Towards a Consistent Interpretation of AIOps Models

Artificial Intelligence for IT Operations (AIOps) has been adopted in organizations in various tasks, including interpreting models to identify indicators of service failures. To avoid misleading practitioners, AIOps model interpretations should be consistent (i.e., different AIOps models on the same task agree with one another on feature importance). However, many AIOps studies violate established practices in the machine learning community when deriving interpretations, such as interpreting models with suboptimal performance, though the impact of such violations on the interpretation consistency has not been studied. In this article, we investigate the consistency of AIOps model interpretation along three dimensions: internal consistency, external consistency, and time consistency. We conduct a case study on two AIOps tasks: predicting Google cluster job failures and Backblaze hard drive failures. We find that the randomness from learners, hyperparameter tuning, and data sampling should be controlled to generate consistent interpretations. AIOps models with AUCs greater than 0.75 yield more consistent interpretation compared to low-performing models. Finally, AIOps models that are constructed with the Sliding Window or Full History approaches have the most consistent interpretation with the trends presented in the entire datasets. Our study provides valuable guidelines for practitioners to derive consistent AIOps model interpretation.

Topsoil Texture Regionalization for Agricultural Soils in Germany—An Iterative Approach to Advance Model Interpretation

Site-specific spatially continuous soil texture data is required for many purposes such as the simulation of carbon dynamics, the estimation of drought impact on agriculture, or the modeling of water erosion rates. At large scales, there are often only conventional polygon-based soil texture maps, which are hardly reproducible, contain abrupt changes at polygon borders, and therefore are not suitable for most quantitative applications. Digital soil mapping methods can provide the required soil texture information in form of reproducible site-specific predictions with associated uncertainties. Machine learning models were trained in a nested cross-validation approach to predict the spatial distribution of the topsoil (0–30 cm) clay, silt, and sand contents in 100 m resolution. The differential evolution algorithm was applied to optimize the model parameters. High-quality nation-wide soil texture data of 2,991 soil profiles was obtained from the first German agricultural soil inventory. We tested an iterative approach by training models on predictor datasets of increasing size, which contained up to 50 variables. The best results were achieved when training the models on the complete predictor dataset. They explained about 59% of the variance in clay, 75% of the variance in silt, and 77% of the variance in sand content. The RMSE values ranged between approximately 8.2 wt.% (clay), 11.8 wt.% (silt), and 15.0 wt.% (sand). Due to their high performance, models were able to predict the spatial texture distribution. They captured the high importance of the soil forming factors parent material and relief. Our results demonstrate the high predictive power of machine learning in predicting soil texture at large scales. The iterative approach enhanced model interpretability. It revealed that the incorporated soil maps partly substituted the relief and parent material predictors. Overall, the spatially continuous soil texture predictions provide valuable input for many quantitative applications on agricultural topsoils in Germany.

Interpreting tree ensemble machine learning models with endoR

Background: Tree ensemble machine learning models are increasingly used in microbiome science as they are compatible with the compositional, high-dimensional, and sparse structure of sequence-based microbiome data. While such models are often good at predicting phenotypes based on microbiome data, they only yield limited insights into how microbial taxa or genomic content may be associated. Results: We developed endoR, a method to interpret a fitted tree ensemble model. First, endoR simplifies the fitted model into a decision ensemble from which it then extracts information on the importance of individual features and their pairwise interactions and also visualizes these data as an interpretable network. Both the network and importance scores derived from endoR provide insights into how features, and interactions between them, contribute to the predictive performance of the fitted model. Adjustable regularization and bootstrapping help reduce the complexity and ensure that only essential parts of the model are retained. We assessed the performance of endoR on both simulated and real metagenomic data. We found endoR to infer true associations with more or comparable accuracy than other commonly used approaches while easing and enhancing model interpretation. Using endoR, we also confirmed published results on gut microbiome differences between cirrhotic and healthy individuals. Finally, we utilized endoR to gain insights into components of the microbiome that predict the presence of human gut methanogens, as these hydrogen-consumers are expected to interact with fermenting bacteria in a complex syntrophic network. Specifically, we analyzed a global metagenome dataset of 2203 individuals and confirmed the previously reported association between Methanobacteriaceae and Christensenellales. Additionally, we observed that Methanobacteriaceae are associated with a network of hydrogen-producing bacteria. Conclusion: Our method accurately captures how tree ensembles use features and interactions between them to predict a response. As demonstrated by our applications, the resultant visualizations and summary outputs facilitate model interpretation and enable the generation of novel hypotheses about complex systems. An implementation of endoR is available as an open-source R-package on GitHub (https://github.com/leylabmpi/endoR).

Integrating Individual Factors to Construct Recognition Models of Consumer Fraud Victimization

Consumer financial fraud has become a serious problem because it often causes victims to suffer economic, physical, mental, social, and legal harm. Identifying which individuals are more likely to be scammed may mitigate the threat posed by consumer financial fraud. Based on a two-stage conceptual framework, this study integrated various individual factors in a nationwide survey (36,202 participants) to construct fraud exposure recognition (FER) and fraud victimhood recognition (FVR) models by utilizing a machine learning method. The FER model performed well (f1 = 0.727), and model interpretation indicated that migration status, financial status, urbanicity, and age have good predictive effects on fraud exposure in the Chinese context, whereas the FVR model shows a low predictive effect (f1 = 0.565), reminding us to consider more psychological factors in future work. This research provides an important reference for the analysis of individual differences among people vulnerable to consumer fraud.

An effective matrix model for dynamical end of the world branes in Jackiw-Teitelboim gravity

We study Jackiw-Teitelboim gravity with dynamical end of the world branes in asymptotically nearly AdS2 spacetimes. We quantize this theory in Lorentz signature, and compute the Euclidean path integral summing over topologies including dynamical branes. The latter will be seen to exactly match with a modification of the SSS matrix model. The resolution of UV divergences in the gravitational instantons involving the branes will lead us to understand the matrix model interpretation of the Wilsonian effective theory perspective on the gravitational theory. We complete this modified SSS matrix model nonperturbatively by extending the integration contour of eigenvalues into the complex plane. Furthermore, we give a new interpretation of other phases in such matrix models. We derive an effective W(Φ) dilaton gravity, which exhibits similar physics semiclassically. In the limit of a large number of flavors of branes, the effective extremal entropy S0,eff has the form of counting the states of these branes.

Automatic Approach for Detecting the Seizure Using RCCN Architecture

Seizure is one of the most common neurodegenerative illnesses in humans, and it can result in serious brain damage, strokes, and tumors. Seizures can be detected early, which can assist prevent harm and aid in the treatment of epilepsy sufferers. A seizure prediction system's goal is to correctly detect the pre-ictal brain state, which occurs before a seizure occurs. Patient-independent seizure prediction models have been recognized as a real-world solution to the seizure prediction problem, since they are designed to provide accurate performance across different patients by using the recorded dataset. Furthermore, building such models to adjust to the significant inter-subject variability in EEG data has received little attention. We present a patient-independent deep learning architectures that can train a global function using data from numerous people with its own learning strategy. On the CHB- MIT-EEG dataset, the proposed models reach state-of-the-art accuracy for seizure prediction, with 95.54 percent accuracy. While predicting seizures, the Siamese model trained on the suggested learning technique is able to understand patterns associated to patient differences in data. Our models outperform the competition in terms of patient-independent seizure prediction, and following model adaption, the same architecture may be employed as a patient-specific classifier. We show that the MFCC feature map used by our models contains predictive biomarkers associated to inter-ictal and pre-ictal brain states, and we are the first study to use model interpretation to explain classifier behaviour for the task of seizure prediction.

Towards More Realistic Simulated Datasets for Benchmarking Deep Learning Models in Regulatory Genomics

Deep neural networks and support vector machines have been shown to accurately predict genomewide signals of regulatory activity from raw DNA sequences. These models are appealing in part because they can learn predictive DNA sequence features without prior assumptions. Several methods such as in-silico mutagenesis, GradCAM, DeepLIFT, Integrated Gradients and GkmExplain have been developed to reveal these learned features. However, the behavior of these methods on regulatory genomic data remains an area of active research. Although prior work has benchmarked these methods on simulated datasets with known ground-truth motifs, these simulations employed highly simplified regulatory logic that is not representative of the genome. In this work, we propose a novel pipeline for designing simulated data that comes closer to modeling the complexity of regulatory genomic DNA. We apply the pipeline to build simulated datasets based on publicly-available chromatin accessibility experiments and use these datasets to benchmark different interpretation methods based on their ability to identify ground-truth motifs. We find that a GradCAM-based method, which was reported to perform well on a more simplified dataset, does not do well on this dataset (particularly when using an architecture with shorter convolutional kernels in the first layer), and we theoretically show that this is expected based on the nature of regulatory genomic data. We also show that Integrated Gradients sometimes performs worse than gradient-times-input, likely owing to its linear interpolation path. We additionally explore the impact of user-defined settings on the interpretation methods, such as the choice of "reference"/"baseline", and identify recommended settings for genomics. Our analysis suggests several promising directions for future research on these model interpretation methods. Code and links to data are available at https://github.com/kundajelab/interpret-benchmark.

Wearable Devices, Smartphones, and Interpretable Artificial Intelligence in Combating COVID-19

Physiological measures, such as heart rate variability (HRV) and beats per minute (BPM), can be powerful health indicators of respiratory infections. HRV and BPM can be acquired through widely available wrist-worn biometric wearables and smartphones. Successive abnormal changes in these indicators could potentially be an early sign of respiratory infections such as COVID-19. Thus, wearables and smartphones should play a significant role in combating COVID-19 through the early detection supported by other contextual data and artificial intelligence (AI) techniques. In this paper, we investigate the role of the heart measurements (i.e., HRV and BPM) collected from wearables and smartphones in demonstrating early onsets of the inflammatory response to the COVID-19. The AI framework consists of two blocks: an interpretable prediction model to classify the HRV measurements status (as normal or affected by inflammation) and a recurrent neural network (RNN) to analyze users’ daily status (i.e., textual logs in a mobile application). Both classification decisions are integrated to generate the final decision as either “potentially COVID-19 infected” or “no evident signs of infection”. We used a publicly available dataset, which comprises 186 patients with more than 3200 HRV readings and numerous user textual logs. The first evaluation of the approach showed an accuracy of 83.34 ± 1.68% with 0.91, 0.88, 0.89 precision, recall, and F1-Score, respectively, in predicting the infection two days before the onset of the symptoms supported by a model interpretation using the local interpretable model-agnostic explanations (LIME).

On BIM Interoperability via the IFC Standard: An Assessment from the Structural Engineering and Design Viewpoint

Building information modelling (BIM) plays a prominent role in a good deal of architecture, engineering and construction (AEC) works, envisaging a full transition to digitalization for the construction industry. This is also due to a number of national and international regulations regarding the design, erection, and management of civil engineering constructions. For this reason, full interoperability of software environments such as computer-aided design (CAD) and computer-aided engineering (CAE) is a necessary requirement, particularly when the exchange of information comes from different disciplines. Users, throughout the years, have faced CAD–CAE interoperability issues despite following the IFC neutral open file format. This inability to share data (CAD to CAD, CAD to CAE) often generates model-interpretation problems as well as a lack of parametric information and a disconnection of elements. This paper addresses issues and mapping mechanisms in the exchange of data for the purpose of defining a baseline for the current status of bidirectional data exchange between AEC CAD/CAE software via the IFC format. A benchmark study, covering three years of software releases is illustrated; the assessment of the software performance was made with reference to criteria associated with the software’s level of suitability for use of the structural models. Four classes of performance, depending on the accuracy of the data transfer and on the associated corrective actions to be taken, were adopted. This confirmed that at the moment, the implementation of the IFC standard by software manufacturers is geared towards an expert class of users. Further efforts are needed in order to ensure its application is adopted by a wider class, thus extending and regulating its use by national, regional, and local authorities.

Benchmarking Geometric Deep Learning for Cortical Segmentation and Neurodevelopmental Phenotype Prediction

The emerging field of geometric deep learning extends the application of convolutional neural networks to irregular domains such as graphs, meshes and surfaces. Several recent studies have explored the potential for using these techniques to analyse and segment the cortical surface. However, there has been no comprehensive comparison of these approaches to one another, nor to existing Euclidean methods, to date. This paper benchmarks a collection of geometric and traditional deep learning models on phenotype prediction and segmentation of sphericalised neonatal cortical surface data, from the publicly available Developing Human Connectome Project (dHCP). Tasks include prediction of postmenstrual age at scan, gestational age at birth and segmentation of the cortical surface into anatomical regions defined by the M-CRIB-S atlas. Performance was assessed not only in terms of model precision, but also in terms of network dependence on image registration, and model interpretation via occlusion. Networks were trained both on sphericalised and anatomical cortical meshes. Findings suggest that the utility of geometric deep learning over traditional deep learning is highly task-specific, which has implications for the design of future deep learning models on the cortical surface. The code, and instructions for data access, are available from https://github.com/Abdulah-Fawaz/Benchmarking-Surface-DL.