Anomaly detection using machine learning models and similarity regularization
Disclosed herein are embodiments for anomaly detection using machine learning models (MLMs) and similarity regularization. An embodiment operates by obtaining data for a first product, a second product, and a target product. The data include a set of sparse data points for the target product. Next, similarity scores between the target product and the first product and the second product may be calculated. The embodiment further operates by generating a target MLM associated with the target product using a regularization penalty. The regularization penalty is based on the similarity scores and distances between a target set of coefficients for the target MLM and coefficients for a first MLM and a second MLM associated with the first product and the second product, respectively. The embodiment may then detect an anomaly associated with the target product by feeding the target MLM with a feature vector associated with the target product.
1. A method for anomaly detection using machine learning models and similarity regularization, comprising:
storing, by at least one processor, a plurality of data points for a plurality of products comprising a first product, a second product, and a target product in a memory, wherein the plurality of data points comprises a sparse set of data points for the target product;
calculating, by the at least one processor, in response to an indication from a trained machine learning (ML) model associated with the first product that a target ML model evaluation for the target product failed to detect an anomaly for the target product, a first similarity score between the first product and the target product and a second similarity score between the second product and the target product;
calibrating, by the at least one processor and in response to the sparse set of data points, the target ML model using a regularization penalty that is based on:
the first similarity score and the second similarity score,
a first distance between a first set of coefficients for the trained ML and a target set of coefficients for the target ML model, and
a second distance between a second set of coefficients for another trained ML model associated with the second product and the target set of coefficients; and
receiving, by the at least one processor, based on feeding a feature vector associated with the target product into the target ML model, an indication that the target ML model detected the anomaly for the target product.
2. The method of claim 1 , wherein the first product, the second product, and the target product are consumer packaged goods (CPGs).
3. The method of claim 1 , wherein the first product, the second product, and the target product are medical devices, and wherein the anomaly is failure of the target product.
4. The method of claim 1 , wherein calculating the first similarity score comprises:
calculating, by the at least one processor, a first cosine similarity between an embedding for the first product and an embedding for the target product;
comparing the first cosine similarity with a similarity threshold; and
in response to the first cosine similarity exceeding the similarity threshold, determining the first product and the target product are similar and setting the first similarity score to 1.
5. The method of claim 4 , wherein calculating the second similarity score comprises:
calculating, by the at least one processor, a second cosine similarity between an embedding for the second product and the embedding for the target product;
comparing the second cosine similarity with the similarity threshold; and
in response to the second cosine similarity falling below the similarity threshold, determining the second product and the target product are dissimilar and setting the second similarity score to 0.
6. The method of claim 1 , wherein the regularization penalty comprises:
a sum of at least a first contrastive loss function associated with the first distance and a second contrastive loss function associated with the second distance.
7. The method of claim 6 , wherein the first contrastive loss function comprises:
a product of Z T,1 d({right arrow over (β)} T , {right arrow over (β)} 1 ), wherein Z T,1 is the first similarity score between the first product and the target product, and wherein d({right arrow over (β)} T , {right arrow over (β)} 1 ) is the first distance.
8. The method of claim 6 , wherein the second contrastive loss function comprises:
a product of (1−Z T,2 )max {0,m−d({right arrow over (β)} T , {right arrow over (β)} 2 )}, wherein Z T,2 is the second similarity score between the second product and the target product, wherein d({right arrow over (β)} T , {right arrow over (β)} 2 ) is the second distance, and wherein m is a minimum margin parameter.
9. A system for anomaly detection using machine learning models and similarity regularization, comprising:
a memory; and
at least one processor coupled to the memory and configured to:
store a plurality of data points for a plurality of products comprising a first product, a second product, and a target product in the memory, wherein the plurality of data points comprises a sparse set of data points for the target product;
calculate, in response to an indication from a trained machine learning (ML) model associated with the first product that a target ML model evaluation for the target product failed to detect an anomaly for the target product, a first similarity score between the first product and the target product and a second similarity score between the second product and the target product;
calibrate, in response to the sparse set of data points, the target ML model using a regularization penalty that is based on:
the first similarity score and the second similarity score,
a first distance between a first set of coefficients for the trained ML and a target set of coefficients for the target ML model, and
a second distance between a second set of coefficients for another trained second ML model associated with the second product and the target set of coefficients; and
receive, based on feeding a feature vector associated with the target product into the target ML model, an indication that the target ML model detected the anomaly for the target product.
10. The system of claim 9 , wherein to calculate the first similarity score the at least one processor is further configured to:
calculate a first cosine similarity between an embedding for the first product and an embedding for the target product;
compare the first cosine similarity with a similarity threshold; and
in response to the first cosine similarity exceeding the similarity threshold, determine the first product and the target product are similar and setting the first similarity score to 1.
11. The system of claim 10 , wherein to calculate the second similarity score the at least one processor is further configured to:
calculate a second cosine similarity between an embedding for the second product and the embedding for the target product;
compare the second cosine similarity with the similarity threshold; and
in response to the second cosine similarity falling below the similarity threshold, determine the second product and the target product are dissimilar and setting the second similarity score to 0.
12. The system of claim 9 , wherein the regularization penalty comprises:
a sum of at least a first contrastive loss function associated with the first distance and a second contrastive loss function associated with the second distance.
13. The system of claim 12 , wherein the first contrastive loss function comprises:
a product of Z T,1 d ({right arrow over (β)} T , {right arrow over (β)} 1 ), wherein Z T,1 is the first similarity score between the first product and the target product, and wherein d({right arrow over (β)} T , {right arrow over (β)} 1 ) is the first distance.
14. The system of claim 12 , wherein the second contrastive loss function comprises:
a product of (1−Z T,2 )max{0, m−d({right arrow over (β)} T , {right arrow over (β)} 2 )}, wherein Z T,2 is the second similarity score between the second product and the target product, wherein d({right arrow over (β)} T , {right arrow over (β)} 2 ) is the second distance, and wherein m is a minimum margin parameter.
15. A non-transitory computer-readable medium having instructions stored thereon for anomaly detection using machine learning models and similarity regularization, the instructions, when executed by at least one computing device, cause the at least one computing device to perform operations comprising:
storing a plurality of data points for a plurality of products comprising a first product, a second product, and a target product in a memory, wherein the plurality of data points comprises a sparse set of data points for the target product;
calculating, in response to an indication from a trained machine learning (ML) model associated with the first product that a target ML model evaluation for the target product failed to detect an anomaly for the target product, a first similarity score between the first product and the target product and a second similarity score between the second product and the target product;
calibrating, in response to the sparse set of data points, the target ML model using a regularization penalty that is based on:
the first similarity score and the second similarity score,
a first distance between a first set of coefficients for the trained ML and a target set of coefficients for the target ML model, and
a second distance between a second set of coefficients for another trained ML model associated with the second product and the target set of coefficients; and
receiving, based on feeding a feature vector associated with the target product into the target ML model, an indication that the target ML model detected the anomaly for the target product.
16. The non-transitory computer-readable medium of claim 15 , wherein calculating the first similarity score further comprises:
calculating a first cosine similarity between an embedding for the first product and an embedding for the target product;
comparing the first cosine similarity with a similarity threshold; and
in response to the first cosine similarity exceeding the similarity threshold, determining the first product and the target product are similar and setting the first similarity score to 1.
17. The non-transitory computer-readable medium of claim 16 , wherein calculating the second similarity score further comprises:
calculating a second cosine similarity between an embedding for the second product and the embedding for the target product;
comparing the second cosine similarity with the similarity threshold; and
in response to the second cosine similarity falling below the similarity threshold, determining the second product and the target product are dissimilar and setting the second similarity score to 0.
18. The non-transitory computer-readable medium of claim 15 , wherein the regularization penalty comprises:
a sum of at least a first contrastive loss function associated with the first distance and a second contrastive loss function associated with the second distance.
19. The non-transitory computer-readable medium of claim 18 , wherein the first contrastive loss function comprises:
a product of Z T,1 d({right arrow over (β)} T , {right arrow over (β)} 1 ), wherein Z T,1 is the first similarity score between the first product and the target product, and wherein d({right arrow over (β)} T , {right arrow over (β)} 1 ) is the first distance.
20. The non-transitory computer-readable medium of claim 18 , wherein the second contrastive loss function comprises:
a product of (1−Z T,2 )max{0, m−d({right arrow over (β)} T , {right arrow over (β)} 2 )}, wherein Z T,2 is the second similarity score between the second product and the target product, wherein d({right arrow over (β)} T , {right arrow over (β)} 2 ) is the second distance, and wherein m is a minimum margin parameter.