Method and apparatus for real-time personalization
A personalization platform is provided, which interprets user behavior and attributes along with the content users are interacting with, to build optimized predictive models of what content the user may want to see next. Those predictive models can be utilized to personalize content in one or more environments, including email, mobile and applications. An extensible and highly flexible framework can be implemented. In some embodiments, when calculating for a single user, recent behavior is scored against an ensemble of analytic models and the totals are amalgamated into a final recommendation. Any combination of analytic models may be explored and tested in a multivariate framework on this final ranking function. Models may be added and removed on a per-application basis.
1. A computer-implemented, network-connected content recommender generating content recommendations for a plurality of content servers hosted by one or more customers, the content recommender comprising:
one or more processors;
a memory storing instruction that, when executed by the one or more processors, cause the recommender to perform operations comprising
receiving a content recommendation request from a querying one of said customer content servers via a plurality of input streams, each input stream including a data repository;
outputting data, from the memory, associated with the content recommendation request; receiving some or all of the data associated with said content recommendation request; generating and outputting a plurality of model-specific recommendation results from the received data wherein the plurality of model-specific recommendation results are from a plurality of models;
combining the plurality of model-specific results to generate an ensemble recommendation result; and
transmitting the ensemble result from the content recommender to said querying customer content server.
2. The content recommender of claim 1 , in which the input streams further comprise:
a people stream input feeding a people data integrity module and a people repository;
a resources stream input feeding a resources data integrity module and a resources repository; and
an events stream input feeding an events data integrity module and an events repository.
3. The content recommender of claim 2 , in which the people repository stores attributes associated with users.
4. The content recommender of claim 2 , in which the resources repository stores records associated with resources that are subject to selection by the content recommender.
5. The content recommender of claim 4 , in which the resources repository further comprises a business rules filter to limit resources returned as candidates.
6. The content recommender of claim 2 , in which the events repository stores actions associated with a user and resource.
7. The content recommender of claim 2 , in which the input streams comprise a resources stream input feeding a resources data integrity module and a resources repository;
the operations further comprising: data scraping responsive to detection by the resources data integrity module of a new resource within the resources stream to collect information about the new resource via the network and store said data within the resources repository.
8. The content recommender of claim 7 , in which the information collected by the data scraping comprises open graph meta tags associated with a new resource.
9. The content recommender of claim 1 , in which the input streams comprise a resources stream input feeding a resources data integrity module and a resources repository;
wherein the operations further comprise accessing an application programming interface implemented by a third party server hosting a resource, in order to obtain information about a new resource for storage in the resource repository.
10. The content recommender of claim 1 , in which the combining model-specific results combines results from a plurality of different vertical specific modules.
11. The content recommender of claim 1 , in which the outputting a model-specific result comprise outputting from one or more behavioral similarity modules, one or more content similarity modules and one or more user-to-user similarity modules.
12. The content recommender of claim 11 , wherein the operations further comprise generating the ensemble result by application of a weighted combination to said model-specific results.
13. The content recommender of claim 1 , in which the operations further comprise dynamically selecting and parameterizing one of a plurality of ensemble candidates to combine model-specific outputs.
14. The content recommender of claim 13 , in which the dynamically selecting comprises a computer-implemented multi-arm bandit model in which each arm comprises an ensemble candidate.
15. The content recommender of claim 13 , in which the dynamically selecting comprises a computer-implemented reinforcement learning system optimizing ensemble selection and parameterization.
16. The content recommender of claim 11 , wherein he operations comprise, training a machine learning module using data associated with the querying third party content server, to optimize weights applied to a subset of analytic modules while determining the ensemble result.
17. The content recommender of claim 1 , wherein the further comprising a data integrity module configured to minimize passing of invalid data to the generating.