Recent advances in Talking Head Generation (THG) have achieved impressive lip synchronization and visual quality through diffusion models; yet existing methods struggle to generate emotionally expressive portraits while preserving speaker identity. We identify three critical limitations in current emotional talking head generation: insufficient utilization of audio's inherent emotional cues, identity leakage in emotion representations, and isolated learning of emotion correlations. To address these challenges, we propose a novel framework dubbed as DICE-Talk, following the idea of disentangling identity with emotion, and then cooperating emotions with similar characteristics. First, we develop a disentangled emotion embedder that jointly models audio-visual emotional cues through cross-modal attention, representing emotions as identity-agnostic Gaussian distributions. Second, we introduce a correlation-enhanced emotion conditioning module with learnable emotion banks that explicitly capture inter-emotion relationships through vector quantization and attention-based feature aggregation. Third, we design an emotion discrimination objective that enforces affective consistency during the diffusion process through latent-space classification. Extensive experiments on MEAD and HDTF datasets demonstrate our method's superiority, outperforming state-of-the-art approaches in emotion accuracy while maintaining competitive lip-sync performance. Qualitative results and user studies further confirm our method's ability to generate identity-preserving portraits with rich, correlated emotional expressions that naturally adapt to unseen identities.
论文检索
输入标题、作者或关键词，从 1,092 篇学术成果中精准定位
The rapid advancement of multimedia technologies and their increasing integration in education have underscored the importance of multimedia learning. Knowledge Tracing (KT) plays a crucial role in enabling adaptive multimedia learning by continuously monitoring students' progress and forecasting their performance throughout the learning process. Question lies at the heart of the KT process, making its representation crucial for building efficient KT models. However, the sparsity and complexity of question data pose significant challenges for existing methods to capture the underlying features of questions, thereby affecting the accuracy of knowledge state predictions. To address this issue, this paper attempts to introduce the diffusion model to the KT field, proposing a novel knowledge tracing model, DiffuQKT. The model presents a diffusion-based generative approach for question representation and enhances the stability of knowledge states through contrastive learning. Specifically, DiffuQKT first constructs question representations based on their concepts, difficulty, and variations, and then, during the forward phase, progressively adds noise to the question representations, disrupting them into a Gaussian distribution. In the reverse phase, DiffuQKT gradually recovers the representations from noise, generating higher-quality question representations for knowledge tracing. Furthermore, to guide more meaningful question generation, we incorporate question concepts and difficulty as conditions during the denoising process. In addition, to improve the robustness of knowledge states against subtle variations in question representations, we employ contrastive learning to stabilize knowledge states across both original and denoised question representations. We conduct extensive experiments on four public datasets, comparing DiffuQKT with 15 baseline methods. The results demonstrate that DiffuQKT significantly outperforms existing models. Moreover, we find that the diffusion-based generative approach for question representation proposed in this paper has the ability to significantly improve the performance of baseline models. The code can be found at https://github.com/lilstrawberry/DiffuQKT.
With the rapid development of the diffusion models, numerous exquisitely generated images have significantly increased the risk of image misuse and abuse. Despite various AI parties and companies having devoted themselves to embedding watermarks into the generated images to curb the potential detriments, the isolated embedding from the generation process makes the watermarks vulnerable to watermark removal networks. To address this issue, we propose a novel generative image watermark scheme, dubbed Diffusion Visible Watermark (DVW), which can generate watermarked images in one step without additional training or fine-tuning of the diffusion models. Specifically, DVW introduces a masked distribution alignment strategy to fuse the watermark distribution with a Gaussian noise distribution. By iterative denoising the fused aligned distribution with the pretraining diffusion models, the watermarked images with coordinated and unified distribution can be generated with natural robustness against removal. In addition, we design and integrate a dynamic transparency module to adaptively control the watermark coverage degree for better visual quality. Comprehensive experiments and analysis are conducted on two representative kinds of diffusion models, GLIDE and StableDiffusion, to prove the superior and generic robustness of our DVW against watermark removal without sacrificing the generation ability of the diffusion models.
Photo-realistic novel view synthesis from multi-view images, such as neural radiance field (NeRF) and 3D Gaussian Splatting (3DGS), has gained significant attention for its superior performance. However, most existing methods rely on low dynamic range (LDR) images, limiting their ability to capture detailed scenes in high-contrast environments. While some prior works address high dynamic range (HDR) scene reconstruction, they typically require multi-view sharp images with varying exposure times captured at fixed camera positions-a process that is time-consuming and impractical. To make data acquisition more flexible, we propose Casual3DHDR, a robust one-stage method that reconstructs 3D HDR scenes from casually-captured auto-exposure (AE) videos, even under severe motion blur and unknown, varying exposure times. Our approach integrates a continuous camera trajectory into a unified physical imaging model, jointly optimizing exposure times, camera poses, and the camera response function (CRF). Extensive experiments on synthetic and real-world datasets demonstrate that Casual3DHDR outperforms existing methods in robustness and rendering quality.
3D Gaussian Splatting (3DGS) data compression is crucial for enabling efficient storage and transmission in 3D scene modeling. However, its development remains limited due to inadequate entropy models and suboptimal quantization strategies for both lossless and lossy compression scenarios, where existing methods have yet to 1) fully leverage hyperprior information to construct robust conditional entropy models, and 2) apply fine-grained, element-wise quantization strategies for improved compression granularity. In this work, we propose a novel Mixture of Priors (MoP) strategy to simultaneously address these two challenges. Specifically, inspired by the Mixture-of-Experts (MoE) paradigm, our MoP approach processes hyperprior information through multiple lightweight MLPs to generate diverse prior features, which are subsequently integrated into the MoP feature via a gating mechanism. To enhance lossless compression, the resulting MoP feature is utilized as a hyperprior to improve conditional entropy modeling. Meanwhile, for lossy compression, we employ the MoP feature as guidance information in an element-wise quantization procedure, leveraging a prior-guided Coarse-to-Fine Quantization (C2FQ) strategy with a predefined quantization step value. Specifically, we expand the quantization step value into a matrix and adaptively refine it from coarse to fine granularity, guided by the MoP feature, thereby obtaining a quantization step matrix that facilitates element-wise quantization. Extensive experiments demonstrate that our proposed 3DGS data compression framework achieves state-of-the-art performance across multiple benchmarks, including Mip-NeRF360, BungeeNeRF, DeepBlending, and Tank&Temples.
Dynamic novel view synthesis (NVS) aims to render time-varying scenes from arbitrary viewpoints, balancing rendering quality and computational efficiency. While recent 4D Gaussian Splatting approaches offer promising real-time performance, they fundamentally overlook critical interdependence between Gaussians by modeling deformations independently. Our information-theoretic analysis reveals substantial mutual information across the Gaussian field, manifesting as appearance-preserving radiance coherence and motion-consistent deformation propagation. This finding establishes that rendering quality emerges from coordinated transformation rather than independent processing. We propose Correlation-aware Dynamic Gaussian Splatting (CaDGS) with our novel Gaussian Correlation Tensor Projection (GCTP) method, which efficiently transforms the complex O(n3) mutual information tensor into a dual-channel O(n2) spatial matrix, preserving the critical topological structure of Gaussian interactions. Combined with our Spatio-Temporal Deformation Consistency (STDC) learning, which enforces volumetric coherence through tensor-guided regularization across multiple scales, CaDGS prevents geometric distortions and texture inconsistencies common in previous approaches. Experimental results demonstrate state-of-the-art performance, achieving 32.4 PSNR on the Neu3D dataset with fewer Gaussians while maintaining rendering speeds of 323 FPS at 1353 × 1014 resolution.
Image de-reflection is a critical task in computer vision. Existing methods for de-reflection using monocular cameras face challenges due to the lack of depth cues to separate the transmission and reflection layers, particularly under strong illumination or multi-layer reflection scenarios. Although recent advances, such as 3D Gaussian Splatting (3DGS), utilize novel view-synthesis capabilities to separate transmitted and reflected layers, they still encounter difficulties in practice with monocular images. In this paper, we simplify the de-reflection task by combining dual-pixel (DP) technology with 3DGS, forming the first unsupervised de-reflection framework. Specifically, we propose the Dual-View Coordinated Reflection Removal (DCRR) Framework, which integrates depth cues from DP sensors with the rendering capabilities of 3DGS. The DCRR utilizes a dual-view approach that estimates the image transmission layer and opacity via differentiable rasterization with 3DGS and reconstructs the reflection layer through a lightweight multi-layer perceptron. We then present the Dual-Pixel-Driven Reflection Gaussian Pruning (DPRGP) to refine the separation process. By using the physical properties of DP sensors, DCRR achieves significant accuracy improvements in complex reflection scenarios. A real-world DP-based dataset that includes paired reflection/reflection-free images has been collected. Extensive experiments demonstrate our competitive performance compared to state-of-the-art de-reflection approaches.
Monocular 3D Gaussian Splatting (3DGS) SLAM methods demonstrate outstanding performance in rapid dense 3D reconstruction. Yet former methods frequently exhibit suboptimal localization and mapping quality when processing indoor objects characterized by weak textures, dark colors, and high reflectivity (e.g., leather furniture), primarily due to insufficient surface feature information, even with the aid of depth sensors. To overcome these limitations, this work pioneers the integration of polarization information into the 3DGS SLAM framework. Specifically, we introduce a polarization integrated SLAM front-end that leverages the abundant planar features inherent in indoor environments. By incorporating a Chroma Boost mechanism, our approach effectively enhances the spectral multi-view consistency during the SLAM process, while the integration of a Gaussian-visible polarization difference improves the robustness of keyframe registration in low-texture scenarios. We further propose a flattened Gaussian regularization coupled with normal consistency constraints to capture the local geometric features of surfaces more accurately. Moreover, a novel integration of Pol-RGB hierarchical density plane segmentation and multi-scale plane self-constraint substantially enhances the quality of scene surface reconstruction, with further azimuth refinement achieved through the angle of linear polarization (AoLP). Extensive experiments demonstrate that, compared with previous SLAM methods, our approach significantly improves surface reconstruction quality.
3D Gaussian Splatting (3DGS) has demonstrated impressive capabilities in novel view synthesis. However, rendering reflective objects remains a significant challenge, particularly in inverse rendering and relighting. We introduce RTR-GS, a novel inverse rendering framework capable of robustly rendering objects with arbitrary reflectance properties, decomposing BRDF and lighting, and delivering credible relighting results. Given a collection of multi-view images, our method effectively recovers geometric structure through a hybrid rendering model that combines forward rendering for radiance transfer with deferred rendering for reflections. This approach successfully separates high-frequency and low-frequency appearances, mitigating floating artifacts caused by spherical harmonic overfitting when handling high-frequency details. We further refine BRDF and lighting decomposition using an additional physically-based deferred rendering branch. Experimental results show that our method enhances novel view synthesis, normal estimation, decomposition, and relighting while maintaining efficient training inference process.
Generative diffusion model approaches have achieved remarkable success in multimodal recommendation by generating latent user interest interaction graphs. However, current diffusion methods based on Gaussian noise introduce uncertain interest bias noise. This noise not only disrupts the original user-item interaction bipartite graph structure but also undermines the model's ability to accurately capture user interest preferences. To address these challenges, we propose Unbiased Interest Generation for Multimodal Recommendation (GenRec). Our approach aims to generate valid latent user interests while non-invasively preserving the original interest graph structure. We innovatively introduce the Multi-Modal Interest Generaction Module. During the forward process, we simulate user interest state transitions using ''forward flipping.'' In the reverse stage, we generate binary interaction graphs following a Bernoulli distribution. To further mitigate the random uncertainty during the generation process, we design a Multi-Modal Interest Debiase Module. By constructing a multimodal interest clustering space and using user interest hashing, we correct and enhance the generated interest graphs. Finally, the Multi-modal High-Order Graph Learning Optimization is employed to capture high-order interaction information between users and items. A contrastive learning loss function is used for model optimization. We conduct extensive experiments on different real-world open datasets from the industrial sector. Compared with the state-of-the-art DiffMM, our method significantly improves NDCG@20 by 7.3% on the TikTok dataset and boosts Recall@20 by 4.2% on the Sports dataset. The experimental results validate the effectiveness of GenRec. The code is publicly available at https://github.com/orangeheyue/GenRec-V1.
Novel view synthesis (NVS) in low-light scenes remains a significant challenge due to degraded inputs characterized by severe noise, low dynamic range (LDR) and unreliable initialization. While recent NeRF-based approaches have shown promising results, most suffer from high computational costs, and some rely on carefully captured or pre-processed data-such as RAW sensor inputs or multi-exposure sequences-which severely limits their practicality. In contrast, 3D Gaussian Splatting (3DGS) enables real-time rendering with competitive visual fidelity; however, existing 3DGS-based methods struggle with low-light sRGB inputs, resulting in unstable Gaussian initialization and ineffective noise suppression. To address these challenges, we propose LL-Gaussian, a novel framework for 3D reconstruction and enhancement from low-light sRGB images, enabling pseudo normal-light novel view synthesis. Our method introduces three key innovations: 1) an end-to-end Low-Light Gaussian Initialization Module (LLGIM) that leverages dense priors from learning-based MVS approach to generate high-quality initial point clouds; 2) a dual-branch Gaussian decomposition model that disentangles intrinsic scene properties (reflectance and illumination) from transient interference, enabling stable and interpretable optimization; 3) an unsupervised optimization strategy guided by both physical constrains and diffusion prior to jointly steer decomposition and enhancement. Additionally, we contribute a challenging dataset collected in extreme low-light environments and demonstrate the effectiveness of LL-Gaussian. Compared to state-of-the-art NeRF-based methods, LL-Gaussian achieves up to 2,000× faster inference and reduces training time to just 2%, while delivering superior reconstruction and rendering quality.
Recently, 3D Gaussian Splatting (3DGS) has achieved remarkable results in 3D reconstruction and view synthesis tasks. However, single-view feed-forward 3DGS still faces significant challenges. Current state-of-the-art (SOTA) single-view 3DGS methods typically employ a small number of layers (1-2 layers) with Gaussian Splatting (GS) representations at the same resolution as the input image to address the irregularity of GS data. However, such shallow and uniform GS primitive distributions is difficult to represent occluded regions and important spatial details. Inspired by multi-plane images, this paper proposes a Multi-Layer Gaussian Splatting (MLGS) representation, which consists of shallow base GS layers for visible content and multiple occlusion GS layers dedicated to reconstructing occluded regions. The proposed MLGS representation explicitly decouples the learning processes of visible and occluded content while enhancing occlusion prediction through the following components. First, spatial stratification of GS is achieved by estimating the depth distribution range of GS primitives across different layers, forcing GS to learn spatial content reconstruction at different depths. Second, a mask-guided mechanism is proposed to effectively isolate occlusion regions and guide inpainting using spatially context-aware features. Finally, a gated convolution block is designed to dynamically modulate feature fusion to enhance reconstruction fidelity. With separate loss supervision for base and occlusion layers, MLGS enables geometrically plausible scene completion. Experiments on RealEstate10K, KITTI, and NYUv2 datasets demonstrate that the proposed method achieves SOTA performance for single-image spatial scene reconstruction.
Dynamic scene reconstruction from monocular video is essential for real-world applications. We introduce DGNS, a hybrid framework integrating Deformable Gaussian Splatting and Dynamic Neural Surfaces, effectively addressing dynamic novel-view synthesis and 3D geometry reconstruction simultaneously. During training, depth maps generated by the deformable Gaussian splatting module guide the ray sampling for faster processing and provide depth supervision within the dynamic neural surface module to improve geometry reconstruction. Conversely, the dynamic neural surface directs the distribution of Gaussian primitives around the surface, enhancing rendering quality. In addition, we propose a depth-filtering approach to further refine depth supervision. Extensive experiments conducted on public datasets demonstrate that DGNS achieves state-of-the-art performance in 3D reconstruction, along with competitive results in novel-view synthesis.
PFDepth: Heterogeneous Pinhole-Fisheye Joint Depth Estimation via Distortion-aware Gaussian-Splatted Volumetric Fusion
PDF ↗In this paper, we present the first pinhole-fisheye framework for heterogeneous multi-view depth estimation, PFDepth. Our key insight is to exploit the complementary characteristics of pinhole and fisheye imagery (undistorted vs. distorted, small vs. large FOV, far vs. near field) for joint optimization. PFDepth employs a unified architecture capable of processing arbitrary combinations of pinhole and fisheye cameras with varied intrinsics and extrinsics. Within PFDepth, we first explicitly lift 2D features from each heterogeneous view into a canonical 3D volumetric space. Then, a core module termed Heterogeneous Spatial Fusion is designed to process and fuse distortion-aware volumetric features across overlapping and non-overlapping regions. Additionally, we subtly reformulate the conventional voxel fusion into a novel 3D Gaussian representation, in which learnable latent Gaussian spheres dynamically adapt to local image textures for finer 3D aggregation. Finally, fused volume features are rendered into multi-view depth maps. Through extensive experiments, we demonstrate that PFDepth sets a state-of-the-art performance on KITTI-360 and RealHet datasets over current mainstream depth networks. To the best of our knowledge, this is the first systematic study of heterogeneous pinhole-fisheye depth estimation, offering both technical novelty and valuable empirical insights.
Applying large vision models (LVMs) and large language models (LLMs) for item embedding is becoming cutting-edge for sequential recommendation, given their success in broad applications. Despite their advantages over traditional approaches, these models suffer more significant performance degradation on tail items against conventional ID-based solutions, which are largely overlooked by recent research. In this paper, we substantiate the above challenges as (1) all-in ground-truth, i.e., the standard cross-entropy (CE) loss focuses solely on the target items while treating all non-ground-truth equally, causing insufficient optimization for tail items, and (2) knowledge transfer tax, i.e., the knowledge encapsulated in LLMs and LVMs dominates the optimization process due to insufficient training for tail items. We propose Rewarding and reembedding, a simple yet efficient method to address the above challenges. Specifically, we reinitialize tail item embedding via a Gaussian distribution to alleviate knowledge transfer tax; besides, a rewarding function is incorporated in the CE loss, which adaptively adjusts item rewards during training to encourage the model to pay more attention to tail items rather than exclusively optimizing for ground-truth. Overall, our method enables a more nuanced optimization and is mathematically comparable to the direct preference optimization (DPO) in LLMs. Our extensive experiments on three public datasets show our method outperforms fourteen baselines in overall performance and improves the performance on tail items by a large margin. Our code is available at https://github.com/Yuhanleeee/R2Rec.
Diffusion models (DMs) have emerged as the new state-of-the-art family of deep generative models. To gain deeper insights into the limitations of diffusion models in recommender systems, we investigate the fundamental structural disparities between images and items. Consequently, items often exhibit distinct anisotropic and directional structures that are less prevalent in images. However, the traditional forward diffusion process continuously adds isotropic Gaussian noise, causing anisotropic signals to degrade into noise, which impairs the semantically meaningful representations in recommender systems. Inspired by the advancements in hyperbolic spaces, we propose a novel Hyperbolic Diffusion Recommender Model (named HDRM). Unlike existing directional diffusion methods based on Euclidean space, the intrinsic non-Euclidean structure of hyperbolic space makes it particularly well-adapted for handling anisotropic diffusion processes. In particular, we begin by constructing a geometrically latent space grounded in hyperbolic geometry, incorporating interpretability measures to define the latent anisotropic diffusion processes. Subsequently, we propose a novel hyperbolic latent diffusion process specifically tailored for users and items. Drawing upon the natural geometric attributes of hyperbolic spaces, we restrict both radial and angular components to facilitate directional diffusion propagation, thereby ensuring the preservation of the original topological structure in user-item interaction graphs. Extensive experiments on three benchmark datasets demonstrate the effectiveness of HDRM. Our code is available at https://github.com/yuanmeng-cpu/HDRM.
Model-Agnostic Social Network Refinement with Diffusion Models for Robust Social Recommendation
PDF ↗Social recommendations (SRs) aim to enhance preference modeling by integrating social networks. However, their effectiveness is mainly constrained by two factors: the noisy social connections that may not reflect shared interests, and the limited number of social connections for most users, which hampers the system's ability to fully leverage social influence. Therefore, it is essential to perform social network refinement by removing noisy connections and adding meaningful ones for robust SRs. Inspired by the denoising capability of generative diffusion models, we propose a Model-Agnostic Social Network Refinement framework with Diffusion Models for Robust Social Recommendation (ARD-SR). Specifically, in the forward process, we corrupt the social network by progressively adding position-specific Gaussian noise calibrated to the user preference similarity, better simulating how the social network responds to noise perturbations. The reverse process learns to denoise, guided by each user's neighborhood preferences from the SR backbone, generating a tailored social network aligned with each user's preference for establishing connections. For effective learning, we design a curriculum-based training mechanism that progressively introduces challenging samples characterized by high sparsity or high noise levels. Finally, ARD-SR and the SR backbone are alternately trained, ensuring a continuous mutual enhancement between the social network refinement and the backbone's user representation learning. To further enhance the quality of the refined social network, (1) we introduce a preference-guided flip operation during inference to improve the input quality; and (2) we modify social connections based on the exponential weighted moving average of ARD-SR's predictions across epochs to reduce fluctuations. Experiments on three datasets show that ARD-SR significantly improves SR performance across multiple SR backbones. The code is released at https://github.com/sunyc123r/ARD-SR.
Many real-world applications contain data in the form of multivariate time series (TS) with the hierarchical structure, where classic methods forecasting each TS independently are inadequate for coherency (i.e., satisfying the hierarchical aggregation constraints). Furthermore, the discrepancies between statistical properties of different levels can be huge, exacerbated by non-Gaussian distributions and non-linear correlations. In this paper, we propose a novel end-to-end hierarchical TS forecasting model, i.e., a Flow-based Reconcile Transformer (FRT). FRT employs a conditional normalizing flow-based autoregressive transformer, to represent complex data distribution, while simultaneously reconciling the forecasts to ensure coherency. Go beyond other state-of-the-art methods, FRT accomplishes forecasting and reconciliation simultaneously, while avoiding any post-processing steps. Moreover, FRT is a deep model that does not rely on any strong assumptions such as unbiased estimates or Gaussian distribution. Our experiments are conducted on four real-world hierarchical datasets from different industrial domains (three public ones and a dataset from the application servers of our company's data center) and the results demonstrate the efficacy of our proposed method. Our method has been implemented extensively within the production environments of a prominent global payment company. It has emerged as a cornerstone for workload forecasting within their data center and plays a critical role in the optimization of cloud computing resource allocation across the entire cluster. This successful deployment of the application demonstrates that our approach achieves precise hierarchical workload prediction, which is of great significance for efficient resource scheduling, enhancing resource utilization, and reducing server resource use and energy consumption in data centers.
Incomplete Multi-View Drug Recommendation via Multi-Level Representation Learning and Curriculum Learning
PDF ↗The drug recommendation task aims to provide effective and safe prescription decision support for clinical treatment based on patients' past Electronic Health Records (EHR). However, the prevalent phenomenon of missing views in multi-source heterogeneous EHR data may cause performance degradation. This is due to the lack of sufficient information and increased learning difficulties, which limit the practical effectiveness of drug recommendation models in medical applications. In this paper, we emphasize the problems of incompleteness in practical drug recommendation and propose the Incomplete Multi-View Drug Recommendation model via Multi-Level Representation Learning and Curriculum Learning named IMDR. In particular, IMDR employs a Multi-Level Representation Learning architecture equipped with a Medical Code-Level Drug Knowledge Infusion Module and a Visit-Level Cross-View Information Module for patient representation learning to overcome the information loss caused by incomplete data. And then, a Gaussian-guided curriculum learning strategy is proposed to assist the learning process of IMDR with a novel difficulty measure to achieve effective progressive learning under missing medical views. Systematic evaluation on two large-scale real-world medical datasets, MIMIC-III and MIMIC-IV, demonstrates that IMDR reduces the Drug-Drug Interaction (DDI) rate by 2.97% compared to existing state-of-the-art drug recommendation baselines, while achieving significant improvements of 3.29% and 1.97% in Jaccard similarity scores and F1 score, respectively. Furthermore, compared to advanced incomplete multi-view learning (IML) models, IMDR's advantages in Jaccard similarity scores and F1 score further expand to 4.03% and 2.41%.
Canonical algorithms for multi-armed bandits typically assume a stationary reward environment where the size of the action space (number of arms) is small. More recently developed methods typically relax only one of these assumptions: existing non-stationary bandit policies are designed for a small number of arms, while Lipschitz, linear, and Gaussian process bandit policies are designed to handle a large (or infinite) number of arms in stationary reward environments under constraints on the reward function. In this manuscript, we propose a novel policy to learn reward environments over a continuous space using Gaussian interpolation. We show that our method efficiently learns continuous Lipschitz reward functions with O*(√T) cumulative regret. Furthermore, our method naturally extends to non-stationary problems with a simple modification. We finally demonstrate that our method is computationally favorable (100-10000x faster) and experimentally outperforms sliding Gaussian process policies on datasets with non-stationarity and an extremely large number of arms.