Modeling and automatic filtering/tracking real-time events using dynamic relational bayesian networks
The present invention includes relation-Bayesian-network-based tracking methods. Various, distinct embodiments of the present invention include tracking methods for: real-time relational tracking of e-commerce segmentation and personalization, in-residence tracking of elderly and disabled people, real-time relational tracking for ground target tracking and real-time relational tracking of satellites and satellite constellations.
1. A computer-implemented method for modeling, filtering/tracking of real-time behavior of a system, comprising the steps of:
identifying variables in a dynamic relational Bayesian network (DRBN) to be filtered, the DRBN including a plurality of submodels, the DRBN being stored in a memory;
assigning a characterization to the variables to be filtered;
creating a reduced graph for each submodel of the DRBN by reducing each observed variable to a singleton domain;
compiling on a computer each reduced graph to create compiled inference structures;
creating a mapping between data instances and the compiled inference structures by constructing an inference cache for each variable of the reduced graph for each data instance to which the reduced graph applies;
performing on a computer exact inference over each data instance using a generalized loopy inference method to communicate estimate changes between instances for each loopy pass, wherein the data instances are time ordered and the order of instance application is reversed for each loopy pass.
2. The method of claim 1 , wherein the characterization is a Boolean characterization.
3. The method of claim 1 , wherein the DRBN includes a plurality of classes of submodels, each class having a plurality of submodels, and wherein each of the submodels for each class is associated with a condition of applicability, and wherein at least some nodes of some submodels are labeled as import and update; and
wherein the method further comprises the steps of
classifying each of the submodels (or classes) as static, dynamic or dependent;
following loopy inference over each new instance, for all inference node caches where the instance is of a submodel for class* labeled dynamic and has a node labeled import and update, locating the instance containing the variable corresponding to the node labeled import;
repeating the locating step n times; and scavenging the instance located in the nth locating step.
4. The method of claim 1 , wherein the instances are historical data instances and the inference is used for parameter estimation.
5. The method of claim 1 , wherein the instances are real time data instances and the inference is used for filtering.
6. The method of claim 1 , wherein the DRBN is a model of a web visit.
7. The method of claim 1 , wherein the DRBN includes a page view class of submodels, and wherein the page view class includes at least one first submodel of a page view of a page showing a specific product and at lease one second submodel of a page view of a page not showing a specific product.
8. The method of claim 1 , wherein the system is in-residence patient behavior.
9. The method of claim 1 , wherein the system is a collection of moving, physical targets.
10. The method of claim 1 , wherein the system is a visitor on a website.