Showing papers on "Interpretability published in 2012"

Using probabilistic estimation of expression residuals (PEER) to obtain increased power and interpretability of gene expression analyses

Abstract: We present PEER (probabilistic estimation of expression residuals), a software package implementing statistical models that improve the sensitivity and interpretability of genetic associations in population-scale expression data. This approach builds on factor analysis methods that infer broad variance components in the measurements. PEER takes as input transcript profiles and covariates from a set of individuals, and then outputs hidden factors that explain much of the expression variability. Optionally, these factors can be interpreted as pathway or transcription factor activations by providing prior information about which genes are involved in the pathway or targeted by the factor. The inferred factors are used in genetic association analyses. First, they are treated as additional covariates, and are included in the model to increase detection power for mapping expression traits. Second, they are analyzed as phenotypes themselves to understand the causes of global expression variability. PEER extends previous related surrogate variable models and can be implemented within hours on a desktop computer.

Making machine learning models interpretable

Abstract: Data of different levels of complexity and of ever growing diversity of characteristics are the raw materials that machine learning practitioners try to model using their wide palette of methods and tools. The obtained models are meant to be a synthetic representation of the available, observed data that captures some of their intrinsic regularities or patterns. Therefore, the use of machine learning techniques for data analysis can be understood as a problem of pattern recognition or, more informally, of knowledge discovery and data mining. There exists a gap, though, between data modeling and knowledge extraction. Models, de- pending on the machine learning techniques employed, can be described in diverse ways but, in order to consider that some knowledge has been achieved from their description, we must take into account the human cog- nitive factor that any knowledge extraction process entails. These models as such can be rendered powerless unless they can be interpreted ,a nd the process of human interpretation follows rules that go well beyond techni- cal prowess. For this reason, interpretability is a paramount quality that machine learning methods should aim to achieve if they are to be applied in practice. This paper is a brief introduction to the special session on interpretable models in machine learning, organized as part of the 20 th European Symposium on Artificial Neural Networks, Computational In- telligence and Machine Learning. It includes a discussion on the several works accepted for the session, with an overview of the context of wider research on interpretability of machine learning models.

Interpretability and importance of functionals in competing risks and multistate models

Abstract: The basic parameters in both survival analysis and more general multistate models, including the competing risks model and the illness–death model, are the transition hazards. It is often necessary to supplement the analysis of such models with other model parameters, which are all functionals of the transition hazards. Unfortunately, not all such functionals are equally meaningful in practical contexts, even though they may be mathematically well defined. We have found it useful to check whether the functionals satisfy three simple principles, which may be used as criteria for practical interpretability. Copyright © 2011 John Wiley & Sons, Ltd.

Hybrid system with genetic algorithm and artificial neural networks and its application to retail credit risk assessment

Abstract: The databases of the banks around the world have accumulated large quantities of information about clients and their financial and payment history. These databases can be used for the credit risk assessment, but they are commonly high dimensional. Irrelevant features in a training dataset may produce less accurate results of classification analysis. Data preprocessing is required to prepare the data for classification to increase the predictive accuracy. Feature selection is a preprocessing technique commonly used on high dimensional data and its purposes include reducing dimensionality, removing irrelevant and redundant features, facilitating data understanding, reducing the amount of data needed for learning, improving predictive accuracy of algorithms, and increasing interpretability of models. In this paper we investigate the extent to which the total data, owned by a bank, can be a good basis for predicting the borrower's ability to repay the loan on time. We propose a feature selection technique for finding an optimum feature subset that enhances the classification accuracy of neural network classifiers. Experiments were conducted on the credit dataset collected at a Croatian bank to assess the accuracy of our technique. We found that the hybrid system with genetic algorithm is competitive and can be used as feature selection technique to discover the most significant features in determining risk of default.

On the Interpretability of Law: Lessons from the Decoding of National Constitutions

Abstract: An implicit element of many theories of constitutional enforcement is the degree to which those subject to constitutional law can agree on what its provisions mean (call this constitutional interpretability). Unfortunately, there is little evidence on baseline levels of constitutional interpretability or the variance therein. This article seeks to fill this gap in the literature, by assessing the effect of contextual, textual and interpreter characteristics on the interpretability of constitutional documents. Constitutions are found to vary in their degree of interpretability. Surprisingly, however, the most important determinants of variance are not contextual (for example, era, language or culture), but textual. This result emphasizes the important role that constitutional drafters play in the implementation of their product.

Labeling Ecological Niche Models

Abstract: The ongoing biodiversity crisis is pushing ecologists and conservation biologists to develop models to foretell the effects of human-induced transformation of natural resources on the distribution of species, although ecology and biogeography still lacks a paradigmatic body of theory to fully understand the drivers of biodiversity patterns. Two decades of research on ecological niche models and species distributions have been characterized by technical development and discussions on a plethora of methods or algorithms to infer and predict species distributions. Here we suggest a metaphorical classification scheme for some of the most popular models based on their complexity, interpretability and suitability for specific applications in ecology and conservation biology. Our purpose is not to compare methods by their capacity to accurately predict the observed distribution of species, nor to criticize how they are commonly used in applied studies. Instead, we believe that a simple classification scheme can potentially highlight how some methods are more suited for specific applications in ecology and conservation biology. Envelope and distance-based models are grouped into the “fish bowl” category, for their transparency and simplicity. Statistical models are classified as “turbine” models, because of their hidden complexity and general applicability. Finally, machine-learning models are classified as “vault” models, for their high complexity and lack of interpretability of fit parameters. We conclude that the diversity of species distribution models used today is expected for a young research field, but the choice of modeling strategy depends on the purpose of the study. We provide some general guidelines for choosing models for studies of conservation planning and climate change mitigation and suggest models of intermediate complexity for conservation planning and forecast of climate change effects on biodiversity as they provide a good balance between interpretability, predictive power and robustness to model over-fit.

Leveraging high-level and low-level features for multimedia event detection

Abstract: This paper addresses the challenge of Multimedia Event Detection by proposing a novel method for high-level and low-level features fusion based on collective classification. Generally, the method consists of three steps: training a classifier from low-level features; encoding high-level features into graphs; and diffusing the scores on the established graph to obtain the final prediction. The final prediction is derived from multiple graphs each of which corresponds to a high-level feature. The paper investigates two graph construction methods using logarithmic and exponential loss functions, respectively and two collective classification algorithms, i.e. Gibbs sampling and Markov random walk. The theoretical analysis demonstrates that the proposed method converges and is computationally scalable and the empirical analysis on TRECVID 2011 Multimedia Event Detection dataset validates its outstanding performance compared to state-of-the-art methods, with an added benefit of interpretability.

Differential Privacy and Statistical Disclosure Risk Measures: An Investigation with Binary Synthetic Data

Abstract: We compare the disclosure risk criterion of e-differential privacy with a criterion based on probabilities that intruders uncover actual values given the released data. To do so, we generate fully synthetic data that satisfy e-differential privacy at different levels of e, make assumptions about the information available to intruders, and compute posterior probabilities of uncovering true values. The simulation results suggest that the two paradigms are not easily reconciled, since differential privacy is agnostic to the specific values in the observed data whereas probabilistic disclosure risk measures depend greatly on them. The results also suggest, perhaps surprisingly, that probabilistic disclosure risk measures can be small even when e is large. Motivated by these findings, we present an alternative disclosure risk assessment approach that integrates some of the strong confidentiality protection features in e-differential privacy with the interpretability and data-specific nature of probabilistic disclosure risk measures.

Learning of Fuzzy Cognitive Maps Using Density Estimate

Abstract: Fuzzy cognitive maps (FCMs) are convenient and widely used architectures for modeling dynamic systems, which are characterized by a great deal of flexibility and adaptability. Several recent works in this area concern strategies for the development of FCMs. Although a few fully automated algorithms to learn these models from data have been introduced, the resulting FCMs are structurally considerably different than those developed by human experts. In particular, maps that were learned from data are much denser (with the density over 90% versus about 40% density of maps developed by humans). The sparseness of the maps is associated with their interpretability: the smaller the number of connections is, the higher is the transparency of the map. To this end, a novel learning approach, sparse real-coded genetic algorithms (SRCGAs), to learn FCMs is proposed. The method utilizes a density parameter to guide the learning toward a formation of maps of a certain predefined density. Comparative tests carried out for both synthetic and real-world data demonstrate that, given a suitable density estimate, the SRCGA method significantly outperforms other state-of-the-art learning methods. When the density estimate is unknown, the new method can be used in an automated fashion using a default value, and it is still able to produce models whose performance exceeds or is equal to the performance of the models generated by other methods.

Reconciling performance and interpretability in customer churn prediction using ensemble learning based on generalized additive models

Abstract: To build a successful customer churn prediction model, a classification algorithm should be chosen that fulfills two requirements: strong classification performance and a high level of model interpretability. In recent literature, ensemble classifiers have demonstrated superior performance in a multitude of applications and data mining contests. However, due to an increased complexity they result in models that are often difficult to interpret. In this study, GAMensPlus, an ensemble classifier based upon generalized additive models (GAMs), in which both performance and interpretability are reconciled, is presented and evaluated in a context of churn prediction modeling. The recently proposed GAMens, based upon Bagging, the Random Subspace Method and semi-parametric GAMs as constituent classifiers, is extended to include two instruments for model interpretability: generalized feature importance scores, and bootstrap confidence bands for smoothing splines. In an experimental comparison on data sets of six real-life churn prediction projects, the competitive performance of the proposed algorithm over a set of well-known benchmark algorithms is demonstrated in terms of four evaluation metrics. Further, the ability of the technique to deliver valuable insight into the drivers of customer churn is illustrated in a case study on data from a European bank. Firstly, it is shown how the generalized feature importance scores allow the analyst to identify the relative importance of churn predictors in function of the criterion that is used to measure the quality of the model predictions. Secondly, the ability of GAMensPlus to identify nonlinear relationships between predictors and churn probabilities is demonstrated.

Representative Factor Generation for the Interactive Visual Analysis of High-Dimensional Data

Abstract: Datasets with a large number of dimensions per data item (hundreds or more) are challenging both for computational and visual analysis. Moreover, these dimensions have different characteristics and relations that result in sub-groups and/or hierarchies over the set of dimensions. Such structures lead to heterogeneity within the dimensions. Although the consideration of these structures is crucial for the analysis, most of the available analysis methods discard the heterogeneous relations among the dimensions. In this paper, we introduce the construction and utilization of representative factors for the interactive visual analysis of structures in high-dimensional datasets. First, we present a selection of methods to investigate the sub-groups in the dimension set and associate representative factors with those groups of dimensions. Second, we introduce how these factors are included in the interactive visual analysis cycle together with the original dimensions. We then provide the steps of an analytical procedure that iteratively analyzes the datasets through the use of representative factors. We discuss how our methods improve the reliability and interpretability of the analysis process by enabling more informed selections of computational tools. Finally, we demonstrate our techniques on the analysis of brain imaging study results that are performed over a large group of subjects.

Algorithms for mining meaningful roles

Abstract: Role-based access control (RBAC) offers significant advantages over lower-level access control policy representations, such as access control lists (ACLs). However, the effort required for a large organization to migrate from ACLs to RBAC can be a significant obstacle to adoption of RBAC. Role mining algorithms partially automate the construction of an RBAC policy from an ACL policy and possibly other information, such as user attributes. These algorithms can significantly reduce the cost of migration to RBAC.This paper proposes new algorithms for role mining. The algorithms can easily be used to optimize a variety of policy quality metrics, including metrics based on policy size, metrics based on interpretability of the roles with respect to user attribute data, and compound metrics that consider size and interpretability. The algorithms all begin with a phase that constructs a set of candidate roles. We consider two strategies for the second phase: start with an empty policy and repeatedly add candidate roles, or start with the entire set of candidate roles and repeatedly remove roles. In experiments with publicly available access control policies, we find that the elimination approach produces better results, and that, for a policy quality metric that reflects size and interpretability, our elimination algorithm achieves significantly better results than previous work.

Factorial LDA: Sparse Multi-Dimensional Text Models

Abstract: Latent variable models can be enriched with a multi-dimensional structure to consider the many latent factors in a text corpus, such as topic, author perspective and sentiment. We introduce factorial LDA, a multi-dimensional model in which a document is influenced by K different factors, and each word token depends on a K-dimensional vector of latent variables. Our model incorporates structured word priors and learns a sparse product of factors. Experiments on research abstracts show that our model can learn latent factors such as research topic, scientific discipline, and focus (methods vs. applications). Our modeling improvements reduce test perplexity and improve human interpretability of the discovered factors.

A model for the integration of e-financial services questionnaires with SERVQUAL scales under fuzzy linguistic modeling

Abstract: Although it is habitual to measure human perceptions with quite accurate instruments, perceptions are characterized by uncertainty and fuzziness. Furthermore, variations in individual perceptions and personality mean that the same words can indicate very different perceptions. In this context, the fuzzy linguistic approach seems to be an appropriate framework for modeling information. In this paper we explore the problem of integrating semantically heterogeneous data (natural language included) from various websites with opinions about e-financial services. We develop an extension of the fuzzy model based on semantic translation (FMST) under the perspective of the service quality (SERVQUAL) stream of research. The model permits us to obtain a more precise representation of the opinions using each type of customers. By integrating all customers into different subsets, a financial entity can easily analyze the SERVQUAL characteristics over time or other dimensions owing to the easy linguistic interpretability and high precision of the results of the model.

Improving construct validity yields better models of systematic inquiry, even with less information

Abstract: Data-mined models often achieve good predictive power, but sometimes at the cost of interpretability. We investigate here if selecting features to increase a model's construct validity and interpretability also can improve the model's ability to predict the desired constructs. We do this by taking existing models and reducing the feature set to increase construct validity. We then compare the existing and new models on their predictive capabilities within a held-out test set in two ways. First, we analyze the models' overall predictive performance. Second, we determine how much student interaction data is necessary to make accurate predictions. We find that these reduced models with higher construct validity not only achieve better agreement overall, but also achieve better prediction with less data. This work is conducted in the context of developing models to assess students' inquiry skill at designing controlled experiments and testing stated hypotheses within a science inquiry microworld.

Selecting Diverse Features via Spectral Regularization

Abstract: We study the problem of diverse feature selection in linear regression: selecting a small subset of diverse features that can predict a given objective. Diversity is useful for several reasons such as interpretability, robustness to noise, etc. We propose several spectral regularizers that capture a notion of diversity of features and show that these are all submodular set functions. These regularizers, when added to the objective function for linear regression, result in approximately submodular functions, which can then be maximized by efficient greedy and local search algorithms, with provable guarantees. We compare our algorithms to traditional greedy and l1-regularization schemes and show that we obtain a more diverse set of features that result in the regression problem being stable under perturbations.

Recognizing uncertainty increases robustness and reproducibility of mass spectrometry-based protein inferences.

Abstract: Parsimony and protein grouping are widely employed to enforce economy in the number of identified proteins, with the goal of increasing the quality and reliability of protein identifications; however, in a counterintuitive manner, parsimony and protein grouping may actually decrease the reproducibility and interpretability of protein identifications. We present a simple illustration demonstrating ways in which parsimony and protein grouping may lower the reproducibility or interpretability of results. We then provide an example of a data set where a probabilistic method increases the reproducibility and interpretability of identifications made on replicate analyses of Human Du145 prostate cancer cell lines.

A Mathematical Model for Interpretable Clinical Decision Support with Applications in Gynecology

Abstract: Background: Over time, methods for the development of clinical decision support (CDS) systems have evolved from interpretable and easy-to-use scoring systems to very complex and non-interpretable mathematical models. In order to accomplish effective decision support, CDS systems should provide information on how the model arrives at a certain decision. To address the issue of incompatibility between performance, interpretability and applicability of CDS systems, this paper proposes an innovative model structure, automatically leading to interpretable and easily applicable models. The resulting models can be used to guide clinicians when deciding upon the appropriate treatment, estimating patient-specific risks and to improve communication with patients. Methods and Findings: We propose the interval coded scoring (ICS) system, which imposes that the effect of each variable on the estimated risk is constant within consecutive intervals. The number and position of the intervals are automatically obtained by solving an optimization problem, which additionally performs variable selection. The resulting model can be visualised by means of appealing scoring tables and color bars. ICS models can be used within software packages, in smartphone applications, or on paper, which is particularly useful for bedside medicine and home-monitoring. The ICS approach is illustrated on two gynecological problems: diagnosis of malignancy of ovarian tumors using a dataset containing 3,511 patients, and prediction of first trimester viability of pregnancies using a dataset of 1,435 women. Comparison of the performance of the ICS approach with a range of prediction models proposed in the literature illustrates the ability of ICS to combine optimal performance with the interpretability of simple scoring systems. Conclusions: The ICS approach can improve patient-clinician communication and will provide additional insights in the importance and influence of available variables. Future challenges include extensions of the proposed methodology towards automated detection of interaction effects, multi-class decision support systems, prognosis and high-dimensional data.

A Review on the Interpretability-Accuracy Trade-Off in Evolutionary Multi-Objective Fuzzy Systems (EMOFS)

Abstract: Interpretability and accuracy are two important features of fuzzy systems which are conflicting in their nature. One can be improved at the cost of the other and this situation is identified as “Interpretability-Accuracy Trade-Off”. To deal with this trade-off Multi-Objective Evolutionary Algorithms (MOEA) are frequently applied in the design of fuzzy systems. Several novel MOEA have been proposed and invented for this purpose, more specifically, Non-Dominated Sorting Genetic Algorithms (NSGA-II), Strength Pareto Evolutionary Algorithm 2 (SPEA2), Fuzzy Genetics-Based Machine Learning (FGBML), (2 + 2) Pareto Archived Evolutionary Strategy ((2 + 2) PAES), (2 + 2) Memetic- Pareto Archived Evolutionary Strategy ((2 + 2) M-PAES), etc. This paper introduces and reviews the approaches to the issue of developing fuzzy systems using Evolutionary Multi-Objective Optimization (EMO) algorithms considering ‘Interpretability-Accuracy Trade-off’ and mainly focusing on the work in the last decade. Different research issues and challenges are also discussed.

Accuracy vs. interpretability of fuzzy rule-based classifiers: an evolutionary approach

Abstract: The paper presents a generalization of the Pittsburgh approach to learn fuzzy classification rules from data. The proposed approach allows us to obtain a fuzzy rule-based system with a predefined level of compromise between its accuracy and interpretability (transparency). The application of the proposed technique to design the fuzzy rule-based classifier for the well known benchmark data sets (Dermatology and Wine) available from the http://archive.ics.uci.edu/ml is presented. A comparative analysis with several alternative (fuzzy) rule-based classification techniques has also been carried out.

Feature selection and granularity learning in genetic fuzzy rule-based classification systems for highly imbalanced data-sets

Abstract: This paper proposes a Genetic Algorithm for jointly performing a feature selection and granularity learning for Fuzzy Rule-Based Classification Systems in the scenario of highly imbalanced data-sets. We refer to imbalanced data-sets when the class distribution is not uniform, a situation that it is present in many real application areas. The aim of this work is to get more compact models by selecting the adequate variables and adapting the number of fuzzy labels for each problem, improving the interpretability of the model. The experimental analysis is carried out over a wide range of highly imbalanced data-sets and uses the statistical tests suggested in the specialized literature.

Interpretability of SAR/QSAR Models of any Complexity by Atomic Contributions.

Abstract: This communication concerns a new approach to interpret a classification or regression structure-property model of any complexity based on fragment descriptors. The idea is to evaluate contributions of each atom into modelled property (y) and to visualize them using a colour code. This results in the coloured image of 2D molecular structure in which molecular fragments responsible for increase or decrease of y could be easily recognized. The method has been illustrated on examples of the regression model for aqueous solubility and classification models for Acetylcholinesterase inhibitors and organoleptic activities. Prediction efficiency and interpretability of QSAR models represent two different objectives which are very difficult to achieve simultaneously. Indeed, the models build with modern machine-learning methods, like neural networks or support vector machines, or consensus solutions based on ensembles of different models are widely used in QSAR as predictive tools. On the other hand, their chemical sense interpretation is difficult, if yet possible. A notable exception represent group/atomic contribution methods in which the property value y predicted for a given molecule is divided into a set of contributions of atoms or substructures . A colour code associated with these contributions is a simple way to identify “useful” or “useless” structural motifs. Recently such a coloration has been suggested for some particular cases: random forest models based on simplex descriptors and for the Na ve Bayes models based on the MNA descriptors integrated in the PASS software . Franke et al. reported pharmacophore point visualization approach based on a perturbation and applied this to pharmacophore fingerprints. It assesses the relative importance of different pharmacophore features rather than that of explicit structural motifs. Baskin et al. proposed a differential approach for interpretation of nonlinear neural network QSAR models. Carlsson et al. proposed an assessment of one most important descriptor based on gradient estimations. In particular case of Faulon’s fragment descriptors , a substructure identified as the “most important” can be highlighted. In this paper, we present a fragment based QSAR partial derivatives approach which could be considered as further extension of ideas reported in . It aims to obtain group or atomic contributions of regression, classification or clustering models of any complexity. Another saying, QSAR partial derivatives assess an impact of local perturbations of fragment descriptors on the outcome of the model. As in references , our study is restricted to fragment descriptors. Each fragment represents a subgraph of a molecular graph, whereas its occurrence xi is considered as the descriptor’s value. There exists a huge variety of fragment descriptors . Here, we used the ISIDA fragment descriptors representing either shortest paths between two selected atoms and different size atom centred fragments (see Computational Methods section). For linear QSAR model,

Improved Structure Optimization for Fuzzy-Neural Networks

Abstract: Fuzzy-neural-network-based inference systems are well-known universal approximators which can produce linguistically interpretable results. Unfortunately, their dimensionality can be extremely high due to an excessive number of inputs and rules, which raises the need for overall structure optimization. In the literature, various input selection methods are available, but they are applied separately from rule selection, often without considering the fuzzy structure. This paper proposes an integrated framework to optimize the number of inputs and the number of rules simultaneously. First, a method is developed to select the most significant rules, along with a refinement stage to remove unnecessary correlations. An improved information criterion is then proposed to find an appropriate number of inputs and rules to include in the model, leading to a balanced tradeoff between interpretability and accuracy. Simulation results confirm the efficacy of the proposed method.

Watch this: A taxonomy for dynamic data visualization

Abstract: Visualizations embody design choices about data access, data transformation, visual representation, and interaction. To interpret a static visualization, a person must identify the correspondences between the visual representation and the underlying data. These correspondences become moving targets when a visualization is dynamic. Dynamics may be introduced in a visualization at any point in the analysis and visualization process. For example, the data itself may be streaming, shifting subsets may be selected, visual representations may be animated, and interaction may modify presentation. In this paper, we focus on the impact of dynamic data. We present a taxonomy and conceptual framework for understanding how data changes influence the interpretability of visual representations. Visualization techniques are organized into categories at various levels of abstraction. The salient characteristics of each category and task suitability are discussed through examples from the scientific literature and popular practices. Examining the implications of dynamically updating visualizations warrants attention because it directly impacts the interpretability (and thus utility) of visualizations. The taxonomy presented provides a reference point for further exploration of dynamic data visualization techniques.

Policy Search in a Space of Simple Closed-form Formulas: Towards Interpretability of Reinforcement Learning

Abstract: In this paper, we address the problem of computing interpretable solutions to reinforcement learning (RL) problems. To this end, we propose a search algorithm over a space of simple closed-form formulas that are used to rank actions. We formalize the search for a high-performance policy as a multi-armed bandit problem where each arm corresponds to a candidate policy canonically represented by its shortest formula-based representation. Experiments, conducted on standard benchmarks, show that this approach manages to determine both efficient and interpretable solutions.