TECHNIQUES
Dimensional Reduction
We plot the data in 2D, so we can understand how the algorithm sees it.
Information Retrieval
We can see where the algorithm struggles most, and focus on that.
Transferable Learning
Our algorithm can apply knowledge from similar tasks.
Scalable Infrastructure
So we can run algorithms on hundreds of millions of samples..
Interaction Capture
Our algorithm (the NEPAR) can capture the interactions of patients,
Powerful Classification
We have a powerfull detection system for the understanding of the patients.
NEPAR
This aim of this study is to propose a new classification framework as Networked Pattern Recognition (NEPAR) for different classification problems. In most research studies, classification focuses on either individual observations, which do not consider the dynamic interactions, which ignores the functional roles of observations. When capturing interactions, they just give a general idea about networks. In this study, we propose a unified approach that combines pattern classification techniques and dynamic interactions for better classification approach. Therefore, the NEPAR and five different classification methods (SVM, NB, LR, DT, and kNN) are developed by adding information from the proposed networks (as seen in Figure 2-3). Figure 1. Combining network metrics and pattern recognition. As seen in Figure 1, information from observations is extracted by building the network, and feature properties for each observation are used to classify the output. For the results, we compare three approaches: (1) classic approach that uses traditional pattern recognition techniques; (2) the networked approach that uses pattern recognition techniques on the network topology; and (3) the unified approach that combines network topology and real data with pattern recognition methods (see Figure 1). Figure 2. Networks for the Pima Indian diabetes dataset and Australian credit card approval dataset. More specifically, a new weighted heterogeneous similarity function is also proposed to estimate relationships among interactive events. In the second phase of the framework, combining pattern-recognition techniques with network-based approaches. In this research, we propose a new hybrid detection framework in the proposed network topology.
Figure 1. Combining network metrics and pattern recognition.
As seen in Figure 1, information from observations is extracted by building the network, and feature properties for each observation are used to classify the output. For the results, we compare three approaches: (1) classic approach that uses traditional pattern recognition techniques; (2) the networked approach that uses pattern recognition techniques on the network topology; and (3) the unified approach that combines network topology and real data with pattern recognition methods (see Figure
Figure 2. Networks of Different Applications.
More specifically, a new weighted heterogeneous similarity function is also proposed to estimate relationships among interactive events. In the second phase of the framework, we combine the pattern-recognition techniques with network-based approaches. In this research, we propose a new hybrid detection framework in the proposed network topology.
As a result, the networks (as seen in Figure 2) are built to see how the events are similar and how they interact with each other. Based on the network metrics such as degree centrality, closeness centrality, betweenness centrality, in-degree centrality, out-degree centrality, load centrality and harmonic centrality, the pattern recognition techniques are applied to detect the credit card approval, breast cancer diagnosing, schizophrenia disease in fMRI, and diabetic disease. In conclusion, the proposed approach was tested and validated using real world case studies.
