Method and system for performing product matching on an e-commerce platform
An apparatus for performing product matching may include: a processor configured to receive a search as product item for searching for a catalog product that matches a target product; classify the target product in a product taxonomy tree including a plurality of taxonomy nodes, by identifying a taxonomy node to which the target product belongs, among the taxonomy nodes; obtain product data associated with the target product from an internal source and an external source and measure data quality; extract attributes from the product data based on a machine learning model; validate the attribute in response to the attribute corresponding to defined mandatory attribute of the taxonomy node to which the target product belongs, and provide a search result based on the validated attributes; re-rank searched results governed by relevancy score and display matched set of products to the customer above defined re-rank confidence score.
1 . An apparatus for performing product matching, the apparatus comprising:
a memory storing instructions; and
a processor configured to execute the instructions to:
receive information of a target product in an e-commerce system;
classify the target product in a product taxonomy tree including a plurality of taxonomy nodes, by identifying a taxonomy node to which the target product belongs, among the taxonomy nodes;
obtain product data associated with the target product from an internal source and an external source, the product data including an image associated with the target product;
extract image and text attributes from the product data based on image extraction and text extraction, wherein the image extraction comprises image segmentation and color extraction process to identify, as the image attributes, colors of a plurality of segmented portions corresponding to the target product in the image, and wherein the text extraction comprises
receiving a text input of the product data,
obtaining an annotation result of the text input,
automatically generating synthetic title templates using a neural machine translation (NMT)-based pipeline, wherein the NMT-based pipeline is configured to identify attribute keys from the text input as input, and output the synthetic title templates that are filled with random values,
inputting the annotation result to a machine learning model to obtain the text attributes as outputs of the machine learning model, the machine learning model being trained on an attribute dictionary storing product identification numbers, brand information, color information, and size information,
wherein the machine learning model is trained by validating the outputs of the machine learning model based on the synthetic title templates;
measure a product data quality of the extracted image and text attributes based on a ratio between a number of the extracted image and text attributes, and a number of common features that belong to both the extracted image and text attributes and mandatory attributes, by identifying the mandatory attributes from the taxonomy node to which the target product belongs, among predetermined mandatory attributes;
validate a first set of image and text attributes, among the extracted image and text attributes, in response to the product data quality of the first set of image and text attributes being greater than or equal to a data quality threshold, and the first set of image and text attributes output by the machine learning model matching the attribute dictionary;
provide the validated first set of image and text attributes for further processing including a search;
invalidate a second set of image and text attributes, among the extracted image and text attributes, in response to the product data quality of the second set of image and text attributes being less than the data quality threshold, and the second set of image and text attributes output by the machine learning model not matching the attribute dictionary;
block the invalidated second set of image and text attributes for the further processing, and store the invalidated second set of image and text attributes in a separate database for inspection;
conduct the search of a product catalog based on the validated first set of image and text attributes without using the invalidated second set of image and text attributes; and
provide a user interface that shows results of the search based on the validated first set of image and text attributes without using the invalidated second set of image and text attributes,
wherein the predetermined mandatory attributes are stored in the memory.
2 . The apparatus of claim 1 , wherein the processor is further configured to:
extract image features and red (R), green (G), and blue (B) values (RGB color values) from the image, as the image attributes extracted from the product data, based on the image segmentation and color extraction process.
3 . The apparatus of claim 1 , wherein the image segmentation and color extraction process comprises:
identifying the colors from the plurality of segmented portions of the image corresponding to the target product;
clustering the identified colors based on red, green, and blue values of the identified colors; and
storing a cluster of the identified colors as the image features.
4 . A method for performing, by at least one computer, the following steps:
receiving a target product in an e-commerce system;
classifying the target product in a product taxonomy tree including a plurality of taxonomy nodes, by identifying a taxonomy node to which the target product belongs, among the taxonomy nodes;
obtaining product data associated with the target product from an internal source and an external source, the product data including an image associated with the target product;
extracting image and text attributes from the product data based on image extraction and text extraction, wherein the extracting the image attributes comprises performing image segmentation and color extraction process to identify, as the image attributes, colors of a plurality of segmented portions corresponding to the target product in the image, and wherein the text extraction comprises
receiving a text input of the product data,
obtaining an annotation result of the text input,
automatically generating synthetic title templates using a neural machine translation (NMT)-based pipeline, wherein the NMT-based pipeline is configured to identify attribute keys from the text input as input, and output the synthetic title templates that are filled with random values, and
inputting the annotation result to a machine learning model to obtain the text attributes as outputs of the machine learning model, the machine learning model being trained on an attribute dictionary storing product identification numbers, brand information, color information, and size information,
wherein the machine learning model is trained by validating the outputs of the machine learning model based on the synthetic title templates;
measuring a product data quality of the extracted image and text attributes based on a ratio between a number of the extracted image and text attributes, and a number of common features that belong to both the extracted image and text attributes and mandatory attributes, by identifying the mandatory attributes from the taxonomy node to which the target product belongs, among predetermined mandatory attributes;
validating a first set of image and text attributes, among the extracted image and text attributes, in response to the product data quality of the first set of image and text attributes being greater than or equal to a data quality threshold, and the first set of image and text attributes output by the machine learning model matching the attribute dictionary;
providing the validated first set of image and text attributes for further processing including a search;
invalidating a second set of image and text attributes, among the extracted image and text attributes, in response to the product data quality of the second set of image and text attributes being less than the data quality threshold, and the second set of image and text attributes output by the machine learning model not matching the attribute dictionary;
block the invalidated second set of image and text attributes for the further processing, and store the invalidated second set of image and text attributes in a separate database for inspection;
conducting a search of a product catalog based on the validated first set of image and text attributes without using the invalidated second set of image and text attributes; and
providing a user interface that shows results of the search based on the validated first set of image and text attributes, without using the invalidated second set of image and text attributes,
wherein the predetermined mandatory attributes are stored in a memory.
5 . The method of claim 4 , further comprising:
extract image features and red (R), green (G), and blue (B) values (RGB color values) from the image, as the image attributes extracted from the product data, based on the image segmentation and color extraction process.
6 . The method of claim 4 , wherein the color extraction process comprises:
identifying colors as red (R), green (G), and blue (B) values (RGB color values) from the plurality of segmented portions of the image corresponding to the target product;
clustering the extracted colors based on the RGB color values of the identified colors; and
storing a cluster of the identified colors as the image features.