A5083237271- Advanced Bandit Algorithms Research
- Recommender Systems and Techniques
- Machine Learning and Algorithms
- Mobile Crowdsensing and Crowdsourcing
- Auction Theory and Applications
- Data Stream Mining Techniques
- Smart Grid Energy Management
- Advanced Causal Inference Techniques
- Explainable Artificial Intelligence (XAI)
- Topic Modeling
- Statistical Methods in Clinical Trials
- Privacy-Preserving Technologies in Data
- Data Mining Algorithms and Applications
- Bayesian Modeling and Causal Inference
- Machine Learning and Data Classification
- Advanced Graph Neural Networks
- Consumer Market Behavior and Pricing
- Reinforcement Learning in Robotics
- SARS-CoV-2 detection and testing
- Advanced Text Analysis Techniques
- Healthcare Operations and Scheduling Optimization
- Data Visualization and Analytics
- Wireless Networks and Protocols
- Economic and Environmental Valuation
- Algorithms and Data Compression
Karnataka Health Promotion Trust
2024
University of Antwerp
2018-2023
Amazon (United Kingdom)
2022
Social networks have been widely studied over the last century from multiple disciplines to understand societal issues such as inequality in employment rates, managerial performance, and epidemic spread. Today, these many more can be at global scale thanks digital footprints that we generate when browsing Web or using social media platforms. Unfortunately, scientists often struggle access data primarily because it is proprietary, even shared with privacy guarantees, either no representative...
Approaches to recommendation are typically evaluated in one of two ways: (1) via a (simulated) online experiment, often seen as the gold standard, or (2) some offline evaluation procedure, where goal is approximate outcome an experiment. Several metrics have been adopted literature, inspired by ranking prevalent field Information Retrieval. (Normalised) Discounted Cumulative Gain (nDCG) such metric that has widespread adoption empirical studies, and higher (n)DCG values used present new...
Recommender systems are typically evaluated in an offline setting. A subset of the available user-item interactions is sampled to serve as test set, and some model trained on remaining data points then its performance predict which were left out. Alternatively, online evaluation setting, multiple versions system deployed various metrics for those recorded. Systems that score better these metrics, preferred. Online effective, but inefficient a number reasons. Offline much more efficient,...
Methods for bandit learning from user interactions often require a model of the reward certain context-action pair will yield – example, probability click on recommendation. This common machine task is highly non-trivial, as data-generating process contexts and actions skewed by recommender system itself. Indeed, when deployed recommendation policy at data collection time does not pick its uniformly-at-random, this leads to selection bias that can impede effective modelling. in turn makes...
The bandit paradigm provides a unified modeling framework for problems that require decision-making under uncertainty. Because many business metrics can be viewed as rewards (a.k.a. utilities) result from actions, algorithms have seen large and growing interest industrial applications, such search, recommendation advertising. Indeed, with the lens comes promise of direct optimisation we care about.
A/B-tests are a cornerstone of experimental design on the web, with wide-ranging applications and use-cases. The statistical $t$-test comparing differences in means is most commonly used method for assessing treatment effects, often justified through Central Limit Theorem (CLT). CLT ascertains that, as sample size grows, sampling distribution Average Treatment Effect converges to normality, making valid sufficiently large sizes. When outcome measures skewed or non-normal, quantifying what...
Modern web-based platforms show ranked lists of recommendations to users, attempting maximise user satisfaction or business metrics. Typically, the goal such systems boils down maximising exposure probability for items that are deemed "reward-maximising" according a metric interest. This general framing comprises streaming applications, as well e-commerce job recommendations, and even web search. Position bias models can be used estimate probabilities each use-case, specifically tailored how...
Ad-load balancing is a critical challenge in online advertising systems, particularly the context of social media platforms, where goal to maximize user engagement and revenue while maintaining satisfactory experience. This requires optimization conflicting objectives, such as satisfaction ads revenue. Traditional approaches ad-load rely on static allocation policies, which fail adapt changing preferences contextual factors. In this paper, we present an approach that leverages off-policy...
Online controlled experiments are a crucial tool to allow for confident decision-making in technology companies. A North Star metric is defined (such as long-term revenue or user retention), and system variants that statistically significantly improve on this an A/B-test can be considered superior. metrics typically delayed insensitive. As result, the cost of experimentation high: need run long time, even then, type-II errors (i.e. false negatives) prevalent.
Modern recommender systems are often modelled under the sequential decision-making paradigm, where system decides which recommendations to show in order maximise some notion of either imminent or long-term reward. Such methods require an explicit model reward a certain context-action pair will yield – for example, probability click on recommendation. This common machine learning task is highly non-trivial, as data-generating process contexts and actions can be skewed by itself. Indeed, when...
Conventional approaches to recommendation often do not explicitly take into account information on previously shown recommendations and their recorded responses. One reason is that, since we know the outcome of actions system did take, learning directly from such logs a straightforward task. Several methods for off-policy or counterfactual have been proposed in recent years, but efficacy task remains understudied. Due limitations offline datasets lack access most academic researchers online...
The contextual bandit paradigm provides a general framework for decision-making under uncertainty. It is theoretically well-defined and well-studied, many personalisation use-cases can be cast as learning problem. Because this allows the direct optimisation of utility metrics that rely on online interventions (such click-through-rate (CTR)), has become an attractive choice to practitioners. Historically, literature topic focused one-sided, user-focused notion utility, overall disregarding...
Online advertising opportunities are sold through auctions, billions of times every day across the web. Advertisers who participate in those auctions need to decide on a bidding strategy: how much they willing bid for given impression opportunity. Deciding such strategy is not straightforward task, because interactive and reactive nature repeated auction mechanism. Indeed, an advertiser does observe counterfactual outcomes amounts that were submitted, successful advertisers will adapt their...
Online experiments such as Randomised Controlled Trials (RCTs) or A/B-tests are the bread and butter of modern platforms on web. They conducted continuously to allow estimate causal effect replacing system variant "A" with "B", some metric interest. These variants can differ in many aspects. In this paper, we focus common use-case where they correspond machine learning models. The online experiment then serves final arbiter decide which model is superior, should thus be shipped. statistical...
Recommender systems are more and often modelled as repeated decision making processes – deciding which (ranking of) items to recommend a given user. Each or rank an item has significant impact on immediate future user responses, long-term satisfaction engagement with the system, possibly valuable exposure for provider. This interactive interventionist view of recommender uncovers plethora unanswered research questions, it complicates typically adopted offline evaluation learning procedures...
Recommender systems are inherently decision-making systems, taking actions that have consequences for the world around them. Some might be desirable (for example, growing user base an online platform), others unintended amplifying inequality among item providers). In order to reason about these consequences, we need resort methods from literature on causal inference. Whilst this research area has seen a interest in recent years, there is abundance of open questions how should model...
Recent work has shown that, despite their simplicity, item-based models optimised through ridge regression can attain highly competitive results on collaborative filtering tasks. As these are analytically computable and thus forgo the need for often expensive iterative optimisation procedures, they an attractive choice practitioners. We study applicability of such closed-form to implicit-feedback when additional side-information or metadata about items is available. Two complementary...
The objective of this tutorial is to give a structured overview the conceptual frameworks behind current state-of-the-art recommender systems, explain their underlying assumptions, resulting methods and shortcomings, introduce an exciting new class approaches that frames task recommendation as counterfactual policy learning problem. can be divided into two modules. In module 1, participants learn about for building real-world systems comprise mainly frameworks, namely: optimal...
Practitioners who wish to build real-world applications that rely on ranking models, need decide which modelling paradigm follow. This is not an easy choice make, as the research literature this topic has been shifting in recent years. In particular, whilst Gradient Boosted Decision Trees (GBDTs) have reigned supreme for more than a decade, flexibility of neural networks allowed them catch up, and works report accuracy metrics are par. Nevertheless, practical systems require considerations...
Recommender systems are typically evaluated using either offline methods, online or through user studies. In this paper we take an episode mining approach to analysing recommender system data and demonstrate how can use SNIPER, a tool for interactive pattern mining, analyse understand the behaviour of systems. We describe required format, present useful scenario interact with answer questions about quality recommendations.