Best Setting of Model Parameters in Applying Topic Modeling on Textual Documents.

Abstract

Probabilistic topic modeling is an active research field in machine learning and has been mainly used as an analytical tool to structure large textual corpora for data mining. It offers a viable approach to structure huge textual document collections into latent topic themes to aid text mining. Latent Dirichlet Allocation (LDA) is the most commonly used topic modelling method across a wide number of technical fields. However, model development can be arduous and tedious, and requires burdensome and systematic sensitivity studies in order to find the best set of model parameters. In this study, we use a heuristic approach to estimate the most appropriate number of topics. Specifically, the rate of perplexity change (RPC) as a function of numbers of topics is proposed as a suitable selector. We test the stability and effectiveness of the proposed method for three markedly different types of grounded-truth datasets: Salmonella next generation sequencing, pharmacological side effects, and textual abstracts on computational biology and bioinformatics (TCBB) from PubMed. Then we describe extensive sensitivity studies to determine best practices for generating effective topic models. To test effectiveness and validity of topic models, we constructed a ground truth data set from PubMed that contained some 40 health related themes including negative controls, and mixed it with a data set of unstructured documents. We found that obtaining the most useful model, tuned to desired sensitivity versus specificity, requires an iterative process wherein preprocessing steps, the type of topic modeling algorithm, and the algorithm's model parameters are systematically varied. Models need to be compared with both qualitative, subjective assessments and quantitative, objective assessments, and care is required that Gibbs sampling in model estimation is sufficient to assure stable solutions. With a high quality model, documents can be rank-ordered in accordance with probability of being associated with complex regulatory query string, greatly lessoning text mining work. Importantly, topic models are agnostic about how words and documents are defined, and thus our findings are extensible to topic models where samples are defined as documents, and genes, proteins or their sequences are words.