Exam AWS Certified Machine Learning Engineer - Associate MLA-C01 topic 1 question 14 discussion

https://docs.aws.amazon.com/en_kr/sagemaker/latest/dg/lightgbm.html

Answer is B

1. Clase desbalanceada: LightGBM (Light Gradient Boosted Machine) es muy eficaz para trabajar con datasets desbalanceados, gracias a su capacidad para ajustar los pesos de clases y usar técnicas como el weighted loss o boosting adaptativo. 2. Interdependencia entre variables: LightGBM puede capturar relaciones no lineales e interacciones entre variables gracias a su estructura basada en árboles de decisión. Esto lo hace más adecuado que modelos lineales como el Linear Learner, que solo captura relaciones lineales. 3. No captura de patrones complejos: La descripción indica que el algoritmo actual no está capturando patrones subyacentes complejos, lo que sugiere la necesidad de un modelo más robusto como LightGBM, capaz de modelar relaciones complejas y no lineales en los datos

Fraud detection is a binary classification problem, and Linear Learner is designed for classification tasks. Although LightGBM can handle binary classification tasks, including fraud detection and it is actually widely used for fraud detection, it is not available as a built-in SageMaker algorithm. Linear Learner is a built-in SageMaker algorithm.

because linear learner is a built-in algorithm while lgbm is not

Linear learner is a built-in algorithm whereas LightGBM can be used via custom container.

Linear learner is a built-in algorithm where LightBM is not

ChatGPT say it's A: LightGBM

Is supported by Sagemaker

Linear Learner is a built-in algorithm provided by SageMaker for supervised learning tasks like regression and classification. LightGBM is not a built-in algorithm in Amazon SageMaker. While it is a strong gradient-boosting algorithm, it would need to be implemented as a custom script in SageMaker, which increases operational overhead.

https://docs.aws.amazon.com/en_kr/sagemaker/latest/dg/lightgbm.html

In an ideal scenario, for a problem with these characteristics (fraud detection, class imbalance, feature interdependencies, complex patterns), a tree-based ensemble method like XGBoost (which is a SageMaker built-in algorithm) would be more suitable. XGBoost can handle non-linear relationships, is robust to class imbalance with proper tuning, and can capture complex patterns in the data. However, given the options provided and the requirement to use a SageMaker built-in algorithm, the Linear learner is the best available choice among these options for this specific fraud detection task.

A. Light BGM : It's suitable model, but not built-in model for SageMaker. Answer B. Linear learner : suitable model, built-in model for SageMaker. C. K-means clustering : groups similar data points, not suitable for classification problems, and it's unsupervised learning algorithm so doesn't fit in this case(fraud detection). D. Neural Topic Model: used for topic modeling and document classification, not suitable for fraud detection

Here's why LightGBM is the most suitable algorithm for this fraud detection task: Handling Class Imbalance: LightGBM is particularly effective at handling imbalanced datasets, which is a key issue mentioned in the problem statement. It has built-in mechanisms to deal with class imbalance. Feature Interdependencies: LightGBM can capture complex feature interactions through its tree-based structure, addressing the issue of feature interdependencies mentioned in the problem. Capturing Underlying Patterns: As an advanced gradient boosting framework, LightGBM is excellent at capturing complex patterns in data, which the current algorithm is struggling with. Suitable for Fraud Detection: LightGBM is widely used in fraud detection tasks due to its high performance and ability to handle large datasets efficiently. Handling Various Data Types: It can work well with the mix of data types likely present in transaction logs, customer profiles, and database tables.

We have an unbalanced dataset, this means we have labelled dataset thus we are going to use a supervised model training. This reduce options to A and B (K-means and NTM are unsupervised). Both LightGBM and Linear Learner provides hyperparameter to manage unbalanced datasets, respectively "scale-pos_weight" and "positive_example_weight_mult". I would go for LightGBM as this algorithm is more suited to handle complex relationship among features, while Linear Learner learns a linear function, or, for classification problems, a linear threshold function, and maps a vector x to an approximation of the label y.

Linear Learner. LightGBM is NOT a built in algorithm which the question asks for.

This is a binary classification problem so LightGBM so be used. Other algorithms are not for binary classification.