Content recommendations based upon historical future data
One or more computing devices, systems, and/or methods for content recommendations using historical future data are provided. A model serving delay time is computed as an average of training delays of events. A historical data time interval is determined based upon the model serving delay time. A model is trained for predicting user content preferences using historic user distribution data and historic content distribution data associated with the historic data time interval. The model is utilized to generate and provide content recommendations to users.
1. A method, comprising:
computing, by a recommendation system hosted by one or more servers, a model serving delay time as an average of training delays of events, wherein a training delay corresponds to a difference between an event serve time of an event and a train time of a model that served a recommendation associated with the event;
in response to determining that a current day occurs within a set of weekdays, determining, by the recommendation system hosted by the one or more servers, a historical data time interval as a day prior to the current day plus the model serving delay time;
using the historical data time interval to identify historic user distribution data and historic content distribution data associated with the historical data time interval;
training, by the recommendation system hosted by the one or more servers, the model for predicting user content preferences using the historic user distribution data and the historic content distribution data associated with the historical data time interval;
utilizing the model to generate a content recommendation for a user;
providing, by the recommendation system hosted by the one or more servers, the content recommendation for display on a client device of the user; and
in response to detecting a first event corresponding to the user interacting with the content recommendation, generating an entry indicating an event type of the first event, a first serve time of the model serving the content recommendation, and a first train time of the model being trained using the historic user distribution data and the historic content distribution data.
2. The method of claim 1 , comprising:
applying a first step size to the historic user distribution data and the historic content distribution data for training the model.
3. The method of claim 2 , wherein the first step size is less than a second step size applied to current user distribution data and current content distribution data used to train the model.
4. The method of claim 1 , comprising:
applying a historical interval ratio to the historic user distribution data and the historic content distribution data to select a subset of the historic user distribution data and the historic content distribution data for training the model.
5. The method of claim 4 , comprising:
determining the historical interval ratio based upon at least one of a current training time of training the model or an amount of data available to train the model.
6. The method of claim 4 , comprising:
determining the historical interval ratio based upon experimental results of training the model using historical user and content distribution data.
7. The method of claim 1 , comprising:
providing second content recommendation for display on the client device of the user.
8. The method of claim 7 , comprising:
in response to detecting a second event associated with the second content recommendation, generating a second entry indicating a second event type of the second event, a second serve time of the model serving the second content recommendation, and a second train time of the model being trained using the historic user distribution data and the historic content distribution data.
9. The method of claim 8 , wherein the second event corresponds to an impression of the user viewing the second content recommendation.
10. The method of claim 1 , comprising:
in response to determining that the current day is a weekend day, determining the historical data time interval as a week prior to the current day plus the model serving delay time.
11. The method of claim 1 , comprising:
in response to determining that the current day is a weekday not within the set of weekdays, determining the historical data time interval as a plurality of days prior to the current day plus the model serving delay time.
12. The method of claim 1 , comprising:
in response to determining that the current day is a weekend day, determining the historical data time interval as a day prior to the current day plus the model serving delay time.
13. A computing device comprising:
a processor; and
memory comprising processor-executable instructions that when executed by the processor cause performance of operations, the operations comprising:
computing, by a recommendation system hosted by one or more servers, a model serving delay time as an average of training delays of events, wherein a training delay corresponds to a difference between an event serve time of an event and a train time of a model that served a recommendation associated with the event;
in response to determining that a current day is a weekend day, determining, by the recommendation system hosted by the one or more servers, a historical data time interval as a week prior to the current day plus the model serving delay time;
using the historical data time interval to identify historic user distribution data and historic content distribution data associated with the historical data time interval:
training, by the recommendation system hosted by the one or more servers, the model for predicting user content preferences using the historic user distribution data and the historic content distribution data associated with the historical data time interval;
utilizing the model to generate a content recommendation for a user;
providing, by the recommendation system hosted by the one or more servers, the content recommendation for display on a client device of the user; and
in response to detecting a first event corresponding to the user interacting with the content recommendation, generating an entry indicating an event type of the first event, a first serve time of the model serving the content recommendation, and a first train time of the model being trained using the historic user distribution data and the historic content distribution data.
14. The computing device of claim 13 , comprising:
in response to determining that the current day occurs within a set of weekdays, determining the historical data time interval as a day prior to the current day plus the model serving delay time.
15. The computing device of claim 14 , comprising:
in response to determining that the current day is a weekday not within the set of weekdays, determining the historical data time interval as a plurality of days prior to the current day plus the model serving delay time.
16. The computing device of claim 15 , wherein the plurality of days correspond to 3 days prior to the current day.
17. A non-transitory machine readable medium having stored thereon processor-executable instructions that when executed cause performance of operations, the operations comprising:
computing, by a recommendation system hosted by one or more servers, a model serving delay time as an average of training delays of events, wherein a training delay corresponds to a difference between an event serve time of an event and a train time of a model that served a recommendation associated with the event;
in response to determining that a current day is a weekend day that does not occur within a set of weekdays, determining, by the recommendation system hosted by the one or more servers, a historical data time interval as a plurality of days prior to the current day plus the model serving delay time;
using the historical data time interval to identify historic user distribution data and historic content distribution data associated with the historical data time interval;
training, by the recommendation system hosted by the one or more servers, the model for predicting user content preferences using the historic user distribution data and the historic content distribution data associated with the historical data time interval;
utilizing the model to generate a content recommendation for a user;
providing, by the recommendation system hosted by the one or more servers, the content recommendation for display on a client device of the user; and
in response to detecting a first event corresponding to the user interacting with the content recommendation, generating an entry indicating an event type of the first event, a first serve time of the model serving the content recommendation, and a first train time of the model being trained using the historic user distribution data and the historic content distribution data.
18. The non-transitory machine readable medium of claim 17 , wherein the plurality of days correspond to 3 days prior to the current day.
19. The non-transitory machine readable medium of claim 17 , wherein the operations comprise:
in response to determining that a second current day occurs within the set of weekdays, determining a second historical data time interval as a day prior to the second current day plus a second model serving delay time.
20. The non-transitory machine readable medium of claim 17 , wherein the operations comprise:
in response to determining that the current day is a weekend day, determining the historical data time interval as a week prior to the current day plus the model serving delay time.