Abstract:In this talk we review the recent work done by our group on data mining (DM) technologies deduced from simulating visual principle. Through viewing a DM problem as a cognition problems and treading a data set as an image with each light point located at a datum position, we developed a series of high efficient algorithms for clustering, classification and regression via mimicking visual principles. In pattern recognition, human eyes seem to possess a singular aptitude to group objects and find important structure in an efficient way. Thus, a DM algorithm simulating visual system may solve some basic problems in DM research. From this point of view, we proposed a new approach for data clustering by modeling the blurring effect of lateral retinal interconnections based on scale space theory. In this approach, as the data image blurs, smaller light blobs merge into large ones until the whole image becomes one light blob at a low enough level of resolution. By identifying each blob with a cluster, the blurring process then generates a family of clustering along the hierarchy. The proposed approach provides unique solutions to many long standing problems, such as the cluster validity and the sensitivity to initialization problems, in clustering. We extended such an approach to classification and regression problems, through combatively employing the Weber’s law in physiology and the cell response classification facts. The resultant classification and regression algorithms are proven to be very efficient and solve the problems of model selection and applicability to huge size of data set in DM technologies. We finally applied the similar idea to the difficult parameter setting problem in support vector machine (SVM). Viewing the parameter setting problem as a recognition problem of choosing a visual scale at which the global and local structures of a data set can be preserved, and the difference between the two structures be maximized in the feature space, we derived a direct parameter setting formula for the Gaussian SVM. The simulations and applications show that the suggested formula significantly outperforms the known model selection methods in terms of efficiency and precision. The advantages of the proposed approaches are: 1) The derived algorithms are computational stable and insensitive to initialization and they are totally free from solving difficult global optimization problems. 2) They facilitate the construction of new checks on DM validity and provide the final DM result a significant degree of robustness to noise in data and change in scale. 3) They are free from model selection in application. 4) The DM results are highly consistent with those perceived by our human eyes. 5) They provide unified frameworks for scale-related DM algorithms recently derived from many other fields such as estimation theory, recurrent signal processing, information theory and statistical mechanics, and artificial neural networks.

Data Modeling: Visual Psychology Approach and L-1/2 Regularization Theory

Visual Complexity of Shapes: a Hierarchical Perceptual Learning Model

On How Data Are Partitioned in Model Development and Evaluation: Confronting the Elephant in the Room to Enhance Model Generalization.

Data Mining Technologies Inspired from Visual Principle

Completeness out of Incompleteness: Inferences from Regularities in Imperfect Information Ensembles.

Efficient Probabilistic Latent Semantic Analysis with Sparsity Control

Title From Information Scaling of Natural Images to Regimes of Statistical Models Permalink

Completeness Out of Incompleteness Inferences and Employment Regarding Regularities in Imperfect Information Ensembles

Regularized Latent Class Analysis with Application in Cognitive Diagnosis

Conceptualization and Modeling of Visual Patterns

Data smoothing regularization, multi-sets-learning, and problem solving strategies.

Exploring Latent Constructs through Multimodal Data Analysis

Massive Data Clustering by Multi-Scale Psychological Observations

A New Approach to Data Clustering Using A Computational Visual Attention Model

Sparsity Induced Locality Preserving Projection Approaches for Dimensionality Reduction

Predicting and Explaining Behavioral Data with Structured Feature Space Decomposition

Embedding Gestalt Laws in Markov Random Fields - a theory for shape modeling and perceptual organization

Embedding Gestalt Laws in Markov Random Fields

Understanding Human Cognition Through Computational Modeling

Learning to Model Task-Oriented Attention

Clustering by Scale-Space Filtering