Correctly Classified Instances Research Articles

Breaking New Ground, Not Bones: A Machine Learning Odyssey in Osteoporosis Treatment Predictions

Abstract Background Delayed or omitted treatment for osteoporosis represents a recognised challenge within orthogeriatric healthcare. Machine-learning (ML) decision support systems offer a potential solution by enabling osteoporosis treatment recommendations without necessitating specialist intervention. This study seeks to evaluate the efficacy of ML in accurately recommending osteoporosis treatment. Methods Datasets from January to March 2023 in University Hospital Waterford were sourced from the IHFD (N=85). Osteoporosis treatment decisions were obtained from discharge letters. The dataset was entered into the WEKA 3.8.6 environment for ML processing. The Zero-Rule algorithm is selected as a baseline cross-validation technique to estimate ML performance. Results The Correctly Classified Instances (CCI) and Incorrectly Classified Instances (ICI) are 74.12% and 25.86%, respectively. Eight algorithms were selected, with 800 results produced and tested with a paired T-Tester. The significance level is set at 0.05, with the “Percent_correct” metric selected as the performance measure to compare the algorithms. The J48-Tree algorithm showed the best accuracy at 86% and is statistically better than the baseline algorithm; hence, it was selected as the trained model. Application of the trained model on the unseen dataset revealed CCI and ICI of 98.24% and 1.7%, respectively, indicating a high prediction accuracy rate. Conclusion In conclusion, this study has shown promising strides in using ML to assist osteoporosis treatment decisions. The resulting model showed notable predictive performance, indicating the viability of ML in decision-making. Future studies should focus on acquiring larger datasets for validation, algorithm refinement and exploring the integration of clinical expertise to bolster the model's applicability and reliability in real-world scenarios.

A Hybrid Model for Predicting Classification Dataset based on Random Forest, Support Vector Machine and Artificial Neural Network

Machine Learning offers a rich array of algorithms, and the performance of these algorithms can vary significantly depending on the specific task. Combining these traditional algorithms can lead to the development of innovative hybrid structures that outperform individual models. One such novel hybrid model is the Hybrid Support Random Forest Neural Network (HSRFNN), which is designed to deliver enhanced performance and accuracy. HSRFNN represents a fusion of Random Forest, Support Vector Machine (SVM), and Artificial Neural Network (ANN) to leverage their respective strengths. This hybrid model consistently outperforms the individual models of Random Forest, SVM, and ANN. In this study, ten diverse datasets sourced from UCI and Kaggle data repositories were considered for evaluation. The accuracy of the HSRFNN model was meticulously compared with the three traditional algorithms, namely Random Forest, Support Vector Machine, and Artificial Neural Network. Various accuracy metrics, such as Correctly Classified Instances (CCI), Incorrectly Classified Instances (ICI), Accuracy (A), and Time Taken to Build Model (TTBM), were used for the comparative analysis. This research strives to demonstrate that HSRFNN, through its hybrid architecture, can offer superior accuracy and performance compared to individual algorithms. The choice of datasets from different sources enhances the generalizability of the results, making HSRFNN a promising approach for a wide range of machine learning tasks. Further exploration and fine-tuning of HSRFNN may unlock its potential for even more challenging and diverse datasets.

Risk transboundary transmission areas and driving factors of brucellosis along the borders between China and Mongolia

ObjectiveBrucellosis is a common and neglected zoonotic infectious disease worldwide caused by Brucella. However, transboundary transmissions among countries, particularly those with high incidences, are seldom investigated. In the present study, by taking China and Mongolia as examples, we aim to identify transboundary transmission risk and driving factors of brucellosis along borders. Methods167 brucellosis outbreak locations along the border between China and Mongolia were collected. Wildlife distribution and cross-border activities were mapped. Maximum entropy approach modeling was conducted to predict the potential risk of prevalence of brucellosis with meteorological factors, geographical environment, economic development, living habits et al. The accuracy of the models was assessed by the area under the receiver operating characteristic (ROC) curve (AUC), Kappa test, and correctly classified instances (CCI). ResultsThe spatial model performed excellent predictive performance with the predictor variables of soils, pastures, goat density, mean precipitation of the wettest month, temperature seasonality, and population density, which with the contribution and permutation important in 27.2 %, 31.9; 23.3 %, 6.8; 18.0 %, 17.2; 11.2 %, 18.1; 10. 3 %, 15.2; 10.0 %, 10.8. The calculated AUC, SD, Kappa, and CCI are 0.870, 0.001, 0.882, and 0.883, respectively. The distribution map of brucellosis showed high-risk areas along the borders. ConclusionsOur study identified high-risk areas and the driving effect of brucellosis along the borders between China and Mongolia. Moreover, there is the possibility of cross-border wildlife activities in high-risk areas, which increases the risk of cross-border brucellosis transmission. The funding provides clues for cooperative prevention and control of brucellosis by reducing transboundary transmission.

Groundwater recharge potential zonation using an ensemble of machine learning and bivariate statistical models

Many regions in Iran are currently experience water crisis, largely driven by frequent droughts and expanding agricultural land combined with over abstraction of groundwater. Therefore, it is extremely important to identify potential groundwater recharge (GWR) zones to help in prevent water scarcity. The key objective of this research is to applying different scenarios for GWR potential mapping by means of a classifier ensemble approach, namely a combination of Maximum Entropy (ME) and Frequency Ratio (FR) models in a semi-arid mountainous, Marboreh Watershed of Iran. To consider the ensemble effect of these models, 15 input layers were generated and used in two models and then the models were combined in seven scenarios. According to marginal response curves (MRCs) and the Jackknife technique, quaternary formations (Qft1 and Qft2) of lithology, sandy-clay-loam (Sa. Cl. L) class of soil, 0–4% class of slope, and agriculture & rangeland classes of land use, offered the highest percolation potential. Results of the FR model showed that the highest weight belonged to Qft1 rocks and Sa. Cl. L textures. Seven scenarios were used for GWR potential maps by different ensembles based on basic mathematical operations. Correctly Classified Instances (CCI), and the AUC indices were applied to validate model predictions. The validation indices showed that scenarios 5 had the best performance. The combination of models by different ensemble scenarios enhances the efficiency of these models. This study serves as a basis for future investigations and provides useful information for prediction of sites with groundwater recharge potential through combination of state-of-the-art statistical and machine learning models. The proposed ensemble model reduced the machine learning and statistical models’ limitations gaps and promoted the accuracy of the model where combining, especially for data-scarce areas. The results of present study can be used for the GWR potential mapping, land use planning, and groundwater development plans.

Empirical Assessment of Machine Learning Techniques for Software Requirements Risk Prediction

Software risk prediction is the most sensitive and crucial activity of Software Development Life Cycle (SDLC). It may lead to the success or failure of a project. The risk should be predicted earlier to make a software project successful. A model is proposed for the prediction of software requirement risks using requirement risk dataset and machine learning techniques. In addition, a comparison is made between multiple classifiers that are K-Nearest Neighbour (KNN), Average One Dependency Estimator (A1DE), Naïve Bayes (NB), Composite Hypercube on Iterated Random Projection (CHIRP), Decision Table (DT), Decision Table/Naïve Bayes Hybrid Classifier (DTNB), Credal Decision Trees (CDT), Cost-Sensitive Decision Forest (CS-Forest), J48 Decision Tree (J48), and Random Forest (RF) achieve the best suited technique for the model according to the nature of dataset. These techniques are evaluated using various evaluation metrics including CCI (correctly Classified Instances), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Relative Absolute Error (RAE), Root Relative Squared Error (RRSE), precision, recall, F-measure, Matthew’s Correlation Coefficient (MCC), Receiver Operating Characteristic Area (ROC area), Precision-Recall Curves area (PRC area), and accuracy. The inclusive outcome of this study shows that in terms of reducing error rates, CDT outperforms other techniques achieving 0.013 for MAE, 0.089 for RMSE, 4.498% for RAE, and 23.741% for RRSE. However, in terms of increasing accuracy, DT, DTNB, and CDT achieve better results.

A Machine Learning Based Intrusion Detection System for Mobile Internet of Things.

Intrusion detection systems plays a pivotal role in detecting malicious activities that denigrate the performance of the network. Mobile adhoc networks (MANETs) and wireless sensor networks (WSNs) are a form of wireless network that can transfer data without any need of infrastructure for their operation. A more novel paradigm of networking, namely Internet of Things (IoT) has emerged recently which can be considered as a superset to the afore mentioned paradigms. Their distributed nature and the limited resources available, present a considerable challenge for providing security to these networks. The need for an intrusion detection system (IDS) that can acclimate with such challenges is of extreme significance. Previously, we proposed a cross layer-based IDS with two layers of detection. It uses a heuristic approach which is based on the variability of the correctly classified instances (CCIs), which we refer to as the accumulated measure of fluctuation (AMoF). The current, proposed IDS is composed of two stages; stage one collects data through dedicated sniffers (DSs) and generates the CCI which is sent in a periodic fashion to the super node (SN), and in stage two the SN performs the linear regression process for the collected CCIs from different DSs in order to differentiate the benign from the malicious nodes. In this work, the detection characterization is presented for different extreme scenarios in the network, pertaining to the power level and node velocity for two different mobility models: Random way point (RWP), and Gauss Markov (GM). Malicious activity used in the work are the blackhole and the distributed denial of service (DDoS) attacks. Detection rates are in excess of 98% for high power/node velocity scenarios while they drop to around 90% for low power/node velocity scenarios.

A Cross-Layer, Anomaly-Based IDS for WSN and MANET.

Intrusion detection system (IDS) design for mobile adhoc networks (MANET) is a crucial component for maintaining the integrity of the network. The need for rapid deployment of IDS capability with minimal data availability for training and testing is an important requirement of such systems, especially for MANETs deployed in highly dynamic scenarios, such as battlefields. This work proposes a two-level detection scheme for detecting malicious nodes in MANETs. The first level deploys dedicated sniffers working in promiscuous mode. Each sniffer utilizes a decision-tree-based classifier that generates quantities which we refer to as correctly classified instances (CCIs) every reporting time. In the second level, the CCIs are sent to an algorithmically run supernode that calculates quantities, which we refer to as the accumulated measure of fluctuation (AMoF) of the received CCIs for each node under test (NUT). A key concept that is used in this work is that the variability of the smaller size population which represents the number of malicious nodes in the network is greater than the variance of the larger size population which represents the number of normal nodes in the network. A linear regression process is then performed in parallel with the calculation of the AMoF for fitting purposes and to set a proper threshold based on the slope of the fitted lines. As a result, the malicious nodes are efficiently and effectively separated from the normal nodes. The proposed scheme is tested for various node velocities and power levels and shows promising detection performance even at low-power levels. The results presented also apply to wireless sensor networks (WSN) and represent a novel IDS scheme for such networks.

Bayesian belief network models to analyse and predict ecological water quality in rivers

Economic growth is often based on the intensification of crop production, energy consumption and urbanization. In many cases, this leads to the degradation of aquatic ecosystems. Modelling water resources and the related identification of key drivers of change are essential to improve and protect water quality in river basins. This study evaluates the potential of Bayesian belief network models to predict the ecological water quality in a typical multifunctional and tropical river basin. Field data, expert knowledge and literature data were used to develop a set of Bayesian belief network models. The developed models were evaluated based on weighted Cohen's Kappa (κw), percentage of correctly classified instances (CCI) and spherical payoff. On top, a sensitivity analysis and practical simulation tests of the two most reliable models were performed. Cross-validation based on κw (Model 1: 0.44±0.08; Model 2: 0.44±0.11) and CCI (Model 1: 36.3±2.3; Model 2: 41.6±2.3) indicated that the performance was reliable and stable. Model 1 comprised of input variables main land use, elevation, sediment type, chlorophyll, flow velocity, dissolved oxygen, and chemical oxygen demand; whereas Model 2 did not include dissolved oxygen and chemical oxygen demand. Although the predictive performance of Model 2 was slightly higher than that of Model 1, simulation outcomes of Model 1 were more coherent. Additionally, more management options could be evaluated with Model 1. As the model's ability to simulate management outcomes is of utmost importance in model selection, Model 1 is recommended as a tool to support decision-making in river management. Model predictions and sensitivity analysis indicated that flow velocity is the major variable determining ecological water quality and suggested that construction of additional dams and water abstraction within the basin would have an adverse effect on water quality. Although a case study in a single river basin is presented, the modelling approach can be of general use on any other river basin.

Modelling critical factors affecting the distribution of the vulnerable endemic Eastern Iberian barbel (Luciobarbus guiraonis) in Mediterranean rivers

Luciobarbus guiraonis (Eastern Iberian barbel) is an endemic fish species restricted to Spain, mainly distributed in the Júcar River Basin District. Its study is important because there is little knowledge about its biology and ecology. To improve the knowledge about the species distribution and habitat requirements, nonlinear modelling was carried out to predict the presence/absence and density of the Eastern Iberian barbel, based on 155 sampling sites distributed throughout the Júcar River Basin District (Eastern Iberian Peninsula). We used multilayer feed-forward artificial neural networks (ANN) to represent nonlinear relationships between L. guiraonis descriptors and variables regarding the physical habitat and biological components (macroinvertebrates, fish, riparian forest). The gradient descent algorithm was implemented to find the optimal model parameters; the importance of the ANN’s input variables was determined by the partial derivatives method. The predictive power of the model was evaluated with the Cohen’s kappa (k), the correctly classified instances (CCI), and the area under the curve (AUC) of the receiver operator characteristic (ROC) plots. The best model predicted presence/absence with a high performance (k= 0.66, CCI= 87% and AUC= 0.85); the prediction of density was moderate (CCI = 62%, AUC=0.71 and k= 0.43). The fundamental variables describing the presence/absence were; solar radiation (the highest contribution was observed between 2000 and 4200 WH/m2), drainage area (with the strongest influence between 3000 and 5.000 km2), and the proportion of exotic fish species (with relevant contribution between 50 and 100%). In the density model, the most important variables were the coefficient of variation of mean annual flows (relative importance of 50.5%) and the proportion of exotic fish species (24.4%). The models provide important information about the relation of L. guiraonis with biotic and abiotic variables, this new knowledge can help develop future studies and management plans for the conservation of this species in the Júcar River Basin District and, potentially, for the conservation of other endemic fish species of Barbus and Luciobarbus in Mediterranean rivers.

Physico-chemical and biological characterization of anopheline mosquito larval habitats (Diptera: Culicidae): implications for malaria control

BackgroundA fundamental understanding of the spatial distribution and ecology of mosquito larvae is essential for effective vector control intervention strategies. In this study, data-driven decision tree models, generalized linear models and ordination analysis were used to identify the most important biotic and abiotic factors that affect the occurrence and abundance of mosquito larvae in Southwest Ethiopia.MethodsIn total, 220 samples were taken at 180 sampling locations during the years 2010 and 2012. Sampling sites were characterized based on physical, chemical and biological attributes. The predictive performance of decision tree models was evaluated based on correctly classified instances (CCI), Cohen’s kappa statistic (κ) and the determination coefficient (R2). A conditional analysis was performed on the regression tree models to test the relation between key environmental and biological parameters and the abundance of mosquito larvae.ResultsThe decision tree model developed for anopheline larvae showed a good model performance (CCI = 84 ± 2%, and κ = 0.66 ± 0.04), indicating that the genus has clear habitat requirements. Anopheline mosquito larvae showed a widespread distribution and especially occurred in small human-made aquatic habitats. Water temperature, canopy cover, emergent vegetation cover, and presence of predators and competitors were found to be the main variables determining the abundance and distribution of anopheline larvae. In contrast, anopheline mosquito larvae were found to be less prominently present in permanent larval habitats. This could be attributed to the high abundance and diversity of natural predators and competitors suppressing the mosquito population densities.ConclusionsThe findings of this study suggest that targeting smaller human-made aquatic habitats could result in effective larval control of anopheline mosquitoes in the study area. Controlling the occurrence of mosquito larvae via drainage of permanent wetlands may not be a good management strategy as it negatively affects the occurrence and abundance of mosquito predators and competitors and promotes an increase in anopheline population densities.

Ground water quality classification by decision tree method in Ardebil region, Iran

A decision tree-based approach is proposed to predict ground water quality based on the United States Salinity Laboratory (USSL) diagram using the data from aquifers in agricultural lands of Ardebil province, northwest of Iran. Several combinations of hydro chemical parameters of groundwater and monthly precipitation with different lag time were considered to find an accurate and economical alternative for groundwater quality classification. The performance evaluation was based on the number of correctly classified instances (CCI) and kappa statistics. The results suggested the suitability of decision tree-based classification approach for the used data sets. The overall average of CCI and kappa statistic for the prediction of groundwater quality classes based on the USSL diagram was 0.88 and 0.83 %, respectively. Principal component analysis (PCA) was also used to determine the important parameters for groundwater quality classification. The results showed that groundwater quality classification by decision tree is more precise and efficient in comparison with PCA. The best alternative could evaluate groundwater quality class with only two parameters: electrical conductivity and cumulative precipitation of 11 months earlier. The developed model is able to predict water quality class by only two variables and this lead to a reduction in the number of variables analyzed on a routine basis, resulting in a significant reduction in laboratory costs and latency times between the sampling moment and the outcome of the laboratory analyses.

Application of genetic algorithm and greedy stepwise to select input variables in classification tree models for the prediction of habitat requirements of Azolla filiculoides (Lam.) in Anzali wetland, Iran

The aim of the present study was to predict to what extent wetland characteristics can affect the habitat requirements of an exotic species, Azolla filiculoides (Lam.) in wetland. Biotic and abiotic variables were collected at the Selkeh wildlife refuge (a protected area in Anzali wetland, northern Iran) over the study period 2007–2008. Classification tree (CT) was used to find the relationship between the wetland characteristics and the cover percentage of A. filiculoides. Genetic algorithm (GA) and greedy stepwise (GS) were combined with CT in order to select the most important variables to explain the coverage of A. filiculoides. The applied method was assessed based on the percentage of correctly classified instances (CCI) and Cohen's kappa statistics (k). Different pruning confidence factors (PCFs) were tested in order to improve the predictive results regarding the complexity and accuracy of the prediction. The results showed that the prediction was reliable in terms of both performance criteria. Yet, after variable selection, the predictive performances of the CT improved. Due to potential collinear variables in the model, the GS method was less efficient than GA. The optimization of GA and GS resulted in an easy interpretation of the selected variables. The methods showed that both structural habitat (e.g. air temperature, humidity and depth) and physico-chemical variables (e.g. nutrients) can affect the habitat requirements of A. filiculoides in the wetland but the dependence of this aquatic fern on structural habitat was well confirmed by the CT before and after variable selection. Application of the given algorithms in combination with CT thus proved to have a better capability in selecting the most important variables explaining the cover of A. filiculoides and can be used by wetland managers in their decision-making for wetland conservation and management programs.

Effect of model formulation on the optimization of a genetic Takagi–Sugeno fuzzy system for fish habitat suitability evaluation

► We assess the effects of model formulation on species distribution modelling . ► We focus on model accuracy, threshold criteria and habitat suitability curves. ► A genetic Takagi–Sugeno fuzzy model is used. ► The product operation leads to the highest accuracy and consistent habitat suitability curves. ► Threshold criteria should be selected according to the model formulation used. Species distribution models (SDMs), which evaluate species–environment relationships, are one of the key topics in ecology and biogeography. These models evaluate the current status of target ecosystems and potential impacts in both time and space. Although species distributions are often calculated based on the composite habitat suitability of several variables, there are no guidelines for calculating them. The present study assessed the effects of model formulation on habitat suitability evaluation and the accuracy of species distribution modelling. We employed a genetic algorithm (GA)-optimized fuzzy habitat preference model (FHPM) for evaluating habitat suitability of topmouth gudgeon ( Pseudorasbora parva ) in the Northwestern part of Kyushu Island in Japan. Four operations were used to calculate the composite habitat suitability from multiple habitat variables: arithmetic mean, geometric mean, product and minimum. To transform model outputs to presence/absence, four threshold criteria were compared based on model accuracy: prevalence, conventional 0.5, minimization of the sensitivity–specificity difference threshold (MDT), and maximization of the sensitivity–specificity sum threshold (MST). The models were first calibrated and validated based on the mean squared error (MSE) between composite habitat suitability and the observed presence–absence of the fish, and then evaluated using confusion matrix-derived measures such as the area under the receiver operating characteristics (ROC) curve (AUC), correctly classified instances (CCI), kappa and true skill statistic (TSS). The results clearly illustrated the effects of model formulation and threshold criteria on habitat suitability curves (HSCs) and accuracy in modelling species distributions. The use of the product model formulation led to the best accuracy in terms of MSE and AUC, and consistency in the shape of HSCs. The two threshold criteria of MST and MDT are also recommended for the consistently higher performance in terms of CCI, kappa and TSS. This case study of topmouth gudgeon illustrates the need for further studies on the model behaviour with regard to data characteristics (i.e., sample size and prevalence) and model structure (i.e., fuzzy sets and parameter settings of the GA).

Application of classification trees and support vector machines to model the presence of macroinvertebrates in rivers in Vietnam

In the present study, classification trees (CTs) and support vector machines (SVMs) were used to study habitat suitability for 30 macroinvertebrate taxa in the Du river in Northern Vietnam. The presence/absence of the 30 most common macroinvertebrate taxa was modelled based on 21 physical-chemical and structural variables. The predictive performance of the CT and SVM models was assessed based on the percentage of Correctly Classified Instances (CCI) and Cohen's kappa statistics. The results of the present study demonstrated that SVMs performed better than CTs. Attribute weighing in SVMs could replace the application of genetic algorithms for input variable selection. By weighing attributes, SVMs provided quantitative correlations between environmental variables and the occurrence of macroinvertebrates and thus allowed better ecological interpretation. SVMs thus proved to have a high potential when applied for decision-making in the context of river restoration and conservation management.

Habitat suitability modelling as a mapping tool for macrobenthic communities: An example from the Belgian part of the North Sea

Being ecologically important and well-known, the spatial distribution pattern of the macrobenthos is often used to support an ecologically sustainable marine management. Though in many cases the macrobenthic spatial distribution is relatively well-known, this information is merely restricted to point observations at the sampling stations: although being increasingly demanded, full coverage spatial distribution maps are generally lacking. This study therefore aimed at demonstrating the usefulness of habitat suitability modelling as a full coverage mapping tool with high relevance for marine management through (1) the construction of a habitat suitability model for the soft sediment macrobenthic communities in the Belgian part of the North Sea (BPNS) and (2) predicting the full coverage spatial distribution of macrobenthic communities within the BPNS. The BPNS was selected as a case study area because of the high data availability on both macrobenthos and environmental characteristics. Discriminant function analysis (DFA) objectively selected median grain size and sediment mud content and omitted bathymetry, slope and distance to the coast to represent the most important environmental variables determining the macrobenthic community distribution. The consequent crossvalidated, empirical habitat suitability model, using both median grain size and mud content, showed an a posteriori average correctly classified instances (CCI) of 79% (community-dependent CCI ranging from 72% to 86%) and a Cohen's kappa of 0.71, pointing towards a very good agreement between model predictions and observations. The application of the habitat suitability model on the full coverage maps of median grain size and sediment mud content, taken from literature, allowed to reliably assess the distribution of the macrobenthic communities within 96.3% of the 53,297 BPNS grid cells with a resolution of 250 m. Next to its applicability to the BPNS, the model is further anticipated to potentially perform well in the full Southern Bight of the North Sea: testing is advised here. Since the habitat suitability is considered far more stable through time compared to the permanently fluctuating macrobenthic communities, information on the habitat suitability of an area is considered highly important for a scientifically sound marine management.

Decision Tree Models for Prediction of Macroinvertebrate Taxa in the River Axios (Northern Greece)

In this study, decision tree models were induced to predict the habitat suitability of six macroinvertebrate taxa: Asellidae, Baetidae, Caenidae, Gammaridae, Gomphidae and Heptageniidae. The modelling techniques were applied on a dataset of 102 samples collected in 31 sites along the river Axios in Northern Greece. The database consisted of eight physical-chemical and seven structural variables, as well as the abundances of 90 macroinvertebrate taxa. A seasonal variable was included allowing the description of potential temporal changes in the macroinvertebrate taxa. Rules relating the presence/absence of six benthic macroinvertebrate taxa with the 15 physical-chemical and structural river characteristics and the seasonal variable were induced using the J48 algorithm. In order to improve the performance and the interpretability of the induced models, three optimisation techniques were applied: tree-pruning, bagging and boosting. The predictive performance of the decision tree models was assessed on the basis of the percentage of Correctly Classified Instances (CCI) and the Cohen’s kappa statistic. The results of the present study demonstrated that although the models had a relatively high predictive performance, noise in the dataset and inappropriate input variables prevented to some extent, the models from making reliable predictions. Although tree-pruning did not improve significantly the reliability of the induced models, it reduced considerably the tree complexity and in this way increased the transparency of the trees. Consequently, the induced models allowed for a correct ecological interpretation. The effect of bagging and boosting on the other hand varied considerably between the different models, as well as within different repetitions of 10-fold cross-validation in an individual model. In some cases the predictive performance was improved, in others stable or even worsened. The effect of bagging and boosting seemed to be strongly dependent on the dataset on which the two techniques were applied. Tree-pruning thus proved to have a high potential when applied in models used for decision-making of river restoration and conservation management.

Genetic algorithms for optimisation of predictive ecosystems models based on decision trees and neural networks

The selection of appropriate input variables in predictive ecological modelling is an important issue since numerous variables can be involved. Most of the input variables cannot be omitted because it results in a significant loss of information. The collection of field data on the other hand is both time-consuming and expensive. Rigorous methods are therefore needed to decide which explanatory variables or combinations of variables should enter the model. Appropriate selection of input variables is not only important for modelling objectives as such, but also to ensure reliable decision-support in river management and policy-making. In this paper, the use of genetic algorithms is explored to automatically select the relevant input variables for classification trees and artificial neural networks (ANNs), predicting the presence or absence of benthic macroinvertebrate taxa. The applied database consisted of measurements from 360 sites in unnavigable watercourses in Flanders, Belgium. The measured variables are a combination of physical–chemical, eco-toxicological and structural ones. The predictive power of the models was assessed on the basis of three performance measures: root mean squared error (RMSE), correctly classified instances (CCI) and Cohen's kappa. The selected genetic algorithm introduced different sets of input variables to the models and compared their predictive power to select the optimal combination of input variables. With this technique, the number of input variables could be reduced from 17 to three to six for classification trees and to five to eleven for ANNs. The prediction success increased significantly based on a statistical test. Overall, a better performance was detected for decision trees. The most appropriate way to assess the performance of the models was a combination of the CCI and Cohen's kappa. By means of this variable selection stage, the key variables that determine the presence or absence of benthic macroinvertebrate taxa in Flanders could also be identified.

Application of Bayesian Belief Networks for the prediction of macroinvertebrate taxa in rivers

Integrated ecological models are of great potential as predictive tools in decision support of river management. Such models need to be transparent and consistent with the existing expert knowledge, and give the river manager adequate information regarding their inherent uncertainty. One way to fulfil these needs is through the use of Bayesian Belief Networks (BBNs). Such networks represent cause-and-effect assumptions between system variables in a graphical structure. To establish the potential of Bayesian Belief Networks in river management, a small-scale study was performed with the aim of assessing the success of prediction of macroinvertebrate taxa in rivers by means of a selected number of environmental variables. Gammaridae and Asellidae were chosen because of their high relative abundances in small and large brooks in contrast to other macroinvertebrate taxa. Based on one-layered BBN networks, the predictive capacity of the models was assessed by means of the number of Correctly Classified Instances (CCI) and Cohen’s Kappa (K). The performance of these models was moderate to good for presence/absence classifications but showed a low to moderate performance when predicting abundance classes. When extending the former BBN network to a two-layered one, enhancing the number of links and variables, no obvious improvement in model performance was detected. The results indicate that thoughtful input variable selection as well as sensitivity analysis will improve the models for practical use in river restoration management.

Optimization of Artificial Neural Network (ANN) model design for prediction of macroinvertebrates in the Zwalm river basin (Flanders, Belgium)

To meet the requirements of the EU Water Framework Directive, models are useful to predict communities in watercourses based on the abiotic characteristics of their aquatic environment. For that purpose back-propagation Artificial Neural Network (ANN) algorithms were used to induce predictive models on a dataset of the Zwalm river basin (Flanders, Belgium). This dataset consisted of 120 samples, collected over a 2-year period. Fifteen environmental variables were measured at each site, as well as the abundance of the aquatic macroinvertebrate taxa. Different neural networks were developed and optimized to obtain the best model configuration for the prediction of the habitat suitability of macroinvertebrate taxa. The best performing number of hidden layers and neurons and training algorithms have been searched for. The different options were theoretically and practically validated and assessed. The theoretical validation was based on cross-validation. For the practical validation, potential applications of the neural network models were analyzed, and the predictive performance of the models was assessed using ecological expert knowledge. The results indicate that the number of times a taxon was found in the whole river basin influences the performance measures and the architecture of the network. Based on the Cohen’s kappa, it could be concluded that ANN models predicting the presence/absence of very rare taxa (e.g. Aplexa) or very common taxa (e.g. Tubificidae) were rather irrelevant, although their correctly classified instances (CCI) was high. Predicting the presence/absence of Asellidae (a moderately present taxon), the highest performances (CCI and Cohen’s kappa) were found for the network model with two hidden layers each having 10 neurons. When calculation time was also taken into account, the network model with one hidden layer having 10 neurons could be preferred. Applying this network architecture, performances were only slightly worse, while calculation time was a lot shorter. One may also conclude that not all network models resulted in a relevant relation between a variable and a specific taxon. For Gammaridae for example, a rather small ANN structure gave a better idea of the impact of dissolved oxygen on its presence than a larger one. More reliable predictions and ecological interpretations for river ecosystem management would thus be possible provided the best configuration could be found.

Use of genetic algorithms to select input variables in decision tree models for the prediction of benthic macroinvertebrates

Predicting freshwater organisms based on machine learning is becoming more and more reliable due to the availability of appropriate datasets, advanced modelling techniques and the continuously increasing capacity of computers. A database consisting of measurements collected at 360 sampling sites in non-navigable watercourses in Flanders was applied to predict the absence/presence of benthic macroinvertebrate taxa by means of decision trees. The measured variables were a combination of physical–chemical (temperature, pH, dissolved oxygen concentration, conductivity, total organic carbon, Kjeldahl nitrogen and total phosphorus), structural (granulometric analysis of the sediment, width, depth and flow velocity of the river) and two ecotoxicological variables. The predictive power of decision trees was assessed on the basis of the number of Correctly Classified Instances (CCI). A genetic algorithm was introduced to compare the predictive power of different sets of input variables for the decision trees. The number of input variables was reduced from 15 to 2–8 variables without affecting the predictive power of the decision trees significantly. Furthermore, reducing the number of input variables allowed to ease the identification of general data trends.