Enhanced PTID prediction and accurate fitment using a PTID understanding model

1 . A system comprising a processor and a non-transitory computer-readable medium storing computing instructions that, when executed on the processor, cause the processor to perform operations comprising:

tokenizing a query into tokens;

transforming the tokens into embeddings;

passing the embeddings through a classification layer to produce log its;

applying a sigmoid function to convert the log its into a probability vector for Part Terminology Identifiers (PTIDs) associated with one or more predetermined domains;

filtering the probability vector to retain PTIDs associated with the one or more predetermined domains based on respective probabilities that are at least equal to a predetermined probability threshold, wherein a sum of the respective probabilities is equal to a predetermined sum value;

ranking the PTIDs based on their respective probabilities; and

applying business logic to the ranked PTIDs using a PTID Understanding Model to determine a final result that is output to a fitment service, wherein the PTID Understanding Model is pretrained to trigger at least one of a fitment determination process or a query results filter process for the one or more predetermined domains when the final result is output as valid,

wherein the business logic comprises:

when a top ranked PTID is invalid and a respective probability is at least equal to a predetermined confidence level, the final result is output as invalid;

when the top ranked PTID is valid and is at least equal to the predetermined confidence level, or when the top ranked PTID is invalid but the respective probability is less than the predetermined confidence level, a sum of all the respective probabilities for valid ranked PTIDs that are included in the probability vector, as filtered, is determined; and

when the sum of all the respective probabilities of the valid ranked PTIDs is at least equal to a predetermined cutoff value, the valid ranked PTIDs are included in the final result, wherein the final result is output as valid.

2 . The system of claim 1 , wherein the predetermined probability threshold is approximately 1%, wherein the predetermined confidence level is approximately 50%, and wherein the predetermined cutoff value is approximately 50%.

3 . The system of claim 1 , wherein the fitment determination process comprises:

determining a fitment widget to display to a user by verifying that the query relates to the one or more predetermined domains;

retrieving a list of recalled PTIDs corresponding to the valid ranked PTIDs and respective product counts from a presentation service, using a part finder application to identify which of the valid ranked PTIDs are fitment-enabled, selecting fitment-enabled valid ranked PTIDs with highest numbers of associated products, and using the part finder application to obtain a fitment form for the selected fitment-enabled valid ranked PTIDs; and

populating the determined fitment widget on a display interface of an e-commerce platform based on the obtained fitment form, prompting the user to input product-specific information.

4 . The system of claim 3 , wherein the query results filter process comprises:

refining query results by adding a fitment filter parameter to a uniform resource locator, passing the uniform resource locator to a search engine, and incorporating the fitment filter parameter as an additional query parameter, wherein the fitment filter parameter is based on the input product-specific information.

5 . The system of claim 1 , wherein the PTID Understanding Model comprises a Bidirectional Encoder Representations from Transformers (BERT) model, and wherein the PTID Understanding Model is a weighted single-label multi-class model.

6 . The system of claim 1 , wherein the operations further comprise:

constructing a training dataset for training the PTID Understanding Model to trigger the at least one of the fitment determination process or the query results filter process for the one or more predetermined domains when the final result is output as valid,

wherein the constructing the training dataset comprises:

aggregating user engagement data from search logs, primarily selecting records in which products were added to a cart, indicating positive engagement;

filtering the user engagement data from the search logs to retain training queries and respective training data associated with the one or more predetermined domains;

standardizing the filtered user engagement data by assigning a predetermined placeholder value to included products without valid PTIDs; and

converting the standardized and filtered user engagement data into “query-PTID” relationships suitable for the training the PTID Understanding Model that is a weighted single-label multi-class model.

7 . The system of claim 6 , wherein the operations further comprise:

preprocessing the training queries by tokenizing and converting the training queries into training embeddings using a Bidirectional Encoder Representations from Transformers (BERT) model included in the PTID Understanding Model;

adding a fully connected layer to the BERT model to convert the training embeddings into training predictions for predetermined training PTIDs; and

outputting predicted training probability distributions for the predetermined training PTIDs to rank the predetermined training PTIDs according to respective predicted probabilities for each of the training queries.

8 . The system of claim 7 , wherein the operations further comprise:

utilizing GPU acceleration for the training the PTID Understanding Model to trigger the at least one of the fitment determination process or the query results filter process for the one or more predetermined domains when the final result is output as valid;

optimizing hyperparameters of the PTID Understanding Model, wherein the hyperparameters comprise learning rate, batch size, and weight decay; and

using a cross-entropy loss function to improve accuracy of PTIDs predicted by the PTID Understanding Model by comparing the predicted training probability distributions with true PTID labels.

9 . The system of claim 8 , wherein the operations further comprise:

selecting a subset of the training queries from the constructed training dataset to create a test sample;

using a remainder of the training queries from the training dataset for the training the PTID Understanding Model to trigger the at least one of the fitment determination process or the query results filter process for the one or more predetermined domains when the final result is output as valid; and

testing the PTID Understanding Model that is pretrained on the test sample to evaluate an accuracy of the PTIDs predicted by the PTID Understanding Model.

10 . A computer-implemented method comprising:

tokenizing a query into tokens;

transforming the tokens into embeddings;

passing the embeddings through a classification layer of a Bidirectional Encoder Representations from Transformers (BERT) model to produce log its;

applying a sigmoid function to convert the log its into a probability vector for Part Terminology Identifiers (PTIDs) associated with one or more predetermined domains;

filtering the probability vector to retain PTIDs based on respective probabilities that are at least equal to a predetermined probability threshold, wherein a sum of the respective probabilities is equal to a predetermined sum value;

ranking the PTIDs based on their respective probabilities;

applying business logic to the ranked PTIDs using a PTID Understanding Model to determine a final result that is output to a fitment service, wherein the PTID Understanding Model is pretrained to trigger at least one of a fitment determination process or a query results filter process for the one or more predetermined domains when the final result is output as valid,

wherein the business logic comprises:

when a top ranked PTID is invalid and a respective probability is at least equal to a predetermined confidence level, the final result is output as invalid;

when the top ranked PTID is valid and is at least equal to the predetermined confidence level, or when the top ranked PTID is invalid but the respective probability is less than the predetermined confidence level, a sum of all the respective probabilities for valid ranked PTIDs that are included in the probability vector, as filtered, is determined; and

when the sum of all the respective probabilities of the valid ranked PTIDs is at least equal to a predetermined cutoff value, the valid ranked PTIDs are included in the final result, wherein the final result is output as valid.

11 . The computer-implemented method of claim 10 , wherein the predetermined probability threshold is approximately 1%, wherein the predetermined confidence level is approximately 50%, and wherein the predetermined cutoff value is approximately 50%.

12 . The computer-implemented method of claim 10 , wherein the fitment determination process comprises:

determining a fitment widget to display to a user by verifying that the query relates to the one or more predetermined domains;

retrieving a list of recalled PTIDs corresponding to the valid ranked PTIDs and respective product counts from a presentation service,

using a part finder application to identify which of the valid ranked PTIDs are fitment-enabled,

selecting fitment-enabled valid ranked PTIDs with highest numbers of associated products, and using the part finder application to obtain a fitment form for the selected fitment-enabled valid ranked PTIDs; and

populating the determined fitment widget on a display interface of an e-commerce platform based on the obtained fitment form, prompting the user to input product-specific information.

13 . The computer-implemented method of claim 12 , wherein the query results filter process comprises:

refining query results by adding a fitment filter parameter to a uniform resource locator, passing the uniform resource locator to a search engine, and incorporating the fitment filter parameter as an additional query parameter, wherein the fitment filter parameter is based on the input product-specific information.

14 . The computer-implemented method of claim 10 , wherein the PTID Understanding Model comprises the BERT model, and wherein the PTID Understanding Model is a weighted single-label multi-class model.

15 . The computer-implemented method of claim 10 , further comprising:

constructing a training dataset for training the PTID Understanding Model to trigger the at least one of the fitment determination process or the query results filter process for the one or more predetermined domains when the final result is output as valid,

wherein the constructing the training dataset comprises:

aggregating user engagement data from search logs, primarily selecting records in which products were added to a cart, indicating positive engagement;

filtering the user engagement data from the search logs to retain training queries and respective training data associated with the one or more predetermined domains;

standardizing the filtered user engagement data by assigning a predetermined placeholder value to included products without valid PTIDs; and

converting the standardized and filtered user engagement data into “query-PTID” relationships suitable for the training the PTID Understanding Model that is a weighted single-label multi-class model.

16 . The computer-implemented method of claim 15 , further comprising:

preprocessing the training queries by tokenizing and converting the training queries into training embeddings using the BERT model included in the PTID Understanding Model;

adding a fully connected layer to the BERT model to convert the training embeddings into training predictions for predetermined training PTIDs; and

outputting predicted training probability distributions for the predetermined training PTIDs to rank the predetermined training PTIDs according to respective predicted probabilities for each of the training queries.

17 . A non-transitory computer-readable medium storing computing instructions that, when executed on a processor, cause the processor to perform operations comprising a computer-implemented method, the computer-implemented method comprising:

tokenizing a query into tokens;

transforming the tokens into embeddings;

passing the embeddings through a classification layer to produce log its;

applying a sigmoid function to convert the log its into a probability vector for Part Terminology Identifiers (PTIDs) associated with one or more predetermined domains;

filtering the probability vector to retain PTIDs based on respective probabilities that are at least equal to a predetermined probability threshold, wherein a sum of the respective probabilities is equal to a predetermined sum value;

ranking the PTIDs based on their respective probabilities;

applying business logic to the ranked PTIDs using a PTID Understanding Model to determine a final result that is output to a fitment service, wherein the PTID Understanding Model is pretrained to trigger at least one of a fitment determination process or a query results filter process for the one or more predetermined domains when the final result is output as valid,

wherein the business logic comprises:

when a top ranked PTID is invalid and a respective probability is at least equal to a predetermined confidence level, the final result is output as invalid;

when the top ranked PTID is valid and is at least equal to the predetermined confidence level, or when the top ranked PTID is invalid but the respective probability is less than the predetermined confidence level, a sum of all the respective probabilities for valid ranked PTIDs that are included in the probability vector, as filtered, is determined; and

when the sum of all the respective probabilities of the valid ranked PTIDs is at least equal to a predetermined cutoff value, the valid ranked PTIDs are included in the final result, wherein the final result is output as valid.

18 . The non-transitory computer-readable medium of claim 17 , wherein the predetermined probability threshold is approximately 1%, wherein the predetermined confidence level is approximately 50%, and wherein the predetermined cutoff value is approximately 50%.

19 . The non-transitory computer-readable medium of claim 18 , wherein the fitment determination process comprises:

determining a fitment widget to display to a user by verifying that the query relates to the one or more predetermined domains;

retrieving a list of recalled PTIDs corresponding to the valid ranked PTIDs and respective product counts from a presentation service,

using a part finder application to identify which of the valid ranked PTIDs are fitment-enabled,

selecting fitment-enabled valid ranked PTIDs with highest numbers of associated products, and using the part finder application to obtain a fitment form for the selected fitment-enabled valid ranked PTIDs; and

populating the determined fitment widget on a display interface of an e-commerce platform based on the obtained fitment form, prompting the user to input product-specific information, and

wherein the query results filter process comprises:

refining query results by adding a fitment filter parameter to a uniform resource locator, passing the uniform resource locator to a search engine, and incorporating the fitment filter parameter as an additional query parameter, wherein the fitment filter parameter is based on the input product-specific information.