Knowledge Engineering for Intelligent Decision Support

Maria Vanina MartinezIJCAI 2017 · DOI

TL;DR

This paper argues that truly intelligent decision support systems must go beyond data and knowledge representation to deeply understand and incorporate human expertise, preferences, and evolving knowledge. The author proposes a framework that combines data-driven AI with knowledge-based approaches, focusing on modeling user preferences, managing conflicting information, and understanding the dynamics of human knowledge.

What Problem Does This Paper Address?

Modern information systems, especially those supporting decision-making, are awash in data. While automated systems can process this data, they often fail to capture the nuanced, implicit knowledge and preferences that humans bring to bear on real-world decisions. This gap leads to systems that are not truly intelligent or personalized. The paper addresses the challenge of building AI systems that can effectively partner with humans, enhancing their decision-making capabilities by understanding and integrating their domain expertise, evolving knowledge, and individual/group preferences. This is crucial as user expectations shift towards highly personalized and adaptive services.

Methods

The author outlines a three-pillar approach to developing intelligent decision support systems:

  • Modeling Preferences: This involves representing and reasoning with individual and group preferences. The work explores extending ontology languages like Datalog+/- to incorporate probabilistic preferences and group decision-making operators, aiming for fairness. Novel reasoning mechanisms are developed to handle the computational complexity arising from richer preference models, utilizing techniques from Order Theory for tractable query answering.
  • Managing Conflicting Information: This pillar focuses on developing "Personalizable Inconsistency Management" policies. Instead of a one-size-fits-all approach, users can define how inconsistencies are handled within their specific context. This includes developing unified formalisms that allow users to inject their domain knowledge into inconsistency resolution and exploring inconsistency-tolerant semantics for query answering in Description Logics and Datalog+/- ontologies, aiming for a balance between answer quality and computational tractability.
  • Understanding Knowledge Dynamics: This addresses how human knowledge, beliefs, and preferences evolve over time due to learning and social interactions. The paper draws on belief revision (BR) research, particularly the AGM model, and explores how to formalize these dynamics within intelligent systems. This includes developing methods for handling belief change operations like contraction and revision, and investigating how incoherence affects query answering.

The overarching methodology emphasizes a synergy between data-driven (machine learning) and knowledge-based (logic, KR) frameworks to structure and formalize data into usable knowledge models.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Human-centric AI is paramount: True intelligence in decision support requires moving beyond raw data to understand the human element – expertise, preferences, and evolving knowledge.
  • Personalization is key: Users expect customized experiences, and AI systems must adapt to individual needs and expectations.
  • Preferences are complex and dynamic: Modeling preferences requires handling incompleteness, inconsistency, and uncertainty, and must account for group dynamics and evolution over time.
  • Inconsistency is a feature, not just a bug: Instead of forcing consistency, systems should allow for personalized management of conflicting information, integrating user-defined policies.
  • Knowledge is not static: Intelligent systems must account for how human knowledge and beliefs change, requiring formalisms for belief revision and knowledge dynamics.
  • Synergy is essential: The most effective intelligent systems will combine data-driven machine learning with knowledge-based reasoning to create robust models.

Key Results

The paper highlights several research contributions, primarily through citations to the author's previous work:

  • Development of Datalog+/- extensions for modeling individual and group preferences, including probabilistic uncertainty.
  • Novel reasoning mechanisms for preference-based query answering that are computationally tractable under certain assumptions.
  • A framework for "Personalizable Inconsistency Management" allowing users to define policies for handling inconsistencies in databases.
  • Exploration of inconsistency-tolerant semantics for Description Logics and Datalog+/- ontologies, aiming for better trade-offs between tractability and answer quality.
  • Approaches to modeling belief dynamics and handling social interactions in belief revision.

Limitations and Future Work

The author acknowledges that the problems addressed are far from solved and require radical breakthroughs. Key future directions include:

  • Explaining Inferences: Developing methods for intelligent agents to adequately and automatically explain their obtained inferences to humans.
  • Human-Machine Synergy: Creating robust mechanisms for human-machine cooperation to enhance decision-making processes.
  • Integrating Human Expertise: The usability of existing frameworks relies on the availability of human expertise as input, which needs to be effectively integrated.
  • Engineering Effort: Significant engineering effort is required to combine machine learning tools with knowledge-based approaches to produce usable knowledge for models.
  • Understanding Social Dynamics: Further research is needed to formalize the impact of social interactions on knowledge, beliefs, and preferences.

Summary updated

BibTeX
@inproceedings{martinez2017knowledge,
  author = {Maria Vanina Martinez},
  title = {{Knowledge Engineering for Intelligent Decision Support}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/736},
}
Figure