公开(公告)号：US11861664B2

公开(公告)日：2024-01-02

申请号：US17955781

申请日：2022-09-29

Applicant: Adobe Inc.

Inventor： Anirban Basu ,  Tathagata Sengupta ,  Kunal Kumar Jain ,  Ashish Kumar

IPC: G06Q30/00 ,  G06Q30/0273 ,  G06F16/9038

CPC classification number: G06Q30/0275 ,  G06F16/9038

Abstract: Keyword bids determined from sparse data are described. Initially, a portfolio optimization platform identifies which keywords included in a portfolio of keywords are low-impression keywords. This platform trains a machine learning model to generate bids for the low-impression keywords with historical data from a search engine. In particular, the platform trains this machine learning model according to an algorithm suited for training with sparse amounts of data, e.g., a temporal difference learning algorithm. In contrast, the platform uses different models, trained according to different algorithms than the low-impression keyword model, to generate bids for keywords determined not to be low-impression keywords. Once the low-impression keyword model is trained offline, the platform deploys the model for use online to generate actual bids for the low-impression keywords and submits them to the search engine. The platform continues to update the low-impression keyword model while deployed according to the sparse-data algorithm.

公开(公告)号：US11494810B2

公开(公告)日：2022-11-08

申请号：US16555380

申请日：2019-08-29

Applicant: Adobe Inc.

Inventor： Anirban Basu ,  Tathagata Sengupta ,  Kunal Kumar Jain ,  Ashish Kumar

IPC: G06Q30/00 ,  G06Q30/02 ,  G06F16/9038

Abstract: Keyword bids determined from sparse data are described. Initially, a portfolio optimization platform identifies which keywords included in a portfolio of keywords are low-impression keywords. This platform trains a machine learning model to generate bids for the low-impression keywords with historical data from a search engine. In particular, the platform trains this machine learning model according to an algorithm suited for training with sparse amounts of data, e.g., a temporal difference learning algorithm. In contrast, the platform uses different models, trained according to different algorithms than the low-impression keyword model, to generate bids for keywords determined not to be low-impression keywords. Once the low-impression keyword model is trained offline, the platform deploys the model for use online to generate actual bids for the low-impression keywords and submits them to the search engine. The platform continues to update the low-impression keyword model while deployed according to the sparse-data algorithm.

公开(公告)号：US20210065250A1

公开(公告)日：2021-03-04

申请号：US16555380

申请日：2019-08-29

Applicant: Adobe Inc.

Inventor： Anirban Basu ,  Tathagata Sengupta ,  Kunal Kumar Jain ,  Ashish Kumar

IPC: G06Q30/02 ,  G06F16/9038

Abstract: Keyword bids determined from sparse data are described. Initially, a portfolio optimization platform identifies which keywords included in a portfolio of keywords are low-impression keywords. This platform trains a machine learning model to generate bids for the low-impression keywords with historical data from a search engine. In particular, the platform trains this machine learning model according to an algorithm suited for training with sparse amounts of data, e.g., a temporal difference learning algorithm. In contrast, the platform uses different models, trained according to different algorithms than the low-impression keyword model, to generate bids for keywords determined not to be low-impression keywords. Once the low-impression keyword model is trained offline, the platform deploys the model for use online to generate actual bids for the low-impression keywords and submits them to the search engine. The platform continues to update the low-impression keyword model while deployed according to the sparse-data algorithm.

公开(公告)号：US20230021653A1

公开(公告)日：2023-01-26

申请号：US17955781

申请日：2022-09-29

Applicant: Adobe Inc.

Inventor： Anirban Basu ,  Tathagata Sengupta ,  Kunal Kumar Jain ,  Ashish Kumar

IPC: G06Q30/02 ,  G06F16/9038

Abstract: Keyword bids determined from sparse data are described. Initially, a portfolio optimization platform identifies which keywords included in a portfolio of keywords are low-impression keywords. This platform trains a machine learning model to generate bids for the low-impression keywords with historical data from a search engine. In particular, the platform trains this machine learning model according to an algorithm suited for training with sparse amounts of data, e.g., a temporal difference learning algorithm. In contrast, the platform uses different models, trained according to different algorithms than the low-impression keyword model, to generate bids for keywords determined not to be low-impression keywords. Once the low-impression keyword model is trained offline, the platform deploys the model for use online to generate actual bids for the low-impression keywords and submits them to the search engine. The platform continues to update the low-impression keyword model while deployed according to the sparse-data algorithm.