bilingual similarities with random walk and generated cluster structures using bilingual similarity matrices. Despite similarities being confirmed, optimization of similarity was not ensured. To provide a solution towards optimization, fuzzy clustering with minimax optimization was designed by Wang and Chen (2017).
In addition, certain data views may be of high dimensionality resulting in increased computational analysis, but relatively low clustering accuracy. A Multi-graph Laplace with each graph corresponding to one view and low rank minimization for correlation constraint was designed by Wang et al. (2016). Adaptive Multi-View Semi-Supervised Nonnegative Matrix Factorization by Wang et al. (2016) was designed with the objective of ensuring accuracy and normalizing the mutual information between different views. Yet another generative method with ensemble manifold regularization was presented by Wang et al. (2015) which considers both generalization and manifold data structure. For certain specific view similarity between clusters are required for improving the clustering efficiency. In other words, cluster similarity with implicit trust is a way that can both simplify the calculation and help to get an accurate data model with multi-view clustering.