In the era where data are often referred to as the new oil, entity resolution (ER) plays a pivotal role in identifying and linking records that represent the same real-world entity across disparate sources. While existing surveys emphasize algorithmic advancements, they largely overlook how database architectures fundamentally shape ER technique selection and performance. This paper presents a comprehensive, database-centric review of ER techniques across SQL and NoSQL Database Management Systems. We systematically classify ER techniques using a three-dimensional framework encompassing data model characteristics, system architecture, and target workloads. Through structured taxonomies, we demonstrate how ER techniques—from traditional rule-based methods to advanced deep learning methods—adapt their implementation and optimization strategies across different paradigms. The review culminates in actionable guidelines that enable practitioners to select appropriate ER techniques aligned with their specific data infrastructure, bridging the gap between algorithmic innovation and practical, database-aware deployment.
Multi-vector dense retrieval methods have set the state-of-the-art retrieval quality on ranking tasks. However, the superior effectiveness of such methods comes with substantial computational overhead, which makes their deployment in large-scale search engines infeasible. As a remedy, we propose a novel approach, Sparcent, that leverages learned sparse retrieval (LSR) to accelerate the query-centroid matching stage in multi-vector dense retrieval. Specifically, we first derive sparse centroid representations for multi-vector term-embedding clusters and then prune them, producing compact centroid vectors that remain both efficient and effective for the indexing via LSR methods. Next, we integrate Sparcent with two well-known multi-vector retrieval systems, PLAID and WARP, already optimized for single-core CPU search. Our experimental results show that Sparcent provides up to 4.6 × speed-up over the original PLAID and 1.3 × speed-up over WARP, respectively, with virtually zero additional memory, while maintaining result effectiveness on both in-domain and out-of-domain retrieval tasks.
In multi-turn conversations, insufficient knowledge reasoning and ambiguity caused by missing entities often hinder the system’s ability to effectively leverage conversation history, resulting in inaccurate responses. To address this issue, this paper proposes a Topic Semantic Relational Knowledge (TSR) model that integrates external knowledge to resolve ambiguity and enhance dialogue comprehension. The TSR model first retrieves historical dialogue utterances based on the semantics of the current conversation, filters them to extract topic-relevant history utterances, and combines them with the current dialogue to form topic semantic information. This information is subsequently used for semantic matching with external knowledge to strengthen reasoning. Additionally, a Global Knowledge Attention Layer (GKA) is introduced to fully exploit knowledge by capturing richer knowledge representations through multilayer feature fusion. Experiments conducted on the QuAC, CoQA, and DoQA datasets demonstrate that the TSR model achieves state-of-the-art performance in both F1 and Human Equivalence Score at the Dialog level (HEQ-D) metrics. The TSR model shows significant improvements in handling multi-turn conversations, leading to more accurate and contextually relevant responses.
Traditional deterministic algorithms for mining high utility itemsets in data streams often run into memory explosion problems and can easily run out of memory with large, dense datasets. Most current metaheuristic methods just adapt generic swarm-intelligence frameworks and miss out on customized mechanisms for sliding window data stream scenarios, which can lead to itemset loss, slow convergence, and lots of redundant fitness evaluations from duplicate candidate itemsets. So, this paper picks cuckoo search as the base metaheuristic population search framework and develops three sets of cooperative specialized strategies for mining high utility itemsets in data streams, leading to the SHUIM-CS algorithm. First, for the non-differentiable nature of the binary discrete itemset search space, a discrete pseudo-gradient position update strategy based on utility differences is designed, with bit flipping used to evolve individuals, cutting down itemset loss in metaheuristic algorithms and speeding up population convergence. Next, a BARS block-based circular buffer index structure is built, integrated with the SWIA sliding window incremental index tree, using arrays to circularly store transaction blocks and update window utility incrementally at a logarithmic scale, easing memory bottlenecks in large, dense data streams. Finally, an HRM result mapping strategy using double-hash probing is proposed to remove duplicate candidate itemsets and cut out redundant fitness calculations. Experiments on multiple real and synthetic datasets compare metaheuristic and deterministic baselines, and ablation experiments verify the effectiveness of the three core strategies. Plus, theoretical analyses on pruning safety, hash mechanism performance, and algorithm complexity are included. The results show SHUIM-CS has clear memory advantages and is way more suitable for large-scale data stream mining when memory is limited.
Live-streaming e-commerce is rapidly reshaping retail supply chains, creating an environment where traditional inventory planning methods fall short. The primary challenge is the high-stakes nature of pre-session inventory allocation; misjudging the demand for a single, hours-long event can lead to significant lost sales from stockouts or capital waste from overstocking. Historical forecasting is often unreliable as the demand is heavily influenced by session-specific factors, including the session hosts, the product assortment, and expected user engagement. This study develops an integrated framework that begins with session-based demand forecasting, utilizing a rich blend of historical data and engagement-proxy features to accurately predict total demand for products within a given live-streaming event. Building on this framework, we propose a dynamic inventory orchestration mechanism. This mechanism uses session-specific forecasting to automate and optimize the crucial pre-session task of allocating inventory from a central supply to the live-streaming event, ensuring the right amount of stock is in the right place at the right time. The model’s core forecasting engine was empirically validated on a synthetic dataset of 500 live-streaming sessions derived from public Taobao and Twitch sources. The arithmetic-mean ensemble of 10 independently trained Gradient Boosting Machines achieved the highest accuracy, with a mean R-squared of 0.9364 and the lowest standard deviation across 20 independent runs. Furthermore, we enhance the inventory orchestration stage with a tabular Q-learning agent that learns optimal ordering policies directly from demand sequences, converging within approximately 900 episodes and achieving a 93.3
