Human-in-the-Loop Interpretability Prior.
Ontology highlight
ABSTRACT:
SUBMITTER: Lage I
PROVIDER: S-EPMC7899143 | biostudies-literature | 2018 Dec
REPOSITORIES: biostudies-literature
Publications
Human-in-the-Loop Interpretability Prior.
Lage Isaac I Ross Andrew Slavin AS Kim Been B Gershman Samuel J SJ Doshi-Velez Finale F
Advances in neural information processing systems 20181201
We often desire our models to be interpretable as well as accurate. Prior work on optimizing models for interpretability has relied on easy-to-quantify proxies for interpretability, such as sparsity or the number of operations required. In this work, we optimize for interpretability by <i>directly</i> including humans in the optimization loop. We develop an algorithm that minimizes the number of user studies to find models that are both predictive and interpretable and demonstrate our approach o ...[more]
Similar Datasets
Project description:Biological interpretability is a key requirement for the output of microarray data analysis pipelines. The most used pipeline first identifies a gene signature from the acquired measurements and then uses gene enrichment analysis as a tool for functionally characterizing the obtained results. Recently Knowledge Driven Variable Selection (KDVS), an alternative approach which performs both steps at the same time, has been proposed. In this paper, we assess the effectiveness of KDVS against standard approaches on a Parkinson's Disease (PD) dataset. The presented quantitative analysis is made possible by the construction of a reference list of genes and gene groups associated to PD. Our work shows that KDVS is much more effective than the standard approach in enhancing the interpretability of the obtained results.
| S-EPMC5003491 | biostudies-literature
Project description:Univariate analyses of metabolomics data currently follow a frequentist approach, using p-values to reject a null hypothesis. We here propose the use of Bayesian statistics to quantify evidence supporting different hypotheses and discriminate between the null hypothesis versus the lack of statistical power. We used metabolomics data from three independent human cohorts that studied the plasma signatures of subjects with myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS). The data are publicly available, covering 84-197 subjects in each study with 562-888 identified metabolites of which 777 were common between the two studies and 93 were compounds reported in all three studies. We show how Bayesian statistics incorporates results from one study as "prior information" into the next study, thereby improving the overall assessment of the likelihood of finding specific differences between plasma metabolite levels. Using classic statistics and Benjamini-Hochberg FDR-corrections, Study 1 detected 18 metabolic differences and Study 2 detected no differences. Using Bayesian statistics on the same data, we found a high likelihood that 97 compounds were altered in concentration in Study 2, after using the results of Study 1 as the prior distributions. These findings included lower levels of peroxisome-produced ether-lipids, higher levels of long-chain unsaturated triacylglycerides, and the presence of exposome compounds that are explained by the difference in diet and medication between healthy subjects and ME/CFS patients. Although Study 3 reported only 92 compounds in common with the other two studies, these major differences were confirmed. We also found that prostaglandin F2alpha, a lipid mediator of physiological relevance, was reduced in ME/CFS patients across all three studies. The use of Bayesian statistics led to biological conclusions from metabolomic data that were not found through frequentist approaches. We propose that Bayesian statistics is highly useful for studies with similar research designs if similar metabolomic assays are used.
| S-EPMC10535181 | biostudies-literature
Project description:Human brucellosis has reemerged in China, with a distinct change in its geographical distribution. The incidence of human brucellosis has significantly risen in inland regions of China. To gain insights into epidemic characteristics and identify factors influencing the geographic spread of human brucellosis, our study utilized the Extreme Gradient Boosting (XGBoost) algorithm and interpretable machine learning techniques. The results showed a consistent upward trend in the incidence of human brucellosis, with a significant increase of 8.20% from 2004 to 2021 (95% CI: 1.70, 15.10). The northern region continued to face a serious human situation, with a gradual upward trend. Meanwhile, the western and southern regions have experienced a gradual spread of human brucellosis, encompassing all regions of China over the past decade. Further analysis using Shapley Additive Explanations (SHAP) demonstrated that higher Gross Domestic Product (GDP) per capita and increased funding for education have the potential to reduce the spread. Conversely, the expansion of human brucellosis showed a positive correlation with bed availability per 1000 individuals, humidity, railway mileage, and GDP. These findings strongly suggest that socioeconomic factors play a more significant role in the spread of human brucellosis than other factors.
| S-EPMC10901783 | biostudies-literature
Project description:BackgroundStratification of patients according to their clinical prognosis is a desirable goal in cancer treatment in order to achieve a better personalized medicine. Reliable predictions on the basis of gene signatures could support medical doctors on selecting the right therapeutic strategy. However, during the last years the low reproducibility of many published gene signatures has been criticized. It has been suggested that incorporation of network or pathway information into prognostic biomarker discovery could improve prediction performance. In the meanwhile a large number of different approaches have been suggested for the same purpose.MethodsWe found that on average incorporation of pathway information or protein interaction data did not significantly enhance prediction performance, but indeed greatly interpretability of gene signatures. Some methods (specifically network-based SVMs) could greatly enhance gene selection stability, but revealed only a comparably low prediction accuracy, whereas Reweighted Recursive Feature Elimination (RRFE) and average pathway expression led to very clearly interpretable signatures. In addition, average pathway expression, together with elastic net SVMs, showed the highest prediction performance here.ResultsThe results indicated that no single algorithm to perform best with respect to all three categories in our study. Incorporating network of prior knowledge into gene selection methods in general did not significantly improve classification accuracy, but greatly interpretability of gene signatures compared to classical algorithms.
| S-EPMC3436770 | biostudies-literature
Project description:BackgroundCircular RNAs (circRNAs) can regulate microRNA activity and are related to various diseases, such as cancer. Functional research on circRNAs is the focus of scientific research. Accurate identification of circRNAs is important for gaining insight into their functions. Although several circRNA prediction models have been developed, their prediction accuracy is still unsatisfactory. Therefore, providing a more accurate computational framework to predict circRNAs and analyse their looping characteristics is crucial for systematic annotation.ResultsWe developed a novel framework, CircDC, for classifying circRNAs from other lncRNAs. CircDC uses four different feature encoding schemes and adopts a multilayer convolutional neural network and bidirectional long short-term memory network to learn high-order feature representation and make circRNA predictions. The results demonstrate that the proposed CircDC model is more accurate than existing models. In addition, an interpretable analysis of the features affecting the model is performed, and the computational framework is applied to the extended application of circRNA identification.ConclusionsCircDC is suitable for the prediction of circRNA. The identification of circRNA helps to understand and delve into the related biological processes and functions. Feature importance analysis increases model interpretability and uncovers significant biological properties. The relevant code and data in this article can be accessed for free at https://github.com/nmt315320/CircDC.git .
| S-EPMC10898045 | biostudies-literature
Project description:Improving the predictive capability and computational cost of dynamical models is often at the heart of augmenting computational physics with machine learning (ML). However, most learning results are limited in interpretability and generalization over different computational grid resolutions, initial and boundary conditions, domain geometries, and physical or problem-specific parameters. In the present study, we simultaneously address all these challenges by developing the novel and versatile methodology of unified neural partial delay differential equations. We augment existing/low-fidelity dynamical models directly in their partial differential equation (PDE) forms with both Markovian and non-Markovian neural network (NN) closure parameterizations. The melding of the existing models with NNs in the continuous spatiotemporal space followed by numerical discretization automatically allows for the desired generalizability. The Markovian term is designed to enable extraction of its analytical form and thus provides interpretability. The non-Markovian terms allow accounting for inherently missing time delays needed to represent the real world. Our flexible modeling framework provides full autonomy for the design of the unknown closure terms such as using any linear-, shallow-, or deep-NN architectures, selecting the span of the input function libraries, and using either or both Markovian and non-Markovian closure terms, all in accord with prior knowledge. We obtain adjoint PDEs in the continuous form, thus enabling direct implementation across differentiable and non-differentiable computational physics codes, different ML frameworks, and treatment of nonuniformly-spaced spatiotemporal training data. We demonstrate the new generalized neural closure models (gnCMs) framework using four sets of experiments based on advecting nonlinear waves, shocks, and ocean acidification models. Our learned gnCMs discover missing physics, find leading numerical error terms, discriminate among candidate functional forms in an interpretable fashion, achieve generalization, and compensate for the lack of complexity in simpler models. Finally, we analyze the computational advantages of our new framework.
| S-EPMC10313723 | biostudies-literature
Project description:BackgroundKernel methods have been proven to be a powerful tool for the integration and analysis of high-throughput technologies generated data. Kernels offer a nonlinear version of any linear algorithm solely based on dot products. The kernelized version of principal component analysis is a valid nonlinear alternative to tackle the nonlinearity of biological sample spaces. This paper proposes a novel methodology to obtain a data-driven feature importance based on the kernel PCA representation of the data.ResultsThe proposed method, kernel PCA Interpretable Gradient (KPCA-IG), provides a data-driven feature importance that is computationally fast and based solely on linear algebra calculations. It has been compared with existing methods on three benchmark datasets. The accuracy obtained using KPCA-IG selected features is equal to or greater than the other methods' average. Also, the computational complexity required demonstrates the high efficiency of the method. An exhaustive literature search has been conducted on the selected genes from a publicly available Hepatocellular carcinoma dataset to validate the retained features from a biological point of view. The results once again remark on the appropriateness of the computed ranking.ConclusionsThe black-box nature of kernel PCA needs new methods to interpret the original features. Our proposed methodology KPCA-IG proved to be a valid alternative to select influential variables in high-dimensional high-throughput datasets, potentially unravelling new biological and medical biomarkers.
| S-EPMC10337068 | biostudies-literature
Project description:BackgroundThe Rasch model allows for linear measurement based on ordinal item responses from rating scales commonly used to assess health outcomes. Such linear measures may be inconvenient since they are expressed as log-odds units (logits) that differ from scores that users may be familiar with. It can therefore be desirable to transform logits into more user-friendly ranges that preserve their linear properties. In addition to user-defined ranges, three general transformations have been described in the literature: the least measurable difference (LMD), the standard error of measurement (SEM) and the least significant difference (LSD). The LMD represents the smallest possible meaningful unit, SEM relates the transformed scale values to measurement uncertainty (one unit on the transformed scale represents roughly one standard error), and LSD represents a lower bound for how coarse the transformed scale can be without loss of valid information. However, while logit transformations are relatively common in the health sciences, use of LMD, SEM and LSD transformations appear to be uncommon despite their potential role.MethodsLogit transformations were empirically illustrated based on 1053 responses to the Epworth Sleepiness Scale. Logit measures were transformed according to the LMD, SEM and LSD, and into 0-10, 0-100, and the original raw score (0-24) ranges. These transformations were conducted using a freely available Excel tool, developed by the authors, that transforms logits into user-defined ranges along with the LMD, SEM and LSD transformations.ResultsResulting LMD, SEM and LSD transformations ranged 0-34, 0-17 and 0-12, respectively. When considering these relative to the three user-defined ranges, it is seen that the 0-10 range is narrower than the LSD range (i.e., loss of valid information), and a 0-100 range gives the impression of better precision than there is, since it is considerably wider than the LMD range. However, the 0-24 transformation appears reasonable since it is wider than the LSD, but narrower than the LMD ranges.ConclusionsIt is suggested that LMD, SEM and LSD transformations are valuable for benchmarking in deciding appropriate ranges when transforming logit measures. This process can be aided by the Excel tool presented and illustrated in this paper.
| S-EPMC9783398 | biostudies-literature
Project description:Deep neural networks display impressive performance but suffer from limited interpretability. Biology-inspired deep learning, where the architecture of the computational graph is based on biological knowledge, enables unique interpretability where real-world concepts are encoded in hidden nodes, which can be ranked by importance and thereby interpreted. In such models trained on single-cell transcriptomes, we previously demonstrated that node-level interpretations lack robustness upon repeated training and are influenced by biases in biological knowledge. Similar studies are missing for related models. Here, we test and extend our methodology for reliable interpretability in P-NET, a biology-inspired model trained on patient mutation data. We observe variability of interpretations and susceptibility to knowledge biases, and identify the network properties that drive interpretation biases. We further present an approach to control the robustness and biases of interpretations, which leads to more specific interpretations. In summary, our study reveals the broad importance of methods to ensure robust and bias-aware interpretability in biology-inspired deep learning.
| S-EPMC10564878 | biostudies-literature
Project description:In recent years, with the continuous improvement of the financial system and the rapid development of the banking industry, the competition of the banking industry itself has intensified. At the same time, with the rapid development of information technology and Internet technology, customers' choice of financial products is becoming more and more diversified, and customers' dependence and loyalty to banking institutions is becoming less and less, and the problem of customer churn in commercial banks is becoming more and more prominent. How to predict customer behavior and retain existing customers has become a major challenge for banks to solve. Therefore, this study takes a bank's business data on Kaggle platform as the research object, uses multiple sampling methods to compare the data for balancing, constructs a bank customer churn prediction model for churn identification by GA-XGBoost, and conducts interpretability analysis on the GA-XGBoost model to provide decision support and suggestions for the banking industry to prevent customer churn. The results show that: (1) The applied SMOTEENN is more effective than SMOTE and ADASYN in dealing with the imbalance of banking data. (2) The F1 and AUC values of the model improved and optimized by XGBoost using genetic algorithm can reach 90% and 99%, respectively, which are optimal compared to other six machine learning models. The GA-XGBoost classifier was identified as the best solution for the customer churn problem. (3) Using Shapley values, we explain how each feature affects the model results, and analyze the features that have a high impact on the model prediction, such as the total number of transactions in the past year, the amount of transactions in the past year, the number of products owned by customers, and the total sales balance. The contribution of this paper is mainly in two aspects: (1) this study can provide useful information from the black box model based on the accurate identification of churned customers, which can provide reference for commercial banks to improve their service quality and retain customers; (2) it can provide reference for customer churn early warning models of other related industries, which can help the banking industry to maintain customer stability, maintain market position and reduce corporate losses.
| S-EPMC10707658 | biostudies-literature