Automatically determining items to include in a variant group
A method including obtaining image data and attribute information of a first item in an item catalog. The method also can include generating candidate variant items from the item catalog for the first item using a combination of (a) a k-nearest neighbors approach to search for first candidate variant items based on text embeddings for the attribute information of the first item, and (b) an elastic search approach to search for second candidate variant items based on image embeddings for the image data of the first item. The method additionally can include performing respective classifications based on respective pairs comprising the first item and each of the candidate variant items to filter the candidate variant items. The method further can include determining a respective distance between the first item and each of the candidate variant items, as filtered. The method additionally can include determining one or more items in the candidate variant items, as filtered, to include in a variant group for the first item, based on a decision function using a predetermined threshold and the respective distance for the each of the candidate variant items, as filtered. Other embodiments are described.
1. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform operations comprising:
obtaining image data and attribute information of a first item in an item catalog;
generating candidate variant items from the item catalog for the first item using a combination of (a) a k-nearest neighbors approach to search for first ones of the candidate variant items based on text embeddings for the attribute information of the first item, and (b) an elastic search approach to search for second ones of the candidate variant items based on image embeddings for the image data of the first item;
performing respective classifications based on respective pairs comprising the first item and each of the first ones and the second ones of the candidate variant items to filter the candidate variant items;
determining a respective distance between the first item and each of the first ones and the second ones of the candidate variant items, as filtered;
determining one or more items in the first ones and the second ones of the candidate variant items, as filtered, to include in a variant group for the first item, based on a decision function using a predetermined threshold and the respective distance for each of the first ones and the second ones of the candidate variant items, as filtered, wherein when the respective distance is below the predetermined threshold, determining the one or more items comprises the one or more items of the first ones and the second ones of the candidate variant items in the variant group for the first item, wherein the variant group for the first item is used at least in part as a same base variant group on a website, and wherein the same base variant group and the variant group for the first item is displayed on a webpage on the website; and
determining whether at least one of one or more respective items included in the same base variant group is a variant of another of the items included in the same base variant group by automatically detecting inconsistencies in product type data listed in the item catalog for the one or more respective items in the same base variant group on the website, wherein the inconsistencies comprise different product types within the same base variant group; and wherein the inconsistencies further comprise hard negatives in which respective pairs of items in the same base variant group have similar titles but are not variants of the same base variant group.
2. The system of claim 1 , wherein the computing instructions, when executed on the one or more processors, further cause the one or more processors to perform operations comprising:
before generating the candidate variant items, performing a pre-processing comprising:
generating the text embeddings from the attribute information of the first item; and
generating the image embeddings from the image data of the first item; and
splitting each non-variant item of the one or more respective items from the same base variant group on the website.
3. The system of claim 2 , wherein:
the text embeddings are generated using an LSTM model.
4. The system of claim 2 , wherein:
the image embeddings are generated using a VGG16 model.
5. The system of claim 1 , wherein:
the k-nearest neighbors approach comprises using a non-metric space library (NMSLIB); and
the inconsistencies further comprise a different brand.
6. The system of claim 1 , wherein:
the combination comprises a union of the first ones of the candidate variant items generated using the k-nearest neighbors approach and the second ones of the candidate variant items generated using the elastic search approach.
7. The system of claim 1 , wherein performing the respective classifications further comprises:
performing the respective classifications using a logistic regression model with L1 regularization based on respective product type classification embeddings for the candidate variant items.
8. The system of claim 1 , wherein determining the respective distance further comprises:
determining the respective distance using a Siamese network.
9. The system of claim 1 , wherein the computing instructions, when executed on the one or more processors, further cause the one or more processors to perform an operation comprising:
determining a coherence score for the variant group to select one or more supergroups from among multiple variant groups, wherein the multiple variant groups comprise the variant group.
10. The system of claim 1 , wherein the computing instructions, when executed on the one or more processors, further cause the one or more processors to perform an operation comprising:
determining a Jaccard similarity score between multiple variant groups, wherein the multiple variant groups comprise the variant group.
11. A method being implemented via execution of computing instructions configured to run at one or more processors stored on one or more non-transitory computer-readable media, the method comprising:
obtaining image data and attribute information of a first item in an item catalog;
generating candidate variant items from the item catalog for the first item using a combination of (a) a k-nearest neighbors approach to search for first ones of the candidate variant items based on text embeddings for the attribute information of the first item, and (b) an elastic search approach to search for second ones of the candidate variant items based on image embeddings for the image data of the first item;
performing respective classifications based on respective pairs comprising the first item and each of the first ones and the second ones of the candidate variant items to filter the candidate variant items;
determining a respective distance between the first item and each of the first ones and the second ones of the candidate variant items, as filtered;
determining one or more items in the first ones and the second ones of the candidate variant items, as filtered, to include in a variant group for the first item, based on a decision function using a predetermined threshold and the respective distance for each of the first ones and the second ones of the candidate variant items, as filtered, wherein when the respective distance is below the predetermined threshold, determining the one or more items comprises the one or more items of the first ones and the second ones of the candidate variant items in the variant group for the first item, wherein the variant group for the first item is used at least in part as a same base variant group on a website, and wherein the same base variant group and the variant group for the first item is displayed on a webpage on the website; and
determining whether at least one of one or more respective items included in the same base variant group is a variant of another of the items included in the same base variant group by automatically detecting inconsistencies in product type data listed in the item catalog for the one or more respective items in the same base variant group on the web site, wherein the inconsistencies comprise different product types within the same base variant group; and wherein the inconsistencies further comprise hard negatives in which respective pairs of items in the same base variant group have similar titles but are not variants of the same base variant group.
12. The method of claim 11 , further comprising:
before generating the candidate variant items, performing a pre-processing comprising:
generating the text embeddings from the attribute information of the first item; and
generating the image embeddings from the image data of the first item; and
splitting each non-variant item of the one or more respective items from the same base variant group on the website.
13. The method of claim 12 , wherein:
the text embeddings are generated using an LSTM model.
14. The method of claim 12 , wherein:
the image embeddings are generated using a VGG16 model.
15. The method of claim 11 , wherein:
the k-nearest neighbors approach comprises using a non-metric space library (NMSLIB); and
the inconsistencies further comprise a different brand.
16. The method of claim 11 , wherein:
the combination comprises a union of the first ones of the candidate variant items generated using the k-nearest neighbors approach and the second ones of the candidate variant items generated using the elastic search approach.
17. The method of claim 11 , wherein performing the respective classifications further comprises:
performing the respective classifications using a logistic regression model with L1 regularization based on respective product type classification embeddings for the candidate variant items.
18. The method of claim 11 , wherein determining the respective distance further comprises:
determining the respective distance using a Siamese network.
19. The method of claim 11 , further comprising:
determining a coherence score for the variant group to select one or more supergroups from among multiple variant groups, wherein the multiple variant groups comprise the variant group.
20. The method of claim 11 , further comprising:
determining a Jaccard similarity score between multiple variant groups, wherein the multiple variant groups comprise the variant group.