Article
Statistics & Probability
Guanhua Fang, Ganggang Xu, Haochen Xu, Xuening Zhu, Yongtao Guan
Summary: In this work, the event occurrences of individuals interacting in a network are studied using a group network Hawkes process (GNHP) model. The model considers the dynamic interactions among individuals and introduces a latent group structure to account for heterogeneous user-specific characteristics. A maximum likelihood approach is proposed to cluster individuals and estimate model parameters simultaneously.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2023)
Article
Physics, Multidisciplinary
Hiroki Watari, Hideki Takayasu, Misako Takayasu
Summary: This study used data to track the order history of individual high-frequency traders (HFTs) in the USD/JPY forex market and revealed their interaction with the order book and strategies for placing limit orders. The results showed that the limit order generation processes of 104 out of 134 HFTs could be modeled by a multivariate Hawkes process. The HFTs were categorized into three groups based on their excitation mechanisms.
Article
Computer Science, Artificial Intelligence
Zhi-yan Song, Jian-wei Liu, Jie Yang, Lu-ning Zhang
Summary: With the development of the Internet and the era of big data, neural point process has become a mainstream solution for modeling asynchronous event sequences. However, due to the complexity and nonlinearity, improving the prediction accuracy of neural point process has been a challenge. In response, researchers have proposed using multi-layer perceptron architectures, which outperform attention-based neural point process. Inspired by this, the Linear Normalization Attention Hawkes Process (LNAHP) is proposed, which reduces the complexity of the model by using multi-head dot-product attention from the transformer instead of linear normalization attention.
NEURAL COMPUTING & APPLICATIONS
(2023)
Article
Business, Finance
Lina Fan, Hao Yang, Jia Zhai, Xiaotao Zhang
Summary: We employ HAR-RV-H model to predict the volatilities of 300 major individual stocks in Chinese stock market during the 2015 market crash. The Hawkes intensity process is calculated using the tick-by-tick data of individual stocks. Our study demonstrates that the Hawkes indicator has predictive power for most individual stocks during the market crash period. By comparing the in- and out-of-sample forecast results for HAR type models, we conclude that the Hawkes indicator can improve both in- and out-of-sample forecasting abilities.
FINANCE RESEARCH LETTERS
(2023)
Article
Computer Science, Artificial Intelligence
Lu-ning Zhang, Jian-wei Liu, Zhi-yan Song, Xin Zuo
Summary: In recent years, the mining of knowledge from asynchronous sequences using Hawkes process has attracted continuous attention. While the Hawkes processes based on neural networks, particularly recurrence neural networks (RNN), have become popular research fields, they still face inherent shortcomings such as gradient vanishing and explosion as well as long-term dependency issues. Transformer-based on self-attention has seen success in various sequential modeling tasks, however, the existing Transformer Hawkes processes do not effectively utilize temporal information in asynchronous events. The introduction of temporal attention in the proposed TAA-THP model shows significant improvements compared to baseline models in various measurements through experiments on synthetic and real data sets.
NEURAL COMPUTING & APPLICATIONS
(2022)
Article
Physics, Multidisciplinary
Cassien Habyarimana, Jane A. Aduda, Enrico Scalas, Jing Chen, Alan G. Hawkes, Federico Polito
Summary: We characterize a Hawkes point process using the kernel based on the probability density function of Mittag-Leffler random variables. The kernel follows a power law decay with an exponent fl + 1 in the range (1, 2]. We prove several analytical results, particularly for the expected intensity and the expected number of events. These results are then used to validate algorithms for numerically inverting the Laplace transform of expected intensity and Monte Carlo simulations of the process. Furthermore, Monte Carlo simulations are applied to derive the full distribution of the number of events. The algorithms used in this study are available at https://github.com/habyarimanacassien/Fractional-Hawkes.
PHYSICA A-STATISTICAL MECHANICS AND ITS APPLICATIONS
(2023)
Article
Automation & Control Systems
Lu-ning Zhang, Jian-wei Liu, Zhi-yan Song, Xin Zuo
Summary: The paper discusses the development of point process models and Transformer Hawkes process models based on neural networks, as well as the proposed new universal Transformer Hawkes process model (UTHP). Through experiments, it is demonstrated that the performance of the UTHP model has shown improvement compared to previous state-of-the-art models.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2021)
Article
Computer Science, Artificial Intelligence
Myrl G. Marmarelis, Greg Ver Steeg, Aram Galstyan
Summary: This paper introduces a multivariate Hawkes process model that uncovers hidden geometric relationships between events by embedding event types into a hidden metric space. The effectiveness of the model is validated through multiple experiments, and it is shown that learning the embedding can reveal significant interactions in various application domains.
JOURNAL OF ARTIFICIAL INTELLIGENCE RESEARCH
(2022)
Article
Economics
Sarita D. Lee, Andy A. Shen, Junhyung Park, Ryan J. Harrigan, Nicole A. Hoff, Anne W. Rimoin, Frederic Paik Schoenberg
Summary: In a study comparing Hawkes and recursive point process models in modeling the Ebola virus disease outbreak in the Democratic Republic of the Congo from 2018-2020, both models showed similar performance with smaller errors in the beginning and waning phases of the epidemic. The Hawkes model, on average, had slightly smaller error sizes compared to the recursive model. The results suggest that both models can be used in near real time during an epidemic to predict future cases and inform management strategies.
JOURNAL OF FORECASTING
(2022)
Article
Statistics & Probability
Biao Cai, Jingfei Zhang, Yongtao Guan
Summary: Learning the latent network structure from large scale multivariate point process data is important in various scientific and business applications. We propose a new class of nonstationary Hawkes processes to characterize the complex processes underlying the data, and use efficient sparse least squares estimation approach to estimate the network structure. We also establish concentration inequalities using a thinning representation.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2022)
Article
Computer Science, Information Systems
Zhihong Cui, Xiangguo Sun, Li Pan, Shijun Liu, Guandong Xu
Summary: Incremental recommendation systems adapt to users' ongoing events and recommend items without retaining the model. However, existing models often ignore underlying factors that may trigger event generation. In this study, we propose the Factors Mixed Hawkes Process (FMHP) model which incorporates intrinsic, external, and historical intensity to evaluate event occurrence. Our experiments on public datasets demonstrate that FMHP outperforms state-of-the-art methods in terms of HR, NDCG, and Recall.
INFORMATION SCIENCES
(2023)
Article
Statistics & Probability
Tomasz R. Bielecki, Jacek Jakubowski, Mariusz Nieweglowski
Summary: This work introduces and studies a new class of Hawkes processes called generalized Hawkes processes, and their special subclass - the generalized multivariate Hawkes processes (GMHPs). GMHPs allow for explicit modeling of simultaneous occurrence of excitation events from different sources in a multivariate process.
Review
Automation & Control Systems
Michele Garetto, Emilio Leonardi, Giovanni Luca Torrisi
Summary: The study proposes a stochastic model based on the Hawkes process to describe the outbreak and mitigation of epidemics, which has been successfully applied to the COVID-19 outbreaks in Italy. The model can accurately capture specific features of the novel coronavirus, such as the impact of undetected and asymptomatic individuals on disease spread, and is important for studying the evolution of epidemics and different response strategies.
ANNUAL REVIEWS IN CONTROL
(2021)
Article
Multidisciplinary Sciences
Archit Verma, Siddhartha G. Jena, Danielle R. Isakov, Kazuhiro Aoki, Jared E. Toettcher, Barbara E. Engelhardt
Summary: This study proposes a spatiotemporal model of dynamic cell signaling based on Hawkes processes, which can capture both the autonomous behavior of single cells and the interactions of cells with their neighbors simultaneously. The model is applicable to tissues composed of heterogeneous cell types and can identify drug-induced signaling deficits and characterize signaling changes across different cell populations.
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
(2021)
Article
Computer Science, Theory & Methods
Leigh Shlomovich, Edward A. K. Cohen, Niall Adams
Summary: This study models aggregated event data using multivariate Hawkes processes to describe mutually-exciting behavior. The parameter estimation of the multivariate aggregated Hawkes processes is performed using a Monte Carlo expectation-maximization algorithm. Simulation results demonstrate that the multivariate MC-EM method outperforms alternative approaches in terms of mean square error in all considered cases.
STATISTICS AND COMPUTING
(2022)
Article
Computer Science, Artificial Intelligence
Raluca Jalaboi, Frederik Faye, Mauricio Orbes-Arteaga, Dan Jorgensen, Ole Winther, Alfiia Galimzianova
Summary: Dermatological diagnosis automation is crucial for addressing the high prevalence of skin diseases and shortage of dermatologists. DermX and DermX+ are two explainable automated dermatological diagnosis methods that achieve near-expert diagnosis performance while providing expert-level explanations.
MEDICAL IMAGE ANALYSIS
(2023)
Article
Health Care Sciences & Services
Raluca Jalaboi, Ole Winther, Alfiia Galimzianova
Summary: ImageQX is a convolutional neural network that can automatically assess and explain image quality, identifying common issues such as bad framing, bad lighting, blur, low resolution, and distance problems. Trained and validated on photographs taken using a mobile skin disease tracking application, ImageQX performs at an expert-level and is easily deployable on mobile devices.
TELEMEDICINE AND E-HEALTH
(2023)
Article
Microscopy
Matthew Helmi Leth Larsen, Frederik Dahl, Lars P. Hansen, Bastian Barton, Christian Kisielowski, Stig Helveg, Ole Winther, Thomas W. Hansen, Jakob Schiotz
Summary: Convolutional neural networks can reconstruct the exit wave function from a short focal series of HRTEM images, achieving a similar fidelity compared to conventional methods. By training a fully convolutional neural network based on the U-Net architecture with simulated exit waves and HRTEM images, we successfully applied it to analyze experimentally obtained images of MoS2 nanoparticles on graphene support and obtain atomically resolved exit wave structures. Furthermore, we demonstrated the feasibility of training the network to reconstruct exit waves for a wide range of two-dimensional materials.
Article
Chemistry, Medicinal
Felix Teufel, Dennis Madsen, Kristine Deibler, Jan C. Refsgaard, Marina A. Kasimova, Christian T. Madsen, Carsten Stahlhut, Mads Gronborg, Ole Winther
Summary: In this study, the use of AlphaFold-Multimer complex structure prediction and transmembrane topology prediction for peptide deorphanization is investigated. It is found that AlphaFold's confidence metrics have strong performance in prioritizing true peptide-receptor interactions.
JOURNAL OF CHEMICAL INFORMATION AND MODELING
(2023)
Article
Biochemistry & Molecular Biology
Arnor Sigurdsson, Ioannis Louloudis, Karina Banasik, David Westergaard, Ole Winther, Ole Lund, Sisse Rye Ostrowski, Christian Erikstrup, Ole Birger Vesterager Pedersen, Mette Nyegaard, Soren Brunak, Bjarni J. Vilhjalmsson, Simon Rasmussen
Summary: We developed a deep learning framework for polygenic risk score (PRS) prediction that can handle large-scale genomics data, support multi-task learning, and automatically integrate clinical and biochemical data. The framework demonstrated competitive performance and improved predictions for complex genetic relationships and non-additive genetic effects and epistasis. The model also outperformed traditional linear PRS methods for Type 1 Diabetes.
NUCLEIC ACIDS RESEARCH
(2023)
Article
Chemistry, Multidisciplinary
Kasper A. Einarson, Kristian M. Bendtsen, Kang Li, Maria Thomsen, Niels R. Kristensen, Ole Winther, Simone Fulle, Line Clemmensen, Hanne H. F. Refsgaard
Summary: This study presents a novel combination of molecular descriptors for predicting the pharmacokinetic parameters of insulin analogs. Machine-learning models were used to predict pharmacokinetic parameters, and the results showed that combining protein and small molecule descriptors was crucial for accurate predictions.
Review
Biochemical Research Methods
Jun Wang, Marc Horlacher, Lixin Cheng, Ole Winther
Summary: RNA localization is important for spatial translation regulation, and this review discusses its molecular mechanisms, experimental techniques, and machine learning-based prediction tools. The three main molecular mechanisms controlling RNA localization to distinct cellular compartments, including directed transport, mRNA degradation protection, and diffusion/local entrapment, are reviewed. Advances in experimental methods provide ample data resources for the design of powerful machine learning models in RNA localization prediction. The review also covers publicly available predictive tools, serving as a guide for users and encouraging the development of more effective prediction models. Lastly, an overview of multimodal learning is presented as a potential new avenue for RNA localization prediction.
BRIEFINGS IN BIOINFORMATICS
(2023)
Article
Multidisciplinary Sciences
Simon Ott, Konstantin Hebenstreit, Valentin Lievin, Christoffer Egeberg Hother, Milad Moradi, Maximilian Mayrhauser, Robert Praas, Ole Winther, Matthias Samwald
Summary: Large language models (LLMs) like GPT-4 have shown impressive performance in various tasks, but they still have limitations in complex reasoning, opaque reasoning processes, fact hallucination, and potential biases. To address these issues, chain-of-thought prompting, a technique that allows models to verbalize reasoning steps in natural language, has been proposed. ThoughtSource is introduced as a meta-dataset and software library for chain-of-thought reasoning, aiming to improve future AI systems by enhancing qualitative understanding, enabling empirical evaluations, and providing training data. The initial release of ThoughtSource includes datasets from scientific/medical, general-domain, and math word question answering.
Article
Genetics & Heredity
Felix Teufel, Magnus Halldor Gislason, Jose Juan Almagro Armenteros, Alexander Rosenberg Johansen, Ole Winther, Henrik Nielsen
Summary: A homology partitioning algorithm called GraphPart is proposed, which divides the data in such a way that closely related sequences always end up in the same partition, while retaining as many sequences as possible. Evaluation on Protein, DNA and RNA datasets shows that GraphPart is capable of preserving a larger number of sequences, while achieving homology separation on a par with reduction approaches.
NAR GENOMICS AND BIOINFORMATICS
(2023)
Article
Biotechnology & Applied Microbiology
Marc Horlacher, Nils Wagner, Lambert Moyon, Klara Kuret, Nicolas Goedert, Marco Salvatore, Jernej Ule, Julien Gagneur, Ole Winther, Annalisa Marsico
Summary: RBPNet is a new deep learning method that predicts CLIP-seq crosslink count distribution from RNA sequence. Training on millions of regions, RBPNet shows high generalization on eCLIP, iCLIP, and miCLIP assays, outperforming state-of-the-art classifiers. RBPNet performs bias correction by modeling the raw signal as a mixture of protein-specific and background signal. By using Integrated Gradients for model interrogation, RBPNet identifies predictive sub-sequences corresponding to known and novel binding motifs and enables variant-impact scoring through in silico mutagenesis. Overall, RBPNet improves the imputation of protein-RNA interactions and enhances mechanistic interpretation of predictions.
Article
Genetics & Heredity
Marco Salvatore, Marc Horlacher, Annalisa Marsico, Ole Winther, Robin Andersson
Summary: Dysfunction of regulatory elements through genetic variants is a central mechanism in disease pathogenesis. Deep learning methods have shown promise in modeling biomolecular data from DNA sequence but require large input data for training. ChromTransfer, a transfer learning method, utilizes a pre-trained model of open chromatin regions to fine-tune on regulatory sequences and demonstrates superior performance in learning cell-type specific chromatin accessibility. It is able to fine-tune on small input data with minimal decrease in accuracy and utilizes sequence features matching binding site sequences of key transcription factors for prediction, making it a promising tool for learning the regulatory code.
NAR GENOMICS AND BIOINFORMATICS
(2023)
Article
Chemistry, Multidisciplinary
Laura Hannemose Rieger, Eibar Flores, Kristian Frellesen Nielsen, Poul Norby, Elixabete Ayerbe, Ole Winther, Tejs Vegge, Arghya Bhowmik
Summary: Enhancing cell lifetime is crucial in battery design and development, and early prediction of cell aging can accelerate the discovery and production of better battery chemistries. This study introduces an early prediction model with reliable uncertainty estimates, which utilizes a small number of initial cycles to predict the entire battery degradation trajectory.
Article
Biochemical Research Methods
Felix Teufel, Jan Christian Refsgaard, Christian Toft Madsen, Carsten Stahlhut, Mads Gronborg, Ole Winther, Dennis Madsen
Summary: DeepPeptide is a deep learning model that predicts cleaved peptides directly from the amino acid sequence, showing improved precision and recall compared to previous methodology. It is capable of identifying peptides in underannotated proteomes.
Article
Chemistry, Physical
Jonas Busk, Mikkel N. Schmidt, Ole Winther, Tejs Vegge, Peter Bjorn Jorgensen
Summary: This research presents a complete framework for training and recalibrating graph neural network ensemble models to accurately predict energy and forces with calibrated uncertainty estimates. The method is demonstrated and evaluated on challenging datasets, achieving good prediction accuracy and uncertainty calibration.
PHYSICAL CHEMISTRY CHEMICAL PHYSICS
(2023)
Article
Computer Science, Artificial Intelligence
Jonas Busk, Peter Bjorn Jorgensen, Arghya Bhowmik, Mikkel N. Schmidt, Ole Winther, Tejs Vegge
Summary: In this study, a message passing neural network model is extended to incorporate both aleatoric and epistemic uncertainty in a unified framework, and the predictive distribution is recalibrated for improved accuracy. The proposed method is shown to accurately predict molecular properties with well calibrated uncertainty estimates in experimental settings.
MACHINE LEARNING-SCIENCE AND TECHNOLOGY
(2022)