Semantic-Guided RL for Interpretable Feature Engineering

Mohamed Bouadi, Arta Alavi, Salima Benbernou, Mourad OuziriarXiv 2024 · DOI · arXiv

TL;DR

This paper introduces SMART, a novel approach that uses semantic knowledge graphs and reinforcement learning to automatically engineer interpretable features for machine learning models. By combining domain knowledge with a DRL-based exploration strategy, SMART significantly improves model performance while ensuring that the generated features are understandable to domain experts.

What Problem Does This Paper Address?

The quality of machine learning (ML) models heavily relies on the input data, and effective Feature Engineering (FE) is crucial for extracting valuable patterns. However, manual FE is time-consuming and requires deep domain expertise. While automated FE (AutoFE) tools exist, a significant challenge remains: generating interpretable features. Existing AutoFE methods often focus on performance and can produce features that are opaque to domain experts, hindering trust and understanding. This paper addresses the gap of generating features that are both high-performing and interpretable, bridging the divide between statistical ML and symbolic AI.

Methods

SMART employs a two-step process: Exploitation and Exploration.

  • Exploitation: This phase leverages a Knowledge Graph (KG) and Description Logics (DL) to infer new, domain-specific features. The KG contains both domain-agnostic knowledge (e.g., units of measurement) and domain-specific concepts (e.g., healthcare, finance). DL reasoners (like HermiT) are used to infer relationships and generate features based on existing data and KG semantics. SWRL rules are also used to define interpretability criteria.
  • Exploration: This phase uses a Deep Reinforcement Learning (DRL) agent, specifically a Deep Q-Network (DQN), to explore the feature engineering search space. The agent learns a policy to generate new features by iteratively applying transformations. The state for the DQN is a semantic vectorization of the input data, and the reward is a combination of model performance improvement and a novel feature interpretability metric. This metric assesses how well a generated feature aligns with concepts in the KG.

Datasets

The paper evaluates SMART on several widely-used real-world datasets, including:

  • Wine Quality White/Red
  • Bikeshare DC
  • Home Credit Default Risk
  • Amazon Employee Access Challenge
  • Higgs Boson
  • Brazilian E-Commerce
  • Medical Appointment
  • Fraud Detection
  • Retail Spending
  • NYC Taxi Ride Data

No new benchmarks were introduced; the focus was on evaluating SMART against existing datasets and baselines.

Key Insights and Ideas

  • Bridging Symbolic and Statistical AI: SMART effectively combines the reasoning capabilities of symbolic AI (DL, KGs) with the learning power of statistical AI (DRL) to achieve both performance and interpretability.
  • KG as a Semantic Backbone: The KG is central to SMART, providing the domain knowledge necessary for both inferring interpretable features and guiding the DRL agent's exploration.
  • Novel Interpretability Metric: The introduction of a metric that quantifies feature interpretability based on its relationship with KG concepts is a significant contribution, addressing a long-standing challenge.
  • Feature-Level Interpretability is Key: The paper emphasizes that interpretable features are crucial for overall model interpretability, even for simple models like regression.
  • DRL for Feature Engineering Search: Using a DQN to navigate the vast and complex feature engineering search space proves effective, learning to generate promising transformations.

Key Results

  • Performance Improvement: SMART outperforms baselines in most cases, improving model performance by an average of 20.94% compared to using raw data. It also shows significant improvements over other advanced AutoFE methods like DIFER, NFS, and mCAFE.
  • Interpretability: The generated features are shown to be more interpretable, aligning better with domain knowledge. For instance, in the Medical Appointment dataset, SMART generated features like "duration between appointment and actual appointment date," which directly correlate with patient no-shows. In the NYC Taxi dataset, features like "distance between pickup and dropoff locations" and "duration between pickup and dropoff time" were generated.
  • Feature Importance: SHAP analysis demonstrates that features generated by SMART are more important for model predictions compared to raw features across various datasets.

Limitations and Future Work

  • Scalability of KG: While the KG can be expanded, its size and the complexity of reasoning over it could become a bottleneck for extremely large or complex domains.
  • Computational Cost: The DRL exploration phase, while more efficient than exhaustive search, can still be computationally intensive.
  • Future Work: The authors suggest exploring more advanced DRL techniques and expanding the KG to cover a wider range of domains and applications. Further research into formalizing and quantifying feature interpretability remains an open area.

Summary updated

Abstract

The quality of Machine Learning (ML) models strongly depends on the input data, as such generating high-quality features is often required to improve the predictive accuracy. This process is referred to as Feature Engineering (FE). However, since manual feature engineering is time-consuming and requires case-by-case domain knowledge, Automated Feature Engineering (AutoFE) is crucial. A major challenge that remains is to generate interpretable features. To tackle this problem, we introduce SMART, a hybrid approach that uses semantic technologies to guide the generation of interpretable features through a two-step process: Exploitation and Exploration. The former uses Description Logics (DL) to reason on the semantics embedded in Knowledge Graphs (KG) to infer domain-specific features, while the latter exploits the knowledge graph to conduct a guided exploration of the search space through Deep Reinforcement Learning (DRL). Our experiments on public datasets demonstrate that SMART significantly improves prediction accuracy while ensuring a high level of interpretability.
BibTeX
@misc{bouadi2024semantic,
  author = {Mohamed Bouadi and Arta Alavi and Salima Benbernou and Mourad Ouziri},
  title = {{Semantic-Guided RL for Interpretable Feature Engineering}},
  year = {2024},
  eprint = {2410.02519},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.02519},
}
Figure