Deepfakes, visual content created or modified by AI algorithms are increasingly being used to spread misinformation. Consequently, social media platforms are now labelling AI-content. To implement deepfake detectors in a way that permits user acceptance and trust in the decision, it is important to provide explanations of the model decision that align with user preferences. In order to better understand these user preferences, we will compare user preferences along the dimensions of acceptance and trust between (C1) just the output label, (C2) model output + feature importance mask over regions of the image that are salient to model output, (C3) model output + natural language explanation of image decision, and (C4) model output + mask + natural language explanation. Preferences will be measured using Likert scale and statistical significance performed via Friedman test. We find that the presence of a natural language component (C3 and C4) produces clearer explanations, and increases user trust and acceptance over the baseline (C1) and mask (C2).
To deter SNS users from posting or reposting misinformation/disinformation, we develop a method to quantify each user’s degree of responsibility when misinformation/disinformation spreads. We utilize the Banzhaf index, a technique for quantitatively evaluating player contributions in participatory games, for this purpose. Although the computation of the Banzhaf index has exponential time complexity with respect to the number of users, we developed an algorithm computable in realistic time by limiting it to tree structures modeling information propagation in social networks. We also discuss challenges for practical implementation of the proposed method and possibilities for extending the current work.
Governance in data ecosystems relies on data policies to control how assets are accessed and used. However, automatically detecting conflicts and relationships between these policies is challenging, as they often rely on implicit knowledge that is not explicitly available. In this work, we propose a Neuro-Symbolic pipeline that discovers these relationships. Our approach uses Large Language Models (LLMs) to systematically generate logical rules that approximate these implicit semantic relationships into explicit logical rules executable by a deterministic reasoning engine. We include a proof-of-concept implementation demonstrating how this approach automatically discovers hidden inter-policy conflicts based on a set of real data policies, ensuring a systematic and auditable governance process.
Private set operations are fundamental building blocks for privacy-preserving data analytics and services. In outsourced computation settings, trusted execution environments (TEEs) enable encrypted data processing, but they remain vulnerable to side channel leakages through data-dependent memory access patterns and program control flow. Doubly oblivious algorithms mitigate against this threat by ensuring both data-independent control flow and memory access behavior. In this work, we investigate the application of a doubly oblivious radix-partitioning technique, originally proposed for oblivious join operations, to the problems of Private Set Intersection (PSI) and Private Set Union (PSU). Particularly, for cases where the operations are between a client’s set and a server’s set, where the server’s set is much bigger than the client’s, we explore the possibility of processing multiple client requests at once to amortize the cost of the expensive operations. We propose algorithms for PSI and PSU using the doubly oblivious radix-partitioning and extend them to handle multi-set operations, which will enable processing multiple requests at once.
Schema matching is a core task in data integration that aims to identify semantic correspondences between elements of different schemas. Recent work shows that Large Language Models (LLMs) achieve promising results even when only the names and descriptions of schema elements are available. An additional advantage of LLMs is that they can provide natural-language explanations for their predictions. However, the informal nature of these explanations prevents automatic verification, requiring their correctness to be assessed by domain experts. In this paper, we conduct an initial investigation into grounding explanations for LLM-generated schema matchings as formal explanation expressions in a logical framework. These explanatory expressions consist of similarity predicates that capture both syntactic and semantic relationships among metadata elements. Predicted correspondences are hence compiled into machine-checkable explanation expressions that serve both as a basis for automatic verification and as concise, standardized artefacts that enable humans to inspect and understand matching decisions. We empirically explore the effectiveness of our approach on two healthcare benchmarks, providing initial evidence of its potential.
The Single Instruction Multiple Data (SIMD) paradigm has become a popular part of the optimization of query processing in state-of-the-art column-stores. Recent additions to the SIMD instruction sets offer matrix operations, thereby going beyond the vector operations offered by previous instruction sets. A main feature which is available in all of these instruction sets is the matrix multiplication. Although algebra of matrices is not commonly used in analytical query processing of relational data, there is a set of well-defined relational operators relying heavily on matrix multiplication. This definition uses relations represented as associative arrays which can serve as a common representation for multiple formats, e.g., graphs and (key, value)-pairs. In this work, we realize a selection of these operator definitions using Intel’s most recent instruction sets for vector and matrix processing (AVX512 and AMX). For this purpose, we identify the code sections where AMX resp. AVX512 is useful, provide a short outline of our implementation, and create test datasets with different properties. Then, we elaborate on the contribution of each code section to the execution time of complete operators, and examine the benefits of hardware-accelerated matrix multiplication for our use-case.
We study the emerging task of cross-schema SQL mapping, where large language models (LLMs) translate SQL queries across databases with different schemas. This task tests whether models can preserve source query intent and structural properties while adapting to unfamiliar target schema elements. To understand the sources of difficulty, we analyze a dataset of over 100K mapped queries from SQL-Exchange, spanning two benchmarks and two LLMs. We define a taxonomy of more than 50 structural and schema-related features and identify three key signals—query complexity, schema mismatch, and mapping flexibility—that strongly influence execution correctness and semantic validity. Our analysis reveals how different forms of source-target schema mismatch drive model errors and highlights concrete directions for building more robust schema-schema mapping systems.
Artificial intelligence is reshaping the labor market, with effects that vary by sector, skill level, and gender.The literature paints an ambivalent picture. Artificial intelligence has favored women’s employment in some contexts, but they remain exposed to biased algorithmic systems, underrepresented in technology-intensive professions, and significantly less likely to adopt generative AI tools.This vision paper summarizes existing literature on the impact of AI on women in the labor market across three interconnected themes: the gendered impact of AI on employment and skill demand; gender bias embedded in AI systems; and the gender gap in generative AI adoption. We conclude by identifying priority areas for future research: establishing causal evidence on AI’s gendered employment effects, validating fair hiring systems, understanding how LLMs reproduce gendered responses and occupational labeling across sectors, and developing upstream approaches to bias mitigation in model training.
Hybrid Transactional and Analytical Processing (HTAP) systems unify transactional and analytical workloads within a single database engine, but they are not designed to efficiently support retrieval-centric workloads that increasingly underpin AI-enabled applications. These workloads include full-text search for keyword-based relevance retrieval and vector similarity search for embedding-based semantic retrieval. Existing solutions typically rely on external systems or tightly coupled storage-level indexes, resulting in data duplication, limited extensibility, and interference with core HTAP workloads.This paper presents TiRex (TiDB Retrieval Execution Engine), a framework that extends HTAP toward HTAIP (Hybrid Transactional, Analytical, and AI Processing) by natively supporting transactional, analytical, and retrieval workloads within a unified execution model. TiRex decouples retrieval-oriented indexing from database storage, constructing full-text and vector indexes asynchronously from transactional logs and persisting them independently. At query time, analytical queries and retrieval workloads are executed uniformly using an MPP execution engine, enabling shared scheduling and parallelism without impacting transactional processing.TiRex introduces a shard-based indexing and scheduling model and provides bounded-staleness consistency semantics tailored to retrieval workloads. Experimental results show that TiRex achieves scalable indexing throughput, low-latency full-text and vector search, and stable performance under mixed workloads. These results demonstrate that HTAIP is a practical system abstraction beyond traditional HTAP to support modern retrieval workloads alongside transactions and analytics.
Temporal order reconstruction is the task of inferring the original chronological order of events when explicit timestamps are partially missing or unreliable. In such cases, interpreting attribute value semantics is critical. Existing approaches have largely relied on declarative (static) rules to represent attribute value orderings, such as an employment status of ‘student’ must precede ‘working’ and then ‘retirement’. However, the time duration (or time gap) between these events varies among entities and their attribute semantics. For example, there exist different levels and interpretations of ‘student’ (under-graduate vs. doctorate). Static rules are unable to adapt to the diverse precedence relationships and attribute value semantics that exist across entities. Varying time gaps between events provide clues to compute temporal orders, yet they are often overlooked. We introduce a Semantic-driven Temporal Order Reconstruction (STOR) framework that uses semi-supervised learning to model varying time gaps among events, and adapts to evolving semantics across entities. STOR uses a graph-based model to learn relationships between events and their attribute values, while adapting to the changing semantics and context. It also integrates time-gap information via Key/Value Relative Position Encoding (RPE) to structurally guide the learning process. In addition, STOR combines pseudo labeling, consistency regularization, and cycle consistency to effectively learn the intrinsic structure of the data. Through this approach, it achieves robust order reconstruction even in low-resource environments. Our preliminary evaluation shows that STOR achieves an 8.7% gain in F1-score compared to supervised baselines with only 5% labeled data. Furthermore, it demonstrates superior robustness in low-resource settings by maintaining a significantly lower performance variance (±0.018 vs. ±0.053) across different initializations. Importantly, the gap-encoding mechanism serves as a strong training scaffold, enabling the model to retain its performance advantage even when timestamps are entirely absent at inference time.
Designing inclusive urban spaces that improve personal well-being requires decision processes that are both data-driven and socially accountable, yet current analytics remain largely expert-centric and often ignore how data are unevenly distributed across neighbourhoods and communities. This paper proposes a relational, human-centred framework and platform vision that treats data distribution as a first-class object for urban design, linking heterogeneous urban datasets to well-being objectives, constraints, and evaluation metrics. We combine data science with computational creativity by modelling creativity as a relational process involving both human and non-human agents, and by supporting co-creative scenario generation and refinement through conversational interaction. To ensure transparency and reproducibility, the framework embeds provenance-aware workflows that capture data lineage, modelling assumptions, and decision rationales. Finally, we outline multi-objective assessment methods that expose trade-offs among well-being, equity, sustainability, and cost under uncertainty, enabling stakeholders to compare feasible interventions rather than rely on single-point recommendations. The result is a foundation for democratizing the design, simulation, and assessment of urban solutions across diverse city contexts, with equitable water distribution as a motivating and high-impact application domain.
Urban traffic noise is a major environmental and public well-being concern, yet traditional physics-based simulation models require extensive data collection and complex calibration. This paper presents an open-data machine learning pipeline for city-scale traffic noise prediction and conducts initial experiments on the transferability of learned models across urban environments. A dataset for noise classification in the cities of Tampere and Oulu was constructed using geospatial features, including road networks, building characteristics, land use, elevation, and traffic counters, together with noise labels obtained from official municipal noise maps. Tree-based and neural network machine learning models were assessed under both random and systematic station-based train-test splits to examine real-world generalization. Also, initial experiments were conducted to explore the feasibility of transferring noise assessment knowledge between cities and traffic stations using instance-based transfer learning (TrAdaBoost), adversarial feature adaptation (DANN), and a tabular foundation model (TabPFN). Preliminary results indicate that selected off-the-shelf domain-adaptation methods struggle to achieve strong performance in our settings, whereas the tabular foundation model successfully leverages small labeled target subsets. The findings demonstrate the potential and limitations of open-data machine learning for urban noise mapping and suggest that task-adaptive foundation models could offer a promising direction for noise assessment in data-sparse cities.
Flash-based SSDs have become the dominant storage medium for modern database management systems. Although SSDs provide high random read throughput, sustained writes can generate greater internal device overhead than reads due to flash management activities, even when individual write requests are acknowledged quickly. This asymmetry creates performance challenges for checkpointing, a core DBMS mechanism that bounds crash recovery time and controls log growth. Specifically, in write-intensive OLTP workloads, checkpointing generates substantial background write traffic, which can become a bottleneck on SSDs. Additionally, a small number of hot pages often remain dirty across many checkpoint intervals and are repeatedly selected for checkpoint flushing, leading to redundant writeback that yields little durable progress while consuming device bandwidth and increasing write-induced interference. These redundant writes exacerbate the performance issues by causing noticeable latency that degrades transactional throughput and overall system efficiency. To better align checkpointing with SSD characteristics, we present a novel hot-page-aware checkpointing technique that reduces redundant writes, ultimately improving SSD endurance and overall DBMS performance. Our lightweight per-page mechanism tracks consecutive checkpoint selections and defers unnecessary flushes. We implement this in MySQL/InnoDB and evaluate it using TPC-C workloads. Experiment results show that it improves transaction throughput by 1.58× compared to Vanilla MySQL.
Automatic index selection is a critical task for optimising query performance in distributed database systems. This problem is NP-hard, and existing approaches based on heuristic methods or optimisation problems struggle to generate good recommendations within a reasonable timeframe for large-scale database applications. While using classical deep reinforcement learning with Deep Q-Networks (DQN) has shown promise, its scalability in high-dimensional state spaces remains a challenge. In this paper, we explore the application of hybrid quantum-classical deep reinforcement learning, specifically a Quantum Deep Q-Network (Q-DQN), to the divergent index tuning problem. Using the TPC-H benchmark on a varying number of database replicas, we compare the performance of classical divergent index tuning against a hybrid quantum-classical algorithm using a Q-DQN implemented on the IBM Qiskit Aer simulator. We find that the hybrid quantum-classical algorithm exhibits remarkably higher sample efficiency, converging in significantly fewer action steps while maintaining a comparable query execution time. Overall, our findings suggest that variational quantum circuits provide an effective representation for deep reinforcement learning, achieving faster policy convergence, and that when appropriately configured, this improved sample efficiency can be obtained without loss in query execution performance.
Vision-based urban street object detection is a crucial component of modern smart city systems leveraging urban visual data and image machine-learning techniques to localize objects of interest, such as homeless encampments (static objects) and freight trucks (dynamic objects). However, issues like high intra-class variations, image-based noise conditions such as occlusion and truncation, and the lack of large, well-labeled visual datasets hinder progress in urban street object detection. Semi-supervised object detection (SSOD) methods offer a promising solution by leveraging small amount of labeled data with large-scale unlabeled data to provide a performance comparable to, and in some cases surpassing, supervised baselines. Despite this potential, most SSOD algorithms have been evaluated primarily on relatively clean benchmark datasets such as MS-COCO and PASCAL VOC, limiting their applicability to real-world urban imagery that exhibits significant visual noise. This study addresses this gap by presenting empirical results from two diverse real-world urban datasets focused on homeless encampments and freight trucks. Using limited labeled data and large-scale unlabeled visual data, we evaluate the performance, robustness, and practical feasibility of SSOD models for smart city applications. Our experiments highlight the generalizability of SSOD for urban street image detection and analyze the relative impact of different noise conditions on SSOD performance.
Transformer-based language models convert DNA sequences into high-dimensional embeddings, enabling powerful genomic analysis. These embeddings retain sufficient genetic information to pose serious privacy risks through reconstruction attacks. In this work, we address the challenge of balancing utility and genome embedding privacy along three dimensions: (1) Systematic Vulnerability Assessment: we analyze single-nucleotide reconstruction attacks across 13+ transformer architectures, including general-purpose models (BERT, GPT-2, RoBERTa, XLNet) and genomic specialized architectures (DNABERT, Nucleotide Transformer, DNAGPT). We demonstrate that fine-tuning can enhance rather than degrade privacy (e.g., XLNet-Large achieves a statistically significant 19.5% privacy improvement while simultaneously improving downstream utility). (2) Advanced Evaluation Metrics: we introduce novel metrics (e.g., privacy change measurement) that capture distinct aspects of information leakage, revealing different vulnerability patterns across all model families. (3) Privacy Preservation: we propose a Vulnerability-Enhanced Selective Privacy Adaptation (VESPA), achieving 57–64% attack reduction while maintaining over 96% utility through Adaptive Dynamic Adversarial Perturbation (ADAP).This work provides a foundation for secure deployment of genomic transformers in clinical and research settings.
With the rapid advancement of AI and digital technologies, content generation and distribution have become increasingly accessible, leading to a surge in redundant and duplicate content across platforms. This paper investigates the problem of near-duplicate subsequence alignment, which aims to identify subsequences (e.g., text spans or video clips) in a dataset that are similar to a given query (e.g., a potentially plagiarized passage or short video). Traditional techniques primarily target textual data and rely on lexical similarity, making them ineffective for detecting semantically similar content with different surface forms (e.g., rephrasing by large language models) or for handling non-textual modalities such as audio and video. To address these limitations, we propose representing data as sequences of high-dimensional vectors by first chunking the content and encoding each chunk into a feature vector. We define a novel edit cost to quantify the similarity between two sequences of vectors. Building on this representation, we design a filter-and-refine-based framework that integrates state-of-the-art approximate nearest neighbor search (ANNS) algorithms to efficiently identify candidate subsequences. We introduce four filtering techniques that balance pruning power and computational cost. Extensive experiments on two real-world datasets demonstrate that our approach significantly outperforms existing methods.
Urban spatial question answering requires translating everyday questions about places, distances, and regions into executable queries over spatial databases. We formulate this task as natural-language to geospatial structured query translation (NL2GeoSQL). In this work, we present CityChat, an LLM-driven NL2GeoSQL framework for urban spatial question answering. CityChat improves reliability by combining schema-aware prompting with template-guided generation that constrains outputs to valid query shapes, and by using execution feedback to repair hard database errors. We evaluate CityChat on a multiple-city Southern California PostGIS workload across multiple LLM backbones. Results show that CityChat substantially increases executability and error recovery, and remains effective even for lightweight local models, offering a practical path toward reliable natural-language spatial querying at the urban scale.
Mobile app reviews provide valuable feedback for developers and policymakers, but their informal and noisy nature poses significant challenges for sentiment analysis systems. In particular, the frequent use of emojis and slang expressions can distort the semantic interpretation of user opinions and negatively impact classification performance. This paper investigates the effect of emoji and slang normalization on BERT-based sentiment analysis of mobile app reviews. Using a dataset of Instagram and WhatsApp reviews labeled through star ratings, we compare a baseline BERT model with enhanced variants incorporating emoji normalization, slang normalization, and their combination. Experimental results show that normalization improves overall classification performance, achieving an increase of approximately 3–4 percentage points in macro F1-score compared to the baseline. The findings suggest that handling informal digital expressions reduces noise and improves model robustness when analyzing user-generated content. This study highlights the importance of preprocessing strategies for more reliable sentiment analysis in real-world app review environments and provides insights for developing more inclusive and adaptable text analytics systems.
Enterprises increasingly generate vast amounts of data across communication and collaboration platforms such as Slack and Microsoft Teams, where employees informally discuss and resolve issues in domains such as IT, HR, Finance, CX, etc. Data from these conversational platforms remains largely untapped for enterprise knowledge automation. These conversational platforms are rich in expert-generated resolutions, which means that vital support knowledge remains locked in different conversational channels. In this paper, we propose Conversational Resolution Extraction (CORE), a pipeline designed to automatically extract, structure, and retrieve support resolutions from conversational enterprise platforms such as Slack, MS Teams. It leverages LLMs to convert chat interactions into retrievable resolution artifacts, named resolution from conversations (REC), and vector-based resolution retrieval using HNSW. We also discuss the design tradeoffs of LLMs and vector databases in resolution extraction systems under the constraints of latency, recall, and consistency in enterprise environments. We assess CORE using key business metrics and found a 52% reduction in Mean Time to Resolution (MTTR) and 19% improvement in discovery of resolution topics. Our work aligns with the workshop’s focus on LLMs and vector databases, which are increasingly central to scalable enterprise knowledge automation.
