The solution is to have a set of features and learn their weights, as in Markov networks. For every person X, we can have the feature X has the flu; for every pair of acquaintances X and Y, the feature X and Y both have the flu; and so on. As in Markov networks, the maximum-likelihood weights are the ones that make each feature occur with the frequency observed in the data. The weight of X has the flu will be high if a lot of people have the flu. The weight of X and Y both have the flu will be high if, when person X has the flu, the odds that acquaintance Y also has the flu are higher than for
...more
