The lack of access to eviction records in many regions poses significant challenges to the effective and efficient implementation of eviction prevention and diversion programs across the USA. To help better identify in-need regions nationwide in the face of existing data access challenges, this study proposes a Machine Learning (ML) driven solution that leverages Airbnb data to monitor eviction status at various spatial resolutions. In particular, this paper first shows the possibility of predicting the eviction status from Airbnb data using various types of fully-supervised ML approaches. After establishing the value of Airbnb data for this task, we then develop an ML-driven solution to learn the eviction status from Airbnb data when ground-truth labels are not available in the training data. Experimental results demonstrate that our method achieves performance comparable to the best-performing fully supervised ML approach. Cross-region tests further examine the generalizability to areas not observed during model training. The code is available at https://github.com/maryam-tabar/Airbnb-Eviction 1 .
Video Individual Counting (VIC), which aims to enumerate unique individuals in a video sequence, is a vital problem with key applications in urban analytics, including transportation management and public safety. Existing methods often struggle with two interconnected challenges: scale ambiguity and motion diversity. The coexistence of diverse motion patterns and object scales challenges prior methods, which lack the flexibility to handle both simultaneously. To overcome these limitations, we propose the D ensity-assisted A daptive A lignment N etwork (DAANet), a novel framework that enables motion-adaptive and scale-aware alignment. DAANet introduces two key modules: (1) a Density-driven Feature Enhancement (DFE) module that exploits density maps to construct spatial priors, and (2) a Hierarchical Deformable Feature Alignment (HDFA) module that allows multi-level features to adaptively align individuals based on their scale and motion characteristics. These components work in concert to overcome the fundamental challenges of VIC, leading to more accurate feature alignments and improved individual count estimates. Experimental results on three public VIC benchmarks show that DAANet achieves state-of-the-art performance, outperforming prior methods and reducing Mean Squared Error (MSE) by up to 31.6%. Source code is publicly available at https://github.com/LeoHuang0511/DAANet.
The use of synthetic data is becoming increasingly widespread in a variety of applications. For hate speech detection, synthetic data offers the chance to overcome limitations of existing models and datasets, such as data scarcity, negative psychological impact on annotators, and privacy issues. However, previous work on the topic has only focused on evaluating synthetic datasets extrinsically, by measuring the performance of models trained on synthetic texts. In this work, in addition to measuring the performance of models trained on synthetic data in cross-distribution setups, we carry out a qualitative analysis, aiming at assessing how closely synthetic data can mimic authentic user-generated and human-annotated data. We focus in particular on possible issues related to (under)representation of specific minority groups as targets of hate in synthetic data. To do this, we augment an existing English hate speech dataset by paraphrasing the original gold data using three large language models, and then evaluate the synthetic data both quantitatively, including testing out-of-distribution model generalization, and qualitatively, examining the realism of synthetic texts, their hate speech content, and whether they represent the same identity groups as the original data. We pair this analysis with a detailed breakdown, for each demographic characteristic, of how the synthetic data creation process affects the representation of each specific type of identity, which is the main contribution of our work. We find that, while the performance of models trained on synthetic data is promising, these texts are easily identified as written by an LLM by humans, hate speech labels are not necessarily preserved, and the paraphrasing process tends to erase representation of minority groups and intersectional identities, affecting certain demographics more than others.
Explaining anomalies in spatiotemporal data requires more than identifying statistical deviations; it often demands coherent narratives that relate events to their spatial, temporal, and social context. In urban settings, where anomalies may reflect public events, behavioral shifts, or environmental factors, the lack of interpretable, contextual explanations limits the effectiveness of automated analysis. We present GroundSTory , an interactive framework that reframes spatiotemporal anomaly explanation as computational storytelling, integrating graph-based representations of urban space, coordinated visual analytics, and large language models (LLMs) to support the construction of grounded, human-readable narrative explanations. Anomalies act as narrative anchors, while localized spatial and temporal relations and contextual signals derived from the graph guide and constrain narrative generation. The LLM functions as a narrative assistant, proposing plausible, evidence-backed hypotheses that analysts can inspect, compare, and validate through visual interaction. We illustrate the approach through qualitative usage scenarios on real-world urban datasets from São Paulo and New York City, showing how interactive narratives surface meaningful links between detected anomalies and contextual factors such as holidays, mass events, and routine activity patterns. A prompt analysis further suggests that structured, graph-derived context tends to affect the stability and perceived quality of generated hypotheses. Overall, our results indicate that combining knowledge-driven representations, interactive visualization, and constrained prompting can support more coherent, plausible, and interpretable narratives for spatiotemporal anomaly analysis, contributing to ongoing research in interactive storytelling and narrative-centered AI systems.
Modeling user intent across heterogeneous domains remains a central challenge in cross-domain sequential recommendation (CDSR). Although cross-domain sequences naturally embed richer behavioral cues that traverse multiple domains, existing approaches often struggle to disentangle their distinctive effects from those of single-domain interactions. To address this limitation, this study proposes a novel user modeling framework that explicitly decomposes user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive representations. The proposed method employs a triple-encoder architecture, where each encoder is specialized to capture one of these preference types. This is achieved through domain masking objectives and adversarial learning with a gradient reversal layer to promote balanced separation. In addition, representational alignment and mutual independence are jointly enforced to achieve a clean disentanglement among the components. To further enhance temporal adaptability, a dynamic gating mechanism is incorporated to modulate the relative influence of each preference type at every timestep, culminating in a unified session-level representation. Extensive experiments conducted on two large-scale real-world benchmarks confirm that the proposed approach not only enhances recommendation accuracy but also yields interpretable insights into the interplay between domain-specific and cross-domain preference signals.
