The rapid growth of model parameters presents a significant challenge when deploying large generative models on GPU. Existing LLM runtime memory management solutions tend to maximize batch size to saturate GPU device utilization. Nevertheless, this practice leads to situations where the KV Cache of certain sequences cannot be accommodated on GPUs with limited memory capacity during the model inference, requiring temporary eviction from GPU memory (referred to as KV Cache spillover). However, without careful consideration of the LLM inference's runtime pattern, current LLM inference memory management solutions face issues like one-size-fits-all spillover handling approach for different platforms, under-utilization of GPU in prefill stage, and suboptimal sequence selection due to direct employment of swap or recomputation. In this paper, we introduce FuseSpill, a holistic KV Cache management solution designed to boost LLM inference on memory-constrained GPU by efficiently handling KV Cache spillover. Specifically, FuseSpill consists of a spillover cost model that analyzes the system cost of spillover handling techniques quantitatively, a KV cache swap orchestrator to further refine the basic swap technique to sophisticated disaggregate KV Cache across heterogeneous devices for decoding iterations, a multi-executor scheduler to effectively coordinate task executors across devices, and a response length predictor to exploit the length-aware sequence selection strategy when KV Cache spillover occurs. The experimental results demonstrate that our implementation outperforms existing solutions, delivering a 20% to 40% increase in throughput while simultaneously reducing the inference latency of the spillover sequences.

Stream Window Join (SWJ), a vital operation in stream analytics, struggles with achieving a balance between accuracy and latency due to out-of-order data arrivals. Existing methods predominantly rely on adaptive buffering, but often fall short in performance, thereby constraining practical applications. We introduce PECJ, a solution that proactively incorporates unobserved data to enhance accuracy while reducing latency, thus requiring robust predictive modeling of stream oscillation. At the heart of PECJ lies a mathematical formulation of the posterior distribution approximation (PDA) problem using variational inference (VI). This approach circumvents error propagation while meeting the low-latency demands of SWJ. We detail the implementation of PECJ, striking a balance between complexity and generality, and discuss both analytical and learning-based approaches. Experimental evaluations reveal PECJ's superior performance. The successful integration of PECJ into a multi-threaded SWJ benchmark testbed further establishes its practical value, demonstrating promising advancements in enhancing data stream processing capabilities amidst out-of-order data.

As the use of online AI inference services rapidly expands in various applications (e.g., fraud detection in banking, product recommendation in e-commerce), real-time feature extraction (RTFE) systems have been developed to compute the requested features from incoming data tuples in ultra-low latency. Similar to relational databases, these RTFE procedures can be expressed using SQL-like languages. However, there is a lack of research on the workload characteristics and specialized benchmarks for RTFE, especially in comparison with existing database workloads and benchmarks (e.g., concurrent transactions in TPC-C). In this paper, we study the RTFE workload characteristics using over one hundred real datasets from open repositories (e.g. Kaggle, Tianchi, UCI ML, KiltHub) and those from 4Paradigm. The study highlights the significant differences between RTFE workloads and existing database benchmarks in terms of application scenarios, operator distributions, and query structures. Based on these findings, we propose to develop a realtime feature extraction benchmark named FEBench based on the four important criteria for a domain-specific benchmark proposed by Jim Gray. FEBench consists of selected representative datasets, query templates, and an online request simulator. We use FEBench to evaluate the effectiveness of feature extraction systems including OpenMLDB and Flink and find that each system exhibits distinct advantages and limitations in terms of overall latency, tail latency, and concurrency performance.

Federated learning (FL), which enables collaborative learning without revealing raw data, is an emerging topic in privacy-preserving machine learning. Based on our experiences in thousands of real-world applications, time-series feature extraction plays a significant role in improving model quality. In this work, we propose a system automatically integrating time series feature extraction for training FL models. Our experiments show that by adopting time series feature extraction, the model accuracy (AUC) is improved by 3% on average, and recall is increased by 10% in recommender systems. We have open-sourced the project https://github.com/4paradigm/tsfe and provided a step by step demonstration on how audiences can use our system to create their own FL pipeline that extracts time series features. Demonstration video at: https://youtu.be/UW27dWT-ays

On-line decision augmentation (OLDA) has been considered as a promising paradigm for real-time decision making powered by Artificial Intelligence (AI). OLDA has been widely used in many applications such as real-time fraud detection, personalized recommendation, etc. On-line inference puts real-time features extracted from multiple time windows through a pre-trained model to evaluate new data to support decision making. Feature extraction is usually the most time-consuming operation in many OLDA data pipelines. In this work, we started by studying how existing in-memory databases can be leveraged to efficiently support such real-time feature extractions. However, we found that existing in-memory databases cost hundreds or even thousands of milliseconds. This is unacceptable for OLDA applications with strict real-time constraints. We therefore propose FEDB ( F eature E ngineering D ata b ase), a distributed in-memory database system designed to efficiently support on-line feature extraction. Our experimental results show that FEDB can be one to two orders of magnitude faster than the state-of-the-art in-memory databases on real-time feature extraction. Furthermore, we explore the use of the Intel Optane DC Persistent Memory Module (PMEM) to make FEDB more cost-effective. When comparing the proposed PMEM-optimized persistent skiplist to the FEDB using DRAM+SSD, PMEM-based FEDB can shorten the tail latency up to 19.7%, reduce the recovery time up to 99.7%, and save up to 58.4% total cost of a real OLDA pipeline.

The fifth‐generation‐enabled Internet of Things (5G‐enabled IoT) has been considered as a key enabler for the automation of almost all industries. In 5G‐enabled IoT, resource‐limited passive devices are expected to join the IoT using the WiFi backscatter communication (WiFi‐BSC) technology. However, WiFi‐BSC deployment is currently limited due to high equipment cost and short transmission range. To address these two drawbacks, in this paper, we propose a low‐cost and long‐range node‐assisted WiFi backscatter communication scheme. In our scheme, a WiFi node can receive backscatter signals using two cheap regular half‐duplex antennas (instead of using expensive full‐duplex technique or collaborating with multiple other nodes), thereby reducing the equipment cost. Besides, WiFi nodes can help relay backscatter signals to remote 5G infrastructure, greatly extending the backscatter’s transmission range. We then develop a theoretical model to analyze the throughput of WiFi‐BSC. Extensive simulations verify the effectiveness of our scheme and the accuracy of our model.

Tabular data is the most common data format adopted by our customers ranging from retail, finance to E-commerce, and tabular data classification plays an essential role to their businesses. In this paper, we present Network On Network (NON), a practical tabular data classification model based on deep neural network to provide accurate predictions. Various deep methods have been proposed and promising progress has been made. However, most of them use operations like neural network and factorization machines to fuse the embeddings of different features directly, and linearly combine the outputs of those operations to get the final prediction. As a result, the intra-field information and the non-linear interactions between those operations (e.g. neural network and factorization machines) are ignored. Intra-field information is the information that features inside each field belong to the same field. NON is proposed to take full advantage of intra-field information and non-linear interactions. It consists of three components: field-wise network at the bottom to capture the intra-field information, across field network in the middle to choose suitable operations data-drivenly, and operation fusion network on the top to fuse outputs of the chosen operations deeply. Extensive experiments on six real-world datasets demonstrate NON can outperform the state-of-the-art models significantly. Furthermore, both qualitative and quantitative study of the features in the embedding space show NON can capture intra-field information effectively.

Automatic machine learning (AutoML) aims at automatically choosing the best configuration for machine learning tasks. However, a configuration evaluation can be very time consuming particularly on learning tasks with large datasets. This limitation usually restrains derivative-free optimization from releasing its full power for a fine configuration search using many evaluations. To alleviate this limitation, in this paper, we propose a derivative-free optimization framework for AutoML using multi-fidelity evaluations. It uses many lowfidelity evaluations on small data subsets and very few highfidelity evaluations on the full dataset. However, the lowfidelity evaluations can be badly biased, and need to be corrected with only a very low cost. We thus propose the Transfer Series Expansion (TSE) that learns the low-fidelity correction predictor efficiently by linearly combining a set of base predictors. The base predictors can be obtained cheaply from down-scaled and experienced tasks. Experimental results on real-world AutoML problems verify that the proposed framework can accelerate derivative-free configuration search significantly by making use of the multi-fidelity evaluations.

To meet the standard of differential privacy, noise is usually added into the original data, which inevitably deteriorates the predicting performance of subsequent learning algorithms. In this paper, motivated by the success of improving predicting performance by ensemble learning, we propose to enhance privacy-preserving logistic regression by stacking. We show that this can be done either by sample-based or feature-based partitioning. However, we prove that when privacy-budgets are the same, feature-based partitioning requires fewer samples than sample-based one, and thus likely has better empirical performance. As transfer learning is difficult to be integrated with a differential privacy guarantee, we further combine the proposed method with hypothesis transfer learning to address the problem of learning across different organizations. Finally, we not only demonstrate the effectiveness of our method on two benchmark data sets, i.e., MNIST and NEWS20, but also apply it into a real application of cross-organizational diabetes prediction from RUIJIN data set, where privacy is of a significant concern.

Feature crossing captures interactions among categorical features and is useful to enhance learning from tabular data in real-world businesses. In this paper, we present AutoCross, an automatic feature crossing tool provided by 4Paradigm to its customers, ranging from banks, hospitals, to Internet corporations. By performing beam search in a tree-structured space, AutoCross enables efficient generation of high-order cross features, which is not yet visited by existing works. Additionally, we propose successive mini-batch gradient descent and multi-granularity discretization to further improve efficiency and effectiveness, while ensuring simplicity so that no machine learning expertise or tedious hyper-parameter tuning is required. Furthermore, the algorithms are designed to reduce the computational, transmitting, and storage costs involved in distributed computing. Experimental results on both benchmark and real-world business datasets demonstrate the effectiveness and efficiency of AutoCross. It is shown that AutoCross can significantly enhance the performance of both linear and deep models.

Taxi-calling apps are gaining increasing popularity for their efficiency in dispatching idle taxis to passengers in need. To precisely balance the supply and the demand of taxis, online taxicab platforms need to predict the Unit Original Taxi Demand (UOTD), which refers to the number of taxi-calling requirements submitted per unit time (e.g., every hour) and per unit region (e.g., each POI). Predicting UOTD is non-trivial for large-scale industrial online taxicab platforms because both accuracy and flexibility are essential. Complex non-linear models such as GBRT and deep learning are generally accurate, yet require labor-intensive model redesign after scenario changes (e.g., extra constraints due to new regulations). To accurately predict UOTD while remaining flexible to scenario changes, we propose LinUOTD, a unified linear regression model with more than 200 million dimensions of features. The simple model structure eliminates the need of repeated model redesign, while the high-dimensional features contribute to accurate UOTD prediction. We further design a series of optimization techniques for efficient model training and updating. Evaluations on two large-scale datasets from an industrial online taxicab platform verify that LinUOTD outperforms popular non-linear models in accuracy. We envision our experiences to adopt simple linear models with high-dimensional features in UOTD prediction as a pilot study and can shed insights upon other industrial large-scale spatio-temporal prediction problems.

Stochastic gradient descent (SGD) is a popular technique for large-scale optimization problems in machine learning. In order to parallelize SGD, minibatch training needs to be employed to reduce the communication cost. However, an increase in minibatch size typically decreases the rate of convergence. This paper introduces a technique based on approximate optimization of a conservatively regularized objective function within each minibatch. We prove that the convergence rate does not decrease with increasing minibatch size. Experiments demonstrate that with suitable implementations of approximate optimization, the resulting algorithm can outperform standard SGD in many scenarios.

Transfer learning as a new machine learning paradigm has gained increasing attention lately. In situations where the training data in a target domain are not sufficient to learn predictive models effectively, transfer learning leverages auxiliary source data from other related source domains for learning. While most of the existing works in this area only focused on using the source data with the same structure as the target data, in this paper, we push this boundary further by proposing a heterogeneous transfer learning framework for knowledge transfer between text and images. We observe that for a target-domain classification problem, some annotated images can be found on many social Web sites, which can serve as a bridge to transfer knowledge from the abundant text documents available over the Web. A key question is how to effectively transfer the knowledge in the source data even though the text can be arbitrarily found. Our solution is to enrich the representation of the target images with semantic concepts extracted from the auxiliary source data through a novel matrix factorization method. By using the latent semantic features generated by the auxiliary data, we are able to build a better integrated image classifier. We empirically demonstrate the effectiveness of our algorithm on the Caltech-256 image dataset.

Advertising in the case of textual Web pages has been studied extensively by many researchers. However, with the increasing amount of multimedia data such as image, audio and video on the Web, the need for recommending advertisement for the multimedia data is becoming a reality. In this paper, we address the novel problem of visual contextual advertising, which is to directly advertise when users are viewing images which do not have any surrounding text. A key challenging issue of visual contextual advertising is that images and advertisements are usually represented in image space and word space respectively, which are quite different with each other inherently. As a result, existing methods for Web page advertising are inapplicable since they represent both Web pages and advertisement in the same word space. In order to solve the problem, we propose to exploit the social Web to link these two feature spaces together. In particular, we present a unified generative model to integrate advertisements, words and images. Specifically, our solution combines two parts in a principled approach: First, we transform images from a image feature space to a word space utilizing the knowledge from images with annotations from social Web. Then, a language model based approach is applied to estimate the relevance between transformed images and advertisements. Moreover, in this model, the probability of recommending an advertisement can be inferred efficiently given an image, which enables potential applications to online advertising.

As more and more multimedia data become available on the Web, mining on those data is playing an increasingly important role in Web applications. In this paper, we investigate the interplay between multimedia data mining and text data mining. Specifically, in an approach we called text-aided image classification (TAIC), we address the problem of image classification with very limited amount of labeled images and a large amount of auxiliary labeled text data. This problem is important in practice, since currently on the Web, labeled text data are usually much more than image data. To solve the problem, based on the “bag-of-words” view and the Naive Bayes classification model, we focus our attention on the estimation of the image feature distribution under given concept. We extend the Naive Bayes algorithm by considering a mapping that maps the most discriminative text features into the image feature space. This feature mapping is estimated based on the text-image cooccurrence data on the Web, acting like a bridge that connects text and image knowledge. With this process, we estimate target image feature distribution from a text model based on sufficient labeled data. Our empirical results on real world data sets show that our method makes a good approximation of the image feature distribution when trained with abundant labeled images. In the case amount of labeled images is very limited, the classification performance is improved by using auxiliary labeled text data, which shows that our method can indeed integrate text and image knowledge in a simple yet effective way.

In this paper, we investigate the problem of mispronunciation detection by considering the influence of speaker and syllables. Machine learning techniques are used to make our method more convenient and flexible for new features, such as syllables normalization. The experimental results on our database, consisting of 9898 syllables pronounced by 100 speakers, show the effectiveness of our method by reducing the average false acceptance rate (FAR) by 42.5% using data set generated by model without adaptation to observation set and reducing average FAR by 32.5% using data set generated by model with adaptation to observation set.

This paper investigates a new machine learning strategy called translated learning. Unlike many previous learning tasks, we focus on how to use labeled data from one feature space to enhance the classification of other entirely different learning spaces. For example, we might wish to use labeled text data to help learn a model for classifying image data, when the labeled images are difficult to obtain. An important aspect of translated learning is to build a "bridge" to link one feature space (known as the "source space") to another space (known as the "target space") through a translator in order to migrate the knowledge from source to target. The translated learning solution uses a language model to link the class labels to the features in the source spaces, which in turn is translated to the features in the target spaces. Finally, this chain of linkages is completed by tracing back to the instances in the target spaces. We show that this path of linkage can be modeled using a Markov chain and risk minimization. Through experiments on the text-aided image classification and cross-language classification tasks, we demonstrate that our translated learning framework can greatly outperform many state-of-the-art baseline methods.