Many sequential decision making problems can be formulated as an adaptive submodular maximization problem. However, most of existing studies in this field focus on pool-based setting, where one can pick items in any order, and there have been few studies for the stream-based setting where items arrive in an arbitrary order and one must immediately decide whether to select an item or not upon its arrival. In this paper, we introduce a new class of utility functions, semi-policywise submodular functions. We develop a series of effective algorithms to maximize a semi-policywise submodular function under the stream-based setting.
Many sequential decision making problems, including pool-based active learning and adaptive viral marketing, can be formulated as an adaptive submodular maximization problem. Most of existing studies on adaptive submodular optimization focus on either monotone case or non-monotone case. Specifically, if the utility function is monotone and adaptive submodular, \cite{golovin2011adaptive} developed a greedy policy that achieves a $(1-1/e)$ approximation ratio subject to a cardinality constraint. If the utility function is non-monotone and adaptive submodular, \cite{tang2021beyond} showed that a random greedy policy achieves a $1/e$ approximation ratio subject to a cardinality constraint. In this work, we aim to generalize the above mentioned results by studying the partial-monotone adaptive submodular maximization problem. To this end, we introduce the notation of adaptive monotonicity ratio $m\in[0,1]$ to measure the degree of monotonicity of a function. Our main result is to show that a random greedy policy achieves an approximation ratio of $m(1-1/e)+(1-m)(1/e)$ if the utility function is $m$-adaptive monotone and adaptive submodular. Notably this result recovers the aforementioned $(1-1/e)$ and $1/e$ approximation ratios when $m = 0$ and $m = 1$, respectively. We further extend our results to consider a knapsack constraint. We show that a sampling-based policy achieves an approximation ratio of $(m+1)/10$ if the utility function is $m$-adaptive monotone and adaptive submodular. One important implication of our results is that even for a non-monotone utility function, we still can achieve an approximation ratio close to $(1-1/e)$ if this function is ``close'' to a monotone function. This leads to improved performance bounds for many machine learning applications whose utility functions are almost adaptive monotone.
In this paper, we study the classic submodular maximization problem subject to a group fairness constraint under both non-adaptive and adaptive settings. It has been shown that the utility function of many machine learning applications, including data summarization, influence maximization in social networks, and personalized recommendation, satisfies the property of submodularity. Hence, maximizing a submodular function subject to various constraints can be found at the heart of many of those applications. On a high level, submodular maximization aims to select a group of most representative items (e.g., data points). However, the design of most existing algorithms does not incorporate the fairness constraint, leading to under- or over-representation some particular groups. This motivates us to study the fair submodular maximization problem, where we aim to select a group of items to maximize a (possibly non-monotone) submodular utility function subject to a group fairness constraint. To this end, we develop the first constant-factor approximation algorithm for this problem. The design of our algorithm is robust enough to be extended to solving the submodular maximization problem under a more complicated adaptive setting. Moreover, we further extend our study to incorporating a global cardinality constraint.
Background: The assessment of left ventricular (LV) function by myocardial perfusion SPECT (MPS) relies on accurate myocardial segmentation. The purpose of this paper is to develop and validate a new method incorporating deep learning with shape priors to accurately extract the LV myocardium for automatic measurement of LV functional parameters. Methods: A segmentation architecture that integrates a three-dimensional (3D) V-Net with a shape deformation module was developed. Using the shape priors generated by a dynamic programming (DP) algorithm, the model output was then constrained and guided during the model training for quick convergence and improved performance. A stratified 5-fold cross-validation was used to train and validate our models. Results: Results of our proposed method agree well with those from the ground truth. Our proposed model achieved a Dice similarity coefficient (DSC) of 0.9573(0.0244), 0.9821(0.0137), and 0.9903(0.0041), a Hausdorff distances (HD) of 6.7529(2.7334) mm, 7.2507(3.1952) mm, and 7.6121(3.0134) mm in extracting the endocardium, myocardium, and epicardium, respectively. Conclusion: Our proposed method achieved a high accuracy in extracting LV myocardial contours and assessing LV function.
To break the bottlenecks of mainstream cloud-based machine learning (ML) paradigm, we adopt device-cloud collaborative ML and build the first end-to-end and general-purpose system, called Walle, as the foundation. Walle consists of a deployment platform, distributing ML tasks to billion-scale devices in time; a data pipeline, efficiently preparing task input; and a compute container, providing a cross-platform and high-performance execution environment, while facilitating daily task iteration. Specifically, the compute container is based on Mobile Neural Network (MNN), a tensor compute engine along with the data processing and model execution libraries, which are exposed through a refined Python thread-level virtual machine (VM) to support diverse ML tasks and concurrent task execution. The core of MNN is the novel mechanisms of operator decomposition and semi-auto search, sharply reducing the workload in manually optimizing hundreds of operators for tens of hardware backends and further quickly identifying the best backend with runtime optimization for a computation graph. The data pipeline introduces an on-device stream processing framework to enable processing user behavior data at source. The deployment platform releases ML tasks with an efficient push-then-pull method and supports multi-granularity deployment policies. We evaluate Walle in practical e-commerce application scenarios to demonstrate its effectiveness, efficiency, and scalability. Extensive micro-benchmarks also highlight the superior performance of MNN and the Python thread-level VM. Walle has been in large-scale production use in Alibaba, while MNN has been open source with a broad impact in the community.
Data heterogeneity is an intrinsic property of recommender systems, making models trained over the global data on the cloud, which is the mainstream in industry, non-optimal to each individual user's local data distribution. To deal with data heterogeneity, model personalization with on-device learning is a potential solution. However, on-device training using a user's small size of local samples will incur severe overfitting and undermine the model's generalization ability. In this work, we propose a new device-cloud collaborative learning framework, called CoDA, to break the dilemmas of purely cloud-based learning and on-device learning. The key principle of CoDA is to retrieve similar samples from the cloud's global pool to augment each user's local dataset to train the recommendation model. Specifically, after a coarse-grained sample matching on the cloud, a personalized sample classifier is further trained on each device for a fine-grained sample filtering, which can learn the boundary between the local data distribution and the outside data distribution. We also build an end-to-end pipeline to support the flows of data, model, computation, and control between the cloud and each device. We have deployed CoDA in a recommendation scenario of Mobile Taobao. Online A/B testing results show the remarkable performance improvement of CoDA over both cloud-based learning without model personalization and on-device training without data augmentation. Overhead testing on a real device demonstrates the computation, storage, and communication efficiency of the on-device tasks in CoDA.
In this paper, we study the constrained stochastic submodular maximization problem with state-dependent costs. The input of our problem is a set of items whose states (i.e., the marginal contribution and the cost of an item) are drawn from a known probability distribution. The only way to know the realized state of an item is to select that item. We consider two constraints, i.e., \emph{inner} and \emph{outer} constraints. Recall that each item has a state-dependent cost, and the inner constraint states that the total \emph{realized} cost of all selected items must not exceed a give budget. Thus, inner constraint is state-dependent. The outer constraint, one the other hand, is state-independent. It can be represented as a downward-closed family of sets of selected items regardless of their states. Our objective is to maximize the objective function subject to both inner and outer constraints. Under the assumption that larger cost indicates larger "utility", we present a constant approximate solution to this problem.
The goal of a typical adaptive sequential decision making problem is to design an interactive policy that selects a group of items sequentially, based on some partial observations, to maximize the expected utility. It has been shown that the utility functions of many real-world applications, including pooled-based active learning and adaptive influence maximization, satisfy the property of adaptive submodularity. However, most of existing studies on adaptive submodular maximization focus on the fully adaptive setting, i.e., one must wait for the feedback from \emph{all} past selections before making the next selection. Although this approach can take full advantage of feedback from the past to make informed decisions, it may take a longer time to complete the selection process as compared with the non-adaptive solution where all selections are made in advance before any observations take place. In this paper, we explore the problem of partial-adaptive submodular maximization where one is allowed to make multiple selections in a batch simultaneously and observe their realizations together. Our approach enjoys the benefits of adaptivity while reducing the time spent on waiting for the observations from past selections. To the best of our knowledge, no results are known for partial-adaptive policies for the non-monotone adaptive submodular maximization problem. We study this problem under both cardinality constraint and knapsack constraints, and develop effective and efficient solutions for both cases. We also analyze the batch query complexity, i.e., the number of batches a policy takes to complete the selection process, of our policy under some additional assumptions.
The idea of social advertising (or social promotion) is to select a group of influential individuals (a.k.a \emph{seeds}) to help promote some products or ideas through an online social networks. There are two major players in the social advertising ecosystem: advertiser and platform. The platform sells viral engagements such as "like"s to advertisers by inserting their ads into the feed of seeds. These seeds receive monetary incentives from the platform in exchange for their participation in the social advertising campaign. Once an ad is engaged by a follower of some seed, the platform receives a fixed amount of payment, called cost per engagement, from the advertiser. The ad could potentially attract more engagements from followers' followers and trigger a viral contagion. At the beginning of a campaign, the advertiser submits a budget to the platform and this budget can be used for two purposes: recruiting seeds and paying for the viral engagements generated by the seeds. Note that the first part of payment goes to the seeds and the latter one is the actual revenue collected by the platform. In this setting, the problem for the platform is to recruit a group of seeds such that she can collect the largest possible amount of revenue subject to the budget constraint. We formulate this problem as a seed selection problem whose objective function is non-monotone and it might take on negative values, making existing results on submodular optimization and influence maximization not applicable to our setting. We study this problem under both non-adaptive and adaptive settings. Although we focus on social advertising in this paper, our results apply to any optimization problems whose objective function is the expectation of the minimum of a stochastic submodular function and a linear function.
We study practical data characteristics underlying federated learning, where non-i.i.d. data from clients have sparse features, and a certain client's local data normally involves only a small part of the full model, called a submodel. Due to data sparsity, the classical federated averaging (FedAvg) algorithm or its variants will be severely slowed down, because when updating the global model, each client's zero update of the full model excluding its submodel is inaccurately aggregated. Therefore, we propose federated submodel averaging (FedSubAvg), ensuring that the expectation of the global update of each model parameter is equal to the average of the local updates of the clients who involve it. We theoretically proved the convergence rate of FedSubAvg by deriving an upper bound under a new metric called the element-wise gradient norm. In particular, this new metric can characterize the convergence of federated optimization over sparse data, while the conventional metric of squared gradient norm used in FedAvg and its variants cannot. We extensively evaluated FedSubAvg over both public and industrial datasets. The evaluation results demonstrate that FedSubAvg significantly outperforms FedAvg and its variants.