Adapting large language models (LLMs) for personalized recommendation requires aligning their general-purpose capabilities with user-specific preferences while effectively leveraging both behavioral and semantic signals. Existing approaches typically integrate these signals at either the input level (e.g., injecting behavioral embeddings into the token space) or the output level (e.g., contrastive alignment of separate encoders), suffering from distribution gaps or lack of end-to-end task supervision. In this work, we introduce L2Rec, which unifies behavioral and semantic understanding at the parameter level of LLMs. Our key insight is that the same set of Transformer parameters can serve as a shared medium for both views: by applying view-specific, personalized low-rank perturbations via a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism, L2Rec enables a single LLM backbone to produce complementary behavioral and semantic adaptations for each user with minimal representation-level misalignment. An adaptive cross-view fusion module further integrates the dual-view outputs into a unified user preference. Experiments on four datasets show that L2Rec consistently outperforms state-of-the-art baselines, and online A/B testing on a large-scale industrial platform validates significant improvements in key engagement metrics.

In music recommendation systems, multimodal interest learning is pivotal, which allows the model to capture nuanced preferences, including textual elements such as lyrics and various musical attributes such as different instruments and melodies. Recently, methods that incorporate multimodal content features through semantic IDs have achieved promising results. However, existing methods suffer from two critical limitations: 1) intra-modal semantic degradation, where residual-based quantization processes gradually decouple discrete IDs from original content semantics, leading to semantic drift; and 2) inter-modal modeling gaps, where traditional fusion strategies either overlook modal-specific details or fail to capture cross-modal correlations, hindering comprehensive user interest modeling. To address these challenges, we propose a novel multimodal recommendation framework with two stages. In the first stage, our Progressive Semantic Residual Quantization (PSRQ) method generates modal-specific and modal-joint semantic IDs by explicitly preserving the prefix semantic feature. In the second stage, to model multimodal interest of users, a Multi-Codebook Cross-Attention (MCCA) network is designed to enable the model to simultaneously capture modal-specific interests and perceive cross-modal correlations. Extensive experiments on multiple real-world datasets demonstrate that our framework outperforms state-of-the-art baselines. This framework has been deployed on one of China's largest music streaming platforms, and online A/B tests confirm significant improvements in commercial metrics, underscoring its practical value for industrial-scale recommendation systems.

Transformer-based generative models have achieved remarkable success across domains with various scaling law manifestations. However, our extensive experiments reveal persistent challenges when applying Transformer to recommendation systems: (1) Transformer scaling is not ideal with increased computational resources, due to structural incompatibilities with recommendation-specific features such as multi-source data heterogeneity; (2) critical online inference latency constraints (tens of milliseconds) that intensify with longer user behavior sequences and growing computational demands. We propose Climber, an efficient recommendation framework comprising two synergistic components: the model architecture for efficient scaling and the co-designed acceleration techniques. Our proposed model adopts two core innovations: (1) multi-scale sequence extraction that achieves a time complexity reduction by a constant factor, enabling more efficient scaling with sequence length; (2) dynamic temperature modulation adapting attention distributions to the multi-scenario and multi-behavior patterns. Complemented by acceleration techniques, Climber achieves a 5.15× throughput gain without performance degradation by adopting a ''single user, multiple item'' batched processing and memory-efficient Key-Value caching.

Representation learning remains one of the most important but challenging tasks within industrial music rec-ommendation systems. In the context of the Matthew effect, item exposure frequency demonstrates substantial inequality, leading to the Harry Potter problem for popular items and the long-tail issue for less interacted items, collectively impairing the adequacy and accuracy of representation learning. In this paper, to alleviate the negative impact of bias on representation learning in music recommendation systems, we propose a unified model based on introducing the unbiased Cascading Multimodal Feature, called CMF4Rec. Specifically, with our cascading feature enhancement module, we implement a dual-stage representation enhancement strategy. In the first stage, the pivotal subsequence is extracted from the coarse-grained similarity sequence derived from cascading multimodal features, which is subsequently ag-gregated to generate the enhanced representation of the candidate item. Moreover, in the feature interaction module, the enhanced representation is crossed with user behaviors to capture the diverse and dynamic interests of users. Furthermore, we employ contrastive learning and design an auxiliary contrastive task to provide high-quality gradients for the main recommendation task. We demonstrate the effectiveness of this model with extensive experiments on public and industrial datasets. Moreover, the deployment of CMF4Rec in a real music recommendation system has also yielded significant improvements.

User behavior has been validated to be effective in revealing personalized preferences for commercial recommendations. However, few user-item interactions can be collected for new users, which results in a nullspace for their interests, ie, the cold-start dilemma. In this paper, a two-tower framework, namely, the model-agnostic interest learning (MAIL) framework, is proposed to address the cold-start recommendation (CSR) problem for recommender systems. In MAIL, one unique tower is constructed to tackle the CSR from a zero-shot view, and the other tower focuses on the general ranking task. Specifically, the zero-shot tower first performs cross-modal reconstruction with dual autoencoders to obtain virtual behavior data from highly aligned hidden features for new users; and the ranking tower can then output recommendations for users based on the completed data by the zero-shot tower. Practically, the ranking tower in MAIL is model-agnostic and can be implemented with any embedding-based deep models. Based on the cotraining of the two towers, the MAIL presents an end-to-end method for recommender systems that shows an incremental performance improvement. The proposed method has been successfully deployed on the live recommendation system of NetEase Cloud Music to achieve a click-through rate improvement of 13% to 15% for millions of users. Offline experiments on real-world datasets also show its superior performance in CSR. Our code is available.

Flash Sale Sites popularize because they save great money for users. Good recommender systems can further save users' time to improve their online shopping experiences. Although there exist a lot of studies on recommender system, very few focus on flash sale sites. Big Data, Context Sensitivity, and Feature Engineering are three key challenges for one to build a good recommender system. This paper proposes two deep learning oriented models: Tensor-AutoRec and Hybrid-AutoRec to cope with the problems within an industrial context. First, these two models can handle storage and speed problem caused by big data. Second, both models incorporate context information, so they can generate more relevant recommendations by adapting to specific contextual situations. Third, our deep learning-based models can be trained end-to-end without tedious feature engineerings. Extensive experiments with a half year real transcation data demonstrate that our models can outperform classical ones in terms of different evaluation metrices. Finally, online A/B testing results showed that our model can improve our old recommendation system over various online performance indicators.

Abstract Point cloud data is one of the most common types of input for geometric processing applications. In this paper, we study the point cloud density adaptation problem that underlies many pre‐processing tasks of points data. Specifically, given a (sparse) set of points Q sampling an unknown surface and a target density function, the goal is to adapt Q to match the target distribution. We propose a simple and robust framework that is effective at achieving both local uniformity and precise global density distribution control. Our approach relies on the Gaussian‐weighted graph Laplacian and works purely in the points setting. While it is well known that graph Laplacian is related to mean‐curvature flow and thus has denoising ability, our algorithm uses certain information encoded in the graph Laplacian that is orthogonal to the mean‐curvature flow. Furthermore, by leveraging the natural scale parameter contained in the Gaussian kernel and combining it with a simulated annealing idea, our algorithm moves points in a multi‐scale manner. The resulting algorithm relies much less on the input points to have a good initial distribution (neither uniform nor close to the target density distribution) than many previous refinement‐based methods. We demonstrate the simplicity and effectiveness of our algorithm with point clouds sampled from different underlying surfaces with various geometric and topological properties.

Creating floor plans for large areas via manual surveying is labor-intensive and error-prone. In this paper, we present a system, Walk&Sketch, that creates floor plans of an indoor environment by a person walking through the environment at a normal strolling pace and taking videos using a consumer RGB-D camera. The method computes floor maps represented by polylines from a 3D point cloud based on precise frame-to-frame alignment. It aligns a reference frame with the floor and computes the frame-to-frame offsets from the continuous RGB-D input. Line segments at a certain height are extracted from the 3D point cloud, and are merged to form a polyline map, which can be further modified and annotated by users. The explored area is visualized as a sequence of polygons, providing users with the information on coverage. Experiments have done in various areas of an office building and have shown encouraging results.

Abstract Although understanding of shape features in the context of shape matching and retrieval has made considerable progress in recent years, the case for partial and incomplete models in presence of pose variations still begs a robust and efficient solution. A signature that encodes features at multi‐scales in a pose invariant manner is more appropriate for this case. The Heat Kernel Signature function from spectral theory exhibits this multi‐scale property. We show how this concept can be merged with the persistent homology to design a novel efficient pose‐oblivious matching algorithm for all models, be they partial, incomplete, or complete. We make the algorithm scalable so that it can handle large data sets. Several test results show the robustness of our approach.

Integration over a domain, such as a Euclidean space or a Riemannian manifold, is a fundamental problem across scientific fields. Many times, the underlying domain is only accessible through a discrete approximation, such as a set of points sampled from it, and it is crucial to be able to estimate integral in such discrete settings. In this paper, we study the problem of estimating the integral of a function defined over a k-submanifold embedded in $d$-dimensional space, from its function values at a set of sample points. Previously, such estimation is usually obtained in a statistical setting, where input data is typically assumed to be drawn from certain probabilistic distribution. Our paper is the first to consider this important problem of estimating integral from point clouds data (PCD) under the more general non-statistical setting, and provide certain theoretical guarantees. Our approaches consider the problem from a geometric point of view. Specifically, we estimate the integral by computing a weighted sum, and propose two weighting schemes: the Voronoi and the Principle Eigenvector schemes. The running time of both methods depends mostly on the intrinsic dimension of the underlying manifold, instead of on the ambient dimensions. We show that the estimation based on the Voronoi scheme converges to the true integral under the so-called (ε, δ)-sampling condition with explicit error bound presented. This is the first result of this sort for estimating integral from general PCD. For the Principle Eigenvector scheme, although no theoretical guarantee is established, we present its connection to the Heat diffusion operator, and illustrate justifications behind its construction. Experimental results show that both new methods consistently produce more accurate integral estimations than common statistical methods under various sampling conditions.

Abstract The gradient of a function defined on a manifold is perhaps one of the most important differential objects in data analysis. Most often in practice, the input function is available only at discrete points sampled from the underlying manifold, and the manifold is approximated by either a mesh or simply a point cloud. While many methods exist for computing gradients of a function defined over a mesh, computing and simplifying gradients and related quantities such as critical points, of a function from a point cloud is non‐trivial. In this paper, we initiate the investigation of computing gradients under a different metric on the manifold from the original natural metric induced from the ambient space. Specifically, we map the input manifold to the eigenspace spanned by its Laplacian eigenfunctions, and consider the so‐called diffusion distance metric associated with it. We show the relation of gradient under this metric with that under the original metric. It turns out that once the Laplace operator is constructed, it is easier to approximate gradients in the eigenspace for discrete inputs (especially point clouds) and it is robust to noises in the input function and in the underlying manifold. More importantly, we can easily smooth the gradient field at different scales within this eigenspace framework. We demonstrate the use of our new eigen‐gradients with two applications: approximating / simplifying the critical points of a function, and the Jacobi sets of two input functions (which describe the correlation between these two functions), from point clouds data.

An integrated framework mainly focusing on stereo matching has been presented in this paper to obtain dense depth maps for a mobile robot that is equipped with a novel omnidirectional stereo vision sensor that is designed to obtain height information. The vision sensor is composed of a common perspective camera and two hyperbolic mirrors, which are separately fixed inside a glass cylinder. As the separation between the two mirrors provides much enlarged baseline, the precision of the system has improved correspondingly. Nevertheless, the large disparity space and image particularities that are different from general stereo vision system result in poor performance using common methods. To satisfy the reliability requirement by mobile robot navigation, we use improved graph cuts method, in which more appropriate three-variable smootheness model is proposed for general priors corresponding to more reasonable piecewise smoothness assumption since the well-known swap move algorithm can be applied to a wider class of functions. We also show the necessary modification to handle panoramic images, including deformed matching template, adaptable template scale. Experiment shows that this proposed vision system is feasible as a practical stereo sensor for accurate depth map generation.