Decision Tree and K-Means Analysis of Raman Spectra for Edible Oils: A Physics-Informed AI Approach

Food fraud involving edible oils is a persistent challenge for regulators, manufacturers, and consumers alike. When oils are adulterated or mislabeled, the health, economic, and trust implications can ripple through supply chains. Researchers have turned to Raman spectroscopy—a non‑destructive technique that captures molecular vibrations—as a way to fingerprint oils. By pairing this optical data with machine‑learning models, the recent arXiv preprint “Decision Tree and K‑Means Analysis of Raman Spectra for Edible Oils: A Physics‑Informed AI Approach” shows how interpretable algorithms can both reveal hidden structure in the spectra and deliver reliable classification, even when oils are embedded in a complex food matrix.

What You Need to Know

The study examined five common edible oils—olive, sunflower, canola, soybean, and palm—both in their pure state and after being mixed into a fried‑potato‑chip matrix. Raman spectra were first visualized with t‑SNE to inspect natural groupings, then clustered using K‑means to assess how well the algorithms could separate the oil types without any label guidance. In the pure‑oil condition, the t‑SNE plot showed distinct, tightly packed clusters for each oil, and K‑means recovered these groups with high purity.

When the same oils were embedded in the chip matrix, spectral features from the starch, moisture, and frying by‑products overlapped with the oil signals, causing the clusters to blur and merge. To overcome this, the team trained a Decision Tree classifier on the Raman features, using the known oil labels as supervision. The tree achieved 100 % classification accuracy for the pure oils and maintained performance above 90 % for the matrix samples, demonstrating that a simple, rule‑based model can capture the discriminative patterns hidden in the noisy spectra.

To further link the data to underlying physics, the researchers applied Non‑Negative Least Squares (NNLS) decomposition, fitting each spectrum as a combination of reference oil spectra and matrix components. The NNLS coefficients provided quantitative estimates of oil fractions that aligned closely with the Decision Tree splits, illustrating how the model’s decisions can be interpreted in terms of real molecular contributions.

Why It Matters

Raman spectroscopy combined with transparent machine‑learning offers a practical path for on‑site oil authentication. Unlike chromatographic methods that require sample preparation and lengthy analysis, a handheld Raman probe paired with a Decision Tree can deliver a verdict in seconds, enabling rapid screening at processing lines, customs checkpoints, or retail outlets. The ability to maintain high accuracy even when oils are part of a processed food product addresses a key limitation of many existing spectroscopy‑based approaches, which often fail when confronted with complex matrices.

From a regulatory perspective, the interpretability of Decision Trees—each node corresponds to a measurable Raman shift or intensity threshold—facilitates auditability and compliance reporting. Stakeholders can trace a classification back to specific vibrational modes (e.g., C=C stretch at ~1650 cm⁻¹ for unsaturated fats) rather than trusting a black‑box score. This transparency builds confidence among food safety agencies and helps manufacturers demonstrate due diligence in labeling claims.

Key Details

• Five edible oils tested: olive, sunflower, canola, soybean, palm.
• Pure‑oil spectra showed strong class separation in t‑SNE and K‑means (clusters >85 % purity).
• Food‑matrix (fried‑potato‑chip) induced spectral overlap, reducing unsupervised cluster purity to ~55 %.
• Decision Tree classifier: 100 % accuracy on pure oils, >90 % accuracy on matrix samples.
• NNLS‑based spectral decomposition provided quantitative oil fraction estimates that matched Decision Tree splits.
• Model interpretable via specific Raman shifts (e.g., 1150 cm⁻¹ for C‑C stretch, 1650 cm⁻¹ for C=C).

What’s Next

Future work could extend the framework to additional oils, adulterants, and varied processing conditions (e.g., baking, deep‑frying at different temperatures). Integrating the Decision Tree with portable Raman hardware and developing a user‑friendly interface would enable real‑world deployment. Moreover, exploring hybrid models that combine the physical constraints of NNLS with ensemble learning may further improve robustness against unknown matrix interferences while preserving the transparency that regulators and industry value.

📌 Source: Arxiv Ml