Analysis of uncertain data: Evaluation of Given Hypotheses Selection of probes for information gathering Anatole Gershman, Eugene Fink, Bin Fu, and Jaime G. Carbonell

Analysis of uncertain data: Evaluation of Given Hypotheses Selection of probes for information gathering Anatole Gershman, Eugene Fink, Bin Fu, and Jaime G. Carbonell

Example The analyst has to distinguish between two hypotheses: Retires Joins Vikings

Example Observations: According to many rumors, quarterback Brett Favre has closed on the purchase of a home in Eden Prairie, MN, where the Minnesota Vikings' team facility is located. Without the tearful public ceremony that accompanied his retirement announcement from the Green Bay Packers just 11 months ago, quarterback Brett Favre has told the New York Jets he is retiring. Minnesota coach Brad Childress, jilted at the altar Tuesday afternoon by Brett Farve telling him he wasn’t going to play for the Vikings in 2009.

Example Observation distributions: Without the tearful public ceremony that accompanied his retirement announcement from the Green Bay Packers just 11 months ago, quarterback Brett Favre has told the New York Jets he is retiring. P(says retire | Retires) = 0.9 P(says retire | Joins Vikings) = 0.6 Bayesian induction: P (Retire|says retire) = P (Retire) ∙ P(says retire|Retire) / (P (Retire) ∙ P(says retire|Retire) + P (Joins Vikings) ∙ P(Joins Vikings|Retire)) = 0.5

General problem We have to distinguish among n mutually exclusive hypotheses, denoted H 1, H 2,…, H n. 0.6 For every hypothesis, we know its prior; thus, we have an array of n of priors, P(H 1 ), P(H 2 ), P(H n ) 0.4

General problem We base the analysis on m observable features, denoted OBS 1, OBS 2, …, OBS m. Each observation is a variable that takes one of several discrete values. OBS 1 For every observation, OBS a, we know the number of its possible values, num[a]. Thus, we have num[1..m] with the number of values for each observation. num[1] = 2 I will RETIRE! I won’t RETIRE! For every hypothesis, we know the related probability distribution of each observation. P(o a,j | H i ) represents the probabilities of possible values of OBS a We know a specific value of each observation val [1..m].

General problem We have to evaluate the posterior probabilities of the n given hypotheses, denoted Post(H 1 ), Post(H 2 ), Post(H n ) 0.5

Extension #1 Prior: Something else H0 (“surprise”) 0.05

Extension #1 After discovering val, Posterior probability of H 0 : Post(H 0 ) = P(H 0 ) ∙ P(val | H 0 ) / P(val) = P(H 0 ) ∙ P(val | H 0 ) / ( P(H 0 ) ∙ P(val | H 0 ) + likelihood(val) ). Bad news: We do not know P(val | H 0 ). Good news: Post(H 0 ) monotonically depends on P(val | H 0 ) ; thus, if we obtain lower and upper bounds for P(val | H 0 ), we also get bounds for Post(H 0 ).

Plausibility principle Unlikely events normally do not happen; thus, if we have observed val, then its likelihood must not be too small. Plausibility threshold: We use a global constant plaus, which must be between 0.0 and 1.0. If we have observed val, we assume that P( val ) ≥ plaus / num. We use it to obtains bounds for P(val | H 0 ), : Lower: (plaus / num − likelihood(val)) / prior[0]. Upper: 1.0.

Plausibility principle We use it to obtains bounds for P(val | H 0 ) : Lower: (plaus / num − likelihood(val)) / P(H 0 ). Upper: 1.0. We substitute these bounds into the dependency of Post(H 0 ) on P(val | H 0 ),, thus obtaining the bounds for Post(H 0 ) : Lower: 1.0 − likelihood(val) ∙ num / plaus. Upper: P(H 0 ) / (P(H 0 ) + likelihood(val)). We have derived bounds for the probability that none of the given hypotheses is correct.

Extension #2 Multiple observations: Which one(s) to Use? Independence assumption: usually does not work. Bayesian analysis: Use their joint distribution? Difficult to get. We identify the highest-utility observation and do not use other observations to corroborate it.

Extension #2 Utility Function Which one is “better”?

Extension #2 Utility Function Shannon’s Entropy (negation)

Extension #2 Utility Function KL-divergence

Extension #2 Utility Function Self-defined function

Analysis of uncertain data: Evaluation of Given Hypotheses Selection of probes for information gathering Anatole Gershman, Eugene Fink, Bin Fu, and Jaime G. Carbonell

Example The analyst has to distinguish between two hypotheses: Retires Joins Vikings

Example I will RETIRE! Ask Probe: Execute external action and observe its response, to gather more information. 0.5

Example Probe: Probe Cost Observation Probability Gain (utility function)

Probe Selection single-obs-gain(probe j ) = visible[i, a, j] · (likelihood(1) · probe-gain(1) + … + likelihood(num[a]) · probe-gain(num[a])) + (1.0 − visible[i, a, j]) · cost[j] gain(prob j ) = max (single-obs-gain(prob j, obs1),…, single-obs-gain(prob j, obsm)) Probe Cost Observation Probability Utility Function

Experiment Task: Evaluating hypothesizes (H1, H2, H3, H4). No Probe, Accuracy of distinguishing between H1 and other hypotheses

Experiment Probe Selection to distinguish H1 and other hypotheses Task: Evaluating hypothesizes (H1, H2, H3, H4).

Experiment Probe Selection to distinguish four hypotheses Task: Evaluating hypothesizes (H1, H2, H3, H4).

Summary Use Bayesian inference to distinguish among mutually exclusive hypotheses.  H 0 hypothesis  Multiple observations Use Probe to gather more information for better analysis  Cost, Utility function, Observation Probability,...

Thank you