Improving mining quality by exploiting data dependency

Abstract. The usefulness of the results produced by data mining methods can be critically impaired by several factors such as (1) low quality of data, including errors due to contamination, or incompleteness due to limited bandwidth for data acquisition, and (2) inadequacy of the data model for capturing complex probabilistic relationships in data. Fortunately, a wide spectrum of applications exhibit strong dependencies between data samples. For example, the readings of nearby sensors are generally correlated, and proteins interact with each other when performing crucial functions. Therefore, dependencies among data can be successfully exploited to remedy the problems mentioned above. In this paper, we propose a unified approach to improving mining quality using Markov networks as the data model to exploit local dependencies. Belief propagation is used to efficiently compute the marginal or maximum posterior probabilities, so as to clean the data, to infer missing values, or to improve the mining results from a model that ignores these dependencies. To illustrate the benefits and great generality of the technique, we present its application to three challenging problems: (i) cost-efficient sensor probing, (ii) enhancing protein function predictions, and (iii) sequence data denoising.