| US 11,790,037 B1 | ||
| Down-sampling of negative signals used in training machine-learned model | ||
| Xiaowen Zhang, Santa Clara, CA (US); Girish Kathalagiri Somashekariah, Santa Clara, CA (US); and Samaneh Abbasi Moghaddam, Santa Clara, CA (US) | ||
| Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US) | ||
| Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US) | ||
| Filed on Mar. 27, 2019, as Appl. No. 16/366,977. | ||
| Int. Cl. G06F 16/00 (2019.01); G06F 9/30 (2018.01); G06F 16/9538 (2019.01); G06F 16/903 (2019.01); G06N 20/00 (2019.01); G06F 18/2113 (2023.01) | ||
| CPC G06F 18/2113 (2023.01) [G06F 9/30069 (2013.01); G06F 16/90335 (2019.01); G06F 16/9538 (2019.01); G06N 20/00 (2019.01)] | 17 Claims |
|
1. A system comprising:
a computer-readable medium having instructions stored thereon, which, when executed by a processor, cause the system to:
in a training phase of a result ranking model:
obtain training data pertaining to sample results and user data corresponding to users, the training data comprising indications as to which of the sample results were explicitly interacted with by which users;
identify one or more of the sample results not having indications that they were explicitly interacted with by users as being implicit label candidates;
downsample the implicit label candidates in accordance with a downsampling scheme, the downsampling scheme specifying that implicit label candidates are reduced to a percentage, specified by a downsampling formula, of the implicit label candidates, without reducing the sample results that were explicitly interacted with by users; and
feed the sample results that were explicitly interacted with by users and the downsampled implicit label candidates into a machine learning algorithm to train the result ranking model to a likelihood score for a candidate result and candidate user data, the likelihood score indicating a likelihood that a user corresponding to the candidate user data will positively interact with the candidate result.
|