Machine Learning Clustering Research Articles

Competition for popularity and interventions on a Chinese microblogging site.

Microblogging sites are important vehicles for the users to obtain information and shape public opinion thus they are arenas of continuous competition for popularity. Most popular topics are usually indicated on ranking lists. In this study, we investigate the public attention dynamics through the Hot Search List (HSL) of the Chinese microblog Sina Weibo, where trending hashtags are ranked based on a multi-dimensional search volume index. We characterize the rank dynamics by the time spent by hashtags on the list, the time of the day they appear there, the rank diversity, and by the ranking trajectories. We show how the circadian rhythm affects the popularity of hashtags, and observe categories of their rank trajectories by a machine learning clustering algorithm. By analyzing patterns of ranking dynamics using various measures, we identify anomalies that are likely to result from the platform provider's intervention into the ranking, including the anchoring of hashtags to certain ranks on the HSL. We propose a simple model of ranking that explains the mechanism of this anchoring effect. We found an over-representation of hashtags related to international politics at 3 out of 4 anchoring ranks on the HSL, indicating possible manipulations of public opinion.

The Current State of the Livestream Shopping Market in the US

The global live-streaming market was significantly impacted by the COVID-19 pandemic in the first half of 2020. In this study, we aim to analyze the current state of the live-streaming shopping market in the US from the outbreak's onset to the present, i.e., 2022. In our research, our primary core approach is around the modeling of machine learning clustering. First, we used R language to get data from Twitter, the mainstream social media in the US, then used common word text mining to analyze 5000 tweets. After several trials and analyses, we confirmed the division of eight subgroups and three large groups (platforms, live beauty streaming, and blockchain virtual currency) after the final data analysis session. These specific groups demonstrate which parts of Livestream shopping are of particular appeal to American customers, providing other marketing ideas for merchants and platform partners.

Assessing county-level vulnerability to the energy transition in the United States using machine learning

Energy sector decarbonization will play a critical role in society's attempt to minimize climate change. This transition will unavoidably involve economic disruption to communities that rely on fossil fuels for jobs and economic activity. In this work, we assess the economic vulnerability of counties across the contiguous United States to the energy transition using a novel methodological approach that combines exposure through reduction in energy employment and socio-economic sensitivity due to a lack of adaptive capacity during the energy transition. Using machine learning clustering, we first identify county-level exposure to negative employment shocks. We then develop and implement the Resilience during the Energy Transition Index (RETI) to measure county-level sensitivity and adaptive capacity and compared this Index across employment clusters. We find that counties in coal mining clusters are significantly – and uniquely – less able to cope with economic distress. We also apply our framework to a case study focused on three of the states in Marcellus-Utica shale plays: Pennsylvania, Ohio and West Virginia. The results of this case study support the argument that financial and infrastructure boosts from non-renewable energy extraction and utilization build resilience only if they are used to building capital for the post‑carbon economy, providing further evidence of the conditions required for fossil fuels to serve as bridge fuels towards a low-carbon economy.

Identifikasi Arsip Digital dengan Pendekatan Machine Learning

The development of archive digitization technology, especially cameras, is an advantage for archive users. On the other hand, the ease of falsifying by digitizing records for a particular purpose is a problem for archivists in authenticating digital records. The purpose of this study is to identify digital sources, especially cameras as one of the tools applied in the archive digitization process. The research methodology combines clustering and classification in machine learning to determine 6 brands of cellphone cameras. The total data used in this experiment is 2400 archived digital images. The experimental results show that the identification rate of classification accuracy is 99%. This shows that this study is very effective in determining the authentication of digital archives, especially determining the source of the camera.

Spatiotemporal analysis of dune stabilization in the Kyzylkum and Karakum sandy deserts

Sand seas and dune bedforms are climatic chronicles that have a wide range of scientific meanings. The Karakum and Kyzylkum sand seas in Central Asia span ∼500,000 km2, among the largest sand seas on Earth; yet, very little is known about their emergence and stabilization. Further, the internationally accessible literature lacks comprehensive studies on the interactions between contemporary climate and dune morphodynamics for this region. In this study, we applied the conventional approach for examining wind data—Fryberger’s drift equations and the latest Copernicus ERA5 wind reanalysis model—to demonstrate that the Karakum and Kyzylkum deserts had (and continue to have) low-energy environments during 1950–2019. Time-series multidimensional raster analysis and data mining indicated that 93% of surveyed desert areas were subjected to low-energy wind conditions during the last 70 years. The decrease in wind energy was further confirmed using the partitioning around medoids (PAM) unsupervised machine-learning clustering method and Getis-Ord Gi* statistics. Such an environmental setting promotes plant growth and biogenic crust development, and strengthens dune stability across the deserts. Furthermore, the results suggested that the spatial patterns of active dune morphologies (e.g., complex barchanoid ridges and overlaying superimpositions) were closely linked to the drift direction, power, and variability of contemporary winds, whereas stabilized aeolian formations (e.g., distinct formations of complex and compound vegetated linear dunes) indicated debatable disassociations. The relationship between the recognized orientation of the surveyed dune types and the computed resultant drift direction did not change significantly, suggesting that over time, the winds in these areas have most likely varied in strength, but not in direction. The calculated drift powers are not sufficient for the formation of registered mega profiles. Rather, they bolster dunes stability, stressing both deserts as indicative of earlier climatic eras that permitted aeolian sand activity and severe shifts that have initiated their steadiness. This is the first comprehensive study to statistically evaluate the links between the contemporary climate and dune morphologies of the Karakum and Kyzylkum deserts. Our study can serve as a foundation for future studies on climate change, dune mapping, and optically stimulated luminescence (OSL) dating, while offering a broad understanding of the region’s paleo and contemporary environmental systems dynamics.

Influence of User Profile Attributes on e-Cigarette-Related Searches on YouTube: Machine Learning Clustering and Classification.

The proliferation of e-cigarette content on YouTube is concerning because of its possible effect on youth use behaviors. YouTube has a personalized search and recommendation algorithm that derives attributes from a user's profile, such as age and sex. However, little is known about whether e-cigarette content is shown differently based on user characteristics. The aim of this study was to understand the influence of age and sex attributes of user profiles on e-cigarette-related YouTube search results. We created 16 fictitious YouTube profiles with ages of 16 and 24 years, sex (female and male), and ethnicity/race to search for 18 e-cigarette-related search terms. We used unsupervised (k-means clustering and classification) and supervised (graph convolutional network) machine learning and network analysis to characterize the variation in the search results of each profile. We further examined whether user attributes may play a role in e-cigarette-related content exposure by using networks and degree centrality. We analyzed 4201 nonduplicate videos. Our k-means clustering suggested that the videos could be clustered into 3 categories. The graph convolutional network achieved high accuracy (0.72). Videos were classified based on content into 4 categories: product review (49.3%), health information (15.1%), instruction (26.9%), and other (8.5%). Underage users were exposed mostly to instructional videos (37.5%), with some indication that more female 16-year-old profiles were exposed to this content, while young adult age groups (24 years) were exposed mostly to product review videos (39.2%). Our results indicate that demographic attributes factor into YouTube's algorithmic systems in the context of e-cigarette-related queries on YouTube. Specifically, differences in the age and sex attributes of user profiles do result in variance in both the videos presented in YouTube search results as well as in the types of these videos. We find that underage profiles were exposed to e-cigarette content despite YouTube's age-restriction policy that ostensibly prohibits certain e-cigarette content. Greater enforcement of policies to restrict youth access to e-cigarette content is needed.

A new computational approach, based on images trajectories, to identify the subjacent heterogeneity of sperm to the effects of ketanserin.

The identification of kinematic subpopulations is of paramount importance to understanding the biological nature of the sperm heterogeneity. Nowadays, the data of motility parameters obtained by a computer-assisted sperm analysis (CASA) system has been used as input to distinct algorithms to identify kinematic subpopulations. In contrast, the images of the trajectories were depicted only as examples of the patterns of motility in each subpopulation. Here, python code was written to reconstruct the images of trajectories, from their coordinates, then the images of trajectories were used as input to a machine learning clustering algorithm of classification, and the subpopulations were described statistically by the motility parameters. Finally, the images of trajectories in each subpopulation were displayed in a way we called Pollock plots. Semen samples of boar sperm were treated with distinct concentrations of ketanserin (an antagonist of the 5-HT2 receptor of serotonin) and untreated samples were used as a control. The motility of sperm in each sample was analyzed at 0 and 30 min of incubation. Six subpopulations were found. The subpopulation 2 presented the highest values of velocities at 0 or 30 min. After 30 min of incubation, the ketanserin increased the values of the curvilinear velocity at high concentrations, whereas the linearity and the straight velocity decreased. Our computational model permits better identification of the kinematic subpopulations than the traditional approach and provides insights onto the heterogeneity of the response to ketanserin; thus, it could significantly impact the research on the relationship between sperm heterogeneity-fertility.

Abstract 6614: Spatial transcriptomic driven mechanistic model to investigate and predict pathogenesis of oligodendroglioma

Abstract Rationale: Oligodendrogliomas represent the second most common glioma in adults after glioblastoma. Recent studies conjecture that oligodendrocyte progenitor cells (OPCs) are precursors for glioma. The etiology of OPC derived oligodendrogliomas, the corresponding genetic program and genetic aberrations, and the underlying mechanisms leading to the pathogenesis of oligodendroglioma is yet to be explored. We hypothesize that genetic mechanisms mediating oligodendrogenesis (the process of OPCs maturing to oligodendrocytes (OL)) are temporally and spatially coordinated. In this study, we spearhead a comprehensive computational framework that utilizes a) Machine learning (ML) to interpret spatial transcriptomic (ST) images of a normal whole mouse brain dataset obtained from Vizgen’s MERFISH Mouse Brain Receptor Map and b) mathematical modeling to perturb the cellular mechanisms in the normal brain to understand the abnormalities leading to the pathogenesis of oligodendroglioma. Methods: Using ML for clustering and trajectory inference, we create a comprehensive spatial catalog of cell types. The cell types are validated using canonical genes from the scRNAseq mouse brain data from CZI Biohub’s Tabula Muris Consortium. From the cell types, we extract potential OL pseudotime differentiation (from immature to mature OLs) trajectories. These trajectories are used to develop a phenotype structured partial differentiation equation (PDE) model of oligodendrogenesis. This can be used to simulate both spatial and phenotypic interactions, differentiation status, and the pathogenesis of oligodendroglioma by mathematically perturbing the process of cellular proliferation and differentiation using genes known to cause oligodendroglioma: PDGFRA, OLIG1, MAS1. Results: We mathematically perturb cell differentiation lineages in normal oligodendrogenesis to simulate, and predict the effects of perturbation and how this leads to abnormal (cancer aiding) cell states. Our PDE simulations reinforce that OPC proliferation and OL differentiation occur more in the white matter in the normal mouse brain, and further postulate that OPC and OL differentiation occur more in gray matter in malignant mouse brains. Conclusions: Our study is one of the earliest attempts to integrate qualitative and scalable ML techniques with quantitative and interpretable mathematical modeling at the cellular, omics, and organ level using ST imaging and sequencing data. We develop and calibrate PDE mathematical models for robustly simulating, predicting, and studying the effects of perturbation from sparse, noisy single cell spatial measurements to forecast abnormal behavior in OLs based on spatial and (pseudo)temporal changes. In future work, we will apply our framework to jointly perturb differentiation trajectories of multiple cell types to understand their cumulative effects in tumor progression. Citation Format: Sandhya Prabhakaran, Heyrim Cho, Maximilian Strobl, Russell Rockne, Alexander R. Anderson. Spatial transcriptomic driven mechanistic model to investigate and predict pathogenesis of oligodendroglioma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 6614.

Deep Learning-Based Job Placement in Distributed Machine Learning Clusters With Heterogeneous Workloads

Nowadays, most leading IT companies host a variety of distributed machine learning (ML) workloads in ML clusters to support AI-driven services, such as speech recognition, machine translation, and image processing. While multiple jobs are executed concurrently in a shared cluster to improve resource utilization, interference among co-located ML jobs can lead to significant performance downgrade. Existing cluster schedulers, such as YARN and Mesos, are interference-agnostic in their job placement, leading to suboptimal resource efficiency and usage. Some literature has studied interference-aware job placement policy, but relies on detailed workload profiling and interference modeling, which is not a general solution. In this work, we present Harmony, a deep learning-driven ML cluster scheduler that places heterogeneous training jobs (either with parameter server architecture or all-reduce architecture) in a manner that minimizes interference and maximizes performance (i.e., training completion time minimization). The design of Harmony is based on a carefully designed deep reinforcement learning (DRL) framework enhanced with reward modeling. The DRL integrates a dynamic sequence-to-sequence model with the state-of-the-art techniques to stabilize training and improve convergence, including actor-critic algorithm, job-aware action space exploration, multi-head attention, and experience replay. In view of a common lack of reward samples corresponding to different placement decisions, we build an auxiliary sequence-to-sequence reward prediction model, which is trained with historical samples and used for producing reward for unseen placement. Experiments using real ML workloads in a Kubernetes cluster of 6 GPU servers show that Harmony outperforms representative schedulers by 16%–42% in terms of average job completion time.

Distinct Phenotypes of Non-Citizen Kidney Transplant Recipients in the United States by Machine Learning Consensus Clustering.

Better understanding of the different phenotypes/subgroups of non-U.S. citizen kidney transplant recipients may help the transplant community to identify strategies that improve outcomes among non-U.S. citizen kidney transplant recipients. This study aimed to cluster non-U.S. citizen kidney transplant recipients using an unsupervised machine learning approach; Methods: We conducted a consensus cluster analysis based on recipient-, donor-, and transplant- related characteristics in non-U.S. citizen kidney transplant recipients in the United States from 2010 to 2019 in the OPTN/UNOS database using recipient, donor, and transplant-related characteristics. Each cluster's key characteristics were identified using the standardized mean difference. Post-transplant outcomes were compared among the clusters; Results: Consensus cluster analysis was performed in 11,300 non-U.S. citizen kidney transplant recipients and identified two distinct clusters best representing clinical characteristics. Cluster 1 patients were notable for young age, preemptive kidney transplant or dialysis duration of less than 1 year, working income, private insurance, non-hypertensive donors, and Hispanic living donors with a low number of HLA mismatch. In contrast, cluster 2 patients were characterized by non-ECD deceased donors with KDPI <85%. Consequently, cluster 1 patients had reduced cold ischemia time, lower proportion of machine-perfused kidneys, and lower incidence of delayed graft function after kidney transplant. Cluster 2 had higher 5-year death-censored graft failure (5.2% vs. 9.8%; p < 0.001), patient death (3.4% vs. 11.4%; p < 0.001), but similar one-year acute rejection (4.7% vs. 4.9%; p = 0.63), compared to cluster 1; Conclusions: Machine learning clustering approach successfully identified two clusters among non-U.S. citizen kidney transplant recipients with distinct phenotypes that were associated with different outcomes, including allograft loss and patient survival. These findings underscore the need for individualized care for non-U.S. citizen kidney transplant recipients.

Exploring the potential role of bikeshare to complement public transit: The case of San Francisco amid the coronavirus crisis

The recent worldwide SARS-CoV-2 (COVID-19) pandemic has reshaped the way people live, how they access goods and services, and how they perform various activities. For public transit, there have been health concerns over the potential spread to transit users and transit service staff, which prompted transportation agencies to make decisions about the service, e.g., whether to reduce or temporarily shut down services. These decisions had substantial negative consequences, especially for transit-dependent travelers, and prompted transit users to explore alternative transportation modes, e.g., bikeshare. However, local governments and the public in general have limited information about whether and to what extent bikeshare provides adequate accessibility and mobility to those transit-dependent residents. To fill this gap, this study implemented spatial and visual analytics to identify how micro-mobility in the form of bikesharing has addressed travel needs and improved the resilience of transportation systems. The study analyzed the case of San Francisco in California, USA, focusing on three phases of the pandemic, i.e., initial confirmed cases, shelter-in-place, and initial changes in transit service. First, the authors implemented unsupervised machine learning clustering methods to identify different bikesharing trip types. Moreover, through spatiotemporally matching bikeshare ridership data with transit service information (i.e., General Transit Feed Specification, GTFS) using the tool called OpenTripPlanner (OTP), the authors studied the travel behavior changes (e.g., the proportion of bikeshare trips that could be finished by transit) for different bikeshare trip types over the three specified phases. This study revealed that during the pandemic, more casual users joined bikeshare programs; the proportion of recreation-related bikeshare trips increased; and routine trips became more prevalent considering that docking-station-based bikeshare trips increased. More importantly, the analyses also provided insights about mode substitution, because the analyses identified an increase in dockless bikeshare trips in areas with no or limited transit coverage.

Carpooling: Who is closest to adopting it? An investigation into the potential car-poolers among private vehicle users: A case of a developing country, India

The demand for transit is linked to a country's economic and population growth. As per the Ministry of Road Transport and Highways (MORTH), from 2009 to 2019, India witnessed an average growth rate of 9.9% per annum in registered vehicles. Even though the vehicle ownership is often viewed as a symbol of economic prosperity, the associated negative externalities (congestion, air pollution, road crashes, noise pollution, etc.) are substantial. To counter this, carpooling has been proposed as a solution globally. However, there is a paucity of literature for analysing the difference in travel characteristics of car-poolers among other private vehicle users. The present study aims to bridge this gap by comparing the demographic and travel characteristics of existing car-poolers with two-wheelers and single-occupant car users in Gurugram in the national capital region of India. Using an unsupervised machine learning clustering method (two-step cluster), the findings of the present study indicate that the profile of single-occupant car (SOC) users and car-poolers are highly distinguished from two-wheeler users. However, car-poolers have similar characteristics to single-occupant car users, except for travel cost, income level, and gender, suggesting that SOC users are potential car-poolers. A comparative study of this kind would help city planners and policymakers identify prospective car-poolers, and formulate policies to encourage SOC users to adopt carpooling and promote urban transportation sustainability.

A Taxonomy of Machine Learning Clustering Algorithms, Challenges, and Future Realms

In the field of data mining, clustering has shown to be an important technique. Numerous clustering methods have been devised and put into practice, and most of them locate high-quality or optimum clustering outcomes in the field of computer science, data science, statistics, pattern recognition, artificial intelligence, and machine learning. This research provides a modern, thorough review of both classic and cutting-edge clustering methods. The taxonomy of clustering is presented in this review from an applied angle and the compression of some hierarchical and partitional clustering algorithms with various parameters. We also discuss the open challenges in clustering such as computational complexity, refinement of clusters, speed of convergence, data dimensionality, effectiveness and scalability, data object representation, evaluation measures, data streams, and knowledge extraction; scientists and professionals alike will be able to use it as a benchmark as they strive to advance the state-of-the-art in clustering techniques.

Machine learning clustering algorithms for the automatic generation of chemical reactor networks from CFD simulations

Predicting the thermal and environmental performances of combustion systems can be difficult and computationally expensive. Chemical Reactor Networks (CRN) represent an appealing solution for performing faster numerical simulations since they can carry out calculations with detailed kinetics in a significantly reduced amount of time. However, the design of such models is challenging, as it usually requires a considerable amount of expertise from the user. In this work, we present a novel, automatic methodology for the design of CRN models, which consists of the post-processing of CFD data via a combination of unsupervised clustering and graph scanning algorithms. The methodology was tested on a semi-industrial furnace which can operate in Moderate or Intense Low oxygen Dilution (MILD) combustion. The furnace operates at a nominal power input of 15 kW and is fed with several CH4-H2 mixtures, with two different air injector diameters. First, RANS simulations of the different cases were performed and validated against experimental data. Subsequently, different CRNs were extracted singularly for each case and simulated using detailed kinetics mechanisms. The different CRNs showed good performances in predicting NO emissions for the entire range of cases, as the results were in reasonable agreement with experimental data. Last, the capability of a single CRN to extrapolate towards different operating conditions was also assessed by using one single CRN for the simulation of different fuel mixtures. The approach showed a good level of generalization since the NO predictions were close to the experimental values.

Seismic identification of carbonate reservoir sweet spots using unsupervised machine learning: A case study from Brazil deep water Aptian pre-salt data

The pre-salt high-yield carbonate reservoirs of Brazil are responsible for about 75% of national oil production. Still, seventeen years since its discovery, detailed seismic characterization of these reservoirs remains challenging, with producing and nonproducing lithologies often exhibiting the same amplitude response. Those reservoirs are complex, composed of high productivity lacustrine carbonates (stromatolites and grainstones), mixed with non-producing mudstones, clays, and igneous intrusions. To address this heterogeneity, we evaluate unsupervised machine learning clustering techniques including Self Organizing Maps (SOM), Generative Topographic Maps (GTM), and K-means and apply them to a suite of instantaneous, geometric, interval velocity, and AVO seismic attributes. We use both static and dynamic reservoir data from 13 wells at the Mero field that penetrated lacustrine carbonates from Barra Velha formation to validate the accuracy of our unsupervised learning models. These wells are excellent blind tests as they were not used for any algorithm training. In this pioneer work, the first done using machine learning in the Mero field, we find that SOM, GTM, and K-means models provide clusters that can be linked to sweet spots associated with fracturing, karstification, and hydrothermal alteration. Additionally, clusters related to sill intrusions, platform facies and distal fine grained non-reservoirs were identified. The results can be extrapolated to the adjacent Central Libra Area (still in exploration phase), where we have identified the same clusters of the sweet spots reservoirs in Mero Field and, thus, they can be considered as future targets. For accuracy assessment, the same attribute combination SOM clustering was performed in another Petrobras’ pre-salt confidential dataset and the results were satisfactory. So, for exploration projects optimization purposes, the new proposed workflows can also be applied to other areas with geological similarity to the Mero Field.

Large Data Limit of the MBO Scheme for Data Clustering

AbstractThe MBO scheme is a highly performant scheme used for data clustering and other applications in machine learning. We report on the first theoretical studies of the scheme in the large data limit. Our results relate (i) the final state of the scheme to minimal surfaces in the data manifold and (ii) the dynamics of the scheme to the steepest descent for surfaces, which is mean curvature flow. The tools employed are variational methods for (i) and viscosity solution techniques for (ii).

Diffusion Basis Spectrum Imaging Identifies Clinically Relevant Disease Phenotypes of Cervical Spondylotic Myelopathy.

Prospective cohort study. Apply a machine learning clustering algorithm to baseline imaging data to identify clinically relevant cervical spondylotic myelopathy (CSM) patient phenotypes. A major shortcoming in improving care for CSM patients is the lack of robust quantitative imaging tools to guide surgical decision-making. Advanced diffusion-weighted magnetic resonance imaging (MRI) techniques, such as diffusion basis spectrum imaging (DBSI), may help address this limitation by providing detailed evaluations of white matter injury in CSM. Fifty CSM patients underwent comprehensive clinical assessments and diffusion-weighted MRI, followed by DBSI modeling. DBSI metrics included fractional anisotropy, axial and radial diffusivity, fiber fraction, extra-axonal fraction, restricted fraction, and nonrestricted fraction. Neurofunctional status was assessed by the modified Japanese Orthopedic Association, myelopathic disability index, and disabilities of the arm, shoulder, and hand. Quality-of-life was measured by the 36-Item Short Form Survey physical component summary and mental component summary. The neck disability index was used to measure self-reported neck pain. K-means clustering was applied to baseline DBSI measures to identify 3 clinically relevant CSM disease phenotypes. Baseline demographic, clinical, radiographic, and patient-reported outcome measures were compared among clusters using one-way analysis of variance (ANOVA). Twenty-three (55%) mild, 9 (21%) moderate, and 10 (24%) severe myelopathy patients were enrolled. Eight patients were excluded due to MRI data of insufficient quality. Of the remaining 42 patients, 3 groups were generated by k-means clustering. When compared with clusters 1 and 2, cluster 3 performed significantly worse on the modified Japanese Orthopedic Association and all patient-reported outcome measures (P<0.001), except the 36-Item Short Form Survey mental component summary (P>0.05). Cluster 3 also possessed the highest proportion of non-Caucasian patients (43%, P=0.04), the worst hand dynamometer measurements (P<0.05), and significantly higher intra-axonal axial diffusivity and extra-axonal fraction values (P<0.001). Using baseline imaging data, we delineated a clinically meaningful CSM disease phenotype, characterized by worse neurofunctional status, quality-of-life, and pain, and more severe imaging markers of vasogenic edema. II.

Intercalation Chemistry of the Disordered Rocksalt Li3V2O5 Anode from Cluster Expansions and Machine Learning Interatomic Potentials

Disordered rocksalt (DRX) Li3V2O5 is a promising anode candidate for rechargeable lithium-ion batteries because of its low voltage, high rate capability, and good cycling stability. Herein, we present a comprehensive study of the intercalation chemistry of the DRX-Li3V2O5 anode using density functional theory (DFT) calculations combined with machine learning cluster expansions and interatomic potentials. The predicted voltage profile of the DRX Li3V2O5 anode at room temperature based on Monte Carlo simulations with a fitted cluster expansion model is in good agreement with experiments. In contrast to previous DFT results, we find that Li ions predominately intercalate into tetrahedral sites during charging, while a majority of Li and V ions at octahedral sites remain stable. In addition, molecular dynamics simulations with a fitted moment tensor potential attribute the fast-charging capability of DRX-Li3V2O5 to the facile diffusivity of Li+ via a tetrahedral–octahedral–tetrahedral pathway. We further suggest tuning the Li:V ratio as a means of trading off increased lithiation capacity and decreased anode voltage in this system. This work provides in-depth insights into the high-performance DRX-Li3V2O5 anode and paves the way for the discovery of other disordered anode materials.

Machine Learning Feature Based Job Scheduling for Distributed Machine Learning Clusters

With the rapid proliferation of Machine Learning (ML) and Deep learning (DL) applications running on modern platforms, it is crucial to satisfy application performance requirements such as meeting deadline and ensuring accuracy. To this end, researchers have proposed several job schedulers for ML clusters. However, none of the previously proposed schedulers consider ML model parallelism, though it has been proposed as an approach to increase the efficiency of running large-scale ML and DL jobs. Thus, in this paper, we propose an ML job Feature based job Scheduling system (MLFS) for ML clusters running both data parallelism and model parallelism ML jobs. MLFS first uses a heuristic scheduling method that considers an ML job&#x2019;s spatial and temporal features to determine task priority for job queue ordering in order to improve job completion time (JCT) and accuracy performance. It uses the data from the heuristic scheduling method for training a deep reinforcement learning (RL) model. After the RL model is well trained, it then switches to the RL method to automatically make decisions on job scheduling. In addition, MLFS has a system load control method that selects tasks from overloaded servers to move to underloaded servers based on task priority, and also intelligently removes the tasks that generate little or no improvement on the desired accuracy performance when the system is overloaded to improve JCT and accuracy by job deadline. Furthermore, we propose Optimal ML iteration stopping method that determines the proper time to stop training ML model when this model reaches the minimum loss value. Our real experiments and large-scale simulation based on real trace show that MLFS reduces JCT by up to 53% and makespan by up to 52%, and improves accuracy by up to 64% when compared with existing ML job schedulers. We also open sourced our code.

Observations of intra-peatland variability using multiple spatially coincident remotely sensed data sources and machine learning

Peatlands are important sites of ecosystem services, particularly as soil carbon stores, and are recognised in many international climate strategies. However, drained peatlands, which have been modified for industrial extraction or agriculture, are responsible for carbon emission. Peatland restoration aims to return these degraded sites to a natural state. Multiple means of remotely monitoring the success of peat restoration are available, ranging from space-based satellite measurements (optical and radar) to airborne geophysical measurements (electro-magnetic and radiometric). This paper integrates multi-band, spatially coincident, remotely sensed data into a single framework, resulting in a comprehensive interpretation of intra-peatland variation of key restoration indicators. It uses a semi-automatic, data driven approach with unsupervised neural network machine learning clustering. A Multi-Cluster Average Standard Deviation metric is introduced which can determine the appropriate number of clusters for any dataset. The method was applied to a site in Ireland, representative of degraded peatlands, where optical satellite and airborne radiometric geophysical measurements were combined. The method was successful at determining the appropriate number of clusters for single and combined datasets, and the resulting cluster signatures provided visually compelling representations of the intra-peatland variation. This resulted in a comprehensive interpretation of intra-peatland variation of several key peatland restoration indicators, namely surface vegetation levels and soil moisture to ∼ 60 cm of the peat surface. The study provides a framework for high spatial and temporal resolution monitoring of peatland restoration using future drone-based platforms.