arXiv subjects

Dong Deng

Publications and source records attributed to Dong Deng.

11 recordsLinked to original sources

arXiv2026

Adaptive Multi-Objective Tiered Storage Configuration for KV Cache in LLM Service

The memory-for-computation paradigm of KV caching is essential for accelerating large language model (LLM) inference service, but limited GPU high-bandwidth memory (HBM) capacity motivates offloading the KV cache to cheaper external storage tiers. While this expands capacity, it introduces the challenge of dynamically managing heterogeneous storage resources to balance cost, throughput, and latency under varying workloads. We formulate this as a multi-objective optimization problem: identifying the Pareto frontier across these metrics within the storage configuration space. Using a high-fidelity end-to-end simulator, we observe that the objective functions are non-analytic and exhibit complex variable coupling, making the Pareto frontier difficult to approximate analytically. To obtain the frontier, we introduce Kareto, a KV-cache Adaptive REsource managemenT Optimizer. Kareto leverages a diminishing-return-guided pruning method to efficiently navigate the large configuration space and approximate the Pareto frontier. Additionally, it incorporates a fine-grained adaptive tuner that uses eviction policies in tier storage and KV block access patterns for group-specific cache management, improving cache efficiency. Experiments on real-world traces show that Kareto adapts to workload and can identify configurations of better cost efficiency, covering static strategies. Compared to the fixed setup with 1024 GB DRAM, Kareto can improve throughput by up to 9.3%, or reduce latency by up to 58.3%, or lower cost by up to 20.2% under respective optimization objectives.

cs.AR↗

arXiv2025

Near-Duplicate Text Alignment under Weighted Jaccard Similarity

Near-duplicate text alignment is the task of identifying, among the texts in a corpus, all the subsequences (substrings) that are similar to a given query. Traditional approaches rely on seeding-extension-filtering heuristics, which lack accuracy guarantees and require many hard-to-tune parameters. Recent methods leverage min-hash techniques under a hash-based framework: group subsequences by their min-hash, and for any query, find all sketches similar to the query's sketch. These methods guarantee to report all subsequences whose estimated unweighted Jaccard similarity with the query exceeds a user-provided threshold and are efficient. However, they fail to account for token importance or frequency, which limits their use in real scenarios where tokens carry weights, such as TF-IDF. To address this, we propose MONO, an approach that supports weighted Jaccard similarity using consistent weighted sampling. MONO achieves optimality within the hash-based framework. For example, when token weights are proportional to frequencies, MONO generates O(n + n log f) groups in expectation for a text of length n, where f is the maximum token frequency. Each group takes O(1) space and represents a few subsequences sharing the same sampling. We prove this bound is tight: any algorithm must produce Omega(n + n log f) groups in expectation in the worst case. Experiments show that MONO outperforms the state of the art by up to 26x in index construction time, reduces index size by up to 30 percent, and improves query latency by up to 3x, while scaling well.

cs.DB↗

arXiv2024

From Commands to Prompts: LLM-based Semantic File System for AIOS

Large language models (LLMs) have demonstrated significant potential in the development of intelligent applications and systems such as LLM-based agents and agent operating systems (AIOS). However, when these applications and systems interact with the underlying file system, the file system still remains the traditional paradigm: reliant on manual navigation through precise commands. This paradigm poses a bottleneck to the usability of these systems as users are required to navigate complex folder hierarchies and remember cryptic file names. To address this limitation, we propose an LLM-based semantic file system ( LSFS ) for prompt-driven file management. Unlike conventional approaches, LSFS incorporates LLMs to enable users or agents to interact with files through natural language prompts, facilitating semantic file management. At the macro-level, we develop a comprehensive API set to achieve semantic file management functionalities, such as semantic file retrieval, file update monitoring and summarization, and semantic file rollback). At the micro-level, we store files by constructing semantic indexes for them, design and implement syscalls of different semantic operations (e.g., CRUD, group by, join) powered by vector database. Our experiments show that LSFS offers significant improvements over traditional file systems in terms of user convenience, the diversity of supported functions, and the accuracy and efficiency of file operations. Additionally, with the integration of LLM, our system enables more intelligent file management tasks, such as content summarization and version comparison, further enhancing its capabilities.

cs.HC↗

arXiv2024

Vector Retrieval with Similarity and Diversity: How Hard Is It?

Dense vector retrieval is an important building block of modern machine learning systems, underlying applications ranging from semantic search to retrieval-augmented generation and knowledge-intensive reasoning. Beyond retrieving items that are individually similar to a query, many applications require a set of results that is also diverse, complementary, and collectively informative. Balancing similarity and diversity is therefore central to effective retrieval, but remains challenging to optimize in a stable and theoretically grounded way. Maximal Marginal Relevance (MMR) is a widely adopted heuristic for this problem, yet its reliance on a manually tuned parameter leads to optimization fluctuations and unpredictable retrieval results. More broadly, existing methods provide limited theoretical insight into how similarity and diversity interact in dense vector spaces, leaving the joint optimization problem insufficiently understood. To address these challenges, this paper introduces a novel approach that characterizes both constraints simultaneously by maximizing the similarity between the query vector and the sum of the selected candidate vectors. We formally define this optimization problem, Vector Retrieval with Similarity and Diversity (VRSD), and prove that it is NP-complete, establishing a rigorous theoretical bound on the inherent difficulty of this dual-objective retrieval. Subsequently, we present a parameter-free heuristic algorithm to solve VRSD. Extensive evaluations on multiple datasets, incorporating both objective geometric metrics and LLM-simulated subjective assessments, demonstrate that our VRSD heuristic consistently outperforms established baselines, including MMR and Determinantal Point Processes (k-DPP).

cs.IR↗

arXiv2024

Neural Locality Sensitive Hashing for Entity Blocking

Locality-sensitive hashing (LSH) is a fundamental algorithmic technique widely employed in large-scale data processing applications, such as nearest-neighbor search, entity resolution, and clustering. However, its applicability in some real-world scenarios is limited due to the need for careful design of hashing functions that align with specific metrics. Existing LSH-based Entity Blocking solutions primarily rely on generic similarity metrics such as Jaccard similarity, whereas practical use cases often demand complex and customized similarity rules surpassing the capabilities of generic similarity metrics. Consequently, designing LSH functions for these customized similarity rules presents considerable challenges. In this research, we propose a neuralization approach to enhance locality-sensitive hashing by training deep neural networks to serve as hashing functions for complex metrics. We assess the effectiveness of this approach within the context of the entity resolution problem, which frequently involves the use of task-specific metrics in real-world applications. Specifically, we introduce NLSHBlock (Neural-LSH Block), a novel blocking methodology that leverages pre-trained language models, fine-tuned with a novel LSH-based loss function. Through extensive evaluations conducted on a diverse range of real-world datasets, we demonstrate the superiority of NLSHBlock over existing methods, exhibiting significant performance improvements. Furthermore, we showcase the efficacy of NLSHBlock in enhancing the performance of the entity matching phase, particularly within the semi-supervised setting.

cs.IR↗

arXiv2019

Technical Report: Optimizing Human Involvement for Entity Matching and Consolidation

An end-to-end data integration system requires human feedback in several phases, including collecting training data for entity matching, debugging the resulting clusters, confirming transformations applied on these clusters for data standardization, and finally, reducing each cluster to a single, canonical representation (or "golden record"). The traditional wisdom is to sequentially apply the human feedback, obtained by asking specific questions, within some budget in each phase. However, these questions are highly correlated; the answer to one can influence the outcome of any of the phases of the pipeline. Hence, interleaving them has the potential to offer significant benefits. In this paper, we propose a human-in-the-loop framework that interleaves different types of questions to optimize human involvement. We propose benefit models to measure the quality improvement from asking a question, and cost models to measure the human time it takes to answer a question. We develop a question scheduling framework that judiciously selects questions to maximize the accuracy of the final golden records. Experimental results on three real-world datasets show that our holistic method significantly improves the quality of golden records from 70% to 90%, compared with the state-of-the-art approaches.

cs.DB↗

arXiv2018

Unsupervised String Transformation Learning for Entity Consolidation

Data integration has been a long-standing challenge in data management with many applications. A key step in data integration is entity consolidation. It takes a collection of clusters of duplicate records as input and produces a single "golden record" for each cluster, which contains the canonical value for each attribute. Truth discovery and data fusion methods, as well as Master Data Management (MDM) systems, can be used for entity consolidation. However, to achieve better results, the variant values (i.e., values that are logically the same with different formats) in the clusters need to be consolidated before applying these methods. For this purpose, we propose a data-driven method to standardize the variant values based on two observations: (1) the variant values usually can be transformed to the same representation (e.g., "Mary Lee" and "Lee, Mary") and (2) the same transformation often appears repeatedly across different clusters (e.g., transpose the first and last name). Our approach first uses an unsupervised method to generate groups of value pairs that can be transformed in the same way (i.e., they share a transformation). Then the groups are presented to a human for verification and the approved ones are used to standardize the data. In a real-world dataset with 17,497 records, our method achieved 75% recall and 99.5% precision in standardizing variant values by asking a human 100 yes/no questions, which completely outperformed a state of the art data wrangling tool.

cs.DB↗

arXiv2017

SilkMoth: An Efficient Method for Finding Related Sets with Maximum Matching Constraints

Determining if two sets are related - that is, if they have similar values or if one set contains the other - is an important problem with many applications in data cleaning, data integration, and information retrieval. A particularly popular metric that has been proposed is to measure the relatedness of two sets by treating the elements as vertices of a bipartite graph and calculating the score of the maximum matching pairing between elements. Compared to other metrics which require exact matchings between elements, this metric uses a similarity function to compare elements between the two sets, making it robust to small dissimilarities in elements and more useful for real-world, dirty data. Unfortunately, the metric suffers from expensive computational cost, taking O(n^3) time, where n is the number of elements in sets, for each set-to-set comparison. Thus for applications which try to search for all pairings of related sets in a brute-force manner, the runtime becomes unacceptably large. To address this challenge, we developed SilkMoth, a system capable of rapidly discovering related set pairs in collections of sets. Internally, SilkMoth creates a signature for each set, with the property that any other set which is related must match the signature. SilkMoth then uses these signatures to prune the search space, so only sets which match the signatures are left as candidates. Finally, SilkMoth applies the maximum matching metric on remaining candidates to verify which of these candidates are truly related sets. Thus, a contribution of this paper is the characterization of the space of signatures which enable this property. We show that selecting the optimal signature in this space is NP-complete, and based on insights from the characterization of the space, we propose two novel filters which help to prune the candidates further before verification.

cs.DB↗

arXiv2017

Error-Tolerant Big Data Processing

Real-world data contains various kinds of errors. Before analyzing data, one usually needs to process the raw data. However, traditional data processing based on exactly match often misses lots of valid information. To get high-quality analysis results and fit in the big data era, this thesis studies the error-tolerant big data processing. As most of the data in real world can be represented as a sequence or a set, this thesis utilizes the widely-used sequence-based and set-based similar functions to tolerate errors in data processing and studies the approximate entity extraction, similarity join and similarity search problems. The main contributions of this thesis include: 1. This thesis proposes a unified framework to support approximate entity extraction with both sequence-based and set-based similarity functions simultaneously. The experiments show that the unified framework can improve the state-of-the-art methods by 1 to 2 orders of magnitude. 2. This thesis designs two methods respectively for the sequence and the set similarity joins. For the sequence similarity join, this thesis proposes to evenly partition the sequences to segments. It is guaranteed that two sequences are similar only if one sequence has a subsequence identical to a segment of another sequence. For the set similarity join, this thesis proposes to partition all the sets into segments based on the universe. This thesis further extends the two partition-based methods to support the large-scale data processing framework, Map-Reduce and Spark. The partition-based method won the string similarity join competition held by EDBT and beat the second place by 10 times. 3. This thesis proposes a pivotal prefix filter technique to solve the sequence similarity search problem. This thesis shows that the pivotal prefix filter has stronger pruning power and less filtering cost compared to the state-of-the-art filters.

cs.DB↗

arXiv2017

A Technical Report: Entity Extraction using Both Character-based and Token-based Similarity

Entity extraction is fundamental to many text mining tasks such as organisation name recognition. A popular approach to entity extraction is based on matching sub-string candidates in a document against a dictionary of entities. To handle spelling errors and name variations of entities, usually the matching is approximate and edit or Jaccard distance is used to measure dissimilarity between sub-string candidates and the entities. For approximate entity extraction from free text, existing work considers solely character-based or solely token-based similarity and hence cannot simultaneously deal with minor variations at token level and typos. In this paper, we address this problem by considering both character-based similarity and token-based similarity (i.e. two-level similarity). Measuring one-level (e.g. character-based) similarity is computationally expensive, and measuring two-level similarity is dramatically more expensive. By exploiting the properties of the two-level similarity and the weights of tokens, we develop novel techniques to significantly reduce the number of sub-string candidates that require computation of two-level similarity against the dictionary of entities. A comprehensive experimental study on real world datasets show that our algorithm can efficiently extract entities from documents and produce a high F1 score in the range of [0.91, 0.97].

cs.DB↗

arXiv2011

PASS-JOIN: A Partition-based Method for Similarity Joins

As an essential operation in data cleaning, the similarity join has attracted considerable attention from the database community. In this paper, we study string similarity joins with edit-distance constraints, which find similar string pairs from two large sets of strings whose edit distance is within a given threshold. Existing algorithms are efficient either for short strings or for long strings, and there is no algorithm that can efficiently and adaptively support both short strings and long strings. To address this problem, we propose a partition-based method called Pass-Join. Pass-Join partitions a string into a set of segments and creates inverted indices for the segments. Then for each string, Pass-Join selects some of its substrings and uses the selected substrings to find candidate pairs using the inverted indices. We devise efficient techniques to select the substrings and prove that our method can minimize the number of selected substrings. We develop novel pruning techniques to efficiently verify the candidate pairs. Experimental results show that our algorithms are efficient for both short strings and long strings, and outperform state-of-the-art methods on real datasets.

cs.DB↗