Reinforcement learning from human feedback (RLHF) has demonstrated effectiveness in aligning large language models (LLMs) with human preferences. However, token-level RLHF suffers from the credit assignment problem over long sequences, where delayed rewards make it challenging for the model to discern which actions contributed to preferred outcomes. This hinders learning efficiency and slows convergence.In this paper, we propose MA-RLHF, a simple yet effective RLHF framework that incorporates macro actions --- sequences of tokens or higher-level language constructs --- into the learning process. By operating at higher level of abstraction, our approach reduces the temporal distance between actions and rewards, facilitating faster and more accurate credit assignment. This results in more stable policy gradient estimates and enhances learning efficiency within each episode, all without increasing computational complexity during training or inference. We validate our approach through extensive experiments across various model sizes and tasks, including text summarization, dialogue generation, question answering, and program synthesis. Our method achieves substantial performance improvements over standard RLHF, with performance gains of up to 30\% in text summarization and code generation, 18\% in dialogue, and 8\% in question answering tasks. Notably, our approach reaches parity with vanilla RLHF $1.7 \sim 2$ times faster in terms of training time and continues to outperform it with further training. We make our code and data publicly available at \url{https://github.com/ernie-research/MA-RLHF}.

Knowledge graph embedding (KGE) models achieved state-of-the-art results on many knowledge graph tasks including link prediction and information retrieval.Despite the superior performance of KGE models in practice, we discover a deficiency in the expressiveness of some popular existing KGE models called Z-paradox.Motivated by the existence of Z-paradox, we propose a new KGE model called MQuinE that does not suffer from Zparadox while preserves strong expressiveness to model various relation patterns including symmetric/asymmetric, inverse, 1-N/N-1/N-N, and composition relations with theoretical justification.Experiments on real-world knowledge bases indicate that Z-paradox indeed degrades the performance of existing KGE models, and can cause more than 20% accuracy drop on some challenging test samples.Our experiments further demonstrate that MQuinE can mitigate the negative impact of Z-paradox and outperform existing KGE models by a visible margin on link prediction tasks.

Federated learning is an emerging technology for training machine learning models across decentralized data sources without sharing data. Vertical federated learning, also known as feature-based federated learning, applies to scenarios where data sources have the same sample IDs but different feature sets. To ensure fairness among data owners, it is critical to objectively assess the contributions from different data sources and compensate the corresponding data owners accordingly. The Shapley value is a provably fair contribution valuation metric originating from cooperative game theory. However, its straight-forward computation requires extensively retraining a model on each potential combination of data sources, leading to prohibitively high communication and computation overheads due to multiple rounds of federated learning. To tackle this challenge, we propose a contribution valuation metric called vertical federated Shapley value (VerFedSV) based on the classic Shapley value. We show that VerFedSV not only satisfies many desirable properties of fairness but is also efficient to compute. Moreover, VerFedSV can be adapted to both synchronous and asynchronous vertical federated learning algorithms. Both theoretical analysis and extensive experimental results demonstrate the fairness, efficiency, adaptability, and effectiveness of VerFedSV.

In the context of privacy-aware social web services, secure causal reasoning on graphs is an essential task, for example, identifying the key influential nodes/users through which information, ideas, trends spread across the network. Despite the significant advancements achieved by neural network-based methods in these domains, existing research often assumes complete observable of graph data, an assumption that conflicts with privacy concerns in social web services. This assumption may not hold in numerous real-world social analysis applications due to privacy restrictions. In such scenarios, some graph links might remain unobserved until the model is trained, posing a critical challenge for privacy preservation. To address this issue, we introduce a novel model named Causal Effect on Coarsened Graphs (Cecoar), which is specifically designed for privacy-aware social web services. This model employs a new type of coarsened graph that accurately represents partially observed nodes and edges, aligning with the privacy constraints of social web services. We have developed three crucial modules to enable causal estimation on coarsened graphs: (1) a graph sampling module, (2) a confounder encoder, and (3) a graph-based encoder with causal prediction layers. Furthermore, we present a comprehensive theoretical analysis of the proposed model, demonstrating its performance and generalization capabilities. We evaluate our model on two benchmark datasets for causal estimation, and the results confirm its effectiveness in handling graph data containing private information. The code is in https://github.com/liu-yang-maker/Cecoar-ICWS24.

Knowledge graph (KG) reasoning utilizes two primary techniques, i.e., rule-based and KG-embedding based. The former provides precise inferences, but inferring via concrete rules is not scalable. The latter enables efficient reasoning at the cost of ambiguous inference accuracy. Neuro-symbolic reasoning seeks to amalgamate the advantages of both techniques. The crux of this approach is replacing the predicted existence of all possible triples (i.e., truth scores inferred from rules) with a suitable approximation grounded in embedding representations. However, constructing an effective approximation of all possible triples' truth scores is a challenging task, because it needs to balance the tradeoff between accuracy and efficiency, while compatible with both the rule-based and KG-embedding models. To this end, we proposed a differentiable framework - DiffLogic. Instead of directly approximating all possible triples, we design a tailored filter to adaptively select essential triples based on the dynamic rules and weights. The truth scores assessed by KG-embedding are continuous, so we employ a continuous Markov logic network named probabilistic soft logic (PSL). It employs the truth scores of essential triples to assess the overall agreement among rules, weights, and observed triples. PSL enables end-to-end differentiable optimization, so we can alternately update embedding and weighted rules. On benchmark datasets, we empirically show that DiffLogic surpasses baselines in both effectiveness and efficiency.

Markov logic network (MLN) is a powerful statistical modeling framework for probabilistic logic reasoning. Despite the elegancy and effectiveness of MLN, the inference of MLN is known to suffer from an efficiency issue. Even the state-of-the-art MLN engines can not scale to medium-size real-world knowledge bases in the open-world setting, i.e., all unobserved facts in the knowledge base need predictions. In this work, by focusing on a certain class of first-order logic rules that are sufficiently expressive, we develop a highly efficient MLN inference engine called MLN4KB that can leverage the sparsity of knowledge bases. MLN4KB enjoys quite strong theoretical properties; its space and time complexities can be exponentially smaller than existing MLN engines. Experiments on both synthetic and real-world knowledge bases demonstrate the effectiveness of the proposed method. MLN4KB is orders of magnitudes faster (more than 103 times faster on some datasets) than existing MLN engines in the open-world setting. Without any approximation tricks, MLN4KB can scale to real-world knowledge bases including WN-18 and YAGO3-10 and achieve decent prediction accuracy without bells and whistles.

Federated learning is an emerging decentralized machine learning scheme that allows multiple data owners to work collaboratively while ensuring data privacy. The success of federated learning depends largely on the participation of data owners. To sustain and encourage data owners' participation, it is crucial to fairly evaluate the quality of the data provided by the data owners as well as their contribution to the final model and reward them correspondingly. Federated Shapley value, recently proposed by Wang et al. [Federated Learning, 2020], is a measure for data value under the framework of federated learning that satisfies many desired properties for data valuation. However, there are still factors of potential unfairness in the design of federated Shapley value because two data owners with the same local data may not receive the same evaluation. We propose a new measure called completed federated Shapley value to improve the fairness of federated Shapley value. The design depends on completing a matrix consisting of all the possible contributions by different subsets of the data owners. It is shown under mild conditions that this matrix is approximately low-rank by leveraging concepts and tools from optimization. Both theoretical analysis and empirical evaluation verify that the proposed measure does improve fairness in many circumstances.

Online mirror descent (OMD) and dual averaging (DA)---two fundamental algorithms for online convex optimization---are known to have very similar (and sometimes identical) performance guarantees when used with a fixed learning rate. Under dynamic learning rates, however, OMD is provably inferior to DA and suffers linear regret, even in common settings such as prediction with expert advice. We modify the OMD algorithm through a simple technique that we call stabilization. We give essentially the same abstract regret bound for OMD with stabilization and for DA by modifying the classical OMD convergence analysis in a careful and modular way that allows for straightforward and flexible proofs. Simple corollaries of these bounds show that OMD with stabilization and DA enjoy the same performance guarantees in many applications---even under dynamic learning rates. We also shed light on the similarities between OMD and DA and show simple conditions under which stabilized-OMD and DA generate the same iterates. Finally, we show how to effectively use dual-stabilization with composite cost functions with simple adaptations to both the algorithm and its analysis.

Greedy coordinate descent (GCD) is an efficient optimization algorithm for a wide range of machine learning and data mining applications. GCD could be significantly faster than randomized coordinate descent (RCD) if they have similar per iteration cost. Nevertheless, in some cases, the greedy rule used in GCD cannot be efficiently implemented, leading to huge per iteration cost and making GCD slower than RCD. To alleviate the cost per iteration, the existing solutions rely on maximum inner product search (MIPS) as an approximate greedy rule. But it has been empirically shown that GCD with approximate greedy rule could suffer from slow convergence even with the state-of-the-art MIPS algorithms. We propose a hybrid coordinate descent algorithm with a simple variable partition strategy to tackle the cases when greedy rule cannot be implemented efficiently. The convergence rate and theoretical properties of the new algorithm are presented. The proposed method is shown to be especially useful when the data matrix has a group structure. Numerical experiments with both synthetic and real-world data demonstrate that our new algorithm is competitive against RCD, GCD, approximate GCD with MIPS and their accelerated variants.

We consider greedy coordinate descent (GCD) for composite problems with sparsity inducing regularizers, including 1-norm regularization and non-negative constraints. Empirical evidence strongly suggests that GCD, when initialized with the zero vector, has an implicit screening ability that usually selects at each iteration coordinates that at are nonzero at the solution. Thus, for problems with sparse solutions, GCD can converge significantly faster than randomized coordinate descent. We present an improved convergence analysis of GCD for sparse optimization, and a formal analysis of its screening properties. We also propose and analyze an improved selection rule with stronger ability to produce sparse iterates. Numerical experiments on both synthetic and real-world data support our analysis and the effectiveness of the proposed selection rule.

Online mirror descent (OMD) and dual averaging (DA)—two fundamental algorithms for online convex optimization—are known to have very similar (and sometimes identical) performance guarantees when used with a fixed learning rate. Under dynamic learning rates, however, OMD is provably inferior to DA and suffers a linear regret, even in common settings such as prediction with expert advice. We modify the OMD algorithm through a simple technique that we call stabilization. We give essentially the same abstract regret bound for OMD with stabilization and for DA by modifying the classical OMD convergence analysis in a careful and modular way that allows for straightforward and flexible proofs. Simple corollaries of these bounds show that OMD with stabilization and DA enjoy the same performance guarantees in many applications—even under dynamic learning rates. We also shed light on the similarities between OMD and DA and show simple conditions under which stabilized-OMD and DA generate the same iterates.

We consider the problem of training one-versus-all (OVA) linear classifiers for multiclass or multilabel classification when the number of labels is large. A naive extension of OVA to this problem, even with hundreds of cores, usually requires hours for training on large real world datasets. We propose a novel algorithm called OVA-Primal++ that speeds up the training of OVA by using a tree-structured training order, where each classifier is trained using its parent's classifier as initialization. OVA-Primal++ is both theoretically and empirically faster than the naive OVA algorithm, and yet still enjoys the same highly parallelizability and small memory footprint. Extensive experiments on multiclass and multilabel classification datasets validate the effectiveness of our method.

We consider the semi-supervised dimension reduction problem: given a high dimensional dataset with a small number of labeled data and huge number of unlabeled data, the goal is to find the low-dimensional embedding that yields good classification results. Most of the previous algorithms for this task are linkage-based algorithms. They try to enforce the must-link and cannot-link constraints in dimension reduction, leading to a nearest neighbor classifier in low dimensional space. In this paper, we propose a new hyperplane-based semi-supervised dimension reduction method-the main objective is to learn the low-dimensional features that can both approximate the original data and form a good separating hyperplane. We formulate this as a non-convex optimization problem and propose an efficient algorithm to solve it. The algorithm can scale to problems with millions of features and can easily incorporate non-negative constraints in order to learn interpretable non-negative features. Experiments on real world datasets demonstrate that our hyperplane-based dimension reduction method outperforms state-of-art linkage-based methods when very few labels are available.

Matrix completion is a widely used technique for personalized recommender system. In this paper, we focus on the idea of Bounded Matrix Completion (BMC) which imposes bounded constraint into the original matrix completion problem. It has been shown that BMC works well for several real world datasets, and an efficient coordinate descent solver called BMA has been proposed in~\cite{bma}. However, we observe that the BMA algorithm sometimes fails to converge to a stationary point, resulting in a relatively poor accuracy in those cases. To overcome this issue, we propose our new approach for solving BMC under the ADMM framework. The proposed algorithm is gauranteed to converge to stationary points. Experimental results on real world datasets show that our algorithm can reach a lower objective value, obtain a higher predict accuracy rate and have better scalability compared with BMA. We also present that our method outperforms the state-of-art standard matrix factorization in most cases.