Technical Summary of AnTenA: Actionable and Explainable Tensor Analysis System with Large Language Models

Problem Statement

Accurately explaining hidden patterns within multi-aspect (tensor) data has traditionally relied on labels and auxiliary metadata. However, this approach faces significant limitations in real-world scenarios: metadata may be inaccurate (nonstandard or inconsistent), insufficient (e.g., static tabular data for time-dependent recordings), or entirely unavailable. Furthermore, existing Explainable AI (XAI) frameworks often focus on explaining the decision-making processes of supervised "black-box" models rather than providing insights into the underlying data structures themselves. In the domain of tensor data mining, latent components derived from decomposition are inherently difficult to interpret because they capture complex multi-way interactions, and relying solely on stale or incomplete metadata can lead to misaligned explanations.

Methodology

The authors propose AnTenA (Actionable and Explainable Tensor Analysis System), a demo system that leverages Large Language Models (LLMs) to generate natural language explanations for latent patterns extracted from tensor decomposition. The system architecture consists of two primary modules:

1. Tensor Analysis Module

This module extracts interpretable latent components using tensor decomposition models.

• Models: The system supports CP (CANDECOMP/PARAFAC) decomposition, known for its simplicity, and NeAT, a neural tensor model that maintains CP structure while capturing non-linear latent patterns via neural networks applied to each component.
• Constraints: To enhance interpretability, the models incorporate non-negativity constraints (ensuring factor values are non-negative so magnitudes are meaningful) and orthogonality constraints (ensuring factors remain distinct).
• Output: The module produces a set of latent components representing co-clustered patterns across different modes of the tensor.

2. Explainer Module

This module utilizes LLMs to translate the mathematical latent components into human-readable narratives.

• Inputs: The LLM receives top-K entities for each mode, their corresponding factor values, decomposition metrics (reconstruction error, Factor Match Score [FMS]), and optional metadata.
• Prompting Strategies:
  • Task-Agnostic Prompts: Instruct the LLM to describe hidden patterns generally, providing a broad overview.
  • Task-Specific Prompts: Guide the LLM to interpret patterns in the context of specific tasks (e.g., suggesting movie recommendations based on MovieLens components or predicting research trends based on DBLP co-clusters).
• Reliability Evaluation (Inference Tasks): To assess the credibility of the LLM's explanations, the system employs two inference tasks motivated by prior work [11]:
  • Forward Inference: The LLM is asked to predict entities that logically belong to a given component but are missing from its top-K list. This tests the model's ability to retrieve plausible entities from a constructed pool (top-50 entities across all components).
  • Backward Inference: The LLM is asked to identify a "false" entity injected into a component. These false entities are selected from the top-K of a distinct component (measured by FMS), testing the model's ability to detect inconsistencies within a small set.

System Demonstration

The paper demonstrates AnTenA using two real-world datasets: MovieLens-Belief-2024 and DBLP (academic publication data). The interface allows users to:

1. Select datasets and models (CP or NeAT) with specific constraints and hyperparameters.
2. Visualize decomposition results, including reconstruction errors (nRE, RMSE) and component similarity (FMS).
3. View heatmaps of top-K factor values and summaries of co-clusters.
4. Observe LLM-generated explanations (both task-agnostic and task-specific) and the results of the forward/backward inference reliability checks.

Key Contributions

• LLM-Based Tensor Explanation: The introduction of a system that uses LLMs to explain extracted latent components from tensor decomposition, moving beyond reliance on potentially flawed auxiliary metadata.
• Enhanced Interpretability: The integration of non-negativity and orthogonality constraints into tensor models (CP and NeAT) to ensure factor values are meaningful and distinct.
• Dual-Prompt Framework: The implementation of both task-agnostic and task-specific prompting to balance general insights with actionable, domain-specific recommendations.
• Reliability Evaluation: The proposal of forward and backward inference tasks specifically designed to evaluate the quality and consistency of LLM-generated explanations in the context of tensor analysis.

Significance and Claims

The paper positions AnTenA as a step toward more intuitive and directly interpretable XAI for unsupervised tensor mining. The authors claim that by leveraging the vast knowledge and reasoning capabilities of LLMs, the system can identify nuances in multi-way interactions that traditional metadata-based methods miss. The inclusion of inference tasks is highlighted as a mechanism to improve the reliability and credibility of the generated explanations. The system is presented as a demo tool that facilitates the analysis of real-world tensors, offering both a general overview of data patterns and specific, actionable insights depending on the user's needs.

Note: The paper does not claim to solve all XAI challenges or propose new theoretical bounds for tensor decomposition; rather, it focuses on the practical application of LLMs to interpret existing decomposition results and validate those interpretations through specific inference tasks.