arXiv subjects

Xuefeng Yao

Publications and source records attributed to Xuefeng Yao.

4 recordsLinked to original sources

arXiv2025

Towards Large-scale Generative Ranking

Generative recommendation has recently emerged as a promising paradigm in information retrieval. However, generative ranking systems are still understudied, particularly with respect to their effectiveness and feasibility in large-scale industrial settings. This paper investigates this topic at the ranking stage of Xiaohongshu's Explore Feed, a recommender system that serves hundreds of millions of users. Specifically, we first examine how generative ranking outperforms current industrial recommenders. Through theoretical and empirical analyses, we find that the primary improvement in effectiveness stems from the generative architecture, rather than the training paradigm. To facilitate efficient deployment of generative ranking, we introduce GenRank, a novel generative architecture for ranking. We validate the effectiveness and efficiency of our solution through online A/B experiments. The results show that GenRank achieves significant improvements in user satisfaction with nearly equivalent computational resources compared to the existing production system.

cs.IR↗

arXiv2023

HeterPS: Distributed Deep Learning With Reinforcement Learning Based Scheduling in Heterogeneous Environments

Deep neural networks (DNNs) exploit many layers and a large number of parameters to achieve excellent performance. The training process of DNN models generally handles large-scale input data with many sparse features, which incurs high Input/Output (IO) cost, while some layers are compute-intensive. The training process generally exploits distributed computing resources to reduce training time. In addition, heterogeneous computing resources, e.g., CPUs, GPUs of multiple types, are available for the distributed training process. Thus, the scheduling of multiple layers to diverse computing resources is critical for the training process. To efficiently train a DNN model using the heterogeneous computing resources, we propose a distributed framework, i.e., Paddle-Heterogeneous Parameter Server (Paddle-HeterPS), composed of a distributed architecture and a Reinforcement Learning (RL)-based scheduling method. The advantages of Paddle-HeterPS are three-fold compared with existing frameworks. First, Paddle-HeterPS enables efficient training process of diverse workloads with heterogeneous computing resources. Second, Paddle-HeterPS exploits an RL-based method to efficiently schedule the workload of each layer to appropriate computing resources to minimize the cost while satisfying throughput constraints. Third, Paddle-HeterPS manages data storage and data communication among distributed computing resources. We carry out extensive experiments to show that Paddle-HeterPS significantly outperforms state-of-the-art approaches in terms of throughput (14.5 times higher) and monetary cost (312.3% smaller). The codes of the framework are publicly available at: https://github.com/PaddlePaddle/Paddle.

cs.DC↗

arXiv2022

Optical Design and Wavelength Calibration of a DMD-based Multi-Object Spectrograph

The multi-object spectrograph (MOS) has been the benchmark for the current generation of astronomical spectrographs, valued for its ability to acquire the spectra of hundreds of objects simultaneously. In the last two decades, the digital micromirror device (DMD) has shown potential in becoming the central component of the MOS, being used as a programmable slit array. We have designed a seeing-limited DMD-based MOS covering a spectral range of 0.4 to 0.7 $μ$m, with a field of view (FOV) of $10.5^\prime \times 13.98^\prime$ and a spectral resolution of $R\sim1000$. This DMD-MOS employs all-spherical refractive optics, and a volume phase holographic (VPH) grism as the dispersive element for high throughput. In this paper, we present the optical design and optimization process of this DMD-MOS, as well as a preliminary wavelength calibration procedure for hyperspectral data reduction. Using simulated data of the DMD-MOS, a procedure was developed to measure hyperspectral imaging distortion and to construct pixel-to-wavelength mappings on the detector. An investigation into the relationships between DMD micromirrors and detector pixels was conducted. This DMD-MOS will be placed on a 0.5 m diameter telescope as an exploratory study for future DMD-based MOS systems.

astro-ph.IM↗

arXiv2017

What is the Optimal Network Deployment for a Fixed Density of Antennas?

In this paper, we answer a fundamental question: when the total number of antennas per square kilometer is fixed, what is the optimal network deployment? A denser network with a less number of antennas per base station (BS) or the opposite case. To evaluate network performance, we consider a practical network scenario with a fixed antennas density and multiuser multiple-input-multiple-output (MU-MIMO) operations for single-antenna users. The number of antennas in each BS is calculated by dividing the antenna density by the BS density. With the consideration of several practical network models, i.e., pilot contamination, a limited user equipment (UE) density and probabilistic line-of-sight (LoS)/non-line-of-sight (NLoS) path loss model, we evaluate the area spectral efficiency (ASE) performance. From our simulation results, we conclude that there exists an optimal BS density for a certain UE density to maximize the ASE performance when the antenna density is fixed. The intuition is that (i) by densifying the network with more BSs, we can achieve a receive power gain due to the smaller distance between the typical UE and its serving BS; (ii) by installing more antennas in each BS, we can achieve a beamforming gain for UEs using MU-MIMO, although such beamforming gain is degraded by pilot contamination; (iii) thus, a trade-off exists between the receive power gain and the beamforming gain, if we fix the antenna density in the network.

cs.NI↗