The rapidly evolving landscape of products, surfaces, policies, and regulations poses significant challenges for deploying state-of-the-art recommendation models at industry scale, primarily due to data fragmentation across domains and escalating infrastructure costs that hinder sustained quality improvements.

Recent advances in recommendation scaling laws have led to foundation models of unprecedented complexity. While these models offer superior performance, their computational demands make real-time serving impractical, often forcing practitioners to rely on knowledge distillation—compromising serving quality for efficiency. To address this challenge, we present SOLARIS (Speculative Offloading of Latent-bAsed Representation for Inference Scaling), a novel framework inspired by speculative decoding. SOLARIS proactively precomputes user-item interaction embeddings by predicting which user-item pairs are likely to appear in future requests, and asynchronously generating their foundation model representations ahead of time. This approach decouples the costly foundation model inference from the latency-critical serving path, enabling real-time knowledge transfer from models previously considered too expensive for online use. Deployed across Meta's advertising system serving billions of daily requests, SOLARIS achieves 0.67% revenue-driving top-line metrics gain, demonstrating its effectiveness at scale.

The determinantal point process (DPP) is an elegant probabilistic model of repulsion with applications in various machine learning tasks including summarization and search. However, the maximum a posteriori (MAP) inference for DPP which plays an important role in many applications is NP-hard, and even the popular greedy algorithm can still be too computationally expensive to be used in large-scale real-time scenarios. To overcome the computational challenge, in this paper, we propose a novel algorithm to greatly accelerate the greedy MAP inference for DPP. In addition, our algorithm also adapts to scenarios where the repulsion is only required among nearby few items in the result sequence. We apply the proposed algorithm to generate relevant and diverse recommendations. Experimental results show that our proposed algorithm is significantly faster than state-of-the-art competitors, and provides a better relevance-diversity trade-off on several public datasets, which is also confirmed in an online A/B test.

Online Video Streaming services such as Hulu hosts tens of millions of premium videos, which requires an effective recommendation system to help viewers discover what they enjoy. In this talk, we will introduce Hulu's recent technical progresses in recommender systems and deep-dive into the topic of generating recommendation reason from knowledge graph. We have two user scenarios: the store-shelf and autoplay. The first requires a list of videos to maximize the chance that a viewer would pick one of them to watch. The second requires a sequence of video recommendations such that the viewer would continuously watch within the current session.

Square-root least absolute shrinkage and selection operator (Lasso), a variant of Lasso, has recently been proposed with a key advantage that the optimal regularization parameter is independent of the noise level in the measurements. In this letter, we introduce a class of nonconvex sparsity-inducing penalties to the square-root Lasso to achieve better sparse recovery performance over the convex counterpart. The resultant formulation is converted to a nonconvex but multiconvex optimization problem, i.e., it is convex in each block of variables. Alternating direction method of multipliers is applied as the solver, according to which two efficient algorithms are devised for row-orthonormal sensing matrix and general sensing matrix, respectively. Numerical experiments are conducted to evaluate the performance of the proposed methods.

In solving the problem of sparse recovery, non-convex techniques have been paid much more attention than ever before, among which the most widely used one is ℓpminimization with p ∈ (0, 1). It has been shown that the global optimality of ℓpminimization is guaranteed under weaker conditions than convex ℓ1minimization, but little interest is shown in the local optimality, which is also significant since practical non-convex approaches can only get local optimums. In this work, we derive a tight condition in guaranteeing the local optimality of ℓpminimization. For practical purposes, we study the performance of an approximated version of ℓpminimization, and show that its global optimality is equivalent to that of ℓpminimization when the penalty approaches the ℓp“norm”. Simulations are implemented to show the recovery performance of the approximated optimization in sparse recovery.

Sparse signal recovery in the static case has been well studied under the framework of Compressive Sensing (CS), while in recent years more attention has also been paid to the dynamic case. In this paper, enlightened by the idea of modified-CS with partially known support, and based on a non-convex optimization approach, we propose the dynamic zero-point attracting projection (DZAP) algorithm to efficiently recover the slowly time-varying sparse signals. Benefiting from the temporal correlation within signal structures, plus an effective prediction method of the future signal based on previous recoveries incorporated, DZAP achieves high-precision recovery with less measurements or larger sparsity level, which is demonstrated by simulations on both synthetic and real data, accompanied by the comparison with other state-of-the-art reference algorithms.

The literature on sparse recovery often adopts the lp “norm” ( p ∈ [0,1]) as the penalty to induce sparsity of the signal satisfying an underdetermined linear system. The performance of the corresponding lp minimization problem can be characterized by its null space constant. In spite of the NP-hardness of computing the constant, its properties can still help in illustrating the performance of lp minimization. In this letter, we show the strict increase of the null space constant in the sparsity level k and its continuity in the exponent p. We also indicate that the constant is strictly increasing in p with probability 1 when the sensing matrix A is randomly generated. Finally, we show how these properties can help in demonstrating the performance of lp minimization, mainly in the relationship between the the exponent p and the sparsity level k.

Existing literatures suggest that sparsity is more likely to be induced with non-convex penalties, but the corresponding algorithms usually suffer from multiple local minima. In this paper, we introduce a class of sparsity-inducing penalties and provide the convergence guarantees of a non-convex approach for sparse recovery using regularized least squares. Theoretical analysis demonstrates that under some certain conditions, if the non-convexity of the penalty is below a threshold (which is in inverse proportion to the distance between the initialization and the sparse signal), the sparse signal can be stably recovered. Numerical simulations are implemented to verify the theoretical results in this paper and to compare the performance of this approach with other references.

Compressive sensing (CS) is a data acquisition technique that measures sparse or compressible signals at a sampling rate lower than their Nyquist rate. Results show that sparse signals can be reconstructed using greedy algorithms, often requiring prior knowledge such as the signal sparsity or the noise level. As a substitute to prior knowledge, cross validation (CV), a statistical method that examines whether a model overfits its data, has been proposed to determine the stopping condition of greedy algorithms. This paper analyses cross validation in a general compressive sensing framework. Furthermore, we provide both theoretical analysis and numerical simulations for a cross-validation modification of orthogonal matching pursuit, referred to as OMP-CV, which has good performance in sparse recovery.

This paper aims to provide theoretical guarantees via non-convex optimization for sparse recovery. It is shown that the sparse signal is the unique local optimal solution within a neighborhood, which contains the least squares solution if the sparsity-inducing penalties are not too non-convex. The idea of projected subgradient method is generalized to solve this non-convex optimization problem. A uniform approximate projection is applied in the projection step to make the algorithm more computationally tractable. The theoretical convergence analysis of the proposed method, approximate projected generalized gradient (APGG), is performed in the noisy scenario. The result reveals that if the non-convexity of the penalties is under a threshold, the bound of the recovery error is linear in both the noise bound and the step size. Numerical simulations are performed to test the performance of APGG and verify its theoretical analysis.

The idea of backtracking has been incorporated into the matching pursuit algorithms in sparse recovery, for example, subspace pursuit (SP) and compressive sampling matching pursuit (CoSaMP), to improve the recovery performance. In each iteration, a supplement set of size K or 2K is added to the candidate set to re-evaluate their reliability and then discard the unreliable indices, where K is the sparsity level of the original sparse signal. Yet the optimal choice of the size of the supplement set is still unclear. This paper aims to provide comprehensive analysis on the optimal choice of the size. The optimality is twofold: performance guarantees and computational complexity. By two theorems,we provide theoretical guarantees for the supplement set of arbitrary size, and computational complexity needed for perfect recovery. Numerical simulations demonstrate that a moderate size, such as 0.25K, results in computational efficiency without loss of recovery quality.

In this paper, we consider orthogonal matching pursuit (OMP) algorithm for multiple measurement vectors (MMV) problem. The robustness of OMPMMV is studied under general perturbations-when the measurement vectors as well as the sensing matrix are incorporated with additive noise. The main result shows that although exact recovery of the sparse solutions is unrealistic in noisy scenario, recovery of the support set of the solutions is guaranteed under suitable conditions. Specifically, a sufficient condition is derived that guarantees exact recovery of the sparse solutions in noiseless scenario.

The Modulated Wideband Converter (MWC) is a recently proposed analog-to-digital converter (ADC) based on Compressive Sensing (CS) theory. Unlike conventional ADCs, its quantization reference voltage, which is important to the system performance, does not equal the maximum amplitude of original analog signal. In this paper, the quantization reference voltage of the MWC is theoretically analyzed and the conclusion demonstrates that the reference voltage is proportional to the square root of q, which is a trade-off parameter between sampling rate and number of channels. Further discussions and simulation results show that the reference voltage is proportional to the square root of Nq when the signal consists of N narrowband signals.