arXiv subjects

Julien Monteil

Publications and source records attributed to Julien Monteil.

18 recordsLinked to original sources

arXiv2026

BACH: A Bayesian Admixture of Contrastive Heads for Multi-Interest Two-Tower Retrieval

Two-tower retrievers compress each user into a single embedding, limiting their ability to serve diverse interests. Multi-interest models give each user several heads scored by a maximum inner product, but their hard-routing training under-utilizes heads (routing collapse) and gives no per-user estimate of how much each interest matters for serving. We present \textbf{BACH} (\emph{Bayesian Admixture of Contrastive Heads}), which casts multi-interest two-tower retrieval as a per-user mixture over the heads, fit by variational inference. The soft mixture trains every head (mitigating collapse), produces a per-user weighting of the interests that is reused at serving, and admits a shared global-codebook variant with precomputable retrieval. On three large-scale benchmarks, MovieLens-20M, Taobao, and Netflix, BACH improves top-of-ranking retrieval over hard-routing multi-interest and single-vector baselines at every head count; we further find that scoring every candidate by its best head, consistent with serving, outperforms the usual target-routed training, and that BACH improves further still.

cs.IR↗

arXiv2026

Online Data Selection for Instruction Tuning via Gaussian Processes

With Large Language Model (LLM) pre-training and fine-tuning shifting its focus from data volume to data quality, quality data selection has emerged as a critical research topic. Existing online data selection methods for LLM training are typically "batch-constrained", limiting optimization to local utility within random batches. To overcome this, we propose GAIA (Global Adaptive Instruction tuning via GAussian processes), a framework that formulates data valuation as a global estimation process. GAIA employs Gaussian Process regression to model continuous utility manifolds across the semantic space, utilizing an adaptive strategy fusion mechanism to dynamically prioritize high-utility samples. By casting the strategy-posterior update as an instance of the classical fixed-share Hedge framework for tracking the best expert, we inherit a dynamic-regret guarantee that characterizes GAIA's robustness under non-stationary quality scores during training. Empirical evaluations on three datasets demonstrate that GAIA significantly outperforms state-of-the-art baselines like \greats, establishing our method as a scalable and robust solution for efficient instruction tuning.

cs.LG↗

arXiv2026

FOSTER: First-order Dataset Distillation for Text-based Sequential Recommendation

Text-based sequential recommender systems, while greatly improving recommendation accuracy by incorporating item contexts, are undeniably more expensive to train. By condensing a large dataset into a compact set of synthetic samples for model training, dataset distillation offers a promising solution. However, its adoption in text-based sequential recommendation is non-trivial given the large pool of discrete items. This challenge is further compounded by language model-based item encoding, which makes bi-level optimization commonly used in dataset distillation prohibitively expensive. To this end, we propose First-order dataset distillation for Text-based Sequential Recommendation (FOSTER), which facilitates effectiveness and efficiency via three novel components: (1) stochastic item subset sampling that replaces costly full-corpus embedding extraction at each distillation step; (2) first-order optimization with trajectory-anchored parameter reset to avoid expensive bi-level gradient computation; and (3) regularization that explicitly promotes co-occurrence between semantically similar items in the synthetic sequences. Extensive experiments on three benchmarks show that FOSTER consistently outperforms existing dataset distillation and coreset selection baselines, approximating full-dataset performance using as few as 20 synthetic interaction sequences.

cs.IR↗

arXiv2025

On the Mechanisms of Collaborative Learning in VAE Recommenders

Variational Autoencoders (VAEs) are a powerful alternative to matrix factorization for recommendation. A common technique in VAE-based collaborative filtering (CF) consists in applying binary input masking to user interaction vectors, which improves performance but remains underexplored theoretically. In this work, we analyze how collaboration arises in VAE-based CF and show it is governed by \emph{latent proximity}: we derive a latent sharing radius that informs when an SGD update on one user strictly reduces the loss on another user, with influence decaying as the latent Wasserstein distance increases. We further study the induced geometry: with clean inputs, VAE-based CF primarily exploits \emph{local} collaboration between input-similar users and under-utilizes \emph{global} collaboration between far-but-related users. We compare two mechanisms that encourage \emph{global} mixing and characterize their trade-offs: \ding{172} $\beta$-KL regularization directly tightens the information bottleneck, promoting posterior overlap but risking representational collapse if too large; \ding{173} input masking induces stochastic \emph{geometric} contractions and expansions, which can bring distant users onto the same latent neighborhood but also introduce neighborhood drift. To preserve user identity while enabling global consistency, we propose an anchor regularizer that aligns user posteriors with item embeddings, stabilizing users under masking and facilitating signal sharing across related items. Our analyses are validated on the Netflix, MovieLens-20M, and Million Song datasets. We also successfully deployed our proposed algorithm on an Amazon streaming platform following a successful online experiment.

cs.LG↗

arXiv2024

Learning Visual Hierarchies in Hyperbolic Space for Image Retrieval

Structuring latent representations in a hierarchical manner enables models to learn patterns at multiple levels of abstraction. However, most prevalent image understanding models focus on visual similarity, and learning visual hierarchies is relatively unexplored. In this work, for the first time, we introduce a learning paradigm that can encode user-defined multi-level complex visual hierarchies in hyperbolic space without requiring explicit hierarchical labels. As a concrete example, first, we define a part-based image hierarchy using object-level annotations within and across images. Then, we introduce an approach to enforce the hierarchy using contrastive loss with pairwise entailment metrics. Finally, we discuss new evaluation metrics to effectively measure hierarchical image retrieval. Encoding these complex relationships ensures that the learned representations capture semantic and structural information that transcends mere visual similarity. Experiments in part-based image retrieval show significant improvements in hierarchical retrieval tasks, demonstrating the capability of our model in capturing visual hierarchies.

cs.CV↗

arXiv2024

Geometric Collaborative Filtering with Convergence

Latent variable collaborative filtering methods have been a standard approach to modelling user-click interactions due to their simplicity and effectiveness. However, there is limited work on analyzing the mathematical properties of these methods in particular on preventing the overfitting towards the identity, and such methods typically utilize loss functions that overlook the geometry between items. In this work, we introduce a notion of generalization gap in collaborative filtering and analyze this with respect to latent collaborative filtering models. We present a geometric upper bound that gives rise to loss functions, and a way to meaningfully utilize the geometry of item-metadata to improve recommendations. We show how these losses can be minimized and gives the recipe to a new latent collaborative filtering algorithm, which we refer to as GeoCF, due to the geometric nature of our results. We then show experimentally that our proposed GeoCF algorithm can outperform other all existing methods on the Movielens20M and Netflix datasets, as well as two large-scale internal datasets. In summary, our work proposes a theoretically sound method which paves a way to better understand generalization of collaborative filtering at large.

cs.IR↗

arXiv2024

Personalised Outfit Recommendation via History-aware Transformers

We present the history-aware transformer (HAT), a transformer-based model that uses shoppers' purchase history to personalise outfit predictions. The aim of this work is to recommend outfits that are internally coherent while matching an individual shopper's style and taste. To achieve this, we stack two transformer models, one that produces outfit representations and another one that processes the history of purchased outfits for a given shopper. We use these models to score an outfit's compatibility in the context of a shopper's preferences as inferred from their previous purchases. During training, the model learns to discriminate between purchased and random outfits using 3 losses: the focal loss for outfit compatibility typically used in the literature, a contrastive loss to bring closer learned outfit embeddings from a shopper's history, and an adaptive margin loss to facilitate learning from weak negatives. Together, these losses enable the model to make personalised recommendations based on a shopper's purchase history. Our experiments on the IQON3000 and Polyvore datasets show that HAT outperforms strong baselines on the outfit Compatibility Prediction (CP) and the Fill In The Blank (FITB) tasks. The model improves AUC for the CP hard task by 15.7% (IQON3000) and 19.4% (Polyvore) compared to previous SOTA results. It further improves accuracy on the FITB hard task by 6.5% and 9.7%, respectively. We provide ablation studies on the personalisation, constrastive loss, and adaptive margin loss that highlight the importance of these modelling choices.

eess.SY↗

arXiv2024

MARec: Metadata Alignment for cold-start Recommendation

For many recommender systems, the primary data source is a historical record of user clicks. The associated click matrix is often very sparse, as the number of users x products can be far larger than the number of clicks. Such sparsity is accentuated in cold-start settings, which makes the efficient use of metadata information of paramount importance. In this work, we propose a simple approach to address cold-start recommendations by leveraging content metadata, Metadata Alignment for cold-start Recommendation. We show that this approach can readily augment existing matrix factorization and autoencoder approaches, enabling a smooth transition to top performing algorithms in warmer set-ups. Our experimental results indicate three separate contributions: first, we show that our proposed framework largely beats SOTA results on 4 cold-start datasets with different sparsity and scale characteristics, with gains ranging from +8.4% to +53.8% on reported ranking metrics; second, we provide an ablation study on the utility of semantic features, and proves the additional gain obtained by leveraging such features ranges between +46.8% and +105.5%; and third, our approach is by construction highly competitive in warm set-ups, and we propose a closed-form solution outperformed by SOTA results by only 0.8% on average.

cs.IR↗

arXiv2019

A city-scale IoT-enabled ridesharing platform

The advent of on-demand mobility systems is expected to have a tremendous potential on the wellness of transportation users in cities. Yet such positive effects are reached when the systems under consideration enable seamless integration between data sources that involve a high number of transportation actors. In this paper we report on the effort of designing and deploying an integrated system, including algorithms and platforms, that can operate in cities, in an Internet of Things (IoT)-aware fashion. The system was evaluated by enabling/disabling the IoT components of the system, highlighting the necessity of real-time data integration for efficient mobility services.

cs.CY↗

arXiv2019

On model selection for scalable time series forecasting in transport networks

The transport literature is dense regarding short-term traffic predictions, up to the scale of 1 hour, yet less dense for long-term traffic predictions. The transport literature is also sparse when it comes to city-scale traffic predictions, mainly because of low data availability. In this work, we report an effort to investigate whether deep learning models can be useful for the long-term large-scale traffic prediction task, while focusing on the scalability of the models. We investigate a city-scale traffic dataset with 14 weeks of speed observations collected every 15 minutes over 1098 segments in the hypercenter of Los Angeles, California. We look at a variety of state-of-the-art machine learning and deep learning predictors for link-based predictions, and investigate how such predictors can scale up to larger areas with clustering, and graph convolutional approaches. We discuss that modelling temporal and spatial features into deep learning predictors can be helpful for long-term predictions, while simpler, not deep learning-based predictors, achieve very satisfactory performance for link-based and short-term forecasting. The trade-off is discussed not only in terms of prediction accuracy vs prediction horizon but also in terms of training time and model sizing.

cs.LG↗

arXiv2019

On the needs for MaaS platforms to handle competition in ridesharing mobility

Ridesharing has been emerging as a new type of mobility. However, the early promises of ridesharing for alleviating congestion in cities may be undermined by a number of challenges, including the growing number of proposed services and the subsequent increasing number of vehicles, as a natural consequence of competition. In this work, we present optimization-based approaches to model cooperation and competition between multiple ridesharing companies, in a real-time on-demand setting. A recent trend relies on solving the integrated combination of Dial-A-Ride Problems (DARP), which compute the cost of assigning incoming requests to vehicle routes, plus Linear Assignment Problems (LAP), which assign vehicles to requests. While the DARPs, are solved at the level of the vehicles of each company, we introduce cooperative and competitive approaches to solve the LAP. The cooperative model, which could make use of Mobility as a Service platforms, is shown to solve the LAP to optimality following closely results from the literature, and limiting the amount of information the companies are required to share. We investigate how a realistic model of competition deviates from this optimality and provide worst case bounds. We evaluate these models with respect to a centralized model on one-week instances of the New York City taxi dataset. Model variants coping with noise in the travel time estimations, bias in the assignment costs, and preferences in the competitive case are also presented and validated. The computational results suggest that cooperation among ridesharing companies can be conducted in such a way to limit the degradation of the level of service with respect to a centralized model. Finally, we argue that the competition can lower the quality of the ridesharing service, especially in the case customer preferences are accommodated.

cs.CY↗

arXiv2019

Personalized Optimization with User's Feedback

This paper develops an online algorithm to solve a time-varying optimization problem with an objective that comprises a known time-varying cost and an unknown function. This problem structure arises in a number of engineering systems and cyber-physical systems where the known function captures time-varying engineering costs, and the unknown function models user's satisfaction; in this context, the objective is to strike a balance between given performance metrics and user's satisfaction. Key challenges related to the problem at hand are related to (1) the time variability of the problem, and (2) the fact that learning of the user's utility function is performed concurrently with the execution of the online algorithm. This paper leverages Gaussian processes (GP) to learn the unknown cost function from noisy functional evaluation and build pertinent upper confidence bounds. Using the GP formalism, the paper then advocates time-varying optimization tools to design an online algorithm that exhibits tracking of the oracle-based optimal trajectory within an error ball, while learning the user's satisfaction function with no-regret. The algorithmic steps are inexact, to account for possible limited computational budgets or real-time implementation considerations. Numerical examples are illustrated based on a problem related to vehicle platooning.

math.OC↗

arXiv2019

Real-time City-scale Ridesharing via Linear Assignment Problems

In this paper, we propose a novel, computational efficient, dynamic ridesharing algorithm. The beneficial computational properties of the algorithm arise from casting the ridesharing problem as a linear assignment problem between fleet vehicles and customer trip requests within a federated optimization architecture. The resulting algorithm is up to four times faster than the state-of-the-art, even if it is implemented on a less dedicated hardware, and achieves similar service quality. Current literature showcases the ability of state-of-the-art ridesharing algorithms to tackle very large fleets and customer requests in almost near real-time, but the benefits of ridesharing seem limited to centralized systems. Our algorithm suggests that this does not need to be the case. The algorithm that we propose is fully distributable among multiple ridesharing companies. By leveraging two datasets, the New York city taxi dataset and the Melbourne Metropolitan Area dataset, we show that with our algorithm, real-time ridesharing offers clear benefits with respect to more traditional taxi fleets in terms of level of service, even if one considers partial adoption of the system. In fact, e.g., the quality of the solutions obtained in the state-of-the-art works that tackle the whole customer set of the New York city taxi dataset is achieved, even if one considers only a proportion of the fleet size and customer requests. This could make real-time urban-scale ridesharing very attractive to small enterprises and city authorities alike. However, in some cases, e.g., in multi-company scenarios where companies have predefined market shares, we show that the number of vehicles needed to achieve a comparable performance to the monopolistic setting increases, and this raises concerns on the possible negative effects of multi-company ridesharing.

math.OC↗

arXiv2018

Using Deep Learning to Extend the Range of Air-Pollution Monitoring and Forecasting

Across numerous applications, forecasting relies on numerical solvers for partial differential equations (PDEs). Although the use of deep-learning techniques has been proposed, actual applications have been restricted by the fact the training data are obtained using traditional PDE solvers. Thereby, the uses of deep-learning techniques were limited to domains, where the PDE solver was applicable. We demonstrate a deep-learning framework for air-pollution monitoring and forecasting that provides the ability to train across different model domains, as well as a reduction in the run-time by two orders of magnitude. It presents a first-of-a-kind implementation that combines deep-learning and domain-decomposition techniques to allow model deployments extend beyond the domain(s) on which the it has been trained.

cs.LG↗

arXiv2018

Bayesian Classifier for Route Prediction with Markov Chains

We present here a general framework and a specific algorithm for predicting the destination, route, or more generally a pattern, of an ongoing journey, building on the recent work of [Y. Lassoued, J. Monteil, Y. Gu, G. Russo, R. Shorten, and M. Mevissen, "Hidden Markov model for route and destination prediction," in IEEE International Conference on Intelligent Transportation Systems, 2017]. In the presented framework, known journey patterns are modelled as stochastic processes, emitting the road segments visited during the journey, and the ongoing journey is predicted by updating the posterior probability of each journey pattern given the road segments visited so far. In this contribution, we use Markov chains as models for the journey patterns, and consider the prediction as final, once one of the posterior probabilities crosses a predefined threshold. Despite the simplicity of both, examples run on a synthetic dataset demonstrate high accuracy of the made predictions.

cs.LG↗

arXiv2018

A Hidden Markov Model for Route and Destination Prediction

We present a simple model and algorithm for predicting driver destinations and routes, based on the input of the latest road links visited as part of an ongoing trip. The algorithm may be used to predict any clusters previously observed in a driver's trip history. It assumes that the driver's historical trips are grouped into clusters sharing similar patterns. Given a new trip, the algorithm attempts to predict the cluster in which the trip belongs. The proposed algorithm has low temporal complexity. In addition, it does not require the transition and emission matrices of the Markov chain to be computed. Rather it relies on the frequencies of co-occurrences of road links and trip clusters. We validate the proposed algorithm against an experimental dataset. We discuss the success and convergence of the algorithm and show that our algorithm has a high prediction success rate.

physics.soc-ph↗

arXiv2018

On $\mathcal{L}_{\infty}$ string stability of nonlinear bidirectional asymmetric heterogeneous platoon systems

This paper is concerned with the study of bidirectionally coupled platoon systems. The case considered is when the vehicles are heterogeneous and the coupling can be nonlinear and asymmetric. For such systems, a sufficient condition for $\mathcal{L}_{\infty}$ string stability is presented. The effectiveness of our approach is illustrated via a numerical example, where it is shown how our result can be recast as an optimization problem, allowing to design the control protocol for each vehicle independently on the other vehicles and hence leading to a bottom-up approach for the design of string stable systems able to track a time-varying reference speed.

eess.SY↗

arXiv2016

$\mathcal{L}_2$ and $\mathcal{L}_{\infty}$ stability analysis of heterogeneous traffic with application to parameter optimisation for the control of automated vehicles

The presence of (partially) automated vehicles on the roads presents an opportunity to compensate the unstable behaviour of conventional vehicles. Vehicles subject to perturbations should (i) recover their equilibrium speed, (ii) react not to propagate but absorb perturbations. In this work, we start with considering vehicle systems consisting of heterogeneous vehicles updating their dynamics according to realistic behavioural car-following models. Definitions of all types of stability that are of interest in the vehicle system, namely input-output stability, scalability, weak and strict string stability, are introduced based on recent studies. Then, frequency domain linear stability analyses are conducted after linearisation of the modelled system of vehicles, leading to conditions for input-output stability, strict and weak string stability over the behavioural parameters of the system, for finite and infinite systems of homogeneous and heterogeneous vehicles. This provides a solid basis that was missing for car-following model-based control design in mixed traffic systems where only a proportion of vehicles can be controlled. After visualisation of the theoretical results in simulation, we formulate an optimisation strategy with LMI constraints to tune the behavioural parameters of the automated vehicles in order to maximise the L1 string stability of the mixed traffic flow while considering the comfort of automated driving. The optimisation strategy systematically leads to increased traffic flow stability. We show that very few automated vehicles are required to prevent the

eess.SY↗