LDAShiny: An R Package for Exploratory Review of Scientific Literature Based on a Bayesian Probabilistic Model and Machine Learning Tools

In this paper we propose an open source application called LDAShiny, which provides a graphical user interface to perform a review of scientific literature using the latent Dirichlet allocation algorithm and machine learning tools in an interactive and easy-to-use way. The procedures implemented are based on familiar approaches to modeling topics such as preprocessing, modeling, and postprocessing. The tool can be used by researchers or analysts who are not familiar with the R environment. We demonstrated the application by reviewing the literature published in the last three decades on the species Oreochromis niloticus. In total we reviewed 6196 abstracts of articles recorded in Scopus. LDAShiny allowed us to create the matrix of terms and documents. In the preprocessing phase it went from 530,143 unique terms to 3268. Thus, with the implemented options the number of unique terms was reduced, as well as the computational needs. The results showed that 14 topics were sufficient to describe the corpus of the example used in the demonstration. We also found that the general research topics on this species were related to growth performance, body weight, heavy metals, genetics and water quality, among others.

Bayesian Inference of Linear Temporal Logic Specifications for Contrastive Explanations

Temporal logics are useful for providing concise descriptions of system behavior, and have been successfully used as a language for goal definitions in task planning. Prior works on inferring temporal logic specifications have focused on "summarizing" the input dataset - i.e., finding specifications that are satisfied by all plan traces belonging to the given set. In this paper, we examine the problem of inferring specifications that describe temporal differences between two sets of plan traces. We formalize the concept of providing such contrastive explanations, then present BayesLTL - a Bayesian probabilistic model for inferring contrastive explanations as linear temporal logic (LTL) specifications. We demonstrate the robustness and scalability of our model for inferring accurate specifications from noisy data and across various benchmark planning domains.

Risk-based strategies for surveillance of tuberculosis infection in cattle for low-risk areas in England and Scotland

SUMMARYDisease surveillance can be made more effective by either improving disease detection, providing cost savings, or doing both. Currently, cattle herds in low-risk areas (LRAs) for bovine tuberculosis (bTB) in England are tested once every 4 years. In Scotland, the default herd testing frequency is also 4 years, but a risk-based system exempts some herds from testing altogether. To extend this approach to other areas, a bespoke understanding of at-risk herds and how risk-based surveillance can affect bTB detection is required. Here, we use a generalized linear mixed model to inform a Bayesian probabilistic model of freedom from infection and explore risk-based surveillance strategies in LRAs and Scotland. Our analyses show that in both areas the primary herd-level risk factors for bTB infection are the size of the herd and purchasing cattle from high-risk areas of Great Britain and/or Ireland. A risk-based approach can improve the current surveillance system by both increasing detection (9% and 7% fewer latent infections), and reducing testing burden (6% and 26% fewer animal tests) in LRAs and Scotland, respectively. Testing at-risk herds more frequently can also improve the level of detection by identifying more infected cases and reducing the hidden burden of the disease, and reduce surveillance effort by exempting low-risk herds from testing.

Bayesian Inference for Single-cell Clustering and Imputing

Single-cell RNA-seq gives access to gene expression measurements for thousands of cells, allowing discovery and characterization of cell types. However, the data is noise-prone due to experimental errors and cell type-specific biases. Current computational approaches for analyzing single-cell data involve a global normalization step which introduces incorrect biases and spurious noise and does not resolve missing data (dropouts). This can lead to misleading conclusions in downstream analyses. Moreover, a single normalization removes important cell type-specific information. We propose a data-driven model, BISCUIT, that iteratively normalizes and clusters cells, thereby separating noise from interesting biological signals. BISCUIT is a Bayesian probabilistic model that learns cell-specific parameters to intelligently drive normalization. This approach displays superior performance to global normalization followed by clustering in both synthetic and real single-cell data compared with previous methods, and allows easy interpretation and recovery of the underlying structure and cell types.