A Neural-Symbolic Approach for User Mental Modeling: A Step Towards Building Exchangeable Identities

Hossein Rahnama, Marjan Alirezaie, Alex PentlandAAAI 2021

Introduction

Modern Human-Computer Interaction (HCI) typically operates on a closed loop where the individual providing data is the sole recipient of the resulting services. This paper proposes a paradigm shift toward "exchangeable identities"—digital surrogates that encapsulate a user's expertise, personality, and reasoning patterns. By combining the pattern-recognition capabilities of deep learning with the structured logic of symbolic reasoning, the authors present a neural-symbolic framework capable of building sophisticated digital mental models.

The Challenge of Exchangeable Identities

Traditional AI models often focus on preference matching (e.g., "users who liked X also liked Y"). However, to enable a "source user"—such as a legal expert or a teacher—to delegate their identity to an AI avatar, the system must understand how that user thinks, not just what they consume.

The concept of exchangeable identities where a source user's data trains a machine to serve target users.

Current machine learning (ML) approaches face significant hurdles: they are often "black boxes" lacking transparency, require massive datasets, and struggle to generalize reasoning across different contexts. The goal is to move from simple data-driven services to a model where a surrogate can interact with "target users" on behalf of the original source.

Methodology: The Neural-Symbolic Double-Loop

The core of the proposal is a double-loop learning system that bridges two AI paradigms: subsymbolic (neural) and symbolic (reasoning).

1. Subsymbolic Layer (Learning)

This layer processes "digital footprints"—raw data from social media, location history, and consumption habits. To address privacy concerns, the framework utilizes Federated Learning, allowing models to be trained locally on a user's device without exposing sensitive raw data to central servers.

2. Symbolic Layer (Reasoning)

The symbolic component utilizes a Digital Mental Model Ontology (DMMO). This structured knowledge base, built on the DOLCE Ultralite (DUL) framework, provides a conceptual map of human traits, emotions, and professional domains.

The double-loop learning process evolving a user's digital mental model over time using multi-modal data.

The Reasoning Workflow

The system evolves the user profile through a multi-stage process:

  • Abductive Reasoning: The system generates hypotheses from ML outputs. For instance, if a user frequently posts about "zero-waste" lifestyles, the system infers they might hold environmentalist values.
  • Deductive Reasoning: It applies general rules to refine these hypotheses. If the user is an environmentalist and frequently attends community events, the system deduces a high "extroversion" score in the Big Five personality traits.
  • Feedback Loop: These high-level inferences are fed back into the neural layers, a process termed "semantic infusion," which improves the model's predictive accuracy and transparency over time.

Insights and Application Domains

The strength of this approach lies in its ability to handle data sparsity. By using abductive reasoning to find the "best explanation," the system can make meaningful inferences even with limited data points. Furthermore, by focusing on the "lens" through which a user views the world, the AI can simulate their perspective for others.

The authors validate this framework through several conceptual domains:

  • Augmented Eternity: Preserving the knowledge and personality of individuals for future generations.
  • Legal Avatars: Scaling legal expertise by allowing AI surrogates to handle preliminary consultations.
  • Education: Helping students overcome "echo chambers" by allowing them to view controversial topics through the "lens" of a different digital identity.
  • Decentralized Recommendations: Navigating a city based on the trusted perspectives of friends' digital surrogates rather than generic algorithms.

Limitations and Future Outlook

While promising, the framework's efficacy is tied to the quality of the underlying ontologies. A poorly defined knowledge model leads to flawed reasoning. Additionally, there is a current scarcity of causal models that can bridge the gap between high-level human reasoning and low-level ML data. Future work will focus on creating "pluggable" ontological patterns for specific industries like healthcare and finance to enhance the scalability of these exchangeable identities.

Summary updated

BibTeX
@inproceedings{rahnama2021neural,
  author = {Hossein Rahnama and Marjan Alirezaie and Alex Pentland},
  title = {{A Neural-Symbolic Approach for User Mental Modeling: A Step Towards Building Exchangeable Identities}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
}
Figure