As global e-commerce platforms continue to expand, companies are entering new markets where they encounter cold-start challenges due to limited human labels and user behaviors. In this paper, we share our experiences in Coupang to provide a competitive cold-start performance of relevance matching for emerging e-commerce markets. Specifically, we present a Cold-Start Relevance Matching (CSRM) framework, utilizing a multilingual Large Language Model (LLM) to address three challenges: (1) activating cross-lingual transfer learning abilities of LLMs through machine translation tasks; (2) enhancing query understanding and incorporating e-commerce knowledge by retrieval-based query augmentation; (3) mitigating the impact of training label errors through a multi-round self-distillation training strategy. Our experiments demonstrate the effectiveness of CSRM-LLM and the proposed techniques, resulting in successful real-world deployment and significant online gains, with a 45.8% reduction in defect ratio and a 0.866% uplift in session purchase rate.

When applying a stochastic algorithm, one must choose an order to draw samples. The practical choices are without-replacement sampling orders, which are empirically faster and more cache-friendly than uniform-iid-sampling but often have inferior theoretical guarantees. Without-replacement sampling is well understood only for SGD without variance reduction. In this paper, we will improve the convergence analysis and rates of variance reduction under without-replacement sampling orders for composite finite-sum minimization.Our results are in two-folds. First, we develop a damped variant of Finito called Prox-DFinito and establish its convergence rates with random reshuffling, cyclic sampling, and shuffling-once, under both generally and strongly convex scenarios. These rates match full-batch gradient descent and are state-of-the-art compared to the existing results for without-replacement sampling with variance-reduction. Second, our analysis can gauge how the cyclic order will influence the rate of cyclic sampling and, thus, allows us to derive the optimal fixed ordering. In the highly data-heterogeneous scenario, Prox-DFinito with optimal cyclic sampling can attain a sample-size-independent convergence rate, which, to our knowledge, is the first result that can match with uniform-iid-sampling with variance reduction. We also propose a practical method to discover the optimal cyclic ordering numerically.

Decentralized computing is widely used for multiagent systems since it works without a central computing node. In this paper, we develop a first-order algorithm for decentralized consensus optimization that is more energy efficient than the current state-of-the-art. Our algorithm is suitable for application scenarios such as networks of wireless sensors and Internet of Things, where some agents have limited (battery) energy. We call our algorithm walk proximal gradient (WPG), which passes a token through a walk (a succession of nodes) in the graph. The agents that are visited during the walk compute the gradients of their private functions and update the token. We analyze WPG where the walk is the repetition of a Hamiltonian cycle and show that the token converges to the consensual solution faster (in terms of energy consumption) than existing gradient-based decentralized methods. We also generalize the analysis to the non-Hamiltonian graphs. Numerical experiments are presented to validate the energy efficiency of our algorithm.

In a change-point detection problem, a sequence of signals switches from one distribution to another at an unknown time step, and the goal is to quickly and reliably detect this change. By providing new insight into signal processing and data analysis, graph signal processing promises various applications including image processing and sensor network analysis, and becomes an emerging field of research. In this work, we formulate the problem of change-point detection on graph. Under the reasonable assumption of normality, we propose a CUSUM-based algorithm for change-point detection with an arbitrary, unknown and perhaps time-varying mean shift after the change-point. We further propose a decentralized, distributed algorithm, which requires no fusion center, to reduce computational complexity, as well as costs and delays of communication. Numerical results on both synthetic and real-world data demonstrate that our algorithms are efficient and accurate.

Sampling and reconstruction is a fundamentally important problem in the field of graph signal processing. Many works have been contributed to reconstructing bandlimited signals from measurements taken on a known subset of vertices. However, in some cases, the vertex defects occur randomly over the graph. In such situation, the existing graph signal reconstruction methods fail to deal with such blind reconstruction problem. In this paper, we formulate the blind reconstruction problem as Mixed-Integer Nonlinear Programming, and propose a Joint Detection and Reconstruction (JDR) method to simultaneously detect the vertices' working states and reconstruct the bandlimited signal. The convergence property of the proposed method is analyzed. In the experimental part, both synthetic dataset and real-world dataset are applied to verify the proposed methods.

Sparse subspace clustering (SSC) is a technique to partition unlabeled samples according to the subspaces they locate in. With the rapid increase of data amount, efficiently downsampling a big dataset, while at the same time keeping the structure of subspaces, becomes an important topic for SSC. In order to reduce the computational cost while preserving clustering accuracy, a new approach of SSC with downsampling (SSCD) is proposed in this paper. In SSCD, the numbers of samples located in respective subspaces are estimated utilizing the ℓ1norm of the sparse representation. Then a downsampling strategy is designed to decimate samples with the probabilities that are in reverse ratio to the amounts of samples in respective subspaces. As a consequence, the samples in different subspaces are expected to be balanced after the downsampling operation. Theoretical analysis proves the correctness of the proposed strategy. Numerical simulations also verify the efficiency of SSCD.