Internal Data Structures Research Articles

SCUBA implements a storage format-agnostic API for single-cell data access in R

While robust tools exist for the analysis of single-cell datasets in both Python and R, interoperability is limited, and analysis tools generally only accept one object class. Considerable programming expertise is required to integrate tools across package ecosystems into a comprehensive analysis, due to their differing languages and internal data structures. This complicates validation of results and leads to inconsistent visualizations between analysis suites. Conversion between object formats is the most common solution, but this is difficult and error-prone due to the rapid pace of development of the analysis suites and their underlying data structures. To address this, we created SCUBA (Single-Cell Unified Backend API), an R package that implements a unified data access API for all common R and Python single-cell object formats. SCUBA extends the data access approach from the widely used Seurat package to SingleCellExperiment and anndata objects. SCUBA also implements new data-specific access functions for all supported object types. Performance scales well across all SCUBA-supported formats. In addition to performance, SCUBA offers several advantages over object conversion for the visualization and further analysis of pre-processed single-cell data. First, SCUBA extracts only data required for the operation at hand, leaving the original object unmodified. This process is simpler, less error prone, and less memory intensive than object conversion, which operates on the entire dataset. Second, code written with SCUBA can use any supported object class as input, with simple and consistent syntax across object formats. This allows a single analysis script or package (like our interactive single-cell browser, scExploreR) to work seamlessly with multiple object types, reducing the complexity of the code and improving both readability and reproducibility. Adoption of SCUBA will ultimately improve collaboration and reproducible research in single-cell analysis by lowering the barriers between package ecosystems.

Dynamic frequent subgraph mining algorithms over evolving graphs: a survey

Frequent subgraph mining (FSM) is an essential and challenging graph mining task used in several applications of the modern data science. Some of the FSM algorithms have the objective of finding all frequent subgraphs whereas some of the algorithms focus on discovering frequent subgraphs approximately. On the other hand, modern applications employ evolving graphs where the increments are small graphs or stream of nodes and edges. In such cases, FSM task becomes more challenging due to growing data size and complexity of the base algorithms. Recently we see frequent subgraph mining algorithms designed for dynamic graph data. However, there is no comparative review of the dynamic subgraph mining algorithms focusing on the discovery of frequent subgraphs over evolving graph data. This article focuses on the characteristics of dynamic frequent subgraph mining algorithms over evolving graphs. We first introduce and compare dynamic frequent subgraph mining algorithms; trying to highlight their attributes as increment type, graph type, graph representation, internal data structure, algorithmic approach, programming approach, base algorithm and output type. Secondly, we introduce and compare the approximate frequent subgraph mining algorithms for dynamic graphs with additional attributes as their sampling strategy, data in the sample, statistical guarantees on the sample and their main objective. Finally, we highlight research opportunities in this specific domain from our perspective. Overall, we aim to introduce the research area of frequent subgraph mining over evolving graphs with the hope that this can serve as a reference and inspiration for the researchers of the field.

DoppelGanger++ in Action: A Database Replay System with Fast Dependency Graph Generation

A Database Replay System (DRS) captures workloads from a production system and subsequently replays them in a testing environment to verify correctness and performance. Prior to the replay process, DRS initially generates a dependency graph from the workload to ensure output determinism and to maximize replay concurrency in the testing system. However, the state-of-the-art inefficiently generates unnecessarily larger dependency graphs, creating a major bottleneck in the end-to-end pipeline. DoppelGanger++ is a new DRS supporting fast dependency graph generation. This demonstration illustrates how it captures and replays workloads, with a focus on efficiently generating compact dependency graphs. Specifically, we showcase the end-to-end database replay workflow using the complete database replay workload, accompanied by a web tool developed for our demo which can animate the dependency graph generation process and visualize important internal data structures.

Self-organizing hierarchical incremental learning framework and universal approximation analysis based on stochastic configuration mechanism

Conventional machine learning algorithms face significant limitations when dealing with high-dimensional data. Besides, deep learning models often require substantial computational resources and have a high processing time despite their excellent performance. Hence, this paper proposes an expanded stochastic configuration network and a self-organizing hierarchical incremental learning (SHIL) framework to overcome these challenges. Specifically, this study introduces a novel supervised hierarchical clustering tree based on the minimum redundancy maximum correlation algorithm, which mines internal data structures to construct diverse hierarchies. Subsequently, by exploiting the parent-child node relationships in the tree structure, SHIL defines the maximum number of nodes as the switching condition between levels uses the supervisory mechanism as the parameter selection criterion, and adopts the tolerance error as the termination criterion for the training. Furthermore, the universal approximation property of the SHIL framework is provided. The proposed SHIL framework is validated on several benchmark datasets, image datasets, and industrial robot cases, with the corresponding experimental results demonstrating that SHIL significantly improves computational efficiency and ensures high accuracy.

Incomplete Multiview Clustering Based on Consensus Information.

In contrast to traditional single-view clustering methods, multiview clustering (MVC) approaches aim to extract, analyze, and integrate structural information from diverse perspectives, providing a more comprehensive understanding of internal data structures. However, with an increasing number of views, maintaining the integrity of view information becomes challenging, giving rise to incomplete MVC (IMVC) methods. While existing IMVC methods have shown notable performance on many incomplete multiview (IMV) databases, they still grapple with two key shortcomings: 1) they treat the information of each view as a whole, disregarding the differences among samples within each view; and 2) they rely on eigenvalue and eigenvector operations on the view matrix, limiting their scalability for large-scale samples and views. To overcome these limitations, we propose a novel multiview clustering with consistent information (IMVC-CI) of sample points. Our method explores the multiview information set of sample points to extract consensus structural information and subsequently restores unknown information in each view. Importantly, our approach operates independently on individual sample points, eliminating the need for eigenvalue and eigenvector operations on the view information matrix and facilitating parallel computation. This significantly enhances algorithmic efficiency and mitigates challenges associated with dimensionality. Experimental results on various public datasets demonstrate that our algorithm outperforms state-of-the-art IMVC methods in terms of clustering performance and computational efficiency. The code for our article has been uploaded to https://github.com/PhdJiayiTang/IMVC-CI.git.

IBUST: An intelligent behavioural trust model for securing industrial cyber-physical systems

To meet the demand of the world’s largest population, smart manufacturing has accelerated the adoption of smart factories—where autonomous and cooperative instruments across all levels of production and logistics networks are integrated through a Cyber-Physical Production System (CPPS). However, these networks are comprised of various heterogeneous devices with varying computational power and memory capabilities. As a result, many secure communication protocols – that demand considerably high computational power and memory – can not be verbatim employed on these networks, and thereby, leaving them more vulnerable to security threats and attacks over conventional networks. These threats can largely be tackled by employing a Trust Management Model (TMM) by exploiting the behavioural patterns of nodes to identify their trust class. In this context, ML-based models are best suited due to their ability to capture hidden patterns in data, learning and improving the pattern detection accuracy over time to counteract and tackle threats of a dynamic nature, which is absent in most of the conventional models. However, among the existing ML-based solutions in detecting attack patterns, many of them are computationally expensive, require a long training time, and a considerably large amount of training data—which are seldom available. An aid to this is the association rule learning (ARL) paradigm, whose models are computationally inexpensive and do not require a long training time. Therefore, this paper proposes an ARL-based intelligent Behavioural Trust Model (iBUST) for securing the CPPS. For this intelligent TMM, a variant of Frequency Pattern Growth (FP-Growth), called enhanced FP-Growth (EFP-Growth) algorithm is developed by altering the internal data structures for faster execution and by developing a modified exponential decay function (MEDF) to automatically calculate minimum supports for adapting trust evolution characteristics. In addition, a new optimisation model for finding optimum parameter values in the MEDF and an algorithm for transmuting a 1D quantitative feature into a respective categorical feature are developed to facilitate the model. Afterwards, the trust class of an object is identified employing the Naïve Bayes classifier. This proposed model is evaluated on a trust evolution-supported experimental environment along with other compared models taking a benchmark dataset into consideration, where it outperforms its counterparts.

Single-source UDA for privacy-preserving intelligent fault diagnosis based on domain augmentation

In practical applications of fault diagnosis, several factors, including fluctuations in load, changes in equipment condition, and environmental noise effects, could cause a classifier that’s been trained on the source domain to be ill-suited for matching data from the target domain. Unsupervised domain adaptation techniques have been developed to tackle this issue, but they typically demand access to fully labeled source domains, ignoring concerns of privacy regarding source domain data. Therefore, we consider a new research scene for source-free unsupervised domain adaptation (SFUDA), which exclusively relies on a source model trained on source domain sample without requiring access to fully labeled source domain data. This paper introduces a SFUDA approach that utilizes knowledge distillation (KD), which involves two stages: (1) generalizing the source model by applying domain augmentation techniques and LS methods that enhance the model’s potential to enhance its generalization capability; (2) adapting the target model using a KD framework to achieve knowledge migration; and in addition, mutual information structure regularization is added to consider the internal data structure, thus enhancing the model’s adaptability. To evaluate the efficacy of our approach, we perform experiments on two datasets—the Case Western Reserve University dataset and the Paderborn University dataset, comprising 24 transfer tasks. Our experiments demonstrate the effectiveness of the domain augmentation technique, mutual information regularization, and the proposed method.

Python meets JIT compilers: A simple implementation and a comparative evaluation

AbstractDeveloping a just‐in‐time (JIT) compiler can be a daunting task, especially for a language as flexible as Python. While PyPy, powered with JIT compilation, can often outperform the official pure interpreter, CPython, by a noteworthy margin, its popularity remains far from comparable to that of CPython due to some issues. Given that an easier‐to‐deploy and better‐compatible JIT compiler would benefit more Python users, we have developed comPyler, a simple JIT compiler functioning as a CPython extension and intended to convert frequently interpreted CPython bytecode into equivalent machine code. Designed with good compatibility in mind, it does not alter CPython's internal data structures or external interfaces. Based on LLVM's mature infrastructure, it can be readily ported to almost all platforms. Compared with CPython, it achieved the highest speedup of 2.205, with an average of 1.093. Despite its relatively limited effect, comPyler incurs low development costs. As a baseline compiler, it also sheds light on the improvement attainable by optimizing solely the overhead of bytecode interpretation. Furthermore, as there is still a dearth of empirical research covering the multitude of JIT compilers available for Python, we have conducted a performance study that examines Jython, IronPython, PyPy, GraalPy, Pyston, Pyjion, and our comPyler. Our research takes into account not only the benchmark speed for various time windows but also the boot latency and memory footprint. Through this comprehensive study, our objective is to assist developers in gaining a better understanding of the effects of distinct JIT compilation techniques and to aid users in making informed decisions when choosing among different Python implementations.

Legolas 2.0: Improvements and extensions to an MHD spectroscopic framework

We report on recent extensions and improvements to the Legolas code, which is an open-source, finite element-based numerical framework to solve the linearised (magneto)hydrodynamic equations for a three-dimensional force- and thermally balanced state with a nontrivial one-dimensional variation. The standard Fourier modes imposed give rise to a complex, generalised non-Hermitian eigenvalue problem which is solved to quantify all linear wave modes of the given system in either Cartesian or cylindrical geometries. The framework now supports subsystems of the eight linearised MHD equations, allowing for pure hydrodynamic setups, only one-dimensional density/temperature/velocity variations, or the option to treat specific closure relations. We discuss optimisations to the internal datastructure and eigenvalue solvers, showing a considerable performance increase in both execution time and memory usage. Additionally the code now has the capability to fully visualise eigenfunctions associated with given wave modes in multiple dimensions, which we apply to standard Kelvin-Helmholtz and Rayleigh-Taylor instabilities in hydrodynamics, thereby providing convincing links between linear stability analysis and the onset of non-linear phenomena.

Self-adaptive seismic data reconstruction and denoising using dictionary learning based on morphological component analysis

Data reconstruction and data denoising are two critical preliminary steps in seismic data processing. Compressed Sensing states that a signal can be recovered by a series of solving algorithms if it is sparse in a transform domain, and has been well applied in the field of reconstruction, when, sparse representation of seismic data is the key point. Considering the complexity and diversity of seismic data, a single mathematical transformation will lead to incomplete sparse expression and bad restoration effects. Morphological Component Analysis (MCA) decomposes a signal into several components with outstanding morphological features to approximate the complex internal data structure. However, the representation ability of combined dictionaries is constrained by the number of dictionaries, and cannot be self-adaptively matched with the data features. Dictionary learning overcomes the limitation of fixed base function by training dictionaries that are fully suitable for processed data, but requires huge amount of time and considerable hardware cost. To solve the above problems, a new dictionary library (K-Singluar Value Decomposition learning dictionary and Discrete Cosine Transform dictionary) is hereby proposed based on the efficiency of fixed base dictionary and the high precision of learning dictionary. The self-adaptive sparse representation is achieved under the Morphological Component Analysis framework and is successfully applied to the reconstruction and denoising of seismic data. Real data tests have proved that the proposed method performs better than single mathematical transformation and other combined dictionaries.

Tinycubes: A modular technology for interactive visual analysis of historical and continuously updated spatiotemporal data

Technological advances allow sending data collected, like IoT data and Social Network data, immediately to computational centers. However, state-of-the-art solutions for interactive visual exploration and analysis of spatiotemporal multidimensional using a single computer are focused on historical datasets. Also, most of these solutions utilize a monolithic architecture, with a restricted set of information extraction techniques and rigid internal data structures, which leads to the creation of forks to handle different usage needs. Tinycubes is a technology that supports the interactive exploration and visual analysis of spatiotemporal multidimensional continuously collected data in real-time combined with data from historical datasets, consistently producing up-to-date information using a single computer. The technology’s core is a novel and efficient data structure based on data cubes that allow data insertion and removal on-the-fly. The Tinycubes Technology is modular, allowing the usage of different information extraction algorithms and efficient data storage methods. Our experiments indicate that Tinycubes will enable the exploration of newly generated data combined with historical data while presenting performance similar or superior to the state-of-the-art solutions that use a single computer.

A Comprehensive Health Electronic Record System with MySQL RDMS, QGIS Database and Mongo DB

The advancement in data storage systems and novel data types has made organizations to stop relying on the use of simple client/server I.T infrastructure and leverage more on multiple categories of database systems to keep heterogeneous data. The study exploits the benefits around deploying hybrid relational database management systems and NoSQL systems while developing better electronic health records (EHR) systems within health facilities alongside facility decision support system (FDSS). More particularly, GIS, MySQL and Mongo DB databases were integrated to enhance EHR systems alongside offering improve clinical decision support. The study adopted experimental design to develop the Electronic Health Records System using GIS, MySQL and Mongo DB software to create the database. Findings revealed that the atomicity, consistency, isolation and durability feature typical of relational database management systems guaranteed data security, integrity, ease of access and efficient transaction processing. Mongo database offered the system a more precise internal data structure and solid scalability along with simplified mapping of application objects to the underlying database design. The GIS database enabled a clear visualization of patients’ geographical locations, medical facilities, and the physical location of the physicians. Integrating these database systems within the health care arena was instrumental in compelling application systems to adhere to the HIPAA EHR standards without compromising performance and scalability.

A knowledge representation model for family relationship to three generation

A system for inputting and storing family information, named “BirThree Enrollment,” was developed to promote a birth and three-generation cohort study (BirThree Cohort Study). In this cohort study, it was necessary to satisfy many operational demands while constantly overwriting and changing input information. Complex kinship information must be quickly and accurately inputed and corrected, and information on those families not yet recruited must be retrieved. For these purposes, many devices are needed, from an input interface to the internal data structure. In the field of genetic statistics, a simple standard expressive form (describe father-child relation and mother-child relation) is used for describing family structure. However, this form doesn't have sufficient information. So we developed a new form in conducting the BirThree Cohort Study. Hence, we expanded the data structure, and constructed the Input control system. Family pedigree information is stored along with initial clinical information, and this enabled the input of all self-reported information to the data base. Operators are able to input this family information before the day is out. As a result, when recruitment is completed, family information will be completed concurrently. Therefore, operators can immediately know certain person's family structure. In this model data correction was improved dramatically, and the system was operated successfully. This study is the first report of the method for storing three generations of family data.

KARGAMobile: Android app for portable, real-time, easily interpretable analysis of antibiotic resistance genes via nanopore sequencing.

Nanopore technology enables portable, real-time sequencing of microbial populations from clinical and ecological samples. An emerging healthcare application for Nanopore includes point-of-care, timely identification of antibiotic resistance genes (ARGs) to help developing targeted treatments of bacterial infections, and monitoring resistant outbreaks in the environment. While several computational tools exist for classifying ARGs from sequencing data, to date (2022) none have been developed for mobile devices. We present here KARGAMobile, a mobile app for portable, real-time, easily interpretable analysis of ARGs from Nanopore sequencing. KARGAMobile is the porting of an existing ARG identification tool named KARGA; it retains the same algorithmic structure, but it is optimized for mobile devices. Specifically, KARGAMobile employs a compressed ARG reference database and different internal data structures to save RAM usage. The KARGAMobile app features a friendly graphical user interface that guides through file browsing, loading, parameter setup, and process execution. More importantly, the output files are post-processed to create visual, printable and shareable reports, aiding users to interpret the ARG findings. The difference in classification performance between KARGAMobile and KARGA is minimal (96.2% vs. 96.9% f-measure on semi-synthetic datasets of 1 million reads with known resistance ground truth). Using real Nanopore experiments, KARGAMobile processes on average 1GB data every 23-48min (targeted sequencing - metagenomics), with peak RAM usage below 500MB, independently from input file sizes, and an average temperature of 49°C after 1h of continuous data processing. KARGAMobile is written in Java and is available at https://github.com/Ruiz-HCI-Lab/KargaMobile under the MIT license.

A dataset of Oceanographic and biogeochemical anomalies in the Caribbean Sea.

This article describes six ocean datasets consistent in anomalies of biogeochemical, physical, sea wave, biological, oceanic and chemical parameters (DACS-BGC, DACS-PHY, DACS-WAVE, DACS-BIO, DACS-OCE and DACS-CHEM) in several time scales from 3-hourly to monthly frequencies, either on the sea surface, downward/upward fluxes between the ocean and the atmosphere and the water column in the Caribbean basin (Gulf of Mexico, Caribbean Sea and Atlantic Ocean) in a geographical domain from latitudes 8 degrees to 35 degrees North and from longitudes 45 degrees to 100 degrees West, obtained, from several satellites, modeling services and observational programs. The datasets were created in NetCDF format conserving their original horizontal resolutions of 1.0, 0.5, 0.26, 0.08333 and 0.04 degrees in gridded structure; only the WAVEWATCH3 dataset has a non-uniform step in latitude and longitude. This internal data structure facilitates its handling due to a wide diversity of existent freeware tools, and it is mainly intended to support researchers to understand the evolution and cycles of physical, biogeochemical, chemical, sea wave, oceanic and biological parameters linked to global climate change.

A method of radar echo extrapolation based on dilated convolution and attention convolution

The neural network method can obtain a higher precision of radar echo extrapolation than the traditional method. However, its application in radar echo extrapolation is still in the initial stage of exploration, and there is still much room for improvement in the extrapolation accuracy. To improve the utilization of radar echo information and extrapolation accuracy, this paper proposes a radar echo extrapolation model (ADC_Net) based on dilated convolution and attention convolution. In this model, dilated convolution, instead of the pooling operation, is used to downsample the feature matrix obtained after the standard convolution operation. In doing so, the internal data structure of the feature matrix is retained, and the spatial features of radar echo data from different scales are extracted as well. Besides, the attention convolution module is integrated in the ADC_Net model to improve its sensitivity to the target features in the feature matrix and suppress the interference information. The proposed model is tested in the extrapolation of radar echo images in the next 90 min from five aspects—extrapolated image, POD index, CSI index, FAR index, and HSS index. The experimental results show that the model can effectively improve the accuracy of radar echo extrapolation.

A Digital Forensic View of Windows 10 Notifications

Windows Push Notifications (WPN) is a relevant part of Windows 10 interaction with the user. It is comprised of badges, tiles and toasts. Important and meaningful data can be conveyed by notifications, namely by so-called toasts that can popup with information regarding a new incoming email or a recent message from a social network. In this paper, we analyze the Windows 10 Notification systems from a digital forensic perspective, focusing on the main forensic artifacts conveyed by WPN. We also briefly analyze Windows 11 first release’s WPN system, observing that internal data structures are practically identical to Windows 10. We provide an open source Python 3 command line application to parse and extract data from the Windows Push Notification SQLite3 database, and a Jython module that allows the well-known Autopsy digital forensic software to interact with the application and thus to also parse and process Windows Push Notifications forensic artifacts. From our study, we observe that forensic data provided by WPN are scarce, although they still need to be considered, namely if traditional Windows forensic artifacts are not available. Furthermore, toasts are clearly WPN’s most relevant source of forensic data.

ASH: A Modern Framework for Parallel Spatial Hashing in 3D Perception.

We present ASH, a modern and high-performance framework for parallel spatial hashing on GPU. Compared to existing GPU hash map implementations, ASH achieves higher performance, supports richer functionality, and requires fewer lines of code (LoC) when used for implementing spatially varying operations from volumetric geometry reconstruction to differentiable appearance reconstruction. Unlike existing GPU hash maps, the ASH framework provides a versatile tensor interface, hiding low-level details from the users. In addition, by decoupling the internal hashing data structures and key-value data in buffers, we offer direct access to spatially varying data via indices, enabling seamless integration to modern libraries such as PyTorch. To achieve this, we 1) detach stored key-value data from the low-level hash map implementation; 2) bridge the pointer-first low level data structures to index-first high-level tensor interfaces via an index heap; 3) adapt both generic and non-generic integer-only hash map implementations as backends to operate on multi-dimensional keys. We first profile our hash map against state-of-the-art hash maps on synthetic data to show the performance gain from this architecture. We then show that ASH can consistently achieve higher performance on various large-scale 3D perception tasks with fewer LoC by showcasing several applications, including 1) point cloud voxelization, 2) retargetable volumetric scene reconstruction, 3) non-rigid point cloud registration and volumetric deformation, and 4) spatially varying geometry and appearance refinement. ASH and its example applications are open sourced in Open3D (http://www.open3d.org).

The structure of protest thinking and the idea of social protest of the youth generations "Y" and "Z" in the Far East and Moscow


 In the course of the study, a reconstruction of the respondents' ideas about five forms of collective protest behaviour was carried out: petition, rally, strike, riot and uprising using a private semantic differential (SD) to assess the protest.
 The obtained raw SD data were processed using the method of factor analysis, as well as the method of calculating semantic universals. The factorisation method allowed to reveal the internal implicit data structure, reducing a large volume of information to several factors. Semantic universals are group-agreed assessments of the stimulus according to the SD scales, that reflects general for all respondents ideas about the stimulus.
 The article presents an analysis of the categorial structures of the protest consciousness of two generations of young people, whose activity plays a special role in protest actions: representatives of the generation "Y" (or millennials) and representatives of the generation "Z" (or centennials).
 The method of factor analysis allowed to identify three integral axes (scales) for evaluating various forms of protest activity that are relevant for the youth groups of generation “Y” and generation “Z”. These scales mediate the formation of youth attitudes towards protest. Knowledge of the categorial structure of the protest consciousness of young people, understanding the structure and content of implicit subjective scales for evaluating various forms of protest by young people, allows us to model situations of the genesis and development of protests, to predict and evaluate social dynamics. Analysis of the ideas about protest forms showed an unconditionally positive attitude of young people to the petition and the rally. These forms of protest activity were consistently evaluated positively in all samples. The image of the strike is mostly positive, but there are peculiarities of perception in different regions. The image of violent radical forms of protest, rebellion and uprising differs markedly from the image of other forms of protest activity.
 First of all, these forms are most perceived as strong and active among all forms of protest. Also, young people highly evaluate their scale (“large”). These forms, moreover, are consistently rated as neither attractive nor beautiful. Also, the youth of all samples consider them extremely aggressive. But the negative connotations do not mean that violent forms of protest are unacceptable. The semantic analysis showed the ambivalence of the content of the images of rebellion and uprising. Respondents in three out of four samples do not consider them alien, senseless or hypocritical. For young people, these forms of protest have great power and scale, are extremely aggressive and not attractive, but are not alien or senseless.

Hippo

Hyper-parameter optimization is crucial for pushing the accuracy of a deep learning model to its limits. However, a hyper-parameter optimization job, referred to as a study, involves numerous trials of training a model using different training knobs, and therefore is very computation-heavy, typically taking hours and days to finish. We observe that trials issued from hyper-parameter optimization algorithms often share common hyper-parameter sequence prefixes. Based on this observation, we propose Hippo, a hyper-parameter optimization system that reuses computation across trials to reduce the overall amount of computation significantly. Instead of treating each trial independently as in existing hyper-parameter optimization systems, Hippo breaks down the hyper-parameter sequences into stages and merges common stages to form a tree of stages (a stage tree). Hippo maintains an internal data structure, search plan, to manage the current status and history of a study, and employs a critical path based scheduler to minimize the overall study completion time. Hippo applies to not only single studies but multi-study scenarios as well. Evaluations show that Hippo's stage-based execution strategy outperforms trial-based methods for several models and hyper-parameter optimization algorithms, reducing end-to-end training time by up to 2.76X (3.53x) and GPU-hours by up to 4.81X (6.77x), for single (multiple) studies.