One-Class Classification Algorithms

One-class classification is a type of machine learning algorithm used to identify instances of a specific class among all other possible instances. It is particularly useful in scenarios where data for one class is abundant, while data for other classes is scarce or non-existent.

Training

Parameters

  • Reference class: The index of the reference class used for training the model. This class represents the “normal” data that the model will learn to recognize.

  • Algorithm: The algorithm to be used for one-class classification. Options include:

    The algorithm’s specific parameters are explained in detail below.

  • Confidence Limit: The percentile value used to set the confidence limit for classification. This parameter is crucial as it determines the threshold for identifying anomalies based on their deviation from the learned map. A lower percentile value will result in a stricter threshold, potentially flagging more data points as anomalies, while a higher percentile value will be more lenient, possibly missing some subtle anomalies.

    A typical value is 95%, which means that the model will classify data points that fall outside the 95:sup:th percentile of the reference class distribution as anomalies. This helps in identifying outliers or unusual patterns in the data, which can be critical for quality control and anomaly detection tasks.

    Another common value is 80, which is often used in scenarios where only a small core of the reference class is considered normal, and deviations from this core are significant. This setting allows for a more sensitive detection of anomalies, ensuring that even slight deviations from the expected pattern are flagged for further investigation.

Self-Organizing Map for Quality Control (SOMQC)

The implementation of Self-Organizing Maps for Quality Control (SOMQC) is based on methodologies detailed in two significant research papers: Self-organising maps for the exploration and classification of thin-layer chromatograms[1] itself based on Self-Organizing Map Quality Control Index[2]. Self-Organizing Maps (SOMs), also known as Kohonen networks, are a type of artificial neural network designed for learning tasks. In the context of quality control, SOMs are particularly effective for dimensionality reduction and clustering, transforming complex, high-dimensional data into a more manageable, typically two-dimensional space, while preserving the intrinsic topological relationships of the original data. This means that data points that are similar in the high-dimensional space remain close to each other in the reduced-dimensional map.

For quality control purposes, SOMs are trained using datasets that represent standard operating conditions or acceptable quality benchmarks. This training enables the network to learn and recognize typical patterns within the data. Once the SOM is trained, it can be utilized to assess new data points. Any significant deviations from the established patterns can be identified as anomalies, signaling potential quality issues or deviations from the norm. This capability makes SOMs an invaluable tool in production environments for monitoring and ensuring high-quality standards. The visual output of SOMs further enhances their utility, providing an intuitive representation of complex data that aids quality control professionals in quickly understanding and responding to insights derived from the model. This approach leverages the detailed methodologies and findings from the referenced papers to enhance the effectiveness of SOMs in quality control applications.

For details on the quality control methods used in SOMQC, refer to the section on Quality Control in Machine Learning Models. For the interpretation of the charts generated by SOMQC, refer to the section on One-Class Classification: Interpreting Training Quality Charts.

Hyperparameters

  • Learning Start: The starting value for scaling the data during the learning process. Scaling is often necessary to normalize the data, ensuring that all features contribute equally to the analysis. This parameter defines the initial scale applied to the data, which can influence how the self-organizing map is formed and how sensitive it is to variations in the input data.

  • Learning End: The ending value for scaling the data, which works in conjunction with the Learning Start parameter to define the range of the scaled data. This parameter helps to control the final scale of the data after the learning process is complete. Together, the Learning Start and Learning End parameters ensure that the data is appropriately normalized, facilitating more accurate and reliable anomaly detection.

Note

The SOM map or grid size is automatically determined based on the number of samples in the reference class. This dynamic sizing ensures that the map is appropriately scaled to capture the underlying patterns in the data without being too sparse or too dense, which could affect the quality of the classification.

Note

Some other hyperparameters are fixed; they are not exposed to the user as they were found to have little impact on the quality of the results. These include:

  • learning rate: linear

  • neighborhood function: gaussian

  • number of iterations or epochs: 100 epochs

One-Class Support Vector Machine (SVM)

One-Class SVM is a one-class classification algorithm based on Support Vector Machines (SVM). It is effective for novelty detection and outlier detection tasks, where the goal is to identify data points that differ significantly from the majority of the training data. For reference, see the original papers on One-Class SVM: Support vector method for novelty detection[3] and Estimating the Support of a High-Dimensional Distribution[4].

One-Class SVM works by learning a decision boundary around the reference class data. The algorithm aims to separate the majority of the data from the origin in the feature space, effectively creating a boundary that encloses the “normal” data. Data points that fall outside this boundary are classified as anomalies.

For details on the quality control methods used in One-Class SVM, refer to the section on Quality Control in Machine Learning Models. For the interpretation of the charts generated by One-Class SVM, refer to the section on One-Class Classification: Interpreting Training Quality Charts.

Warning

Because of the nature of Jackknife with One-Class SVM, many models will be trained and then saved. This will lead it larger disk space usage. This will impact the size of exported Comparison files.

Hyperparameters

  • Kernel: The kernel type to be used in the algorithm. The kernel defines the transformation of the input data into a higher-dimensional space, where the separation between classes is more distinct. Possible values are:

    • Linear: A linear kernel, which is suitable for linearly separable data.

    • Polynomial: A polynomial kernel, which allows for non-linear decision boundaries.

    • RBF (Radial Basis Function): A Gaussian kernel, which is effective for capturing complex relationships in the data.

    • Sigmoid: A sigmoid kernel, which can be used for certain types of data distributions.

    With kernel types ‘rbf’, ‘poly’, and ‘sigmoid’, one additional parameter is available:

    • Gamma: Gamma defines how far the influence of a single training example reaches, with low values meaning ‘far’ and high values meaning ‘close’. Possible values are:

      • \(\frac{1}{n_{\text{features}} \times X_\text{Var}}\): A gamma value that scales based on the number of features and the variance of the data.

      • \(\frac{1}{n_{\text{features}}}\): A gamma value that is inversely proportional to the number of features, providing a baseline scaling.

      • a custom value: A user-defined gamma value that can be set to tailor the model’s sensitivity to the data distribution.