SteveW
There seems to be a space between learning a network from abundant data and building expert models based on knowledge elicitation where there is no data. Specifically, we have enough tacit knowledge to build the topology of a network but very few observations (and hence a sparse JPD). What is the best use of data in this instance? I see a few options:1) pretend the data don't exist and parameterize using expert elicitation - but the experts are aware of the data and want to make use of it in an appropriate way2) Build the network manually and estimate parameters from the sparse data anyway, understanding that the model will be uncertain and perhaps unreliable but it's the best we can do3) Build and parameterize an expert model and then update beliefs based on the sparse data - this seems like the most appropriate but updating based on a few observations from data the experts were already aware of and likely incorporated in their prior beliefs seems inappropriate4. Use exploratory analyses such univariate regressions (even if based on only a few points) to explore relationships between nodes and use regression equations to generate CPTs. This is what experts want to do but it implies a precision to the analysis that is misleading and just feels wrong.Opinions? Alternatives?thanks
Quote 0 0
Mark
Hi Steve,You can mix structure learning from data and expert modelling. In BayesiaLab, there are many tools allowing us to drive the learning process: - you can fix arcs before use a learning algorithm like Taboo or EQ - you can use local structural coefficients to increase or decrease the importance of a node - you can use the virtual number of states to modify the complexity of a node and make it more or less "attractive" for the other nodes - the global structural coefficient allows the learning algorithm to modify the global complexity of the network - the constraints on the arcs prevent the learning algorithm to add some arcs according to the given rules
Quote 0 0