Recent Advances in Bayesian Structural Time Series (PPT)

Steven Scott

Director of Data Science at ShareThis

Bayesian structural time series models are a powerful and flexible class of time series models. They combine additive models for trends, seasonality, regression effects, holidays, and potentially other components in an easy to use framework that allows advanced Bayesian Markov chain Monte Carlo techniques to account for issues like variable and model selection through spike and slab priors. By including additional latent variables the model can be extended to incorporate non-Gaussian error distributions such as student T, Poisson, and binomial observation errors, or student T errors in the state space. The original implementation of Bayesian structural time series was in the bsts R package, which continues to e widely used. Bsts has recently been ported to Python. The R and python packages are not perfect substitutes. The R package supplies many more state models for univariate time series modeling, while the python package contains more recent work involving multivariate series.

Speaker bio:
Steve Scott is the director of Data Science at Share This, a Palo Alto based startup focused on extracting information from a network of internet publishers. Dr Scott is a Bayesian statistician with a 25 year career spanning academics and industry. He received his PhD from the Harvard Statistics Department. He has served on the faculty of the USC Marshall School of Business, and has held director level positions at Capital One, Google, and multiple startups. Dr Scott's published work focuses on applied Bayesian computation, where he's made important contributions around hidden Markov models, multinomial logistic regression, support vector machines, sparse regression, time series modeling, and Bayesian approaches to reinforcement learning and multi-armed bandit problems.

Forecasting Cyber Events

Paul Yoo

Deputy Director, Birkbeck Institute for Data Analytics, University of London, UK

Traditional methods to cyber defence have been reactive monitoring system's traffic or behaviours to find sequences and patterns that match a previously observed signature of attack, incident, threat and action. More recently, machine-learning-based anomaly-detection methods have gained popularity, however, the detection at best could be done in real-time only. Forecasting events related to a cyber-attack prior to or during the attack is an under-explored area in cyber defence. This talk introduces a novel machine-learning based proactive cyber defence strategy that leverages upon various big (literature) data and logs in order to forecast cyber events prior to any malicious activity occurs so that allow defence and security to better prepare for, anticipate and counter future cyber threats, thereby reducing the impact of an attack and its likelihood of success.

Speaker bio: Paul Yoo is currently the Deputy Director for Birkbeck Institute for Data Analytics (BIDA) within Birkbeck College which is part of the University of London, United Kingdom. Prior to this, he held academic/research posts in Cranfield (Defence Academy of the UK), Sydney (USyd) and South Korea (KAIST). In his career, he has amassed more than ninety prestigious journal and conference publications, has been awarded more than US$ 2.3 million in project funding, and a number of prestigious international and national awards for his work in advanced data analytics, machine learning and secure systems research, notably IEEE Outstanding Leadership Award, Rozetta Award, Emirates Foundation Research Award, and the ICT Fund Award. Most recently, he won the prestigious Samsung award for research to protect IoT devices using machine-learning approach and Research England’s Global Challenge Research Fund (GCRF). Paul currently serves as an Associate Editor for ACM Computing Surveys and IEEE Transactions on Sustainable Computing. He had served as an Editor for IEEE COMML (big data and machine learning areas) from 2014 to 2019. He is also a Founder and Chair of the BIDA's Threat Intelligence lab. This lab has set out on a journey to bridge the gap between the advancement of machine learning and the progression of cyber security with the objective of creating the next generation intelligent cyber defence and security research environment for future cyber warfare, including the applications of cyber physical systems (e.g., autonomous vehicle). The lab is now a home for 11 PhD students.

Customized Transformers for Time Series Forecasting(PPT)

Qingsong Wen

Leader at Alibaba DAMO Academy

Although Transformer-based methods have significantly improved state-of-the-art results for time series forecasting, they are not only computationally expensive but more importantly, are unable to capture the global view of time series (e.g., overall trend) and handle complicated periodical patterns (e.g., multiple periods, variable periods). To address these challenges, we design two customized Transformers for time series forecasting. The first is termed as Frequency Enhanced Decomposed Transformer (FEDformer, in ICML’22) by combining Transformer with the seasonal-trend decomposition technique, in which the decomposition method captures the global profile of time series while Transformers capture more detailed structures. Furthermore, we exploit sparse representation in Fourier transform in the FEDformer to enhance the performance as well as reduce complexity to a linear complexity. The second is termed as Quaternion Transformer (Quatformer, in KDD’22) with three major components: 1) learning-to-rotate attention (LRA) based on quaternions which introduces learnable period and phase information to depict intricate periodical patterns; 2) trend normalization to normalize the series representations in hidden layers of the model considering the slowly varying characteristic of trend; 3) decoupling LRA using global memory to achieve linear complexity without losing prediction accuracy. Compared with state-of-the-art methods, extensive experiments on multiple time series benchmark datasets show that FEDformer and Quatformer can significantly improve the forecasting performance.

Speaker bio:
Dr. Qingsong Wen is currently a Staff Engineer / Team Leader at DAMO Academy-Decision Intelligence Lab, Alibaba Group (U.S.), working in the areas of intelligent time series analysis, data-driven intelligence decisions, machine learning, and signal processing. Before that, he worked at Futurewei, Qualcomm, and Marvell in the areas of big data and signal processing, and received his M.S. and Ph.D. degrees in Electrical and Computer Engineering from Georgia Institute of Technology, Atlanta, USA. He has published over 30 top-ranked conference and journal papers, and won the First Place in 2022 ICASSP Grand Challenge (AIOps in Networks) Competition. He is an Associate Editor for Neurocomputing, Guest Editor for Pattern Recognition, Guest Editor for Applied Energy, and regularly served as an SPC/PC member of the major DM/ML/AI conferences including KDD, ICDM, AAAI, IJCAI, etc.

Activity recognition from trajectory data

Md Mashud Rana

Research Scientist at Data61, CSIRO

The technological innovations in tracking systems have made it very easy to collect data from moving objects over space and time. The spatio-temporal tracking data (a.k.a as trajectory) collected by such systems can provide valuable information for industrial management and decision making purposes. For example, understanding the activities of objects from trajectory data in a dynamic and collaborative environment can help manufacturing companies to optimize the productivity, efficiency, and safety. This talk will present a semi-supersized method for activity recognition of objects from trajectory data. The developed method first partitions the long trajectory into a set of segments by applying an information theoretic criterion (minimum description length) and generates pseudo-labels for the segments based on constraint hierarchical clustering utilising only few labelled examples. It then trains a convolutional neural network model (using the segments as inputs and pseudo-labels as output) to predict the activities of an object during each segment duration. The main advantage of the developed method is that it requires only few labelled segments (as low as one per activity type) and hence reduces the burden of manual labelling of all trajectory segments by the domain experts. The method has been tested on multiple trajectories collected using motion captured systems at a tricycle assembly factory and shown to achieve highly accurate prediction accuracy.

Speaker bio:
Dr Md Mashud Rana is a Research Scientist at the Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia, where he focuses on translating machine learning research into applications and creating impact across domains. Previously, he worked at the Centre for Translational Data Science, University of Sydney, and Australian Energy Research Institute, University of New South Wales as a Postdoctoral Research Fellow.  Mashud completed his PhD from the School of Computer Science, University of Sydney, and Master of Science from University of Reading, UK. His research interests include applied machine learning / deep learning, time series analytics, spatial- temporal models,  and semi-supervised learning. He has published over 35 peer-reviewed journal articles, book chapters and conference papers (Google scholar ).  He also serves as a PC member at flagship CS conferences (AAAi, ICDM, PAKDD, IJCNN, ICONIP) and as a reviewer at IEEE TNNLS, Neurocomputing, Knowledge-Based Systems, Applied Energy.

Demand Forecasting in Retail

Arash Sangari

Senior Director at Walmart

Speaker bio:
Arash Sangari is Senior Director of Data Science in Walmart Global Tech. His team with ‪Prakhar Mehrotra‬ (VP of ML) is focused on research and development of consumer demand forecasting models for Walmart stores and e-comm channel to serve variety of business use-cases in the largest retailer in the world. He has extensive experiences in leveraging various machine learning approaches in building robust, accurate and scalable demand forecasting models for major retailers in US. He has +15 years of experience in applying machine learning methods in solving high impact problems in academia and industry, with a track record of 13 published patents and several journal papers in multi-disciplinary research areas.