Presentations, Tutorials and Archived Program

In informal logic and argumentation theory the use of graphs to lay out arguments has a long history, beginning arguably with Wigmore’s maps of legal arguments. They have become popular in recent decades, particularly with computer software which makes their creation and management easier. Advocates of argument map software recognize some relation to Bayesian networks (BNs), but claim their tools are easier to understand and use. In response to the obvious value of representing the degrees of strength of relationships between premises and conclusions that BNs provide, they have started using naive Bayes models for analysing arguments. We analyse this in the context of a few simple arguments, including breast cancer diagnosis and the authorship of the Shakespeare corpus, showing that in these and many other cases naive Bayes collapses under the weight of the argument. We review a number of key issues in analysing arguments with BNs, including Jeffrey conditionalization and representing competing hypotheses, and provide reasons for preferring BNs for argumentation to either naive Bayes or qualitative maps. more >

Carbon farming programs typically aim to maximise landholder participation rates to achieve desired environmental outcomes. This is also critical for programs aiming to tackle both climate change and biodiversity loss simultaneously, as landholder participation in those schemes directly determines the level of carbon sequestered and the potential biodiversity gains. In this study we developed a Bayesian Belief Network (BBN) of landholder participation in biodiverse carbon planting schemes to determine factors most likely to influence program participation. Based on a review of the literature, we developed a conceptual model. The model was refined through interviews with participating landholders, other key stakeholders and, finally, parameterised using expert-elicited knowledge. Our results indicate that participation rates are most influenced by program attractiveness and the identified values of co-benefits (such as biodiversity conservation) rather than financial incentives. Scenario evaluation revealed that providing a combination of biodiversity incentives with more flexible permanence options could increase the program adoption rate. Stacking/ bundling credits combined with contract agreements may also increase the participation rate. These findings can assist policy development by focusing on the aspects of policy design most likely to increase participation rates. more >

In order to provide instruction on what operational conditions need to be set in a validation protocol for membrane bioreactors (MBRs) to ensure appropriately conservative performance testing, factors affecting log removal values (LRV) of microbial indicators by MBRs need to be identified. Microbial indicators of concern in this study include FRNA bacteriophage and somatic coliphage as indicators for virus, Escherichia coli (E. coli) as an indicator for fecal bacteria, and clostridium perfringens as an indicator for protozoa. A large sampling programme was conducted covering 5 full-scale MBRs with various design and operational conditions across Australia. During the full-scale site investigations, the density of microbial indicators in influent, mixed liquor and permeate of the MBRs were assayed, and corresponding sets of operational parameters were also collected. A Bayesian Belief Network (BBN) was constructed from a dataset of 312 cases from which 80% of the data was used to learn the structure and estimate the parameters, and 20% of the data was used for final testing. The network structure was defined by automated structure learning using the bnlearn 3.8.1 package in R. Domain expert knowledge was also introduced in the structure learning in the form of whitelist and backlist. The variables were discretised by density approximation and validation was performed in BayesiaLab 5.4.3. The ROC score was from 70 to 92% for the indicators. The BBN was then used to identify factors influencing LRV of microbial indicators by MBRs by using "Influence path to target" and "Correlation with target node" functions in the BayesiaLab 5.4.3. The results show that membrane age, membrane pore size, permeability and solid retention time are important factors affecting LRV of microbial indicators by MBRs. To the best of the authors’ knowledge, this was the first time that BBN was used for such application. The results have been used to form the basis of the validation protocol for MBRs in Australian water recycling schemes that will ensure consistent accreditation and appreciation of risk. more >

Ultrafiltration is one the most common unit operations used in the tertiary treatment of wastewater. From a risk management viewpoint, it is considered an important step because it can be credited with reduction values for pathogens including viruses. Monitoring of ultrafiltration membranes includes measurement of indicator parameters such as turbidity, direct integrity test, permeability and challenge testing. By observing these parameters an operator should be able to infer the condition of the membrane and determine if corrective actions are required. However, there is a lack of available models able to deal with multiple indicators in a probabilistic fashion. Bayesian networks show great potential for addressing this where inference of a condition or state is required from multiple indicator parameters.

The aim of our study was to investigate the use of Bayesian networks for monitoring the condition of ultrafiltration membrane module based on permeability, direct integrity tests and challenge testing using two microbial indicators (one spore former bacterium and one virus). The network was constructed from expert knowledge in BayesiaLab 5.4.3 considering five candidate variables. The variables were discretised using equal frequency method in three states for all variables except from the microbial indicators node which only had two states. Prospective parameters were determined by EM algorithm and net evaluation was performed using 5-fold cross validation. Our analysis indicated that the model had an overall accuracy of 80% with and AUC score greater than 0.90 for five different seeds. The contingency table fit which measures the quality of the representation of the joint probability distribution with respect to the fully connected network was 83%. These results indicated that using a Bayes Net model membrane condition could be reliably inferred from any of the three indicators and that for the bacterial indicator a damaged membrane had a greater impact than for virus. This can be explained by the smaller size of the virus compared to the nominal pore size of the membrane. more >

The causal discovery of Bayesian networks with the presence of latent variables is a popular topic in artificial intelligence, as sources and volumes of data continue to grow with the popularity of Bayesian modelling methods. Causal discovery is based upon searching the space of causal models for those which can best explain a pattern of probabilistic dependencies shown in the data. Frequently, however, some of those dependencies are generated by causal structures involving variables which have not been measured, i.e., latent variables. Some such patterns of dependency “reveal" themselves, in that no model based solely upon the observed variables can explain them as well as a model using a latent variable. Here we present an algorithm for finding such patterns systematically, so that they may be applied in latent variable discovery in a more rigorous fashion. more >

Learning a causal network over a set of variables from data is NP-hard and is exponential in the number of variables. State-of-the-art causal discovery algorithms do not scale up well in high-dimensional datasets. Recently, a technique called Markov blanket causal discovery was proposed to increase efficiencies of state-of-the-art algorithms and hence scale up to larger networks. This presentation provides an introduction to causal discovery problems using the Markov blanket technique and minimum message length (MML) to search for the most optimal causal network of a given dataset. more >

Volcanic eruptions are preceded by magma rising to the surface and on its way interacting with surrounding rocks and fluids. These interactions lead to earthquakes, the release of magmatic gases, and surface deformations. All of these are observed in volcano monitoring and interpreted as precursors to a volcanic eruption.

In New Zealand, GeoNet is the official source of geological hazard information. GeoNet continuously monitors New Zealand’s active volcanoes. The volcano monitoring team regularly estimates probabilities of an up-coming eruption for active volcanoes. However, the team currently has no quantitative models to estimate the probability of eruption.

For more than 10 years, the scientific literature has proposed the use of Bayesian Networks as a framework for combining different volcanic observations and linking them to the underlying driving processes to assist in estimating eruption probabilities. However, a recent Science paper on Monitoring Volcanoes pointed out the gap between research developments and their implementation as operational forecasting tools.

We aim to bridge this gap. Working with members of the GeoNet volcano monitoring team, we have adapted the published Bayesian Network structure from a volcano in Guadeloupe to the New Zealand White Island volcano. We will hold a workshop using structured expert elicitation based on the Classical Cooke method to quantify the conditional probability tables of our Bayesian Network. We have invited key members of the volcano monitoring team and experts from three New Zealand universities to participate in the workshop. The workshop further provides the opportunity to investigate what makes a good calibration question for weighting expert responses in the context of a Bayesian Network. On this we collaborate with the University of Bristol, UK.

This presentation provides an overview of our project with details on our White Island Bayesian Network and the method for structured expert elicitation. more >

Many local governments and network utilities, such as water providers, face the highly complex process of preparing long term capital works programs and associated financial models to deliver infrastructure to service the future urban growth in their regions. These capital works usually span 20 years or more and the delivery costs typically run into the billions.

The optimal sequencing of projects within these capital works programs can result in significant long term savings to these organisations. To achieve this requires reliable and accurate urban growth models.

The sourcing of credible data and generation of these urban growth projection models is a difficult task to accomplish. It is very time consuming, utilises broad assumptions that often results in models that are inaccurate and prone to human error.

Sizztech has developed a geospatial urban growth modelling tool that overcomes these challenges. At the core of this automated tool is Bayesian Network technology. The modeller utilises Bayesian Network models to provide a predictive algorithm that determines when land parcels will be developed.

The presentation will start with an overview on constructing an urban growth model. The tool’s urban growth modelling functions and Bayesian Networks integration will be discussed. The structure of the Bayesian Network utilised by the tool will be examined. Finally, the presentation will review the tool's methodology to source and deliver the factors that influence the Bayesian Network model. more >

Land surveying offers a rich area of problems. These problems involve issues such as multiple sources of information, constraints, measurement errors, different precision of measurements, multiple measurements, etc. However, land surveying is a field that has been studied for many years, and over these years surveyors have developed techniques to deal with these issues. So, this provides us with an opportunity to compare different approaches for solving these problems, and to learn from the surveyors techniques ways to deal with these issues in Bayesian Network models.

In this talk we will show how several of these problems can be represented by Multivariate Gaussian Bayesian Networks. We will give details of how the message passing algorithm in Gaussian Bayesian Networks can be extended to Multivariate Gaussian Networks. We will illustrate how some Gaussian Bayesian Networks can be simplified if we use Multivariate Gaussian Bayesian Networks. Finally, we will compare the results given by using message passing in these networks with the results given by the techniques surveyors have developed. more >

Structural Equation Modeling (SEM) is a statistical technique for testing and estimating causal relations using a combination of statistical data and qualitative causal assumptions. What we call Probabilistic Structural Equation Models (PSEMs) in BayesiaLab are conceptually similar to traditional SEMs. However, PSEMs are based on a Bayesian network structure as opposed to a series of equations.
In general, specifying and estimating a traditional SEM requires a high degree of statistical expertise. Additionally, the multitude of manual steps involved can make the entire SEM workflow extremely time-consuming. The PSEM workflow in BayesiaLab, on the other hand, is accessible to non-statistician subject matter experts. Perhaps more importantly, it can be faster by several orders of magnitude. Finally, once a PSEM is validated, it can be utilized like any other Bayesian network. This means that the full array of analysis, simulation, and optimization tools is available to leverage the knowledge represented in the PSEM.

We demonstrate a prototypical PSEM application: key drivers analysis and product optimization based on consumer survey data. We examine how consumers perceive product attributes, and how these perceptions relate to the consumers' purchase intent for specific products.
Given the inherent uncertainty of survey data, we also wish to identify higher-level variables, i.e. "latent" variables that represent concepts, which are not directly measured in the survey. We do so by analyzing the relationships between the so-called "manifest" variables, i.e. variables that are directly measured in the survey. Including such concepts helps in building more stable and reliable models than what would be possible using manifest variables only.
Our overall objective is making surveys clearer to interpret by researchers and making them "actionable" for managerial decision makers. The ultimate goal is to use the generated PSEM for prioritizing marketing and product initiatives to maximize purchase intent. more >

This talk will discuss applications of Bayesian methods to complex large-scale problems in bioinformatics and ecology. A first application involves identifying new genes and new classes of functional elements in genomes, by segmenting whole genome alignments and simultaneously classifying segments into putative functional classes. A second application involves modelling the spread of the Brisbane fire ant invasion. Both applications involve large data sets and complex hierarchical Bayesian models. more >

Phyllospora comosa is a key habitat forming alga endemic to south eastern Australia. Over the last few decades it has disappeared in the urban Sydney area as a result of anthropogenic inputs. As climate change and urbanization continue to alter the ocean environment, the future survival of this species is of concern. A Bayesian belief network is being built to model coastal ocean conditions in order to better understand what influence future changes will have on Phyllospora. This model uses data from a wide variety of published work as well as lab and field experiments. The model also compares the potential for the success of future mitigation options.

(A second model is also in the works looking at the impact on the associated invertebrate community. Final abstract on what I will present at the conference will be sent through within the next month) more >

Even-though the last years have seen a "blossoming" of initiatives and meta-analyses aiming to clarify the link between biodiversity-ecosystem functions-ecosystem services, there is still a need for the integration of research. Few attempts have merged existing knowledge and make it useful for decision-makers. This contribution aims to show new ways of collaboratively representing and modelling these linkages to enable decision-making based on current available knowledge. This is managed through: (i) conducting a systematic-review with contributions describing the links between biodiversity-ecosystem functions-ecosystem services- human well-being; (ii) integrating the information on interactions in a conceptual mind-map; (iii) translating an example of some of these interactions to a Bayesian Network in an effort to demonstrate how already existing and future information could be used and represented when data uncertainty exists. Bayesian Networks allow for a better management of the uncertainty involved in the representation of complex models as well as providing the possibility of creating future scenarios where assumptions might be tested. These models can improve as new data and evidences become available, making predictions more robust through time as more evidence becomes available. Thus, this contribution shows how available knowledge and data can be collated to improve our understanding of complex issues. more >

Fire is a key ecological disturbance, shaping biome distribution and community composition globally and affecting faunal distributions via several different pathways. Fire may direct kill individuals, but also causes vegetation change and so indirectly influences the availability of food and shelter. Importantly, these fire-effects occur within a broader environmental context: conditions such as vegetation type and moisture can influence both fire and species distributions. Thus faunal responses to fire often vary between fire events or geographic regions, and time-since-fire is generally a poor predictor of species distributions at a landscape-scale. We investigated whether a Bayesian Network approach could provide an effective way of modelling these interdependent relationships, as regression models are constrained by the need to avoid correlated predictor variables and over-parametrisation. We used motion-sensing cameras to collect native mammal and invasive predator occurrence data at 113 sites across 56,000 ha of continuous eucalypt forest in the Otway Ranges, south-eastern Australia. There was a rainfall and elevation gradient across the study region, with the north-east being substantially flatter, lower and drier than the south-west; time-since-fire ranged from 6 months to 74 years. We developed a conceptual model of the system based on expert opinion and then used Uninet to populate this model as a non-parametric Bayesian Network from our field data. Species occurrences fell into two broad groups, with animals such as swamp wallabies Wallabia bicolor preferring drier vegetation types, less complex habitat and younger forest, but bush rats Rattus fuscipes and long-nosed bandicoots Perameles nasuta being more likely to occur at long-unburnt sites with wet vegetation and high habitat complexity. Fire influenced fauna directly and also through its effects on habitat complexity, woody debris availability and predator occurrence. Our study demonstrates that non-parametric Bayesian Networks are an effective technique for explicitly modelling the complex and context-dependent influence of fire history on faunal distributions. more >

Quantifying the key drivers of fire regimes is fundamental to determining the ability of management to manipulate the risk posed by wildfire to people and property. Annual area burned by wildfire is driven by four hypothetical switches – biomass, fuel moisture, fire weather and ignitions (Archibald et al. 2009; Bradstock 2010). Each of these is, in turn, influenced by a variety of anthropogenic and natural factors, only some of which can be altered by fire management actions. Many of these factors are not independent, having direct or indirect influences on each other.

Traditional statistical methods are incapable of adequately capturing this complexity. Bayesian networks (BNs) provide a methodological framework for dealing with this problem.

In this study, we derive a Bayesian Network for southern Australia to examine the relative importance of the four switches in southern Australia. Specifically, we ask two questions:
1. Which switch or switches have the greatest influence on annual area burned?
2. To what extent can humans influence the annual area burned and therefore alter risk?

The BN model was developed using published relationships and where necessary expert elicitation. Data for the model was collected for 60 000 5km square grid cells in southern Australia. Our approach provides a means for assessing the full complexity of the fire environment relationship. The model structure is readily transferred to other fire prone regions and could be used for a global assessment of the drivers of fire regimes. more >

Insect development and activity are weather dependent processes. Flight behaviour, in particular, is strongly mediated by meteorological conditions and the form of this relationship is normally species specific. To predict the activity of forest insects, their flight patterns, and the subsequent likelihood that they may establish a colony on a recently harvested log during a particular time period, requires an understanding of the relationship between key meteorological conditions and flight behaviour. We test key parameters (temperature, humidity, wind speed, rainfall) and their influence on the flight abilities of key forest insect pests (Hylurgus ligniperda, Hylastes ater, and Arhopalus ferus) over a twelve week period at a temporal resolution of 1 hour. We will demonstrate a Bayesian network developed to predict the likelihood of flight activity on the basis of available weather forecast data. We structured Bayesian network models from field data using a TAN naive Bayes approach, and parameterised the probability tables by using case-file training algorithms. Best-performing models were calibrated against the training data with 5.19% overall confusion error, broken down as 10.8% Type I (false positive) error and 4.1% Type II (false negative) error. We discuss some of the theoretical challenges of using Bayesian networks to analyse flight activity data. more >

The New Zealand Forest Owners Association (NZFOA), in conjunction with the Ministry for Primary Industries (MPI), is in the process of upgrading their current Forest Health Surveillance system to provide better early warning of invasions from a variety of pests that could do harm to New Zealand's plantation forests and, in turn, economy. As a part of this project, we have developed a model of pest exposure, which spans the export of a pest from a source country to its escape at any possible location within New Zealand. The model is capable of modelling the exposure of a variety of pests across a number of items (including cargo, vehicles and passengers) that can enter the country.

While still in the prototype stage, the BN has already assisted in clarifying the reasoning involved in assessing import risk, highlighted areas of focus for data collection, identified key aspects of the problem structure and provided a framework for dealing with highly uncertain or missing information on pests and pathway factors. more >

The Pilbara region of Western Australia has 598 islands, many of which have not been properly surveyed. Many of the islands are the last refuges for threatened and endemic species facing numerous threatening processes on the mainland. The islands are also important sites for recreation, cultural activities, and industrial development, mainly for oil and gas extraction. Quarantine and surveillance are arguably two of the more effective tools against the spread of invasive species to islands. Strategies for quarantine and surveillance rely on risk assessment. Effective quarantine would be focused on the most likely source populations of invasive species, and effective surveillance would be designed to cover higher-risk islands more often. Risk assessments are complicated, however, by lack of information on source populations and the diverse vectors by which invasive species reach islands. We used Bayesian Networks (BNs) to assess each island's risk from establishment of invasive species to prioritize quarantine and surveillance activities on the Pilbara Islands. The BNs were based on collated data on island desirability for visitation, recreational visitor load, infrastructure, habitat mapping, and animal dispersal pathways via human movement, wind, ocean currents, and flooded river plumes. Where variables could not be informed by prior scientific research we asked island managers to supply data via expert elicitation. Ultimately, the biosecurity risk profiles generated by the BNs will be incorporated into new decision-support software that is being designed to prioritize a wide variety of island conservation actions. more >

The ease with which Bayesian network (BN) models can be built had led to some common problems in their construction and interpretation. In this keynote, I address common quandaries and their solutions in BN modeling. Common problems of BN model construction include: assuming that every conditional probability table (CPT) must span [0,100]; use of vague node names and unmeasurable node states; too many parent nodes and little consideration for latent variables; ignoring outlier CPT values; not linking correlated covariates; no peer review of expert-structured BNs; no tests of model calibration or validation; confusing calibration with validation; and "holes" left in CPTs from machine-learning algorithms. Common problems of BN model interpretation include: unclear initial objectives and "model creep;" assuming sensitivity structure from model depth; conflating correlation with causation; conflating proportion with probability; and conflating expectation with probability. Solutions to these various problems pertain to learning correct BN model construction techniques, understanding basics of Bayesian statistical inference, experience with expert knowledge elicitation and necessity of expert peer review and reconciliation, and just being clear with model objectives and variables. more >

We describe the design of Omnigram Explorer (OMG), an open-source tool for the interactive exploration of relationships between variables in Bayesian networks (and other complex systems). OMG is designed to help researchers gain a holistic, qualitative understanding of the relationships between variables, specifically providing interactive, visual support for observational sensitivity analysis. OMG is especially useful for high-lighting dependencies between variables and small groups of variables. It's designed for exploratory analysis of BNs (and other models) and for communicating the salient features of models to non-specialists. more >

The use of Bayesian networks is increasing in many areas of science, research and policy, with decision networks (DN) not far behind. However, end-user and consumer-oriented modelling has seen much more limited adoption. We haven't yet reached the point of a personalised agent with a DN brain acting as our assistant, filtering information and helping (or even doing) our decision-making --- and it looks as though this may still be some years away. Notwithstanding the brief flurry of interest surrounding BN-based spam email management a few years ago, even unambitious personalised consumer applications of BN agents have yet to surface. Yet the potential for BNs and DNs in this space is substantial.

We at Bayesian Intelligence are therefore aiming to develop a series of personal BN-based applications that can assist users with the choices they face in their day-to-day lives. We have developed an initial prototype that demonstrates this idea by helping women with the difficult choice they may face with breast cancer treatment. The prototype allows an individual to assess her chances of developing breast cancer, and makes suggestions tailored to that person's situation and values. The prototype is designed to be very user friendly, hiding away the technical details of the model, while at the same time providing the user with as much useful information and feedback about her options as possible. more >