Online behavioral predictor
In some embodiments, a set of user groups can be defined, with each group relating to a different webpage experience, user action, etc. Requests are assigned to one of the groups based on actual webpage presentation features and/or user actions. A group-specific model is generated for each group and translates user information to a preliminary result (e.g., a purchasing probability). A model combination includes a weighted combination of a set of available group-specific models. User information is processed using the model combination to generate a model result. The model result is evaluated to determine whether a requested webpage is to be customized in a particular manner and/or an opportunity is to be offered.
1. A computer-implemented method, comprising:
defining a set of user groups, wherein each user group is defined according to historical data corresponding to a common experience during a website visit, and wherein each user group is associated with an action on a webpage;
generating a set of group-specific models, each group-specific model corresponding to a particular user group of the set of user groups, wherein each group-specific model is generated by training that group-specific model using the historical data of the particular user group, and wherein each group-specific model is configured to generate a prediction of a likelihood that a user in the user group will complete an action associated with the user group;
generating a model combination function configured to generate a model result from predictions generated by a plurality of group-specific models of the set of group-specific models, wherein the model combination function is generated by:
assigning weights to the predictions generated by the plurality of group-specific models, wherein a weight assigned to each prediction is based on a data size of the historical data used to train the group-specific model that generated that prediction; and wherein predictions generated from a group-specific model with a larger data size of historical data are assigned a higher weight and predictions generated from a group-specific model with a smaller data size of historical data are assigned a lower weight;
deriving a weighted sum from the weights and the corresponding predictions;
receiving a new request for the webpage from a new user;
identifying new user information associated with the new request;
determining, from the new user information, user characteristics corresponding to the new user;
generating a new user prediction in real-time by generating new predictions using the user characteristics as input to the plurality of group-specific models and applying the model combination function to the new predictions, wherein the new user prediction includes a probability that the new user will complete an action on the webpage, and wherein the action is associated with the user group;
customizing the webpage for the new user to increase a likelihood that the new user will complete the action on the webpage according to an evaluation of the new user prediction; and
transmitting the customized webpage in response to the new request.
2. The computer-implemented method of claim 1 , wherein the new user information includes data from a cookie associated with the request for the webpage.
3. The computer-implemented method of claim 1 , wherein the prediction indicates whether an online purchase will be made during a session associated with a future request.
4. The computer-implemented method of claim 1 , wherein the webpage includes a chat session.
5. The computer-implemented method of claim 1 , wherein customizing the webpage for the new user includes modifying the webpage to include a chat session.
6. The computer-implemented method of claim 1 , wherein customizing the webpage for the new user is based on an availability of a resource.
7. The computer-implemented method of claim 1 , wherein customizing the webpage for the new user is based on the new user prediction being greater than a threshold value.
8. A system, comprising:
one or more data processors; and
a non-transitory computer readable storage medium containing instructions which when executed on the one or more data processors, cause the one or more data processors to perform operations including:
defining a set of user groups, wherein each user group is defined according to historical data corresponding to a common experience during a website visit, and wherein each user group is associated with an action on a webpage;
generating a set of group-specific models, each group-specific model corresponding to a particular user group of the set of the set of user groups, wherein each group-specific model is generated by training that group-specific model using the historical data of the particular user group, and wherein each group-specific model is configured to generate a prediction of a likelihood that a user in the user group will complete an action associated with the user group;
generating a model combination function configured to generate a model result from predictions generated by a plurality of group-specific models of the set of group-specific models, wherein the model combination function is generated by:
assigning weights to the predictions generated by the plurality of group-specific models, wherein a weight assigned to each prediction is based on a data size of the historical data used to train the group-specific model that generated the prediction; and wherein predictions generated from a group-specific model with a larger data size of historical data are assigned a higher weight and predictions generated from a group-specific model with a smaller data size of historical data are assigned a lower weight; and
deriving a weighted sum from the weights and the corresponding predictions;
receiving a new request for the webpage from a new user;
identifying new user information associated with the new request;
determining, from the new user information, user characteristics corresponding to the new user;
generating a new user prediction in real-time by generating new predictions using the user characteristics as input to the plurality of group-specific models and applying the model combination function to the new predictions, wherein the new user prediction includes a probability that the new user will complete an action on the webpage, and wherein the action is associated with the user group;
customizing the webpage for the new user to increase a likelihood that the new user will complete the action on the webpage according to an evaluation of the new user prediction; and
transmitting the customized webpage in response to the new request.
9. The system of claim 8 , wherein the new user information includes data from a cookie associated with the request for the webpage.
10. The system of claim 8 , wherein the prediction indicates whether an online purchase will be made during a session associated with a future request.
11. The system of claim 8 , wherein the webpage includes a chat session.
12. The system of claim 8 , wherein customizing the webpage for the new user includes modifying the webpage to include a chat session.
13. The system of claim 8 , wherein customizing the webpage for the new user is based on an availability of a resource.
14. The system of claim 8 , wherein customizing the webpage for the new user is based on the new user prediction being greater than a threshold value.
15. A computer-program product tangibly embodied in a non-transitory machine-readable storage medium, including instructions configured to cause one or more data processors to perform operations including:
defining a set of user groups, wherein each user group is defined according to historical data corresponding to a common experience during a website visit, and wherein each user group is associated with an action on a webpage;
generating a set of group-specific models, each group-specific model corresponding to a particular user group of the set of user groups, wherein each group-specific model is generated by training that group-specific model using the historical data of the particular user group, and wherein each group-specific model is configured to generate a prediction of a likelihood that a user in the user group will complete an action associated with the user group;
generating a model combination function configured to generate a model result from predictions generated by a plurality of group-specific models of the set of group-specific models, wherein the model combination function is generated by:
assigning weights to the predictions generated by the plurality of group-specific models, wherein a weight assigned to each prediction is based on a data size of the historical data used to train the group-specific model that generated that prediction; and wherein predictions generated from a group-specific model with a larger data size of historical data are assigned a higher weight and predictions generated from a group-specific model with a smaller data size of historical data are assigned a lower weight
deriving a weighted sum from the weights and the corresponding predictions;
receiving a new request for the webpage from a new user;
identifying new user information associated with the new request;
determining, from the new user information, user characteristics corresponding to the new user;
generating a new user prediction in real by generating new predictions using the user characteristics as input to the plurality of group-specific models and applying the model combination function to the new predictions, wherein the new user prediction includes a probability that the new user will complete an action on the webpage, and wherein the action is associated with the user group;
customizing the webpage for the new user to increase a likelihood that the new user will complete the action on the webpage according to an evaluation of the new user prediction; and
transmitting the customized webpage in response to the new request.
16. The computer-program product of claim 15 , wherein the new user information includes data from a cookie associated with the request for the webpage.
17. The computer-program product of claim 15 , wherein the prediction indicates whether an online purchase will be made during a session associated with a future request.
18. The computer-program product of claim 15 , wherein the webpage includes a chat session.
19. The computer-program product of claim 15 , wherein customizing the webpage for the new user includes modifying the webpage to include a chat session.
20. The computer-program product of claim 15 , wherein customizing the webpage for the new user is based on an availability of a resource.