Supervised learning of hidden Markov models for sequence discrimination

Conference · Mon Dec 01 04:00:00 EST 1997

OSTI ID:549015

Mamitsuka, Hiroshi ^[1]

---

1. NEC Corp., Kanagawa (Japan)

We present two supervised learning algorithms for hidden Markov models (HMMs) for sequence discrimination. When we model a class of sequences with an HMM, conventional learning algorithms for HMMs have trained the HMM with training examples belonging to the class, i.e. positive examples alone, while both of our methods allow us to use negative examples as well as positive examples. One of our algorithms minimizes a kind of distance between a target likelihood of a given training sequence and an actual likelihood of the sequence, which is obtained by a given HMM, using an additive type of parameter updating based on a gradient-descent learning. The other algorithm maximizes a criterion which represents a kind of ratio of the likelihood of a positive example to the likelihood of the total example, using a multiplicative type of parameter updating which is more efficient in actual computation time than the additive type one. We compare our two methods with two conventional methods on a type of cross-validation of actual motif classification experiments. Experimental results show that in terms of the average number of classification errors, our two methods out-perform the two conventional algorithms. 14 refs., 4 figs., 1 tab.

• Research Organization: Association for Computing Machinery, New York, NY (United States); Sloan (Alfred P.) Foundation, New York, NY (United States)
• OSTI ID: 549015
• Report Number(s): CONF-970137--
• Country of Publication: United States
• Language: English

Related Subjects

55 BIOLOGY AND MEDICINE
BASIC STUDIES
99 GENERAL AND MISCELLANEOUS
ACCURACY
ALGORITHMS
AMINO ACID SEQUENCE
COMPUTER CODES
DNA
DNA SEQUENCING
EFFICIENCY
ERRORS
MARKOV PROCESS
MATHEMATICAL MODELS
MAXIMUM-LIKELIHOOD FIT
MOLECULAR BIOLOGY
NUCLEOTIDES
PARAMETRIC ANALYSIS
PROTEINS
STOCHASTIC PROCESSES
VALIDATION