Modern recommendation systems can be broadly divided into two key stages: the ranking stage, where the system predicts various user engagements (e.g., click-through rate, like rate, follow rate, watch time), and the value model stage, which aggregates these predictive scores through a function (e.g., a linear combination defined by a weight vector) to measure the value of each content by a single numerical score. Both stages play roughly equally important roles in real industrial systems; however, how to optimize the model weights for the second stage still lacks systematic study. This paper focuses on optimizing the second stage through auto-tuning technology. Although general auto-tuning systems and solutions - both from established production practices and open-source solutions - can address this problem, they typically require weeks or even months to identify a feasible solution. Such prolonged tuning processes are unacceptable in production environments for recommendation systems, as suboptimal value models can severely degrade user experience. An effective auto-tuning solution is required to identify a viable model within 2-3 days, rather than the extended timelines typically associated with existing approaches. In this paper, we introduce a practical auto-tuning system named HyperZero that addresses these time constraints while effectively solving the unique challenges inherent in modern recommendation systems. Moreover, this framework has the potential to be expanded to broader tuning tasks within recommendation systems.

The significance of estimating video watch time has been highlighted by the rising importance of (short) video recommendation, which has become a core product of mainstream social media platforms. Modeling video watch time, however, has been challenged by the complexity of user-video interaction, such as different user behavior modes in watching the recommended videos and varying watching probability over the video progress bar. Despite the importance and challenges, existing literature on modeling video watch time mostly focuses on relatively black-box mechanical enhancement of the classical regression/classification losses, without factoring in user behavior in a principled manner. In this paper, we for the first time take on a user-centric perspective to model video watch time, from which we propose a white-box statistical framework that directly translates various user behavior assumptions in watching (short) videos into statistical watch time models. These behavior assumptions are portrayed by our domain knowledge on users' behavior modes in video watching. We further employ bucketization to cope with user's non-stationary watching probability over the video progress bar, which additionally helps to respect the constraint of video length and facilitate the practical compatibility between the continuous regression event of watch time and other binary classification events. We test our models extensively on two public datasets, a large-scale offline industrial dataset, and an online A/B test on a short video platform with hundreds of millions of daily-active users. On all experiments, our models perform competitively against strong relevant baselines, demonstrating the efficacy of our user-centric perspective and proposed framework.

In recommendation systems, the timely delivery of new content to their relevant audiences is critical for generating a growing and high quality collection of content for all users. The nature of this problem requires retrieval models to be able to make inferences in real time and with high relevance. There are two specific challenges for cold start contents. First, the information loss problem in a standard Two Tower model, due to the limited feature interactions between the user and item towers, is exacerbated for cold start items due to training data sparsity. Second, the huge volume of user-generated content in industry applications today poses a big bottleneck in the end-to-end latency of recommending new content. To overcome the two challenges, we propose a novel architecture, the Item History Model (IHM). IHM directly injects user-interaction information into the item tower to overcome information loss. In addition, IHM incorporates an inductive structure using attention-based pooling to eliminate the need for recurring training, a key bottleneck for the real-timeness. On both public and industry datasets, we demonstrate that IHM can not only outperform baselines in recommending cold start contents, but also achieves SoTA real-timeness in industry applications.

Abstract Modern machine learning (ML) systems commonly use stochastic gradient descent (SGD) to train ML models. However, SGD relies on random data order to converge, which usually requires a full data shuffle. For in-DB ML systems and deep learning systems with large datasets stored onblock-addressable secondary storagesuch as HDD and SSD, this full data shuffle leads to low I/O performance—the data shuffling time can be even longer than the training itself, due to massive random data accesses. To balance the convergence rate of SGD (which favors data randomness) and its I/O performance (which favors sequential access), previous work has proposed several data shuffling strategies. In this paper, we first perform an empirical study on existing data shuffling strategies, showing that these strategies suffer from either low performance or low convergence rate. To solve this problem, we propose a simple but noveltwo-leveldata shuffling strategy named , which canavoida full data shuffle while maintainingcomparableconvergence rate of SGD as if a full shuffle were performed. We further theoretically analyze the convergence behavior of and empirically evaluate its efficacy in both in-DB ML and deep learning systems. For in-DB ML systems, we integrate into PostgreSQL by introducing three newphysicaloperators with optimizations. For deep learning systems, we extend single-process to multi-process for the parallel/distributed environment and integrate it into PyTorch. Our evaluation shows that can achieve comparable convergence rate with the full-shuffle-based SGD for both linear models and deep learning models. For in-DB ML with linear models, is 1.6 $$\times $$ × $$-$$ - 12.8 $$\times $$ × faster than two state-of-the-art systems, Apache MADlib and Bismarck, on both HDD and SSD. For deep learning models on ImageNet, is 1.5 $$\times $$ × faster than PyTorch with full data shuffle.

Decentralized online learning (online learning in decentralized networks) has been attracting more and more attention, since it is believed that decentralized online learning can help data providers cooperatively better solve their online problems without sharing their private data to a third party or other providers. Typically, the cooperation is achieved by letting the data providers exchange their models between neighbors, e.g., recommendation model. However, the best regret bound for a decentralized online learning algorithm is 𝒪( n √ T ), where n is the number of nodes (or users) and T is the number of iterations. This is clearly insignificant, since this bound can be achieved without any communication in the networks. This reminds us to ask a fundamental question: Can people really get benefit from the decentralized online learning by exchanging information? In this article, we studied when and why the communication can help the decentralized online learning to reduce the regret. Specifically, each loss function is characterized by two components: the adversarial component and the stochastic component. Under this characterization, we show that decentralized online gradient enjoys a regret bound \( {\mathcal {O}(\sqrt {n^2TG^2 + n T \sigma ^2})} \) , where G measures the magnitude of the adversarial component in the private data (or equivalently the local loss function) and σ measures the randomness within the private data. This regret suggests that people can get benefits from the randomness in the private data by exchanging private information. Another important contribution of this article is to consider the dynamic regret—a more practical regret to track users’ interest dynamics. Empirical studies are also conducted to validate our analysis.

Bi-typed multi-relational heterogeneous graph (BMHG) is one of the most common graphs in practice, for example, academic networks, e-commerce user behavior graph and enterprise knowledge graph. It is a critical and challenge problem on how to learn the numerical representation for each node to characterize subtle structures. However, most previous studies treat all node relations in BMHG as the same class of relation without distinguishing the different characteristics between the intra-type relations and inter-type relations of the bi-typed nodes, causing the loss of significant structure information. To address this issue, we propose a novelDualHierarchicalAttentionNetworks (DHAN) based on the bi-typed multi-relational heterogeneous graphs to learn comprehensive node representations with the intra-type and inter-type attention-based encoder under a hierarchical mechanism. Specifically, the former encoder aggregates information from the same type of nodes, while the latter aggregates node representations from its different types of neighbors. Moreover, to sufficiently model node multi-relational information in BMHG, we adopt a newly proposed hierarchical mechanism. By doing so, the proposed dual hierarchical attention operations enable our model to fully capture the complex structures of the bi-typed multi-relational heterogeneous graphs. Experimental results on various tasks against the state-of-the-arts sufficiently confirm the capability of DHAN in learning node representations on the BMHGs.

Recent years have witnessed an exponential growth of model scale in deep learning-based recommender systems---from Google's 2016 model with 1 billion parameters to the latest Facebook's model with 12 trillion parameters. Significant quality boost has come with each jump of the model capacity, which makes us believe the era of 100 trillion parameters is around the corner. However, the training of such models is challenging even within industrial scale data centers. We resolve this challenge by careful co-design of both optimization algorithm and distributed system architecture. Specifically, to ensure both the training efficiency and the training accuracy, we design a novel hybrid training algorithm, where the embedding layer and the dense neural network are handled by different synchronization mechanisms; then we build a system called Persia (short for parallel recommendation training system with hybrid acceleration) to support this hybrid training algorithm. Both theoretical demonstrations and empirical studies with up to 100 trillion parameters have been conducted to justify the system design and implementation of Persia. We make Persia publicly available (at github.com/PersiaML/Persia) so that anyone can easily train a recommender model at the scale of 100 trillion parameters.

Stochastic gradient descent (SGD) is the cornerstone of modern ML systems. Despite its computational efficiency, SGD requires random data access that is inherently inefficient when implemented in systems that rely on block-addressable secondary storage such as HDD and SSD, e.g., in-DB ML systems and TensorFlow/PyTorch over large files. To address this impedance mismatch, various data shuffling strategies have been proposed to balance the convergence rate of SGD (which favors randomness) and its I/O performance (which favors sequential access).

Multi-task learning (MTL) has been widely utilized in various industrial scenarios, such as recommender systems and search engines. MTL can improve learning efficiency and prediction accuracy by exploiting commonalities and differences across tasks. However, MTL is sensitive to relationships among tasks and may have performance degradation in real-world applications, because existing neural-based MTL models often share the same network structures and original input features. To address this issue, we propose a novel multi-task learning model based on Prototype Feature Extraction (PFE) to balance task-specific objectives and inter-task relationships. PFE is a novel component to disentangle features for multiple tasks. To better extract features from original inputs before gating networks, we introduce a new concept, namely prototype feature center, to disentangle features for multiple tasks. The extracted prototype features fuse various features from different tasks to better learn inter-task relationships. PFE updates prototype feature centers and prototype features iteratively. Our model utilizes the learned prototype features and task-specific experts for MTL. We implement PFE on two public datasets. Empirical results show that PFE outperforms state-of-the-art MTL models by extracting prototype features. Furthermore, we deploy PFE in a real-world recommender system (one of the world’s top-tier short video sharing platforms) to showcase that PFE can be widely applied in industrial scenarios.

The ever-growing demand and complexity of machine learning are putting pressure on hyper-parameter tuning systems: while the evaluation cost of models continues to increase, the scalability of state-of-the-arts starts to become a crucial bottleneck. In this paper, inspired by our experience when deploying hyper-parameter tuning in a real-world application in production and the limitations of existing systems, we propose Hyper-Tune, an efficient and robust distributed hyper-parameter tuning framework. Compared with existing systems, Hyper-Tune highlights multiple system optimizations, including (1) automatic resource allocation, (2) asynchronous scheduling, and (3) multi-fidelity optimizer. We conduct extensive evaluations on benchmark datasets and a large-scale real-world dataset in production. Empirically, with the aid of these optimizations, Hyper-Tune outperforms competitive hyper-parameter tuning systems on a wide range of scenarios, including XGBoost, CNN, RNN, and some architectural hyper-parameters for neural networks. Compared with the state-of-the-art BOHB and A-BOHB, Hyper-Tune achieves up to 11.2X and 5.1X speedups, respectively.

Matching is almost the first and most fundamental step in recommender systems, that is to quickly select hundreds or thousands of related entities from the whole commodity pool. Among all the matching methods, item-to-item (I2I) graph based matching is a handy and highly effective approach and is widely used in most applications, owing to the essential relationships of entities described in a powerful I2I graph. Yet, the I2I graph is not a ready-made product in a data source. To obtain it from users' behaviors, a common practice in the industry is to construct the graph based on the similarity of item embeddings or co-occurrence frequency directly. However, these methods tend to lose the complicated correlations (high-ordered or nonlinear) inside decision-making actions and cannot achieve the global optimal solution. Moreover, the correlations between items are usually contained in users' short-term actions, which are full of noise information (e.g. spurious association, missing connection). It is vitally important to filter out noise while generating the graph. In this paper, we propose a novel framework called Purified Graph Generation (PGG) dedicated to learn faithful I2I graph from sparse and noisy behavior data. We capture the 'confidence value' between user and item to get rid of exception action during decision making, and leverage it to re-sample purified sets that are fed into an unsupervised I2I graph structure learning framework called GPBG. Extensive experimental results from both simulation and real data demonstrate that our method could significantly benefit the performance of I2I graph compared to the typical baselines.

Recent years have witnessed a growing list of systems for distributed data-parallel training. Existing systems largely fit into two paradigms, i.e., parameter server and MPI-style collective operations. On the algorithmic side, researchers have proposed a wide range of techniques to lower the communication via "system relaxations": quantization, decentralization, and communication delay. However, most, if not all, existing systems only rely on standard synchronous and asynchronous stochastic gradient (SG) based optimization, therefore, cannot take advantage of all possible optimizations that the machine learning community has been developing recently. Given this emerging gap between the current landscapes of systems and theory, we build Bagua, a MPI-style communication library, providing a collection of primitives, that is both flexible and modular to support state-of-the-art system relaxation techniques of distributed training. Powered by this design, Bagua has a great ability to implement and extend various state-of-the-art distributed learning algorithms. In a production cluster with up to 16 machines (128 GPUs), Bagua can outperform PyTorch-DDP, Horovod and BytePS in the end-to-end training time by a significant margin (up to 2X) across a diverse range of tasks. Moreover, we conduct a rigorous tradeoff exploration showing that different algorithms and system relaxations achieve the best performance over different network conditions.

We demonstrate ease.ml, a multi-tenant machine learning service we host at ETH Zurich for various research groups. Unlike existing machine learning services, ease.ml presents a novel architecture that supports multi-tenant, cost-aware model selection that optimizes for minimizing total regrets of all users. Moreover, it provides a novel user interface that enables declarative machine learning at a higher level: Users only need to specify the input/output schemata of their learning tasks and ease.ml can handle the rest. In this demonstration, we present the design principles of ease.ml, highlight the implementation of its key components, and showcase how ease.ml can help ease machine learning tasks that often perplex even experienced users.

We present ease.ml, a declarative machine learning service platform. With ease.ml, a user defines the high-level schema of an ML application and submits the task via a Web interface. The system then deals with the rest, such as model selection and data movement. The ultimate question we hope to understand is that, as a "service provider" that manages a shared cluster of machines running machine learning workloads, what is the resource sharing strategy that maximizes the global satisfaction of all our users?

This paper considers the feature selection scenario where only a few features are accessible at any time point. For example, features are generated sequentially and visible one by one. Therefore, one has to make an online decision to identify key features after all features are only scanned once or twice. The optimization based approach is a powerful tool for the online feature selection.

We consider the problem of learning incoherent sparse and low-rank patterns from multiple tasks. Our approach is based on a linear multi-task learning formulation, in which the sparse and low-rank patterns are induced by a cardinality regularization term and a low-rank constraint, respectively. This formulation is non-convex; we convert it into its convex surrogate, which can be routinely solved via semidefinite programming for small-size problems. We propose to employ the general projected gradient scheme to efficiently solve such a convex surrogate; however, in the optimization formulation, the objective function is non-differentiable and the feasible domain is non-trivial. We present the procedures for computing the projected gradient and ensuring the global convergence of the projected gradient scheme. The computation of projected gradient involves a constrained optimization problem; we show that the optimal solution to such a problem can be obtained via solving an unconstrained optimization subproblem and an Euclidean projection subproblem. We also present two projected gradient algorithms and analyze their rates of convergence in details. In addition, we illustrate the use of the presented projected gradient algorithms for the proposed multi-task learning formulation using the least squares loss. Experimental results on a collection of real-world data sets demonstrate the effectiveness of the proposed multi-task learning formulation and the efficiency of the proposed projected gradient algorithms.

We consider the problem of learning incoherent sparse and low-rank patterns from multiple tasks. Our approach is based on a linear multi-task learning formulation, in which the sparse and low-rank patterns are induced by a cardinality regularization term and a low-rank constraint, respectively. This formulation is non-convex; we convert it into its convex surrogate, which can be routinely solved via semidefinite programming for small-size problems. We propose to employ the general projected gradient scheme to efficiently solve such a convex surrogate; however, in the optimization formulation, the objective function is non-differentiable and the feasible domain is non-trivial. We present the procedures for computing the projected gradient and ensuring the global convergence of the projected gradient scheme. The computation of projected gradient involves a constrained optimization problem; we show that the optimal solution to such a problem can be obtained via solving an unconstrained optimization subproblem and an Euclidean projection subproblem. In addition, we present two projected gradient algorithms and discuss their rates of convergence. Experimental results on benchmark data sets demonstrate the effectiveness of the proposed multi-task learning formulation and the efficiency of the proposed projected gradient algorithms.