Authentication of tequilas using pattern recognition and supervised classification

Guadalupe Pérez‐Caballero, J.m. Andrade, Pablo M. Olmos, Yusdivia Molina-Román, Ivan Jimenez, J.j. Durán, Carlos Fernández-Lozano, Floriberto Miguel-CruzTrAC 2017 · DOI

TL;DR

This paper tackles the growing problem of tequila counterfeiting by developing and comparing various chemometric methods for authenticating and classifying different types of tequila. They found that non-linear classification methods, particularly Random Forest and Support Vector Machines, offer high accuracy in distinguishing between white, rested, aged, and extra-aged tequilas using UV-Vis spectral data.

What Problem Does This Paper Address?

The global popularity of tequila has unfortunately led to an increase in counterfeiting. This not only harms legitimate producers but can also pose health risks to consumers. Distinguishing authentic tequilas from fakes, and even classifying the different types (based on aging), is a significant analytical challenge. Traditional methods can be time-consuming or require specialized equipment. This paper aims to provide a fast, cheap, and reliable analytical approach using readily available spectroscopy and advanced chemometric tools.

Methods

The core of this research lies in applying and comparing various chemometric techniques to UV-Vis spectral data of tequilas. The study focuses on supervised classification methods, meaning the algorithms are trained on labeled data (known types of tequila). The methods explored include:

  • K-Nearest Neighbors (KNN): A simple, non-parametric method that classifies a sample based on the majority class of its nearest neighbors.
  • Potential Functions: A probabilistic method for class modeling.
  • PCA-DA (Principal Component Analysis - Discriminant Analysis): Combines dimensionality reduction with classification.
  • SIMCA (Soft Independent Modeling of Class Analogy): Builds individual PCA models for each class.
  • PLS-DA (Partial Least Squares - Discriminant Analysis): A multivariate regression technique adapted for classification.
  • CART (Classification and Regression Trees): Creates a tree-like structure of decisions for classification.
  • Random Forest (RF): An ensemble method that combines multiple decision trees for improved accuracy and robustness.
  • Support Vector Machines (SVM): A powerful method for finding optimal hyperplanes to separate classes, including non-linear separation using kernels.
  • SVM-RFE (Support Vector Machine - Recursive Feature Elimination): A variation of SVM that also performs feature selection.

The study also employed unsupervised methods like PCA and Hierarchical Clustering for initial data exploration and outlier detection.

Datasets

The study utilized a dataset of 170 certified tequila samples, which they claim is the largest set of certified tequilas considered in a single report to date. These samples were categorized as:

  • 65 white tequilas
  • 53 rested tequilas
  • 39 aged tequilas
  • 13 extra-aged tequilas

The data used for analysis consisted of UV-Vis spectra measured in the 190-700 nm range.

Key Insights and Ideas

  • UV-Vis Spectroscopy is Promising: The study demonstrates that simple UV-Vis spectroscopy, when combined with appropriate chemometric tools, can effectively differentiate between tequila types.
  • Non-Linear Methods Outperform Linear Ones: For this complex dataset, non-linear classification algorithms like Random Forest and SVM generally achieved higher accuracy than linear methods.
  • Importance of Spectral Regions: The research highlights that specific regions of the UV-Vis spectrum, particularly in the visible range (333-380 nm), are crucial for distinguishing between tequila types. These regions are thought to be related to compounds that migrate from oak barrels during aging.
  • Outlier Detection is Crucial: Identifying and understanding "outlying" samples (those that don't fit their labeled class) is important. These might indicate mislabeling or genuine anomalies, requiring further investigation.
  • Chemometrics is Key: The raw spectral data alone is not enough; advanced chemometric techniques are essential for extracting meaningful patterns and building predictive models.

Key Results

  • High Classification Accuracy: Non-linear methods achieved excellent results. Random Forest and SVM demonstrated accuracies higher than 0.98 for calibration and 0.94 for validation.
  • KNN and CART Also Performed Well: KNN and CART also showed strong performance, with accuracies above 0.98 and 0.94 respectively.
  • Potential Functions and SIMCA Were Less Successful: These methods struggled with class overlap and sample variability, resulting in lower accuracies and higher unclassified sample rates.
  • Identification of Misclassified Samples: The study successfully identified several samples that were mislabeled or showed unexpected spectral characteristics, underscoring the utility of these methods for quality control.

Limitations and Future Work

  • Limited Number of Extra-Aged Samples: The study grouped aged and extra-aged tequilas due to the small number of extra-aged samples (only 13). Future work could focus on obtaining more data for this category.
  • Chemical Interpretation of Spectral Regions: While specific spectral regions were identified as important, their exact chemical interpretation remains complex and requires further investigation.
  • Need for Portable Solutions: The authors suggest that further development could lead to portable, easy-to-implement devices for on-site authentication, aligning with green chemistry principles.
  • Further Validation: While the study used a significant number of certified samples, broader validation across more producers and regions would strengthen the findings.

Summary updated

BibTeX
@inproceedings{perezcaballero2017authentication,
  author = {Guadalupe Pérez‐Caballero and J.M. Andrade and Pablo M. Olmos and Yusdivia Molina-Román and Ivan Jimenez and J.J. Durán and Carlos Fernández-Lozano and Floriberto Miguel-Cruz},
  title = {{Authentication of tequilas using pattern recognition and supervised classification}},
  booktitle = {TrAC},
  year = {2017},
  doi = {10.1016/j.trac.2017.07.008},
}
Figure