Graph-Based AI Models

Why learn graph-based AI?

As data scientists, we often encounter situations where data is scarce and incomplete, posing a significant challenge for decision makers. In these cases, making informed decisions based on limited data becomes crucial. While supervised machine learning techniques, such as deep learning based on artificial neural networks (ANNs), provide answers derived from vast amounts of data, they may not always be the best solution for handling real-world situations with incomplete data.

This is where graph-based AI techniques, like Bayesian networks (BNs), come into play. These techniques incorporate non-sample or prior human expertise, which is highly relevant in numerous real-world scenarios where data is limited. Throughout this course, we'll explore the powerful impact of combining human expertise with sample data in graph-based AI models.

One effective method for collecting the essential information required to enrich graph-based AI models is to conduct interviews with experts who can provide insights on the impact of various parameters. Thanks to this data+expert knowledge approach, graph-based AI techniques have several key advantages over other techniques:

• Transparency and explainability: Unlike black box models, such as ANNs, graph-based AI models offer a graphical representation of the underlying structure, making it easier to explain the results and communicate with experts. This is particularly beneficial when working with individuals who may not be experts in the modeling technique but possess extensive domain knowledge in fields like healthcare, supply chain management, or project management.

• Flexibility: Graph-based AI models can be built using both expert knowledge and raw data, even when prior knowledge is incomplete. This allows us to start with expert insights, refine the model using data, and then iterate by incorporating additional expert feedback. This iterative process enables the development of more robust models that are fine-tuned to address specific challenges and considerations.

How to use Bayesian networks in the real world

Imagine a large construction company managing multiple projects concurrently. The company aims to improve its project management practices and maintain high operational performance. In this context, the company seeks to use graph-based AI techniques to create a model that combines both expert knowledge and data from previous projects to identify potential issues, or "drift factors," that may negatively impact project performance.

Step 1: Expert interviews

The company interviews experienced project managers and consultants who have dealt with various project management challenges. These experts identify potential drift factors, such as delays in material delivery, communication breakdowns, or insufficient resource allocation. They also establish causal relationships between best practices and each drift factor. Furthermore, experts estimate how the probability of occurrence of these drift factors is related to project management best practices.

Step 2: Data collection

To quantify the consequences of these drift factors on project performance, the company collects data from historical project management software databases. The data includes information about the occurrence (or absence) of a drift factor and its corresponding impact on a project's performance.

Step 3: Building the graph-based AI model

Using the expert knowledge from the interviews and the data collected, the company constructs a BN model that represents the causal relationships between project management maturity criteria and the identified drift factors. The model incorporates the probabilities of the drift factors' occurrences based on maturity levels.

Step 4: Model validation and refinement

The company validates and refines the BN model using additional data from new projects. This iterative process helps improve the model's accuracy and reliability.

Step 5: Implementation and decision-making

With the graph-based AI model in place, the company can now monitor ongoing projects and use the model to predict the potential occurrence of drift factors based on the projects' current maturity levels. The model also helps estimate the impact of these drift factors on project performance, allowing decision-makers to implement corrective or preventive actions before facing harmful consequences.

By leveraging the power of graph-based AI techniques, the construction company can now make more informed decisions, improve its project management practices, and ultimately enhance its projects' operational performance.

Strengths and weakness of Bayesian networks

BNs are particularly well-suited to address problems with the following characteristics:

• Causal relationships: BNs excel in modeling situations where causes must be correlated with consequences. By representing these causal relationships in a BN, it is possible to examine how changes in input variables may affect the output variable, enabling more informed decision-making.

• Limited and changing data: BNs are an ideal choice when the amount of available data is not extensive and when the data is subject to change over time, leading to a high level of uncertainty. BNs can incorporate and update prior knowledge and probabilities as new data becomes available, allowing for continuous model improvement and adaptation to changing conditions. This makes BNs an effective tool for modeling situations where data is scarce or dynamic.

• Combination of data and expert knowledge: BNs are uniquely capable of incorporating both data and human expertise into the model. This is especially useful in real-world scenarios where expert judgment is valuable, but data may be incomplete or insufficient. By combining data-driven insights with expert knowledge, BNs can create a more comprehensive understanding of the problem at hand, leading to better decision-making and more accurate predictions.

Some additional strengths of BNs include:

• Scenario representation: BNs can represent different scenarios, allowing users to explore various possibilities and make informed decisions based on the model's predictions.

• Dynamic learning: BNs have the ability to increase their accuracy over time by incorporating more data, enabling continuous improvement and adaptation to new information.

• Retro-propagation analysis: BNs can perform retro-propagation analysis to investigate the causes of non-performance, helping users to identify areas for improvement or potential risks.

However, BNs also have some weaknesses:

• Subjectivity: The model may be more subjective than others due to the reliance on experts' judgment. While incorporating expert knowledge can be beneficial, it also introduces the potential for biases and inaccuracies, as human judgment can be flawed or influenced by personal experiences and beliefs.

• Conditional probabilities: The outcome of BNs is expressed in terms of conditional probabilities, which can be difficult for some users to interpret. This may require additional effort to translate the results into actionable insights for stakeholders who may not be familiar with probability theory.

In summary, BNs are a powerful modeling technique for problems that involve causal relationships, limited and changing data, and the need to combine data with expert knowledge. Despite their weaknesses, BNs can still provide valuable insights and support decision-making processes in various domains and industries.