Customizable objective function for hidden Markov Models Malhis, Nawar
Identifying sequences with frequent patterns is a major data-mining problem in computational biology, in this work, our focus is on utilizing a HMM like model for extracting sequences with interesting frequent patterns from a set of unlabeled sequences in a heavy noise environment. First we show that the likelihood objective function for HMMs is very sensitive to noise which limits its use to labeled sequences. Then we introduce an alternative model that we call Hidden States Model, HSM, which is a HMM with customizable objective function, OF. We show empirically (on synthetic data to facilitate precise performance evaluation) how OFs can be customized to target specific information such as patterns frequency and size. Results Show HSM vastly outperformed HMM in extracting sequences with interesting patterns for both unmutated and mutated data.
Item Citations and Data
Attribution-NonCommercial-NoDerivs 2.5 Canada