Iterative Bayesian fuzzy clustering toward flexible icon-based assistive software for the disabled
Introduction
Due to the maturity of relevant social systems and a great deal of technological advancement, physically disabled people are now encouraged to participate in social activity [1]. In order to further encourage their active participation in society, steady development of various assistive systems has been brought to the forefront. Unfortunately, many of those systems require separate software, which can be a source of limitations in terms of both usability and accessibility. Thus, there is a growing need for unified assistive software in order to maximize the availability of assistive systems for the disabled; such software is thus required to integrate as many desirable functions as possible so as to provide maximum functionality to the end-user. Additionally, a multi-functional graphical user interface is an essential element of assistive software. This interface is required to preserve simplicity and thus alleviate the need for sophisticated computer expertise. There is therefore a compromise between the number of available functions and the simplicity of the interface. Clustering techniques, in which function subgroups are displayed, are feasible.
There are various forms of similarity-based objective function approaches, including fuzzy c-means (FCM), possibilistic c-means (PCM), fuzzy possibilistic c-means (FPCM), and possibilistic fuzzy c-means (PFCM) [25], [31]. These approaches, however, often suffers from a local minimum problem. They may also require cumbersome processes involving determination of several free parameters – an unwelcome, difficulty in practical design. On the other hand, adaptive resonance theory (ART) and competitive learning [11] utilize neural networks based on competitive learning with dissimilarity measures. The well-known Kohonen self-organizing feature map (KSOFM) [22] utilizes competitive learning [4]. Unlike the aforementioned objective function-based fuzzy clustering methods, these latter techniques perform the task of grouping directly without an objective function. Moreover, performance is increased when the fuzzy concept is incorporated into them [5], [6], [21], [27], [30], [31], [32], [33], [34]. We draw particular attention to the fact that the merits of ART, KSOFM, and FCM have been combined in integrated adaptive fuzzy clustering (IAFC) [18], [19], [20], in which the notion of a vigilance test and a modified fuzzy competitive learning rule [18] is adopted. However, the vigilance test in the IAFC algorithm [18], [19], [20] does not perform a rational outlier detection [28].
In this paper, we propose a new iterative Bayesian fuzzy clustering (IBFC) technique for data grouping. It is an improved version of the IAFC algorithm. By combining a modified fuzzy competitive learning structure with a Bayesian decision rule strategy, we show that the new clustering technique has both the flexibility of a fuzzy technique in handling noise and an optimal decision capability for rational outlier detection. The proposed fuzzy structure with Bayesian interpretation can be extended to handle complex data sets, by which we mean to indicate a data structure in which the shape of a cluster is more than non-ellipsoidal, or in which a large variety of local clusters are required to describe some pattern of data. To cope with this problem, we extend AIBFC by adopting a bottom-up approach in hierarchical clustering. We thus present an advanced algorithm, called agglomerative iterative Bayesian fuzzy clustering (AIBFC), in which local clusters obtained by AIBFC are continually merged until we get the desired number of clusters.
This algorithm is then applied to assistive software for the disabled in order to reach a compromise between the number of available functions and the simplicity of the interface. Specifically, we propose an icon recommendation system in which the behavior data are grouped by the proposed AIBFC and thereby an appropriate icon subgroup is recommended. We call this system flexible icon-based assistive software (FIAS).
The reminder of this paper is organized as follows. In Section 2, the IBFC (iterative Bayesian fuzzy clustering) algorithm is presented. In Section 3, a Bayesian interpretation of IBFC and discussion of the control parameter are given. In Section 4, an AIBFC (Agglomerative IBFC) strategy based on IBFC is proposed. In order to demonstrate the effectiveness of AIBFC for clustering data of complex structure, we provide a comparative evaluation of benchmark data sets in Section 5. In Section 6, we describe the background on the design of assistive software for the disabled and then present the icon recommendation process in FIAS (flexible icon-based assistive software). To assess practical aspects of FIAS, we also present the results of a questionnaire survey of real end-users. The conclusion is provided in Section 7.
Section snippets
Problem under consideration
The problem under consideration is the iterative categorization, via single control parameter, of a given unlabeled data set X = {x1, … , xN} into a finite number of clusters. As an iterative process, this categorization problem is a learning process; the single control parameter defines the effective learning range in the data space. Thus, we may consider a system for which the inputs X = {x1, … , xN} and the single control parameter τgenerate both C (the number of clusters) and their centroids {v1, … , vC
A new look on IBFC in the framework of Bayesian decision theory
As mentioned in Section 1, the IAFC algorithm provides a Bayesian interpretation of the learning procedure [28]. However, it lacks a solid theoretical basis. It was found that the assumption of a highly complex conditional probability density function in [28] lends to a trivial interpretation. Thus, some mismatch between the vigilance test procedure and the learning rule may take place which often results in IAFC generating erroneous results. We find that such partial Bayesian interpretation of
Merging based on a posterior probability
Suppose that the desired number of clusters is given (as is often the case in various clustering algorithms such as hierarchical clustering [12] or FCM, PCM, FPCM, and PFCM [25], [31]). When the data set to be grouped has a simple structure, these above techniques are known to be effective in clustering. However, when the data set has a complex structure, the application of such a clustering algorithms may be problematic [3], [21], [35]. We report that IBFC may also need some refined procedures
Experimental results of AIBFC
In this section, the performance of the proposed AIBFC is demonstrated with six synthetic data sets and with two well-known benchmark data sets from UCI machine-learning repository.
In order to show the effectiveness of the proposed AIBFC algorithm for clustering data of complicated structure, we consider six synthetic data sets as shown in Fig. 4. Synthetic databases DB#1, DB#2 and DB#3 are frequently used to test kernel-based clustering methods found in [3], [13], [17], [23]. Note that KSOM, K
Flexible icon-based assistive software (FIAS) for the disabled
In this section, an application of AIBFC to icon recommendation system is presented. In Section 6.1, we conduct a field study for designing flexible icon-based assistive software (FIAS), and then in Section 6.2, we present AIBFC-based icon recommendation process of FIAS. Finally, user evaluation will be presented in Section 6.3.
Conclusion
This paper first presents a new iterative Bayesian fuzzy clustering (IBFC) algorithm and has shown that the algorithm is successfully applied for grouping and recommending of icons germane to assistive software for the physically disabled. Based on a Bayesian interpretation of IBFC, we have shown that the decision and vigilance test of IBFC coincides with the Bayesian minimum risk classification rule with two loss coefficients. Also, we have shown that the sole vigilance parameter in the
Acknowledgement
This research was supported by the MKE (The Ministry of Knowledge Economy), Korea, under the ITRC (Information Technology Research Center) support program supervised by the IITA (Institute for Information Technology Advancement)” (IITA-2008-C1090-0803-0006)
References (35)
- et al.
Massively parallel architecture for a self-organizing neural pattern recognition machine
Computer Vision, Graphics, and Image Processing
(1987) - et al.
Fuzzy ART: fast stable learning and categorization of analog patterns by an adaptive resonance system
Neural Networks
(1991) - et al.
An adaptive integrated fuzzy clustering model for pattern recognition
Fuzzy Sets and Systems
(1994) - et al.
Evaluation of the performance of clustering algorithms in kernel-induced feature space
Pattern Recognition
(2005) - et al.
Fuzzy Kohonen clustering network
Pattern Recognition
(1994) - et al.
Clustering algorithm for intuitionistic fuzzy sets
Information Sciences
(2008) - Z.Z. Bien, Learning system techniques for human-friendly man–machine interaction, in: Keynote Speech in the 2nd...
- (2006)
- et al.
A novel kernel method for clustering
IEEE Transactions on Pattern Analysis and Machine Intelligence
(2005) - et al.
Fuzzy competitive learning
Neural Networks
(1994)
On the concept of fuzzy probabilistic controllers
Fuzzy Sets and Systems
Pattern Classification
Neuro-Fuzzy and Soft Computing
Applied Multivariate Statistical Analysis
Discussion: on the very real distinction between fuzzy and statistical methods
Technometrics
Cited by (6)
Online recognition of human activities and adaptation to habit changes by means of learning automata and fuzzy temporal windows
2013, Information SciencesCitation Excerpt :They propose the creation of multiple hidden Markov models for each variation of an action, so that the system is able to determine which tasks are currently active even if the user has not yet completed the activity. In [23], Lee et al. propose an algorithm that incorporates a modified fuzzy competitive learning structure with a Bayesian decision rule that is capable of ignoring unintended behavior of the user. An extensive study about the benefits and drawbacks of different probabilistic and statistical techniques for human activities recognition is presented in [19].
Identification of overlapping and non-overlapping community structure by fuzzy clustering in complex networks
2011, Information SciencesCitation Excerpt :Therefore, a generalization is to treat this similar relation as a fuzzy, indeterminate relation [7]. Since the fuzzy set theory was proposed in 1965 by Zadeh [36,37], fuzzy clustering as a fundamental tool has been applied to many fields extensively [3–5,9,19]. In this paper, we provide a novel method based on fuzzy clustering to detect community structure in complex networks.
Fuzzy clustering systems in analyzing high dimensional database
2015, ACM International Conference Proceeding SeriesBayesian Fuzzy Clustering
2015, IEEE Transactions on Fuzzy SystemsDevelopment of a hybrid artificial neural network model and its application to data regression
2012, Intelligent Automation and Soft ComputingAn effective fuzzy clustering algorithm in web mining
2010, Journal of Information and Computational Science