Self-organizing Map Classification Research Articles

TUNING HYPERPARAMETERS OF SELF-ORGANIZING MAPS IN COMBINATION WITH K-NEAREST NEIGHBORS FOR IOT MALWARE DETECTION

In the Internet of Things, sensor devices often generate massive sensory data across multiple domains and applications. Identifying IoT malware from a huge amount of such IoT data is often a challenging task. In our previous studies, analytic techniques were applied to reduce dimensionality and discover valuable information from the original data. Particularly, the Self-organizing Maps (SOM)-based classifier with an AutoEncoder is used to create an end-to-end IoT malware detection model. However, the SOM-based classifier has a constraint that new instances may be incorrectly classified if they are mapped into unlabelled neurons in the SOM map. To address this issue, in this study, a novel hybrid between SOM-based classifier and well-known classification algorithms like K-Nearest Neighbors, Support Vector Machine, Softmax, Random Forest. In this hybrid, classification methods will help to correctly assign labels for instances mapped into the unlabeled neurons. In addition, this article investigates hyperparameter optimization methods for optimizing SOM hyperparameters. Our proposed methods were tested on the NBaIoT dataset with various experimental settings. Experimental results illustrate that SOMKNN often performs better than stand-alone techniques, including the SOM classifier.

Learning Vector Quantization, Hebbian Learning, and Self-Organizing Map for Classification

Deep Learning has been rapidly developed. Almost all proposed methods already have very high accuracy. Most of these methods still use techniques from the past with some modifications to adapt to existing modules. Sometimes it is necessary to understand past methods to produce new methods. Therefore, this research examines past models that have the potential to improve the performance of existing deep learning models. The methods to be examined include Learning Vector Quantization (LVQ), Hebbian learning, and Self-Organizing Map (SOM). The iris dataset available on Scikit-learn (SKlearn) is used here for testing in cases of supervised learning and unsupervised learning (especially SOM). The results show that LVQ has a good accuracy of 93%, while Hebbian learning has an accuracy of 56%. SOM fluctuates between 88% and 93%. Although the accuracy of SOM does not exceed LVQ, this model does not require labels in its training process.

Unsupervised Classification of Water Quality Using Artificial Intelligence: The Case of the Moulouya Wadi's Surface Waters (NE, Morocco)

Due to its cruciality, water requires a high care to its physicochemical and microbiological properties to ensure the quality of several utilizations. The particles it carries are likely to be ingested, breathed, or come into contact with the skin. For the classification of the quality of surface water in the Moulouya River (NE, Morocco), this study presents many unsupervised classification methods. The overall quality of surface water in the Moulouya River in northeast Morocco was assessed using nine physicochemical parameters (pH, T°C, EC, O2-diss, NH4+, NO3-, SO42-, PO43-, and biological oxygen demand after 5 days (BOD5)) from March to August 2014. Over a 600-kilometer stretch, twenty-two sites were examined, from the river's source in the High Atlas to its mouth in the Mediterranean. During the first stage, three quality classes (excellent, good and poor) were defined by the calculation of the water quality index (WQI) and water quality evaluation system (QES-Water). In the second stage, the K-means algorithm, the fuzzy C-means algorithm and the self-organizing maps (SOM) of Kohonen were applied to the nine physicochemical parameters used as input variables for the model. The classification method used is capable of projecting high-dimensional data into a lower dimension, typically 2D. This nonlinear projection can be useful in classes’ analysis and their discovery. In terms of performance, the SOM classification showed very close results compared to the K-means and the fuzzy C-means algorithms, with only an insignificant difference across the three models, with SOM maps having a slight advantage.

Examining village characteristics for forest management using self- and geographic self-organizing maps: A case from the Baekdudaegan mountain range network in Korea

Understanding the village characteristics linked to forest networks is essential for the scientific management of forest resources. Forests are complex socio-ecological systems. This study classifies the resources and characteristics of forest networks and neighboring villages using unsupervised learning algorithms: self-organizing maps (SOM) and geographic-self-organizing maps (Geo-SOMs). Considering ecological, economic, and sociocultural indicators, 18 covariates of 379 villages in two forest networks of the Baekdudaegan Mountain Range in South Korea were analyzed. The data visualizing map size was fixed based on changes in quantization and topographic errors of the same grid maps, and the number of clusters was determined by comparing K-means and hierarchical clustering techniques. An optimal map size of 17 × 12 grids and six clusters was used for further classification of the input data for both SOM and Geo-SOM analyses. The common characteristics of villages were identified using SOM classification, whereas geographically bounded characteristics were identified using Geo-SOM. The approach introduced in this study can be applied to socio-ecological classification and the design of sustainable forest management policies that link the remote sensing and geographic information systems.

Iris Recognition Using Hybrid Self-Organizing Map Classifier and Daugman’s Algorithm

One of the neural network algorithms that can be used in iris recognition is self-organizing map (SOM). This algorithm has a weakness in determining the initial weight of the network, which is generally carried out randomly, which can result in a decrease in accuracy when an incorrect determination is made. The solution that is often used is to apply a hybrid process in determining the initial weight of the SOM network. This study takes an approach using the cosine similarity equation to determine the initial weight of the network SOM in order to increase recognition accuracy. In addition, the localization process needs to be carried out to limit the area of the iris image being studied so that it is easy for the recognition process to be carried out. The method proposed in this study for iris recognition, namely hybrid SOM and Daugman’s algorithm, has been tested on several people by capturing the iris of the eye using a digital camera. The captured eyes have been localized first using the Daugman’s algorithm, and then the image features were extracted using the GLCM and LBP methods. In the final stage of the study, an iris recognition comparison test was performed, and the results obtained an accuracy of 85.50% using the proposed method and an accuracy of 73.50% without performing a hybrid process on the SOM network.

Self-Organizing Maps for the Classification of Spatial and Temporal Variability of Tornado-Favorable Parameters

Abstract A nuanced analysis of the spatial and temporal distribution of supercell tornadoes and the characteristics of the near-storm environments associated with those tornadoes is critical to improving our understanding of the range of environments that can be considered tornado favorable. This work classifies both supercell tornado probabilities and their associated environmental parameters on hourly and daily time scales based on geographical regions: regional probability of tornado events and the probability of deviation above or below the median tornadic near-storm environmental parameter values are estimated by kernel density estimation and classified by self-organizing maps (SOMs). The SOM classification for tornado probability allows for further examination of the deviation of the environmental parameters from the median for each probability cluster. Regions that have similar tornado probabilities but differ in the deviation of the environmental parameters (“parameter anomalies”) are also highlighted using SOMs. The anomaly patterns for different regions and parameters generally evolve along either seasonal or diurnal scales, but rarely both, highlighting the need for flexible models of tornado potential based on the near-storm environment. The spatial and temporal variability of parameter anomalies add complexity to traditional forecasting approaches that depend upon a fixed set of environmental parameter thresholds. This work highlights the need to develop region-specific and potentially time-specific environmental baseline evaluation to improve forecast and warning skill.

Cloud Detection Using an Ensemble of Pixel-Based Machine Learning Models Incorporating Unsupervised Classification

Remote sensing imagery, such as that provided by the United States Geological Survey (USGS) Landsat satellites, has been widely used to study environmental protection, hazard analysis, and urban planning for decades. Clouds are a constant challenge for such imagery and, if not handled correctly, can cause a variety of issues for a wide range of remote sensing analyses. Typically, cloud mask algorithms use the entire image; in this study we present an ensemble of different pixel-based approaches to cloud pixel modeling. Based on four training subsets with a selection of different input features, 12 machine learning models were created. We evaluated these models using the cropped LC8-Biome cloud validation dataset. As a comparison, Fmask was also applied to the cropped scene Biome dataset. One goal of this research is to explore a machine learning modeling approach that uses as small a training data sample as possible but still provides an accurate model. Overall, the model trained on the sample subset (1.3% of the total training samples) that includes unsupervised Self-Organizing Map classification results as an input feature has the best performance. The approach achieves 98.57% overall accuracy, 1.18% cloud omission error, and 0.93% cloud commission error on the 88 cropped test images. By comparison to Fmask 4.0, this model improves the accuracy by 10.12% and reduces the cloud omission error by 6.39%. Furthermore, using an additional eight independent validation images that were not sampled in model training, the model trained on the second largest subset with an additional five features has the highest overall accuracy at 86.35%, with 12.48% cloud omission error and 7.96% cloud commission error. This model’s overall correctness increased by 3.26%, and the cloud omission error decreased by 1.28% compared to Fmask 4.0. The machine learning cloud classification models discussed in this paper could achieve very good performance utilizing only a small portion of the total training pixels available. We showed that a pixel-based cloud classification model, and that as each scene obviously has unique spectral characteristics, and having a small portion of example pixels from each of the sub-regions in a scene can improve the model accuracy significantly.

Unsupervised machine learning for time-lapse seismic studies and reservoir monitoring

Time-lapse (4D) seismic analysis plays a vital role in reservoir management and reservoir simulation model updates. However, 4D seismic data are subject to interference and tuning effects. Being able to resolve and monitor thin reservoirs of different quality can aid in optimizing infill drilling or in locating bypassed hydrocarbons. Using 4D seismic data from the Maui field in the offshore Taranaki Basin of New Zealand, we generate typical seismic attributes sensitive to reservoir thickness and rock properties. We find that spectral instantaneous attributes extracted from time-lapse seismic data illuminate more detailed reservoir features compared with those same attributes computed on broadband seismic data. We have developed an unsupervised machine-learning workflow that enables us to combine eight spectral instantaneous seismic attributes into single classification volumes for the baseline and monitor surveys using self-organizing maps (SOMs). Changes in the SOM natural clusters between the baseline and monitor surveys suggest production-related changes that are caused primarily by water replacing gas as the reservoir is being swept under a strong water drive. The classification volumes also facilitate monitoring water saturation changes within thin reservoirs (ranging from very good to poor quality) as well as illuminating thin baffles. Thus, these SOM classification volumes indicate internal reservoir heterogeneity that can be incorporated into reservoir simulation models. Using meaningful SOM clusters, geobodies are generated for the baseline and monitor SOM classifications. The recoverable gas reserves for those geobodies are then computed and compared with production data. The SOM classifications of the Maui 4D seismic data seem to be sensitive to water saturation change and subtle pressure depletions due to gas production under a strong water drive.

Which seismic attributes are best for subtle fault detection?

Subtle fault detection plays a vital role in reservoir development studies because faults may form baffles or conduits that significantly control how a petroleum reservoir is swept. Small-throw faults are often overlooked in interpreting seismic amplitude data. However, seismic attributes can aid in mapping small faults. Over the years, dozens of seismic attributes have been developed that offer additional features for interpreters with associated caveats. Using the Maui 3D seismic data acquired in the Offshore Taranaki Basin, New Zealand, we have generated seismic attributes that are typically useful for fault detection. We find that multiattribute analysis provides greater geologic information than would be obtained by the analysis of individual attribute volumes. We extract the geologic content of multiple attributes in two ways: interactive corendering of different seismic attributes and the unsupervised machine learning algorithm self-organizing maps (SOM). Corendering seismic attributes that are mathematically independent but geologically interrelated provides a well-integrated structural image. We suggest eight combinations of 16 various attributes useful for a human interpreter with interest in fault and fracture detection. Current interpretation display capabilities constrain corendering to only four attribute volumes. Therefore, we use principal component analysis and SOM techniques to efficiently integrate the geologic information contained within many attributes. This approach gathers the data into one classification volume based on the interrelationships between seismic attributes. We show that our resulting SOM classification volume better highlights small faults that are difficult to image using conventional seismic interpretation techniques. We find that SOM works best when a fault exhibits anomalous features for multiple attributes within the same voxel. However, human interpreters are more adept at recognizing spatial patterns within various attributes and can place them in an appropriate geologic context.

ANN Classification and Modified Otsu Labeling on Retinal Blood Vessels

Background: Diagnosis of ophthalmologic and cardiovascular systems most often rely on the prerequisite step of segmentation of retinal blood vessels. Analysis of vascular structures in the retinal fundus images can aid in the early screening or detection of many ophthalmological diseases like glaucoma, diabetic retinopathy, vein occlusions, hemorrhages etc. In most cases, optic nerve gets damaged causing a blind spot. In this paper, a method of blood vessel segmentation using improved SOM (iSOM) and ANN classifier is presented. Methods: Morphological operations are carried out to enhance the input image. Clustering of pixels is done using improved Kohonen Self- Organizing Map (SOM) based on texture feature wherein a new node is introduced and new learning methodology is adopted using constrained weight updation. Finally, modified Otsu method is designed to label the output neuron class as vessel and non -vessel. Results: Segmentation is tested on public image sets, High Resolution Fundus (HRF) images and DRIONS-DB databases for Accuracy, Recall rate, Precision, F-Score, AUC and JC. The results achieve an appreciable level of accuracy (~97%) as compared to other similar methods of classification. The average time taken is less in estimating the neuron class and is about 12.1 sec per image when evaluated on Intel Core i5 CPU running at 2.30 GHz coupled with 4 GB RAM. The mean squared error for the segmented images is found to be in the range of 4-5%. Conclusion: Segmentation of retinal blood vessels based on artificial neural networks employing iSOM preserves the topology consuming less time for constrained weight updation achieving better results than SOM. A new model to detect vessels can be developed by concatenating iSOMs in parallel for multi class functions.

Clustering Ensemble Model Based on Self-Organizing Map Network.

This paper proposes a clustering ensemble method that introduces cascade structure into the self-organizing map (SOM) to solve the problem of the poor performance of a single clusterer. Cascaded SOM is an extension of classical SOM combined with the cascaded structure. The method combines the outputs of multiple SOM networks in a cascaded manner using them as an input to another SOM network. It also utilizes the characteristic of high-dimensional data insensitivity to changes in the values of a small number of dimensions to achieve the effect of ignoring part of the SOM network error output. Since the initial parameters of the SOM network and the sample training order are randomly generated, the model does not need to provide different training samples for each SOM network to generate a differentiated SOM clusterer. After testing on several classical datasets, the experimental results show that the model can effectively improve the accuracy of pattern recognition by 4%∼10%.

Stratification of cardiopathies using photoplethysmographic signals

Abstract Background and Objectives Cardiovascular health is monitored by many different approaches in order to diagnose and prevent critical patient conditions. The wide used of Photoplethysmography in medicine can facilitate heart status monitoring by heart rate variability (HRV) calculations extracted from blood volume variation measurements. The purpose of this work is to develop a computational process which relies on the evaluation of HRV as a prognostic marker in the stratification of the clinical heart conditions: Idiopathic Dilated Cardiomyopathy (IDC), Chagas Cardiomyopathy (CC) and Ischemic Cardiomyopathy (IC). Methods The database created for the study was developed with the assistance of Dr. Carlos Alberto Studart Gomes Hospital in the city of Fortaleza, CE. The detection of cardiac anomaly is analyzed by computational intelligence and pattern recognition algorithms. The classifiers used are the Multilayer Perceptron (MLP), Self-Organizing Map (SOM), Kernel kmeans (Kk-means), and k-Nearest Neighbors (k-NN). Results The results indicate SOM and Kk-means classifiers are the most suitable algorithms to be applied because they hold the highest true rates (Acc = 85.44% to 100%) and sensitivity (Se = 92.85% to 100%). Conclusions Among the heart conditions considered in this paper, Chagas Cardiomyopathy analysis presented the best results for any algorithm because the differences in the patterns of this disease are easier to identify due to arrhythmogenic characteristics. IDC and IC have similar statistical variables versus healthy subjects, which may complicate the pattern recognition performed by classifiers. The combination of extracted PPG signal features with machine learning may result in a high stratification performance and be of assistance to doctors for early diagnosis of disease.

Novel Optimization Based Hybrid Self-Organizing Map Classifiers for Iris Image Recognition

The concern over security in all fields has intensified over the years. The prefatory phase of providing security begins with authentication to provide access. In many scenarios, this authentication is provided by biometric systems. Moreover, the threat of pandemic has made the people to think of hygienic systems which are noninvasive. Iris image recognition is one such noninvasive biometric system that can provide automated authentication. Self-organizing map is an artificial neural network which helps in iris image recognition. This network has the ability to learn the input features and perform classification. However, from the literature it is observed that the performance of this classifier has scope for refinement to yield better classification. In this paper, heterogeneous methods are adapted to improve the performance of the classifier for iris image recognition. The heterogeneous methods involve the application of Gravity Search Optimization, Teacher Learning Based Optimization, Whale Optimization and Gray Wolf Optimization in the training process of the self-organizing map classifier. This method was tested on iris images from IIT-Delhi database. The results of the experiment show that the proposed method performs better.

Potential impacts of specific global warming levels on extreme rainfall events over southern Africa in CORDEX and NEX‐GDDP ensembles

AbstractThis study examines the potential impacts of specific global warming levels (GWLs) on extreme rainfall events in southern Africa and over 12 major cities in the region. We analysed two regional climate simulation datasets: the Coordinated Regional Climate Downscaling Experiment (CORDEX) and the NASA Earth eXchange Global Daily Downscaled Projections (NEX‐GDDP, hereafter referred to as NEX), which used different methods (dynamical and statistical, respectively) in downscaling several global climate model (GCM) simulations from CMIP5. Twenty simulations were used from each dataset. The simulated rainfall indices for the reference period (1971–2000) were compared with eight observation datasets. The projected changes in rainfall indices were examined at four GWLs under the RCP8.5 climate forcing scenario. The projections from the CORDEX and NEX ensembles were compared and the self‐organizing map (SOM) analysis was used to classify the projected changes from the simulations into 12 groups, based on their similarities. There is a good agreement among the eight observation datasets, albeit with some differences. The two simulation datasets (CORDEX and NEX) capture the climatology of rainfall indices over southern Africa and, in some cases, the simulation errors fall within the observation uncertainties. CORDEX and NEX datasets agree on fewer rain days but more intense rainfall events over southern Africa in the future and indicate that the magnitude of these changes increases with the GWLs. They also agree on a future increase in the frequency and intensity of extreme rainfall events north of 20°S and a decrease south of 20°S. However, CORDEX indicates that the rainfall deficit from the fewer rain events may exceed the rainfall surplus from the more intense rain events, but NEX indicates the opposite. In the SOM classification, while some simulations project a decrease in tropical temperate trough (TTT) rainfall, some project an increase, and some indicate an equatorward shift of the TTT rainfall. The simulations project a decrease in annual rainfall over five cities, an intensification of extreme rainfall over five cities and more frequent extreme rainfall events over four cities. The results of this study have application in mitigating the future impacts of extreme rainfall events over southern Africa, especially in the cities.

Relating anomalous large-scale atmospheric circulation patterns to temperature and precipitation anomalies in the East Asian monsoon region

Abstract The anomalous large-scale atmospheric circulation patterns concerned with surface air temperature and precipitation anomalies within East Asian Monsoon region for winter and summer months from 1979 to 2017 are analyzed by employing the self-organizing map (SOM) neural network. The asymmetric 3 × 4 SOM neural network is firstly constructed with the geopotential height anomalies at 500-hPa level as the only input variable. Then 12 characteristic anomalous large-scale atmospheric circulation patterns (nodes) are identified. The composites of temperature and precipitation anomalies as well as vertical wind anomalies at 500-hPa level assigned to each node are generated and visualized. The spatial distributions of anomalous geopotential height (anticyclones or cyclones) are highly consistent with that of surface air temperature anomalies in both winter and summer. Most precipitation extremes in winter are attributed to the joint effect of the horizontal and vertical atmospheric motions, while strong air convection is prone to inducing extreme precipitation events in summer. Based on the SOM classification, the differences of atmospheric circulations, temperature and precipitation anomalies have also been partly identified in strong and weak phases of East Asia Monsoon.

Projected changes in drought characteristics over the Western Cape, South Africa

AbstractThe devastating socioeconomic impacts of recent droughts in the Western Cape have intensified the quest for future drought mitigation measures. While ongoing global warming may increase atmospheric evaporative demand and worsen drought conditions, most studies on drought in the Western Cape have overlooked the role of potential evapotranspiration (PET). The present study examines the role of PET on future drought characteristics, focusing on four river catchments. Two drought indices, the Standardized Precipitation Index (SPI) and the Standardized Precipitation Evapotranspiration Index (SPEI), were analysed. The capability of the Co‐ordinated Regional Downscaling Experiment simulations to reproduce the drought characteristics was evaluated by comparing present‐day climate simulations with observations. The impacts of different global warming levels (GWLs) on the SPI and SPEI projections were assessed using self‐organizing map (SOM) classifications. The results project that a robust drying signal across the Western Cape, but the magnitudes of the projections, which vary across the river catchments, increase with the GWLs. The changes in the drought intensity and frequency are weaker when using the SPI than the SPEI, suggesting that SPI projections may underestimate the influence of global warming on drought because they do not account for the influence of PET. The SOM classification confirms the differences between the two drought indices reveals the drought patterns that are not seen in the drought ensemble means. Nevertheless, the study underscores the need to mitigate future drought impacts over the Western Cape, but with careful consideration on the drought index used in characterizing the droughts.

Combining domain filling with a self-organizing map to analyze multi-species hydrocarbon signatures on a regional scale.

For the period of the Barnett Coordinated Campaign, October 16-31, 2013, hourly concentrations for 46 volatile organic compounds (VOCs) were recorded at 14 air monitoring stations within the Barnett Shale of North Texas. These measurements are used to identify and analyze multi-species hydrocarbon signatures on a regional scale through the novel combination of two techniques: domain filling with Lagrangian trajectories and the machine learning unsupervised classification algorithm called a self-organizing map (SOM). This combination of techniques is shown to accurately identify concentration enhancements in the lightest measured alkane species at and downwind of the locations of active-permit oil and gas facilities, despite the model having no a priori knowledge of these source locations. Site comparisons further identify the SOM's ability to distinguish between signatures with differing influences from oil- and gas-related processes and from urban processes. A random forest (a machine learning supervised classification) analysis is conducted to further probe the sensitivities of the SOM classification in response to changes in any hydrocarbon species' concentration values. The random forest analysis of four representative classes finds that the SOM classification is appropriately more sensitive to changes in certain urban-related species for urban-related classes, and to changes in oil- and gas-related species for oil- and gas-related classes.

Projecting drought characteristics over East African basins at specific global warming levels

Drought poses a threat to socio-economic activities across most East African river basins. While there are indications that global warming may continue to enhance evaporation from basins and intensify droughts, most drought projections over East Africa are based on rainfall alone. The present study combines rainfall and Potential Evapotranspiration (PET) to examine the impacts of global warming on droughts in East Africa, with a focus on four major river basins. We characterised the drought by means of two drought indices: the Standardized Precipitation Evapotranspiration Index (SPEI) and the Standardized Precipitation Index (SPI). While the SPEI is based on the Climate Water Balance (CWB) (precipitation minus evapotranspiration), SPI is based on only rainfall. Multi-ensemble simulations from the CORDEX (the Coordinated Regional Climate Downscaling Experiment) datasets were analysed for the study. The capability of the simulations to reproduce the relevant climate variables (temperature, precipitation, PET and water balance, SPEI and SPI) over the East Africa region and specifically across four river basins (Rufiji, Tana, Juba and Upper-Nile) was evaluated. The impacts of climate change on the drought indices were considered at four specific Global Warming Levels (GWLs) (i.e. 1.5 °C, 2.0 °C, 2.5 °C, and 3.0 °C) under the Representative Concentration Pathway (RCP) 8.5 scenario and a Self-Organizing Map (SOM) was used to classify the SPEI and SPI projections into 12 groups, based on their similarities.Our results show that the CORDEX models give a realistic simulation of the relevant climate variables (including the drought indices) over East Africa and the river basins. For most of the variables, the spatial correlations between the observed and the simulated variables are high (0.6 < r > 0.9). In the annual cycle over these basins, the observations fall within the simulation spreads, and the simulation ensemble mean closely follows the observations. However, the ensemble mean does not accurately reproduce the spatial distribution of the drought frequency in the region. The CORDEX simulations project an increase in the intensity and frequency of SPEI droughts over East Africa. The magnitude of the projection, which varies over the region, increases with increasing GWLs. In contrast to the SPEI projection, however, the SPI projects only a weak change for intensity and frequency of droughts, and the magnitude does not vary with the GWLs. The SOM classification shows that 90% of the simulations project more severe and more frequent droughts with SPEI than they do with SPI. Over the river basins, the SPEI projections are more robust than the SPI. Hence, using only rainfall for drought projections may underestimate the severity of future droughts, while incorporating PET may enhance the robustness of the projections. The results of the study have application in mitigating the impacts of droughts on socio-economic activities across the East African river basins.

A self-organizing predictive map for non-life insurance

This article explores the capacity of self-organizing maps (SOMs) for analysing non-life insurance data. Contrary to feed forward neural networks, also called perceptron, a SOM does not need any a priori information on the relevancy of variables. During the learning procedure, the SOM algorithm selects the most relevant combination of explanatory variables and reduces by this way the collinearity bias. However, the specific features of insurance data require adapting the classic SOM framework to manage categorical variables and the low frequency of claims. This work proposes several extensions of SOMs in order to study the claims frequency of a portfolio of motorcycle insurances. Results are next compared to these computed with variants of the k-mean clustering algorithm.

Distinguishing Chemically Similar Polyamide Materials with ToF-SIMS Using Self-Organizing Maps and a Universal Data Matrix

Time-of-flight secondary ion mass spectrometry (ToF-SIMS) is advancing rapidly, providing instruments with growing capabilities and resolution. The data sets generated by these instruments are likewise increasing dramatically in size and complexity. Paradoxically, methods for efficient analysis of these large, rich data sets have not improved at the same rate. Clearly, more effective computational methods for analysis of ToF-SIMS data are becoming essential. Several research groups are customizing standard multivariate analytical tools to decrease computational demands, provide user-friendly interfaces, and simplify identification of trends and features in large ToF-SIMS data sets. We previously applied mass segmented peak lists to data from PMMA, PTFE, PET, and LDPE. Self-organizing maps (SOMs), a type of artificial neural network (ANN), classified the polymers based on their molecular composition and primary ion probe type more effectively than simple PCA. The effectiveness of this approach led us to question whether it would be useful in distinguishing polymers that were very similar. How sensitive is the technique to changes in polymer chemical structure and composition? To address this question, we generated ToF-SIMS ion peak signatures for seven nylon polymers with similar chemistries and used our up-binning and SOM approach to classify and cluster the polymers. The widely used linear PCA method failed to separate the samples. Supervised and unsupervised training of SOMs using positive or negative ion mass spectra resulted in effective classification and separation of the seven nylon polymers. Our SOM classification method has proven to be tolerant of minor sample irregularities, sample-to-sample variations, and inherent data limitations including spectral resolution and noise. We have demonstrated the potential of machine learning methods to analyze ToF-SIMS data more effectively than traditional methods. Such methods are critically important for future complex data analysis and provide a pipeline for rapid classification and identification of features and similarities in large data sets.