Olivier Regnier-Coudert

Machine learning for improved pathological staging of prostate cancer: A performance comparison on a range of classifiers

OBJECTIVES Prediction of prostate cancer pathological stage is an essential step in a patients pathway. It determines the treatment that will be applied further. In current practice, urologists use the pathological stage predictions provided in Partin tables to support their decisions. However, Partin tables are based on logistic regression (LR) and built from US data. Our objective is to investigate a range of both predictive methods and of predictive variables for pathological stage prediction and assess them with respect to their predictive quality based on U.K. data. METHODS AND MATERIAL The latest version of Partin tables was applied to a large scale British dataset in order to measure their performances by mean of concordance index (c-index). The data was collected by the British Association of Urological Surgeons (BAUS) and gathered records from over 1700 patients treated with prostatectomy in 57 centers across UK. The original methodology was replicated using the BAUS dataset and evaluated using concordance index. In addition, a selection of classifiers, including, among others, LR, artificial neural networks and Bayesian networks (BNs) was applied to the same data and compared with each other using the area under the ROC curve (AUC). Subsets of the data were created in order to observe how classifiers perform with the inclusion of extra variables. Finally a local dataset prepared by the Aberdeen Royal Infirmary was used to study the effect on predictive performance of using different variables. RESULTS Partin tables have low predictive quality (c-index=0.602) when applied on UK data for comparison on patients with organ confined and extra prostatic extension conditions, patients at the two most frequently observed pathological stages. The use of replicate lookup tables built from British data shows an improvement in the classification, but the overall predictive quality remains low (c-index=0.610). Comparing a range of classifiers shows that BNs generally outperform other methods. Using the four variables from Partin tables, naive Bayes is the best classifier for the prediction of each class label (AUC=0.662 for OC). When two additional variables are added, the results of LR (0.675), artificial neural networks (0.656) and BN methods (0.679) are overall improved. BNs show higher AUCs than the other methods when the number of variables raises CONCLUSION The predictive quality of Partin tables can be described as low to moderate on U.K. data. This means that following the predictions generated by Partin tables, many patients would received an inappropriate treatment, generally associated with a deterioration of their quality of life. In addition to demographic differences between U.K. and the original U.S. population, the methodology and in particular LR present limitations. BN represents a promising alternative to LR from which prostate cancer staging can benefit. Heuristic search for structure learning and the inclusion of more variables are elements that further improve BN models quality.

Using a Markov network as a surrogate fitness function in a genetic algorithm

Surrogate models of fitness have been presented as a way of reducing the number of fitness evaluations required by an evolutionary algorithm. This is of particular interest with expensive fitness functions where the cost of building the model is outweighed by the saving of using fewer function evaluations. In this paper we show how a Markov network model can be used as a surrogate fitness function in a genetic algorithm. We demonstrate this applied to a number of well-known benchmark functions and although the results are good in terms of function evaluations the model-building overhead requires a substantially more expensive fitness function to be worthwhile. We move on to describe a fitness function for feature selection in Case-Based Reasoning, which is considerably more expensive than the other benchmark functions we used. We show that for this problem using the surrogate offers a significant decrease in total run time compared to a GA using the true fitness function.

An application of a GA with Markov network surrogate to feature selection

Surrogate models of fitness have been presented as a way of reducing the number of fitness evaluations required by evolutionary algorithms. This is of particular interest with expensive fitness functions where the time taken for building the model is outweighed by the savings of using fewer function evaluations. In this article, we show how a Markov network model can be used as a surrogate fitness function for a genetic algorithm in a new algorithm called Markov Fitness Model Genetic Algorithm (MFM-GA). We thoroughly investigate its application to a fitness function for feature selection in Case-Based Reasoning (CBR), using a range of standard benchmarks from the CBR community. This fitness function requires considerable computation time to evaluate and we show that using the surrogate offers a significant decrease in total run-time compared to a GA using the true fitness function. This comes at the cost of a reduction in the global best fitness found. We demonstrate that the quality of the solutions obtained by MFM-GA improves significantly with model rebuilding. Comparisons with a classic GA, a GA using fitness inheritance and a selection of filter selection methods for CBR shows that MFM-GA provides a good trade-off between fitness quality and run-time.

An Island Model Genetic Algorithm for Bayesian network structure learning

Bayesian Networks (BNs) are graphical probabilistic models that represent relationships that may exist between variables of a dataset. BN can be applied to data in a variety of different ways. Yet, using a BN requires knowing its structure. BN structure learning represents a challenge as the number of possible structures is very large. Search and score approaches have been used to address the problem. One of them, a Genetic Algorithm based on the K2 search (K2GA) has shown that BNs can be learned from many datasets. However, the computational cost which is involved is high while structures obtained from benchmark data often exhibit significant differences from known correct structures. In this paper, we investigate the use of K2GA within an Island Model (IM) implementation and compare the quality of the BN structures obtained with those of the traditional K2GA. Experiments are run on five datasets created from BNs with known structures. Results show that the use of IM improves the quality of the structures obtained. BNs present better fitnesses, but also sets of edges more consistent with the known true structures. We conclude that migration between islands helps maintaining diversity within each population.

Privacy-preserving approach to bayesian network structure learning from distributed data

In many situations, data is scattered across different sites, making the modeling process difficult or sometimes impossible. Some applications could benefit from collaborations between organisations but data security or privacy policies often act as a barrier to data mining on such contexts. In this paper, we present a novel approach to learning Bayesian Networks (BN) structures from multiple datasets, based on the use of Ensembles and an Island Model Genetic Algorithm (IMGA). The proposed design ensures no data is shared during the process and can fit many applications.

RK-EDA: a novel random key based estimation of distribution algorithm.

The challenges of solving problems naturally represented as permutations by Estimation of Distribution Algorithms (EDAs) have been a recent focus of interest in the evolutionary computation community. One of the most common alternative representations for permutation based problems is the Random Key (RK), which enables the use of continuous approaches for this problem domain. However, the use of RK in EDAs have not produced competitive results to date and more recent research on permutation based EDAs have focused on creating superior algorithms with specially adapted representations. In this paper, we present RK-EDA; a novel RK based EDA that uses a cooling scheme to balance the exploration and exploitation of a search space by controlling the variance in its probabilistic model. Unlike the general performance of RK based EDAs, RK-EDA is actually competitive with the best EDAs on common permutation test problems: Flow Shop Scheduling, Linear Ordering, Quadratic Assignment, and Travelling Salesman Problems.

Factoradic Representation for Permutation Optimisation

It is known that different classes of permutation problems are more easily solved by selecting a suitable representation. In particular, permutation representations suitable for Estimation of Distribution algorithms (EDAs) are known to present several challenges. Therefore, it is of interest to investigate novel representations and their properties. In this paper, we present a study of the factoradic representation which offers new modelling insights through the use of three algorithmic frameworks, a Genetic Algorithm (GA) and two EDAs. Four classic permutation benchmark problems are used to evaluate the factoradic-based algorithms in comparison with published work with other representations. Our experiments demonstrate that the factoradic representation is a competitive approach to apply to permutation problems. EDAs and more specifically, univariate EDAs show the most robust performance on the benchmarks studied. The factoradic representation also leads to better performance than adaptations of EDAs for continuous spaces, overall similar performance to integer-based EDAs and occasionally matches results of specialised EDAs, justifying further study.

Competing mutating agents for bayesian network structure learning

Search and score techniques have been widely applied to the problem of learning Bayesian Networks (BNs) from data. Many implementations focus on finding an ordering of variables from which edges can be inferred. Although varying across data, most search spaces for such tasks exhibit many optima and plateaus. Such characteristics represent a trap for population-based algorithms as the diversity decreases and the search converges prematurely. In this paper, we study the impact of a distance mutation operator and propose a novel method using a population of agents that mutate their solutions according to their respective positions in the population. Experiments on a set of benchmark BNs confirm that diversity is maintained throughout the search. The proposed technique shows improvement on most of the datasets by obtaining BNs of similar of higher quality than those obtained by Genetic Algorithm methods.

Landscape analysis for hyperheuristic Bayesian Network structure learning on unseen problems

Bayesian network (BN) structure learning is an NP hard problem. Search and score algorithms are one of the main approaches proposed for learning BN structure from data. Previous research has shown that the relative performances of such algorithms are problem dependent and that fitness landscape analysis can be used to characterize the difficulty of the search for different scoring functions. In this paper, we construct a classifier based on fitness landscape analysis and receiver operating characteristic curves. The classifier labels search landscapes with the most suitable scoring function. We train the classifier on a number of standard benchmark functions. The classifier forms the basis for a selective hyperheuristic algorithm. This uses an initial landscape analysis stage to select a scoring function using the classifier. The hyperheuristic algorithm is tested on a distribution of unseen problems based on mutations of the standard benchmarks. Our results establish that the hyperheuristic performs better than a uniformly random scoring function selection approach that omit the landscape analysis stage. Therefore the effects on performance of problem-dependency can be significantly reduced.

BPGA-EDA for the multi-mode resource constrained project scheduling problem

The Multi-mode Resource Constrained Project Scheduling Problem (MRCPSP) has been of research interest for over two decades. The problem is composed of two interacting sub problems: mode assignment and activity scheduling. These problems cannot be solved in isolation because of the interaction that exists between them. Many evolutionary algorithms have been applied to this problem most commonly the Genetic Algorithm (GA). It has been common practice to improve the performance of the GA with some local search techniques. The Bi-population Genetic Algorithm (BPGA) is one of the most competitive GAs for solving the MRCPSP. In this paper, we improve the BPGA by hybridising it with an Estimation of Distribution Algorithm that focuses on improving how modes are generated. We also suggest improvement to the existing experimental methodology.