arXiv subjects

Zhaonian Zou

Publications and source records attributed to Zhaonian Zou.

11 recordsLinked to original sources

arXiv2026

Advancing Analytic Class-Incremental Learning through Vision-Language Calibration

Class-incremental learning (CIL) with pre-trained models (PTMs) faces a critical trade-off between efficient adaptation and long-term stability. While analytic learning enables rapid, recursive closed-form updates, its efficacy is often compromised by accumulated errors and feature incompatibility. In this paper, we first conduct a systematic study to dissect the failure modes of PTM-based analytic CIL, identifying representation rigidity as the primary bottleneck. Motivated by this insight, we propose VILA, a novel dual-branch framework that advances analytic CIL via a two-level vision-language calibration strategy. Specifically, we coherently fuse plastic, task-adapted features with a frozen, universal visual anchor at the feature level through geometric calibration, and leverage cross-modal semantic priors at the decision level to rectify prediction bias. This confluence maintains analytic-learning's extreme efficiency while overcoming its inherent brittleness. Extensive experiments across eight benchmarks demonstrate that VILA consistently yields superior performance, particularly in fine-grained and long-sequence scenarios. Our framework harmonizes high-fidelity prediction with the simplicity of analytic learning. Our code is available at https://github.com/byzhaoAI/VILA.

cs.LG↗

arXiv2025

MCE: Towards a General Framework for Handling Missing Modalities under Imbalanced Missing Rates

Multi-modal learning has made significant advances across diverse pattern recognition applications. However, handling missing modalities, especially under imbalanced missing rates, remains a major challenge. This imbalance triggers a vicious cycle: modalities with higher missing rates receive fewer updates, leading to inconsistent learning progress and representational degradation that further diminishes their contribution. Existing methods typically focus on global dataset-level balancing, often overlooking critical sample-level variations in modality utility and the underlying issue of degraded feature quality. We propose Modality Capability Enhancement (MCE) to tackle these limitations. MCE includes two synergistic components: i) Learning Capability Enhancement (LCE), which introduces multi-level factors to dynamically balance modality-specific learning progress, and ii) Representation Capability Enhancement (RCE), which improves feature semantics and robustness through subset prediction and cross-modal completion tasks. Comprehensive evaluations on four multi-modal benchmarks show that MCE consistently outperforms state-of-the-art methods under various missing configurations. The final published version is now available at https://doi.org/10.1016/j.patcog.2025.112591. Our code is available at https://github.com/byzhaoAI/MCE.

cs.CV↗

arXiv2025

GNN-based Anchor Embedding for Efficient Exact Subgraph Matching

Subgraph matching query is a fundamental problem in graph data management and has a variety of real-world applications. Several recent works utilize deep learning (DL) techniques to process subgraph matching queries. Most of them find approximate subgraph matching results without accuracy guarantees. Unlike these DL-based inexact subgraph matching methods, we propose a learning-based exact subgraph matching framework, called \textit{graph neural network (GNN)-based anchor embedding framework} (GNN-AE). In contrast to traditional exact subgraph matching methods that rely on creating auxiliary summary structures online for each specific query, our method indexes small feature subgraphs in the data graph offline and uses GNNs to perform graph isomorphism tests for these indexed feature subgraphs to efficiently obtain high-quality candidates. To make a tradeoff between query efficiency and index storage cost, we use two types of feature subgraphs, namely anchored subgraphs and anchored paths. Based on the proposed techniques, we transform the exact subgraph matching problem into a search problem in the embedding space. Furthermore, to efficiently retrieve all matches, we develop a parallel matching growth algorithm and design a cost-based DFS query planning method to further improve the matching growth algorithm. Extensive experiments on 6 real-world and 3 synthetic datasets indicate that GNN-AE is more efficient than the baselines, especially outperforming the exploration-based baseline methods by up to 1--2 orders of magnitude.

cs.SI↗

arXiv2025

Can Uncertainty Quantification Improve Learned Index Benefit Estimation?

Index tuning is crucial for optimizing database performance by selecting optimal indexes based on workload. The key to this process lies in an accurate and efficient benefit estimator. Traditional methods relying on what-if tools often suffer from inefficiency and inaccuracy. In contrast, learning-based models provide a promising alternative but face challenges such as instability, lack of interpretability, and complex management. To overcome these limitations, we adopt a novel approach: quantifying the uncertainty in learning-based models' results, thereby combining the strengths of both traditional and learning-based methods for reliable index tuning. We propose Beauty, the first uncertainty-aware framework that enhances learning-based models with uncertainty quantification and uses what-if tools as a complementary mechanism to improve reliability and reduce management complexity. Specifically, we introduce a novel method that combines AutoEncoder and Monte Carlo Dropout to jointly quantify uncertainty, tailored to the characteristics of benefit estimation tasks. In experiments involving sixteen models, our approach outperformed existing uncertainty quantification methods in the majority of cases. We also conducted index tuning tests on six datasets. By applying the Beauty framework, we eliminated worst-case scenarios and more than tripled the occurrence of best-case scenarios.

cs.DB↗

arXiv2025

Vbox: Efficient Black-Box Serializability Verification

Verifying the serializability of transaction histories is essential for users to know if the DBMS ensures the claimed serializable isolation level without potential bugs. Black-box serializability verification is a promising approach. Existing verification methods often have one or more limitations such as incomplete detection of data anomalies, long verification time, high memory usage, or dependence on specific concurrency control protocols. In this paper, a new black-box serializability verification method called \textsf{Vbox} is proposed. \textsf{Vbox} is powered by a number of new techniques, including the support for predicate database operations, comprehensive applications of transactions' time information in the verification process, and a simplified satisfiability (SAT) problem formulation and its efficient solver. In this paper, \textsf{Vbox} is verified to be correct, efficient, and capable of detecting more data anomalies, while not relying on any specific concurrency control protocols.

cs.PL↗

arXiv2023

BM2CP: Efficient Collaborative Perception with LiDAR-Camera Modalities

Collaborative perception enables agents to share complementary perceptual information with nearby agents. This would improve the perception performance and alleviate the issues of single-view perception, such as occlusion and sparsity. Most existing approaches mainly focus on single modality (especially LiDAR), and not fully exploit the superiority of multi-modal perception. We propose a collaborative perception paradigm, BM2CP, which employs LiDAR and camera to achieve efficient multi-modal perception. It utilizes LiDAR-guided modal fusion, cooperative depth generation and modality-guided intermediate fusion to acquire deep interactions among modalities of different agents, Moreover, it is capable to cope with the special case where one of the sensors, same or different type, of any agent is missing. Extensive experiments validate that our approach outperforms the state-of-the-art methods with 50X lower communication volumes in both simulated and real-world autonomous driving scenarios. Our code is available at https://github.com/byzhaoAI/BM2CP.

cs.CV↗

arXiv2023

Query Generation based on Generative Adversarial Networks

Many problems in database systems, such as cardinality estimation, database testing and optimizer tuning, require a large query load as data. However, it is often difficult to obtain a large number of real queries from users due to user privacy restrictions or low frequency of database access. Query generation is one of the approaches to solve this problem. Existing query generation methods, such as random generation and template-based generation, do not consider the relationship between the generated queries and existing queries, or even generate semantically incorrect queries. In this paper, we propose a query generation framework based on generative adversarial networks (GAN) to generate query load that is similar to the given query load. In our framework, we use a syntax parser to transform the query into a parse tree and traverse the tree to obtain the sequence of production rules corresponding to the query. The generator of GAN takes a fixed distribution prior as input and outputs the query sequence, and the discriminator takes the real query and the fake query generated by the generator as input and outputs a gradient to guide the generator learning. In addition, we add context-free grammar and semantic rules to the generation process, which ensures that the generated queries are syntactically and semantically correct. We conduct experiments to evaluate our approach on real-world dataset, which show that our approach can generate new query loads with a similar distribution to a given query load, and that the generated queries are syntactically correct with no semantic errors. The generated query loads are used in downstream task, and the results show a significant improvement in the models trained with the expanded query loads using our approach.

cs.DB↗

arXiv2020

Budget Constrained Interactive Search for Multiple Targets

Interactive graph search leverages human intelligence to categorize target labels in a hierarchy, which are useful for image classification, product categorization, and database search. However, many existing studies of interactive graph search aim at identifying a single target optimally, and suffer from the limitations of asking too many questions and not being able to handle multiple targets. To address these two limitations, in this paper, we study a new problem of budget constrained interactive graph search for multiple targets called kBM-IGS-problem. Specifically, given a set of multiple targets T in a hierarchy, and two parameters k and b, the goal is to identify a k-sized set of selections S such that the closeness between selections S and targets T is as small as possible, by asking at most a budget of b questions. We theoretically analyze the updating rules and design a penalty function to capture the closeness between selections and targets. To tackle the kBM-IGS-problem, we develop a novel framework to ask questions using the best vertex with the largest expected gain, which makes a balanced trade-off between target probability and benefit gain. Based on the kBM-IGS framework, we first propose an efficient algorithm STBIS to handle the SingleTarget problem, which is a special case of kBM-IGS. Then, we propose a dynamic programming based method kBM-DP to tackle the MultipleTargets problem. To further improve efficiency, we propose two heuristic but efficient algorithms kBM-Topk and kBM-DP+. kBM-Topk develops a variant gain function and selects the top-k vertices independently. kBM-DP+ uses an upper bound of gains and prunes disqualified vertices to save computations. Experiments on large real-world datasets with ground-truth targets verify both the effectiveness and efficiency of our proposed algorithms.

cs.DB↗

arXiv2018

QuickIM: Efficient, Accurate and Robust Influence Maximization Algorithm on Billion-Scale Networks

The Influence Maximization (IM) problem aims at finding k seed vertices in a network, starting from which influence can be spread in the network to the maximum extent. In this paper, we propose QuickIM, the first versatile IM algorithm that attains all the desirable properties of a practically applicable IM algorithm at the same time, namely high time efficiency, good result quality, low memory footprint, and high robustness. On real-world social networks, QuickIM achieves the $Ω(n + m)$ lower bound on time complexity and $Ω(n)$ space complexity, where $n$ and $m$ are the number of vertices and edges in the network, respectively. Our experimental evaluation verifies the superiority of QuickIM. Firstly, QuickIM runs 1-3 orders of magnitude faster than the state-of-the-art IM algorithms. Secondly, except EasyIM, QuickIM requires 1-2 orders of magnitude less memory than the state-of-the-art algorithms. Thirdly, QuickIM always produces as good quality results as the state-of-the-art algorithms. Lastly, the time and the memory performance of QuickIM is independent of influence probabilities. On the largest network used in the experiments that contains more than 3.6 billion edges, QuickIM is able to find hundreds of influential seeds in less than 4 minutes, while all the state-of-the-art algorithms fail to terminate in an hour.

cs.SI↗

arXiv2017

Diversified Coherent Core Search on Multi-Layer Graphs

Mining dense subgraphs on multi-layer graphs is an interesting problem, which has witnessed lots of applications in practice. To overcome the limitations of the quasi-clique-based approach, we propose d-coherent core (d-CC), a new notion of dense subgraph on multi-layer graphs, which has several elegant properties. We formalize the diversified coherent core search (DCCS) problem, which finds k d-CCs that can cover the largest number of vertices. We propose a greedy algorithm with an approximation ratio of 1 - 1/e and two search algorithms with an approximation ratio of 1/4. The experiments verify that the search algorithms are faster than the greedy algorithm and produce comparably good results as the greedy algorithm in practice. As opposed to the quasi-clique-based approach, our DCCS algorithms can fast detect larger dense subgraphs that cover most of the quasi-clique-based results.

cs.DB↗

arXiv2015

SimRank Computation on Uncertain Graphs

SimRank is a similarity measure between vertices in a graph, which has become a fundamental technique in graph analytics. Recently, many algorithms have been proposed for efficient evaluation of SimRank similarities. However, the existing SimRank computation algorithms either overlook uncertainty in graph structures or is based on an unreasonable assumption (Du et al). In this paper, we study SimRank similarities on uncertain graphs based on the possible world model of uncertain graphs. Following the random-walk-based formulation of SimRank on deterministic graphs and the possible worlds model of uncertain graphs, we define random walks on uncertain graphs for the first time and show that our definition of random walks satisfies Markov's property. We formulate the SimRank measure based on random walks on uncertain graphs. We discover a critical difference between random walks on uncertain graphs and random walks on deterministic graphs, which makes all existing SimRank computation algorithms on deterministic graphs inapplicable to uncertain graphs. To efficiently compute SimRank similarities, we propose three algorithms, namely the baseline algorithm with high accuracy, the sampling algorithm with high efficiency, and the two-phase algorithm with comparable efficiency as the sampling algorithm and about an order of magnitude smaller relative error than the sampling algorithm. The extensive experiments and case studies verify the effectiveness of our SimRank measure and the efficiency of our SimRank computation algorithms.

cs.DB↗