arXiv subjects

Xiao Yan

Publications and source records attributed to Xiao Yan.

At least 73 records · Page 4Linked to original sources

arXiv2023

SPT: Fine-Tuning Transformer-based Language Models Efficiently with Sparsification

Transformer-based large language models (e.g., BERT and GPT) achieve great success, and fine-tuning, which tunes a pre-trained model on a task-specific dataset, is the standard practice to utilize these models for downstream tasks. However, Transformer fine-tuning has long running time and high memory consumption due to the large size of the models. We propose the SPT system to fine-tune Transformer-based models efficiently by introducing sparsity. We observe that the memory consumption of Transformer mainly comes from storing attention weights for multi-head attention (MHA), and the majority of running time is spent on feed-forward network (FFN). Thus, we design the sparse MHA module, which computes and stores only large attention weights to reduce memory consumption, and the routed FFN module, which dynamically activates a subset of model parameters for each token to reduce computation cost. We implement SPT on PyTorch and customize CUDA kernels to run sparse MHA and routed FFN efficiently. Specifically, we use product quantization to identify the large attention weights and compute attention via sparse matrix multiplication for sparse MHA. For routed FFN, we batch the tokens according to their activated model parameters for efficient computation. We conduct extensive experiments to evaluate SPT on various model configurations. The results show that SPT consistently outperforms well-optimized baselines, reducing the peak memory consumption by up to 50% and accelerating fine-tuning by up to 2.2x.

cs.DC↗

arXiv2023

Let AI Entertain You: Increasing User Engagement with Generative AI and Rejection Sampling

While generative AI excels in content generation, it does not always increase user engagement. This can be attributed to two main factors. First, generative AI generates content without incorporating explicit or implicit feedback about user interactions. Even if the generated content seems to be more informative or well-written, it does not necessarily lead to an increase in user activities, such as clicks. Second, there is a concern with the quality of the content generative AI produces, which often lacks the distinctiveness and authenticity that human-created content possesses. These two factors can lead to content that fails to meet specific needs and preferences of users, ultimately reducing its potential to be engaging. This paper presents a generic framework of how to improve user engagement with generative AI by leveraging user feedback. Our solutions employ rejection sampling, a technique used in reinforcement learning, to boost engagement metrics. We leveraged the framework in the context of email notification subject lines generation for an online social network, and achieved significant engagement metric lift including +1% Session and +0.4% Weekly Active Users. We believe our work offers a universal framework that enhances user engagement with generative AI, particularly when standard generative AI reaches its limits in terms of enhancing content to be more captivating. To the best of our knowledge, this represents an early milestone in the industry's successful use of generative AI to enhance user engagement.

cs.HC↗

arXiv2023

Generative and Contrastive Paradigms Are Complementary for Graph Self-Supervised Learning

For graph self-supervised learning (GSSL), masked autoencoder (MAE) follows the generative paradigm and learns to reconstruct masked graph edges or node features. Contrastive Learning (CL) maximizes the similarity between augmented views of the same graph and is widely used for GSSL. However, MAE and CL are considered separately in existing works for GSSL. We observe that the MAE and CL paradigms are complementary and propose the graph contrastive masked autoencoder (GCMAE) framework to unify them. Specifically, by focusing on local edges or node features, MAE cannot capture global information of the graph and is sensitive to particular edges and features. On the contrary, CL excels in extracting global information because it considers the relation between graphs. As such, we equip GCMAE with an MAE branch and a CL branch, and the two branches share a common encoder, which allows the MAE branch to exploit the global information extracted by the CL branch. To force GCMAE to capture global graph structures, we train it to reconstruct the entire adjacency matrix instead of only the masked edges as in existing works. Moreover, a discrimination loss is proposed for feature reconstruction, which improves the disparity between node embeddings rather than reducing the reconstruction error to tackle the feature smoothing problem of MAE. We evaluate GCMAE on four popular graph tasks (i.e., node classification, node clustering, link prediction, and graph classification) and compare with 14 state-of-the-art baselines. The results show that GCMAE consistently provides good accuracy across these tasks, and the maximum accuracy improvement is up to 3.2% compared with the best-performing baseline.

cs.LG↗

arXiv2023

Multi-domain Recommendation with Embedding Disentangling and Domain Alignment

Multi-domain recommendation (MDR) aims to provide recommendations for different domains (e.g., types of products) with overlapping users/items and is common for platforms such as Amazon, Facebook, and LinkedIn that host multiple services. Existing MDR models face two challenges: First, it is difficult to disentangle knowledge that generalizes across domains (e.g., a user likes cheap items) and knowledge specific to a single domain (e.g., a user likes blue clothing but not blue cars). Second, they have limited ability to transfer knowledge across domains with small overlaps. We propose a new MDR method named EDDA with two key components, i.e., embedding disentangling recommender and domain alignment, to tackle the two challenges respectively. In particular, the embedding disentangling recommender separates both the model and embedding for the inter-domain part and the intra-domain part, while most existing MDR methods only focus on model-level disentangling. The domain alignment leverages random walks from graph processing to identify similar user/item pairs from different domains and encourages similar user/item pairs to have similar embeddings, enhancing knowledge transfer. We compare EDDA with 12 state-of-the-art baselines on 3 real datasets. The results show that EDDA consistently outperforms the baselines on all datasets and domains. All datasets and codes are available at https://github.com/Stevenn9981/EDDA.

cs.IR↗

arXiv2023

Effective and Efficient PageRank-based Positioning for Graph Visualization

Graph visualization is a vital component in many real-world applications (e.g., social network analysis, web mining, and bioinformatics) that enables users to unearth crucial insights from complex data. Lying in the core of graph visualization is the node distance measure, which determines how the nodes are placed on the screen. A favorable node distance measure should be informative in reflecting the full structural information between nodes and effective in optimizing visual aesthetics. However, existing node distance measures yield sub-par visualization quality as they fall short of these requirements. Moreover, most existing measures are computationally inefficient, incurring a long response time when visualizing large graphs. To overcome such deficiencies, we propose a new node distance measure, PDist, geared towards graph visualization by exploiting a well-known node proximity measure, personalized PageRank. Moreover, we propose an efficient algorithm Tau-Push for estimating PDist under both single- and multi-level visualization settings. With several carefully-designed techniques, Tau-Push offers non-trivial theoretical guarantees for estimation accuracy and computation complexity. Extensive experiments show that our proposal significantly outperforms 13 state-of-the-art graph visualization solutions on 12 real-world graphs in terms of both efficiency and effectiveness (including aesthetic criteria and user feedback). In particular, our proposal can interactively produce satisfactory visualizations within one second for billion-edge graphs.

cs.SI↗

arXiv2022

DGI: Easy and Efficient Inference for GNNs

While many systems have been developed to train Graph Neural Networks (GNNs), efficient model inference and evaluation remain to be addressed. For instance, using the widely adopted node-wise approach, model evaluation can account for up to 94% of the time in the end-to-end training process due to neighbor explosion, which means that a node accesses its multi-hop neighbors. On the other hand, layer-wise inference avoids the neighbor explosion problem by conducting inference layer by layer such that the nodes only need their one-hop neighbors in each layer. However, implementing layer-wise inference requires substantial engineering efforts because users need to manually decompose a GNN model into layers for computation and split workload into batches to fit into device memory. In this paper, we develop Deep Graph Inference (DGI) -- a system for easy and efficient GNN model inference, which automatically translates the training code of a GNN model for layer-wise execution. DGI is general for various GNN models and different kinds of inference requests, and supports out-of-core execution on large graphs that cannot fit in CPU memory. Experimental results show that DGI consistently outperforms layer-wise inference across different datasets and hardware settings, and the speedup can be over 1,000x.

cs.LG↗

arXiv2022

Supercooled Droplet Icing and Self-Jumping on Micro/nanostructured Surfaces: Role of Vaporization Momentum

Phase change under reduced environmental pressures is key to understanding liquid discharge and propulsion processes for aerospace applications. A representative case is the sessile water droplets exposed to high vacuum, which experience complex phase change and transport phenomena that behave so differently than that under the atmosphere. Here, we demonstrate a previously unexplored aspect of the mechanism governing icing droplet self-launching from superhydrophobic surfaces when exposed to low pressures (~100 Pa). In contrast to the previously reported recalescence-induced local overpressure underneath the droplet that propels icing droplet self-jumping, we show that the progressive recalescence over the free surface plays a significant role in droplet icing and jumping. The joint contribution of the top-down vaporization momentum and bottom-up local overpressure momentum leads to vaporization-compression-detaching dynamics of the freezing droplets. We delineate the jumping velocity of the icing droplet by analyzing droplet vaporization mediated by freezing and substrate structuring, and reveal jumping direction coupled with the spatially probabilistic ice nucleation. Our study provides new insights into phase change of supercooled droplets at extreme conditions seen in aerospace and vacuum industries.

physics.flu-dyn↗

arXiv2022

A high accuracy and low complexity quality control method for image compression

For large-scale still image coding tasks, the processing platform needs to ensure that the coded images meet the quality requirement. Therefore, the quality control algorithms that generate adaptive QP towards a target quality level for image coding are of significant research value. However, the existing quality control methods are limited by low accuracy, excessive computational cost, or temporal information dependence. In this paper, we propose a concise λ domain linear distortion model and an accurate model parameters estimation method based on the original data. Since the model parameters are obtained from the original data, the proposed method is decoupled from the RDO process and can be applied to different image encoders. Experiments show that the proposed quality control algorithm achieves the highest control accuracy and the lowest delay in the literature at the same time. The application of Alibaba's e-commerce platform also shows that the proposed algorithm can significantly reduce the overall bitrate while greatly reducing the bad case ratio.

cs.MM↗

arXiv2022

Automatic Meta-Path Discovery for Effective Graph-Based Recommendation

Heterogeneous Information Networks (HINs) are labeled graphs that depict relationships among different types of entities (e.g., users, movies and directors). For HINs, meta-path-based recommenders (MPRs) utilize meta-paths (i.e., abstract paths consisting of node and link types) to predict user preference, and have attracted a lot of attention due to their explainability and performance. We observe that the performance of MPRs is highly sensitive to the meta-paths they use, but existing works manually select the meta-paths from many possible ones. Thus, to discover effective meta-paths automatically, we propose the Reinforcement learning-based Meta-path Selection (RMS) framework. Specifically, we define a vector encoding for meta-paths and design a policy network to extend meta-paths. The policy network is trained based on the results of downstream recommendation tasks and an early stopping approximation strategy is proposed to speed up training. RMS is a general model, and it can work with all existing MPRs. We also propose a new MPR called RMS-HRec, which uses an attention mechanism to aggregate information from the meta-paths. We conduct extensive experiments on real datasets. Compared with the manually selected meta-paths, the meta-paths identified by RMS consistently improve recommendation quality. Moreover, RMS-HRec outperforms state-of-the-art recommender systems by an average of 7% in hit ratio. The codes and datasets are available on https://github.com/Stevenn9981/RMS-HRec.

cs.IR↗

arXiv2022

Manu: A Cloud Native Vector Database Management System

With the development of learning-based embedding models, embedding vectors are widely used for analyzing and searching unstructured data. As vector collections exceed billion-scale, fully managed and horizontally scalable vector databases are necessary. In the past three years, through interaction with our 1200+ industry users, we have sketched a vision for the features that next-generation vector databases should have, which include long-term evolvability, tunable consistency, good elasticity, and high performance. We present Manu, a cloud native vector database that implements these features. It is difficult to integrate all these features if we follow traditional DBMS design rules. As most vector data applications do not require complex data models and strong data consistency, our design philosophy is to relax the data model and consistency constraints in exchange for the aforementioned features. Specifically, Manu firstly exposes the write-ahead log (WAL) and binlog as backbone services. Secondly, write components are designed as log publishers while all read-only analytic and search components are designed as independent subscribers to the log services. Finally, we utilize multi-version concurrency control (MVCC) and a delta consistency model to simplify the communication and cooperation among the system components. These designs achieve a low coupling among the system components, which is essential for elasticity and evolution. We also extensively optimize Manu for performance and usability with hardware-aware implementations and support for complex search semantics.

cs.DB↗

arXiv2021

Demand Forecasting in Bike-sharing Systems Based on A Multiple Spatiotemporal Fusion Network

Bike-sharing systems (BSSs) have become increasingly popular around the globe and have attracted a wide range of research interests. In this paper, the demand forecasting problem in BSSs is studied. Spatial and temporal features are critical for demand forecasting in BSSs, but it is challenging to extract spatiotemporal dynamics. Another challenge is to capture the relations between spatiotemporal dynamics and external factors, such as weather, day-of-week, and time-of-day. To address these challenges, we propose a multiple spatiotemporal fusion network named MSTF-Net. MSTF-Net consists of multiple spatiotemporal blocks: 3D convolutional network (3D-CNN) blocks, eidetic 3D convolutional long short-term memory networks (E3D-LSTM) blocks, and fully-connected (FC) blocks. Specifically, 3D-CNN blocks highlight extracting short-term spatiotemporal dependence in each fragment (i.e., closeness, period, and trend); E3D-LSTM blocks further extract long-term spatiotemporal dependence over all fragments; FC blocks extract nonlinear correlations of external factors. Finally, the latent representations of E3D-LSTM and FC blocks are fused to obtain the final prediction. For two real-world datasets, it is shown that MSTF-Net outperforms seven state-of-the-art models.

cs.CV↗

arXiv2021

ProSTformer: Pre-trained Progressive Space-Time Self-attention Model for Traffic Flow Forecasting

Traffic flow forecasting is essential and challenging to intelligent city management and public safety. Recent studies have shown the potential of convolution-free Transformer approach to extract the dynamic dependencies among complex influencing factors. However, two issues prevent the approach from being effectively applied in traffic flow forecasting. First, it ignores the spatiotemporal structure of the traffic flow videos. Second, for a long sequence, it is hard to focus on crucial attention due to the quadratic times dot-product computation. To address the two issues, we first factorize the dependencies and then design a progressive space-time self-attention mechanism named ProSTformer. It has two distinctive characteristics: (1) corresponding to the factorization, the self-attention mechanism progressively focuses on spatial dependence from local to global regions, on temporal dependence from inside to outside fragment (i.e., closeness, period, and trend), and finally on external dependence such as weather, temperature, and day-of-week; (2) by incorporating the spatiotemporal structure into the self-attention mechanism, each block in ProSTformer highlights the unique dependence by aggregating the regions with spatiotemporal positions to significantly decrease the computation. We evaluate ProSTformer on two traffic datasets, and each dataset includes three separate datasets with big, medium, and small scales. Despite the radically different design compared to the convolutional architectures for traffic flow forecasting, ProSTformer performs better or the same on the big scale datasets than six state-of-the-art baseline methods by RMSE. When pre-trained on the big scale datasets and transferred to the medium and small scale datasets, ProSTformer achieves a significant enhancement and behaves best.

physics.soc-ph↗

arXiv2021

CSMCNet: Scalable Video Compressive Sensing Reconstruction with Interpretable Motion Estimation

Most deep network methods for compressive sensing reconstruction suffer from the black-box characteristic of DNN. In this paper, a deep neural network with interpretable motion estimation named CSMCNet is proposed. The network is able to realize high-quality reconstruction of video compressive sensing by unfolding the iterative steps of optimization based algorithms. A DNN based, multi-hypothesis motion estimation module is designed to improve the reconstruction quality, and a residual module is employed to further narrow down the gap between re-construction results and original signal in our proposed method. Besides, we propose an interpolation module with corresponding training strategy to realize scalable CS reconstruction, which is capable of using the same model to decode various compression ratios. Experiments show that a PSNR of 29.34dB can be achieved at 2% CS ratio (compressed by 98%), which is superior than other state-of-the-art methods. Moreover, the interpolation module is proved to be effective, with significant cost saving and acceptable performance losses.

eess.IV↗

arXiv2021

Combating Ambiguity for Hash-code Learning in Medical Instance Retrieval

When encountering a dubious diagnostic case, medical instance retrieval can help radiologists make evidence-based diagnoses by finding images containing instances similar to a query case from a large image database. The similarity between the query case and retrieved similar cases is determined by visual features extracted from pathologically abnormal regions. However, the manifestation of these regions often lacks specificity, i.e., different diseases can have the same manifestation, and different manifestations may occur at different stages of the same disease. To combat the manifestation ambiguity in medical instance retrieval, we propose a novel deep framework called Y-Net, encoding images into compact hash-codes generated from convolutional features by feature aggregation. Y-Net can learn highly discriminative convolutional features by unifying the pixel-wise segmentation loss and classification loss. The segmentation loss allows exploring subtle spatial differences for good spatial-discriminability while the classification loss utilizes class-aware semantic information for good semantic-separability. As a result, Y-Net can enhance the visual features in pathologically abnormal regions and suppress the disturbing of the background during model training, which could effectively embed discriminative features into the hash-codes in the retrieval stage. Extensive experiments on two medical image datasets demonstrate that Y-Net can alleviate the ambiguity of pathologically abnormal regions and its retrieval performance outperforms the state-of-the-art method by an average of 9.27\% on the returned list of 10.

cs.IR↗

arXiv2021

A Power and Area Efficient Lepton Hardware Encoder with Hash-based Memory Optimization

Although it has been surpassed by many subsequent coding standards, JPEG occupies a large share of the storage load of the current data hosting service. To reduce the storage costs, DropBox proposed a lossless secondary compression algorithm, Lepton, to further improve the compression rate of JPEG images. However, the bloated probability models defined by Lepton severely restrict its throughput and energy efficiency. To solve this problem, we construct an efficient access probability-based hash function for the probability models, and then propose a hardware-friendly memory optimization method by combining the proposed hash function and the N-way Set-Associative unit. After that, we design a highly parameterized hardware structure for the probability models and finally implement a power and area efficient Lepton hardware encoder. To the best of our knowledge, this is the first hardware implementation of Lepton. The synthesis result shows that the proposed hardware structure reduces the total area of the probability models by 70.97%. Compared with DropBox's software solution, the throughput and the energy efficiency of the proposed Lepton hardware encoder are increased by 55.25 and 4899 times respectively. In terms of manufacturing cost, the proposed Lepton hardware encoder is also significantly lower than the general-purpose CPU used by DropBox.

cs.MM↗

arXiv2021

Self-Enhanced GNN: Improving Graph Neural Networks Using Model Outputs

Graph neural networks (GNNs) have received much attention recently because of their excellent performance on graph-based tasks. However, existing research on GNNs focuses on designing more effective models without considering much about the quality of the input data. In this paper, we propose self-enhanced GNN (SEG), which improves the quality of the input data using the outputs of existing GNN models for better performance on semi-supervised node classification. As graph data consist of both topology and node labels, we improve input data quality from both perspectives. For topology, we observe that higher classification accuracy can be achieved when the ratio of inter-class edges (connecting nodes from different classes) is low and propose topology update to remove inter-class edges and add intra-class edges. For node labels, we propose training node augmentation, which enlarges the training set using the labels predicted by existing GNN models. SEG is a general framework that can be easily combined with existing GNN models. Experimental results validate that SEG consistently improves the performance of well-known GNN models such as GCN, GAT and SGC across different datasets.

cs.LG↗

arXiv2020

The item selection problem for user cold-start recommendation

When a new user just signs up on a website, we usually have no information about him/her, i.e. no interaction with items, no user profile and no social links with other users. Under such circumstances, we still expect our recommender systems could attract the users at the first time so that the users decide to stay on the website and become active users. This problem falls into new user cold-start category and it is crucial to the development and even survival of a company. Existing works on user cold-start recommendation either require additional user efforts, e.g. setting up an interview process, or make use of side information [10] such as user demographics, locations, social relations, etc. However, users may not be willing to take the interview and side information on cold-start users is usually not available. Therefore, we consider a pure cold-start scenario where neither interaction nor side information is available and no user effort is required. Studying this setting is also important for the initialization of other cold-start solutions, such as initializing the first few questions of an interview.

cs.IR↗

arXiv2020

CS-MCNet:A Video Compressive Sensing Reconstruction Network with Interpretable Motion Compensation

In this paper, a deep neural network with interpretable motion compensation called CS-MCNet is proposed to realize high-quality and real-time decoding of video compressive sensing. Firstly, explicit multi-hypothesis motion compensation is applied in our network to extract correlation information of adjacent frames(as shown in Fig. 1), which improves the recover performance. And then, a residual module further narrows down the gap between reconstruction result and original signal. The overall architecture is interpretable by using algorithm unrolling, which brings the benefits of being able to transfer prior knowledge about the conventional algorithms. As a result, a PSNR of 22dB can be achieved at 64x compression ratio, which is about 4% to 9% better than state-of-the-art methods. In addition, due to the feed-forward architecture, the reconstruction can be processed by our network in real time and up to three orders of magnitude faster than traditional iterative methods.

eess.IV↗