In recent years, the large amount of labeled data available has also helped tend research toward using minimal domain knowledge, e.g., in deep neural network research. However, in many situations, data is limited and of poor quality. Can domain knowledge be useful in such a setting? In this paper, we propose domain adapted neural networks (DANN) to explore how domain knowledge can be integrated into model training for deep networks. In particular, we incorporate loss terms for knowledge available as monotonicity constraints and approximation constraints. We evaluate our model on both synthetic data generated using the popular Bohachevsky function and a real-world dataset for predicting oxygen solubility in water. In both situations, we find that our DANN model outperforms its domain-agnostic counterpart yielding an overall mean performance improvement of 19.5% with a worst- and best-case performance improvement of 4% and 42.7%, respectively.

Incorporating Prior Domain Knowledge into Deep Neural Networks

Multi-label audio tagging is the task of predicting the types of sounds occurring in an audio clip. Recently, large-scale audio datasets such as Google’s AudioSet, have allowed researchers to use deep learning techniques for this task but this comes at the cost of label noise in the datasets. Audio datasets such as the AudioSet are usually built following a hierarchical structure known as ontology which captures the relationships between different sound events with domain knowledge. However, existing methods for audio tagging failed to utilize this domain knowledge about label relationships in their models, resulting in models being sensitive to label noise. We therefore present MT-GCN, a Multi-task Learning based Graph Convolutional Network that learns domain knowledge from ontology. The relationships between sound events in our proposed method are described by a graph. We propose two ontology-based graph construction methods, and conduct extensive experiments on the FSDKaggle2019 dataset. The experimental results show that our approach outperforms the baseline methods by a significant margin.

Mt-Gcn For Multi-Label Audio Tagging With Noisy Labels

Acute myocardial infarction (AMI) is a serious cardiovascular disease, followed by a high readmission rate within 30-days of discharge. Accurate prediction of AMI readmission is a crucial way to identify the high-risk group and optimize the distribution of medical resources.
 In this study, we propose a stacking-based model to predict the risk of 30-day unplanned all-cause hospital readmissions for AMI patients based on clinical data. Firstly,
we conducted an under-sampling method of neighborhood cleaning rule (NCR) to alleviate the class imbalance and then utilized a feature selection method of SelectFromModel (SFM) to select effective features. Secondly, we adopted a self-adaptive approach to select base classifiers from eight candidate models according to their performances in datasets. Finally, we constructed a three-layer stacking model in which layer 1 and layer 2 were base-layer and level 3 was meta-layer. The predictions of the base-layer were used to train the meta-layer in order to make the final forecast. The results show that the proposed model exhibits the highest AUC (0.720), which is higher than that of decision tree (0.681), support vector machine (0.707), random forest (0.701), extra trees (0.709), adaBoost (0.702), bootstrap aggregating (0.704), gradient boosting decision tree (0.710) and extreme gradient enhancement (0.713). It is evident that our model could effectively predict the risk of 30-day all cause hospital readmissions for AMI patients and provide decision support for the administration.

/pdf/a-stacking-based-model-for-predicting-30-day-all-cause-382dl4zxyt.pdf

A stacking-based model for predicting 30-day all-cause hospital readmissions of patients with acute myocardial infarction

We present GetFair, a novel framework for tuning fairness of classification models. The fair classification problem deals with training models for a given classification task where data points have sensitive attributes. The goal of fair classification models is to not only generate accurate classification results but also to prevent discrimination against subpopulations (i.e., individuals with a specific value for the sensitive attribute). Existing methods for enhancing fairness of classification models, however, are often specifically designed for a particular fairness metric or a classifier model. They may also not be suitable for scenarios with incomplete training data or where optimizing for multiple fairness metrics is important. GetFair represents a general solution to this problem. The GetFair approach works in the following way: First, a given classifier is trained on training data without any fairness objective. This is followed by a reinforcement learning inspired tuning procedure which updates the parameters of the learned model on a given fairness objective. This disentangles classifier training from fairness tuning, making our framework more general and allowing for the adoption of any parameterized classifier model. Because fairness metrics are designed as reward functions during tuning, GetFair generalizes across any fairness metric. We demonstrate the generalizability of GetFair via evaluation over a benchmark suite of datasets, classification models, and fairness metrics. In addition, GetFair can also be deployed in settings where the training data is incomplete or the classifier needs to be tuned on multiple fairness metrics. GetFair not only contributes a flexible method to the repertoire of tools available to improve the fairness of classification models, it also seamlessly adapts to settings where existing fair classification methods may not be suitable or applicable.

GetFair: Generalized Fairness Tuning of Classification Models

DNA is quintessential to carry out basic functions by organisms as it encodes information necessary for metabolomics and proteomics, among others. In particular, it is common nowadays to use DNA for profiling living organisms based on their phenotypic traits. These traits are the outcomes of the genetic makeup constrained by the interaction between living organisms and their surrounding environment over time. For environmental conditions, however, the conventional assumption is that they are too random and ephemeral to be encoded in the DNA of an organism. Here, we demonstrate that, to the contrary, genomic DNA may also encode sufficient information about some environmental features of an organism’s habitat for a machine learning model to reveal them, although there seem to be exceptions, i.e. some environmental features do not appear to be coded in DNA, unless our methods miss that information. Nevertheless, we demonstrate that these features can be used to train better models for better predictions of other environmental factors. These results lead directly to the question of whether over evolutionary history, DNA itself is actually also a repository of information related to the environment where the lineage has developed, perhaps even more cryptically than the way it encodes phenotypic information.

Profiling Environmental Conditions from DNA.

Traditionally, machine learning extracts knowledge solely based on data. However, huge volume of knowledge is available in other sources which can be included into machine learning models. Still, domain knowledge is rarely used in machine learning. We propose a framework that integrates domain knowledge in form of hierarchies into machine learning models, namely logistic regression. Integration of the hierarchies is done by using stacking (stacked generalization). We show that the proposed framework yields better results compared to standard logistic regression model. The framework is tested on the binary classification problem for predicting 30-days hospital readmission. Results suggest that the proposed framework improves AUC (area under the curve) compared to logistic regression models unaware of domain knowledge by 9% on average.

Framework for integration of domain knowledge into logistic regression

A fair classifier chain for multi‐label bank marketing strategy classification

This paper addresses a novel topic in ski injury research - how a ski group size indicates the risk of ski injury. There is evidence in research literature that people ski in groups. However, the relationship between group size and the risk of injury has remained unexplored. Based on ski lift entrance data, we use the density-based clustering algorithm OPTICS to identify groups of skiers and discuss the advantages of using this algorithm. We show that the ski group size can be used to improve the identification of skiers who experience ski injury. The results of the identification of ski groups at Mt. Kopaonik Ski Resort in Serbia show that skiing alone is most susceptible to ski injury, while skiing in couples or in bigger groups reduces the risk of injury by 46%. In addition, it is confirmed that ski injuries are an early failure event phenomena. Based on the CHAID decision tree analysis, spending a small amount of time at the ski resort and skiing alone are associated with the: 6 times greater ski injury risk than the average ski injury risk.

Boris Delibašić

Papers

Framework for integration of domain knowledge into logistic regression

A fair classifier chain for multi‐label bank marketing strategy classification

A study on ski groups size and their relationship to the risk of injury

Changing criteria weights to achieve fair VIKOR ranking: a postprocessing reranking approach

Controlling highway toll stations using deep learning, queuing theory, and differential evolution