Introduction: Bridging the Symbolic-Connectionist Gap
In the quest for robust AI, two paradigms often clash: symbolic reasoning (expert systems) and connectionist learning (data-driven neural networks). Knowledge Graphs (KGs) are excellent at storing relational facts—like (Aspirin, treats, Headache)—but they struggle with conditional logic. Conversely, expert systems use "if-then" rules that are highly interpretable but brittle and difficult to scale. Meanwhile, modern data-driven models are powerful but act as "black boxes" and require massive amounts of labeled data, which is often scarce in high-stakes fields like medicine.
The authors address the cognitive inference problem: how can we reason about a conclusion (e.g., a medical diagnosis) based on a set of observations (symptoms) by leveraging both structured facts and logical rules, even when training data is limited?
The COGINFER Framework
The researchers introduce the Cognitive Knowledge Graph (CogKG), a unified representation that combines relational facts with logical if-then rules. Their accompanying framework, COGINFER, bridges the gap between symbolic reasoning and deep learning, allowing for high-performance, interpretable inference in both zero-shot (unsupervised) and data-rich (supervised) scenarios.
Cognitive Knowledge Graph (CogKG)
This is a directed hypergraph where entities are nodes. Unlike a standard KG, it uses two types of edges:
- Relation Edges: Standard triples representing facts.
- Rule Hyperedges: These connect a set of "premise" entities to a set of "conclusion" entities, effectively embedding "if-then" logic directly into the graph structure.
Inference Modes
The COGINFER reasoning engine operates in two distinct modes:
- Unsupervised Mode: It uses Certainty Factors (CF) to propagate belief through the graph. If a rule's premise isn't perfectly met, it uses KG embeddings (via TransE) to perform "link prediction," essentially "guessing" missing connections based on relational patterns.
- Supervised Mode: When labeled data is available, the framework transforms rules and facts into "explainable neurons." It constructs a simple, single-layer neural network where the input neurons have explicit semantic meanings (specific rules or facts). The model then optimizes the weights of these connections via gradient descent, making it a "white-box" model where you can see exactly which rules the AI prioritizes.
Datasets and Evaluation
The approach was evaluated on two clinical diagnosis benchmarks:
- Muzhi: A Chinese medical dataset containing 710 samples across 4 diseases and 66 symptoms, aligned with the Chinese Medical Knowledge Graph (CMeKG).
- MDD: An English medical dialogue dataset featuring 2,151 samples, 12 diseases, and 93 symptoms, aligned with the SNOMED-CT knowledge base.
The researchers mined and expert-validated 92 rules for Muzhi and 182 rules for MDD to populate the CogKG.
Key Insights
- Rules as Special Cases: The classic MYCIN expert system is a special case of COGINFER. By setting the link prediction threshold to its maximum, COGINFER reverts to strict symbolic logic.
- Solving the "Cold Start": Because COGINFER inherits symbolic knowledge, it performs remarkably well even with zero training examples. As data becomes available, it refines the importance of these rules rather than discarding them.
- White-Box Interpretability: Because each "neuron" represents a specific expert rule or KG fact, the weight matrix is directly readable. This allows clinicians to visualize which symptoms or rules the model prioritizes for specific diagnoses.
Results and Performance
- Unsupervised Superiority: In settings with no training data, COGINFER outperformed pure rule-based systems (MYCIN) and pure KG-based systems (PURELINK). On the MDD dataset, it achieved a Hits@1 of 0.606, significantly higher than MYCIN's 0.278.
- Supervised Performance: When trained, COGINFER surpassed traditional models like Logistic Regression and Random Forests, achieving an F1-score of 0.857 on the MDD dataset.
- Ablation Success: KG-generated features were found to be more influential than rule-generated features, likely due to the higher density of relational facts.
Limitations and Future Work
The effectiveness of CogKG depends heavily on the quality of the underlying KG and expert rules. The "knowledge acquisition bottleneck" remains a challenge, though association rule mining helps. Future work will focus on more efficient subgraph sampling to handle larger-scale graphs and rulesets.
Summary updated