Position debiased network site searches
A position debiased search system can avoid bias towards top-ranked search results using a position-trained machine-trained model. Past positions for listings can be input into the model with added noise and low-ranked results to train the model to generate rankings that do not exhibit position bias. A network site can implement the position debiased search system to generate network site results that can generate accurate user results in real time as users browse the network site.
1. A method comprising:
generating historical search result data of a network site comprising a plurality of past results presented on the network site comprising a set of seen search results and a set of unseen search results that a user did not navigate to, one or more indications of which of the plurality of past results were selected by network site users, and position data indicating a display position for each of the plurality of past results;
generating training data for a machine learning model by replacing a portion of the position data with arbitrary data comprising zeros and augmenting seen results that a user did navigate to with randomly sampled unseen results;
training the machine learning model using the position data with the portion of the position data replaced with arbitrary data and the augmented seen results to generate a position debiased machine learning scheme, wherein generating the position debiased machine learning scheme further comprises initially training the machine learning model on past position values in the historical search result data followed by retraining the machine learning model using the arbitrary data instead of the past position values;
receiving a search request from a network site user of the network site; and
generating search results for the network site user using the position debiased machine learning scheme.
2. The method of claim 1 , further comprising:
causing display, on a client device of the network site user, of one or more of the search results generated by the position debiased machine learning scheme.
3. The method of claim 1 , wherein the machine learning model is a deep neural network model and the position debiased machine learning scheme is a position debiased deep neural network.
4. The method of claim 1 , wherein the arbitrary data is arbitrary in that it is not past position values from the historical search result data.
5. The method of claim 1 , wherein the plurality of past results includes a portion of low-positioned results.
6. The method of claim 5 , wherein the machine learning model is trained by sampling the low-positioned results.
7. The method of claim 5 , wherein the portion of low-positioned results are non-displayed past results.
8. The method of claim 5 , wherein the portion of low-positioned results are search results that were not displayed on a first page of search results.
9. A system comprising:
one or more processors of a machine; and
a memory storing instructions that, when executed by the one or more processors, cause the machine to perform operations comprising:
generating historical search result data of a network site comprising a plurality of past results presented on the network site comprising a set of seen search results and a set of unseen search results that a user did not navigate to, one or more indications of which of the plurality of past results were selected by network site users, and position data indicating a display position for each of the plurality of past results;
generating training data for a machine learning model by replacing a portion of the position data with arbitrary data comprising zeros and augmenting seen results that a user did navigate to with randomly sampled unseen results;
training the machine learning model using the position data with the portion of the position data replaced with arbitrary data and the augmented seen results to generate a position debiased machine learning scheme, wherein generating the position debiased machine learning scheme further comprises initially training the machine learning model on past position values in the historical search result data followed by retraining the machine learning model using the arbitrary data instead of the past position values;
receiving a search request from a network site user of the network site; and
generating search results for the network site user using the position debiased machine learning scheme.
10. The system of claim 9 , the operations further comprising:
causing, display on a client device of the network site user, of one or more of the search results generated by the position debiased machine learning scheme.
11. The system of claim 9 , wherein the machine learning model is a deep neural network model and the position debiased machine learning scheme is a position debiased deep neural network.
12. The system of claim 9 , wherein the arbitrary data is arbitrary in that it is not past position values from the historical search result data.
13. The system of claim 9 , wherein the plurality of past results includes a portion of low-positioned results.
14. A non-transitory machine-readable storage device embodying instructions that, when executed by a machine, cause the machine to perform operations comprising:
generating historical search result data of a network site comprising a plurality of past results presented on the network site comprising a set of seen search results and a set of unseen search results that a user did not navigate to, one or more indications of which of the plurality of past results were selected by network site users, and position data indicating a display position for each of the plurality of past results;
generating training data for a machine learning model by replacing a portion of the position data with arbitrary data comprising zeros and augmenting seen results that a user did navigate to with randomly sampled unseen results;
training the machine learning model using the position data with the portion of the position data replaced with arbitrary data and the augmented seen results to generate a position debiased machine learning scheme, wherein generating the position debiased machine learning scheme further comprises initially training the machine learning model on past position values in the historical search result data followed by retraining the machine learning model using the arbitrary data instead of the past position values;
receiving a search request from a network site user of the network site; and
generating search results for the network site user using the position debiased machine learning scheme.