This example finds the three hidden states for the table test_loan_prediction in HMMEvaluator Example: Loan Default Prediction, which is an unsupervised loan sequence trained with two models (1 and 2).
Input
- InitialStateProbs: pi_loan
- TransitionProbs: A_loan
- EmissionProbs: B_loan
The preceding models are output by HMMUnsupervised Example: Loan Default Prediction.
SQL Call
SELECT * FROM HMMDecoder ( ON pi_loan AS InitialStateProbs PARTITION BY model_id ORDER BY model_id, state ON A_loan AS TransitionProbs PARTITION BY model_id ON B_loan AS EmissionProbs PARTITION BY model_id ON test_loan_prediction AS ObservationTable PARTITION BY model_id ORDER BY model_id, seq_id, seq_vertex_id ASC USING InitStateModelColumn ('model_id') InitStateColumn ('state') InitStateProbColumn ('probability') TransAttributeColumn ('model_id') TransFromStateColumn('from_state') TransToStateColumn('to_state') TransProbColumn ('probability') EmitModelColumn ('model_id') EmitStateColumn ('state') EmitObsColumn ('observed') EmitProbColumn ('probability') ModelColumn ('model_id') SeqColumn ('seq_id') ObsColumn ('observed_id ') Accumulate ('seq_vertex_id ') ) AS dt ORDER BY 1, 2, 5;
Output
For the same sequence, the hidden states are different for each model.
model_id seq_id observed_id state seq_vertex_id -------- ------ ----------- ----- ------------- 1 17 1 2 0 1 17 1 2 1 1 17 1 2 2 1 17 1 2 3 1 17 1 2 4 1 17 1 2 5 1 17 1 2 6 1 17 2 2 7 1 17 1 2 8 1 17 1 2 9 1 17 1 2 10 1 17 1 2 11 1 17 1 2 12 1 17 1 2 13 1 17 3 1 14 1 17 4 0 15 1 17 5 1 16 1 17 6 1 17 1 17 6 1 18 1 17 6 1 19 1 17 6 1 20 1 17 6 1 21 1 17 6 1 22 1 17 6 1 23 1 17 6 1 24 1 17 6 0 25 2 17 1 2 0 2 17 1 2 1 2 17 1 2 2 2 17 1 2 3 2 17 1 2 4 2 17 1 2 5 2 17 1 2 6 2 17 2 1 7 2 17 1 2 8 2 17 1 2 9 2 17 1 2 10 2 17 1 2 11 2 17 1 2 12 2 17 1 2 13 2 17 3 1 14 2 17 4 0 15 2 17 5 1 16 2 17 6 1 17 2 17 6 1 18 2 17 6 1 19 2 17 6 1 20 2 17 6 1 21 2 17 6 1 22 2 17 6 1 23 2 17 6 1 24 2 17 6 0 25
Download a zip file of all examples and a SQL script file that creates their input tables from the attachment in the left sidebar.