Atopic Dermatitis, commonly known as eczema, manifests as red, itchy, and occasionally painful rashes, affecting both children and adults to varying degrees.
This section examines the many facets of atopic dermatitis, where genetic, environmental, and immunological factors converge to influence its development and progression. To better understand this complex disease, we use Hellixia to generate causal Bayesian networks, which provide a structured framework for deciphering cause-and-effect relationships.
But first, we'll start with a semantic analysis of the domain to get an overview of the main concepts, variables, and relationships in the field of atopic dermatitis.
We create the node "Atopic Dermatitis" and then go through our usual workflow for creating a semantic network and then a hierarchical semantic network (see previous sections, e.g., Hamlet) to perceive the semantic landscape surrounding atopic dermatitis and lay the foundations for a deeper understanding of its underlying dynamics.
We finally select the factors only (i.e., we focus on the higher level of this hierarchical network), and use the Hellixia Report Analyzer to generate a concise summary of the Relationship Analysis Report.
Having gained an overall understanding of the domain through semantic networks, we now move on to the construction of Causal Bayesian Networks using Hellixia's new capabilities that will be released in BayesiaLab 11.2.
We start by creating a node called "Atopic Dermatitis Mechanism", then select the Causal Network Generator feature.
After one or two minutes (the prompt is indeed quite complex), we obtain a fully specified Causal Bayesian Network (graph and probabilities). This network is characterized by causally oriented arcs, each accompanied by a concise explanation of the causal relationship and an estimate of the causal effect, scaled between -100 (shown in red) and 100 (shown in blue). To translate these causal effects into conditional probability tables, we use a new BayesiaLab formula, DualNoisyOr(), specially designed to integrate positive and negative effects between Boolean variables.
Naturally, the networks generated by Hellixia MUST undergo rigorous evaluation by Subject-Matter Experts. This verification is crucial not only from a qualitative point of view to ensure that the network accurately represents real causal relationships but also from a quantitative point of view to confirm the relevance of the suggested causal effects.
Let's delve further into this domain by exploring the underlying causes of "Microbial Infection." To do this, we select the respective node in the network and proceed to the Causal Network Generator.
Displayed below is the generated causal network, showcasing the expanded view with detailed aspects of Microbial Infection. The yellow nodes are common to both the original and expanded networks, the grey nodes represent the original network nodes only, and the red nodes indicate the newly added dimensions specific to microbial infection.
We finally use the Hellixia Report Analyzer to generate a concise summary of the (Causal) Relationship Analysis Report.
We will now adopt a different workflow to construct a Causal Network for Atopic Dermatitis. We start by using Hellixia's Dimension Elicitor to identify relevant dimensions. With these nodes generated, we diverge from our usual practice of generating embeddings for semantic networks. Instead, we utilize Hellixia's new Causal Relationships Finder feature to automatically create a Causal Network based on our set of selected nodes.
We select a range of keywords to guide the Dimension Elicitor process in Hellixia, encompassing various aspects of the domain under study. These keywords include 'Accelerators,' 'Catalyzers,' 'Causes,' 'Drivers,' 'Mechanisms,' 'Consequences,' 'Symptoms,' 'Inhibitors,' 'Moderators,' 'Preventers,' and 'Treatments.'
We run the Causal Relationships Finder on the nodes elicited for the Atopic Dermatitis Mechanism. This tool examines potential causal connections among these nodes and, if required, generates latent variables to enhance the network's explanatory power.
Similar to the Causal Network Generator, the tool does more than identify causal links; it also quantifies the causal effects, which are represented on a scale ranging from -100 (indicated in red) to 100 (indicated in blue).
We conclude this section by utilizing the Hellixia Report Analyzer, which efficiently generates a concise summary of the (causal) Relationship Analysis Report for this latest network.
Welcome to our Causal Bayesian Networks section, where we leverage Hellixia as a Subject Matter Assistant for constructing Causal Bayesian Networks. These networks feature directional arcs that convey causality. In contrast to Causal Semantic Networks, which primarily offer qualitative insights by highlighting semantic causal relationships between variables, Causal Bayesian Networks offer a dual approach, encompassing both qualitative and quantitative aspects. They serve not only to improve our understanding of a domain, but also to enable probabilistic and causal inference.
In the complex and dynamic field of air transport, it is crucial for airlines to understand and mitigate flight delays. With the advent of sophisticated analytical tools like Hellixia, we now have the opportunity to delve deeper into the causal factors behind these delays. This article explores the innovative application of Hellixia in the creation of Causal Bayesian Networks (CBN), a method that transcends traditional data analysis to uncover the root causes of flight delays.
Using Hellixia for this purpose represents a significant advance in the field of causal analysis. By building causal Bayesian networks, we can map the complex web of factors contributing to delays, from weather conditions to logistical challenges.
In the following sections, we'll look at how Hellixia facilitates the construction of these causal networks, and the insights they provide into the management and prevention of flight delays.
First, we will perform a semantic analysis of the domain to obtain an overview of the key concepts and variables within the aircraft delay domain.
For our analysis of "Delays in scheduled flight departures", we start by building a semantic network, followed by a hierarchical semantic network, similar to our previous workflows (for example, as demonstrated with Hamlet). This process is essential for mapping the semantic landscape surrounding flight delays, providing a solid foundation for understanding the underlying dynamics of this issue.
We begin our analysis by creating a node entitled "Delays in scheduled flight departures" and proceed to use Hellixia's Dimension Elicitor, using two distinct groups of keywords: 'Ancestors' and 'Descendants'. This approach allows us to explore in depth the factors leading to and resulting from flight delays.
We carefully examine the dimensions provided by Hellixia, removing any that seem extraneous or irrelevant to our analysis. Next, we exclude 'Delays in scheduled flight departures' and run the Embedding Generator on the remaining nodes. This step is crucial to understanding the semantic relationships linked to their names and comments.
We have two large sets of nodes: one representing "Ancestors" (42 nodes) and the other "Descendants" (69 nodes). Our approach is to learn a separate network for each group. To do this, we define specific constraints that prohibit relationships between nodes that do not belong to the same class.
We then run the Maximum Weight Spanning Tree algorithm to find the most significant semantic relationships between nodes.
To improve visibility, we change the node styles to Badges, clearly displaying the comment associated with node. Next, we run the Dynamic Grid Layout to position the nodes on the graph. It's important to note that this algorithm is not deterministic, resulting in random orientations - vertical, horizontal or mixed. As a result, we may have to apply this layout several times to get a configuration that matches your preferences.
Next, we switch to Validation Mode and opt for the Skeleton View. In this context, since our network doesn't represent causal relationships, this view is particularly useful as it retains only the connections between nodes, omitting direction indicators.
Next, we run Variable Clustering. This step categorizes variables that are similar, grouping them based on the semantic relationships identified between them.
We can now proceed with the creation of two hierarchical semantic networks.
Opening Class Editor: We begin by accessing the Class Editor and then running the Class Description Generator. This generates descriptive names for the factors we're examining.
Exporting Descriptions: Next, we use the Export Descriptions function to save the newly created factor descriptions.
Returning to Modeling Mode: We then switch back to Modeling Mode and conduct Multiple Clustering to create latent variables.
Running the Structural Learning Algorithm (Taboo): We run the Taboo algorithm for structural learning, ensuring that the Delete Unfixed Arcs option is selected.
Renaming Latent Variables with Exported Descriptions: We utilize the descriptions we previously exported as a Dictionary to rename the latent variables, adding clarity to our model.
Switching to Validation Mode and Running Node Force: Finally, we go back to Validation Mode and run the Node Force analysis, which helps us understand the dynamics and strength of the connections within our network.
Having established a global understanding of the domain via semantic networks, we're now ready to move forward with the construction of causal Bayesian networks, taking advantage of the latest capabilities introduced in Hellixia as part of BayesiaLab version 11.2.
We initiate the process by creating a node named Delays in Scheduled Flight Departures and then proceed to use the Causal Network Generator feature.
After one or two minutes, due to the complexity of the prompt, we manage to generate a small but fully specified causal Bayesian network (graph and probabilities). This network features directed arcs to signify causal relationships, with each arc accompanied by a succinct explanation of its causal link and an estimate of the effect, scaled from -100 (shown in red) to 100 (shown in blue).
To differentiate nodes by depth using different colors, we first run the Edit Class function. Next, we select Generate a Predefined Class - Depth. Next, we select the four depth classes that have been created and apply the Colors - Associate Random Colors with Classes function to assign distinct colors to each class.
Nodes marked with an icon representing a function are parameterized using BayesiaLab's new DualNoisyOr() formula. This formula integrates both positive and negative interactions between Boolean variables (the causal effects returned by Hellixia).
By selecting the Create Corresponding Structural Priors option in the Causal Network Generator wizard, we now have access to Structural Priors. The value of each prior is derived from the absolute value of the causal effect returned by Hellixia. In addition, the explanation provided for each prior corresponds to the description of its causal relationship. These structural priors can then be used later for network learning when relevant data becomes available.
To finalize this first causal network, we employ the Hellixia Image Generator to create unique icons for each node, based on the comment.
Let's move on to the creation of a more complex causal network by setting Complexity to High.
The next crucial step is an in-depth examination of this automatically generated network. For example, we observe that Fueling Delays is identified as a direct cause. Interestingly, Aircraft Turnaround Time is also identified as a direct cause. This leads us to speculate that Fueling Delays could be a direct cause of Aircraft Turnaround Time, which would have an indirect effect on flight delays.
To verify this hypothesis, we select the two nodes, Fueling Delays and Aircraft Turnaround Time, and apply the Hellixia Pairwise Causal Link feature. This will help us ascertain the nature of the causal relationship between these variables.
Hellixia validates the existence of this causal relationship and accordingly updates the conditional probability distribution of Aircraft Turnaround Time. This update incorporates a DualNoisyOr() function with a coefficient of 0.75, reflecting the quantified impact of Fueling Delays on Aircraft Turnaround Time.
Following this update, our next step involves removing the direct link from Fueling Delays to Delays in Scheduled Flight Departures. Subsequently, we need to adjust the DualNoisyOr() formula to accurately reflect this change in the network's structure.
Driven by curiosity to delve deeper, we select the relevant node to explore the causes of causes. For this, we once again make use of the Causal Network Generator, but on Fueling Delays.
Upon reviewing the newly added nodes and relationships, we identified that three relationships were incorrectly marked as negative, contrary to the descriptions in their respective link comments. To rectify this, we change the color of these links to accurately reflect their positive nature and accordingly update the DualNoisyOr() formula of Operational Efficiency.
To conclude our analysis, we're going to build a final causal network, this time using the Causal Relationships Finder function. Unlike the Causal Network Generator, which added new nodes for creating the network, this feature works directly with selected nodes. To begin with, we use the Dimension Elicitor tool to identify the 5 main Causes and 5 main Effects associated with Delays in Scheduled Flight Departures.
We proceed by selecting the 10 causes and effects, along with the Delays in Scheduled Flight Departures node. With these nodes selected, we then run the Hellixia Causal Relationships Finder to create the network.
As a result, we obtain the bow-tie network structure below.
This brings us to the end of our article. For further insights, we invite you to view the recorded webinar on this topic, which was conducted in January 2024.