Composed Image Retrieval (CIR) enables editable image search by integrating a query pair—a reference image ref and a textual modification mod—to retrieve a target image tar that reflects the intended change. While existing CIR methods have shown promising performance using well-annotated triplets ⟨ref,mod,tar⟩, almost all of them implicitly assume these triplets are accurately associated with each other. In practice, however, this assumption is often violated due to the limited knowledge of annotators, inevitably leading to incorrect textual modifications and resulting in a practical yet less-touched problem: noisy triplet correspondence (NTC). To tackle this challenge, we propose a Task-oriented Modification Enhancement framework (TME) to learn robustly from noisy triplets, which comprises three key modules: Robust Fusion Query (RFQ), Pseudo Text Enhancement (PTE), and Task-Oriented Prompt (TOP). Specifically, to mitigate the adverse impact of noise, RFQ employs a sample selection strategy to divide the training triplets into clean and noisy sets, thus enhancing the reliability of the training data for robust learning. To further leverage the noisy data instead of discarding it, PTE unifies the triplet noise as an adapter mismatch problem, thereby adjusting mod to align with ref and tar in the mismatched triplet. Finally, TOP replaces ref in the clean set with a trainable prompt, which is then concatenated with mod to form a query independent of the visual reference, aiming to mitigate visually irrelevant noise. Extensive experiments on two domain-specific datasets demonstrate the robustness and superiority of TME, particularly in noisy scenarios.

Trajectory data analysis benefits numerous real-world applications and has attracted substantial attention from the research community. With the rapid proliferation of IoT devices and the emergence of edge computing, there has been an increasing demand for efficient trajectory data analytics in edge environments. However, most existing trajectory analysis systems are designed for cloud-based architectures, which face significant limitations in edge settings. These include resource constraints, dynamic network conditions, and inefficient query handling, leading to sub-optimal performance in edge scenarios. To fill this gap, we propose TrajEdge, an efficient and lightweight framework for trajectory data analysis in edge environments. Implementing TrajEdge requires overcoming obstacles posed by limited resources and the dynamic nature of edge networks. To achieve this, we design a novel trajectory composite compression algorithm that delivers high compression ratios, significantly reducing storage pressure on edge devices. Additionally, we introduce three coflow control strategies optimized for varying network conditions, enabling higher system throughput. To further enhance the efficiency of trajectory queries, we develop a spatiotemporal-aware trie-based peer-to-peer (P2P) index. Experimental evaluations on two real-world datasets and one larger synthetic dataset demonstrate that TrajEdge achieves remarkable performance improvements: more than 200 × gains in storage and query efficiency, up to 64% increases in network throughput, compression ratios of up to 95%, and exceptional scalability compared to the state-of-the-art systems. Our source code is available at https://github.com/ZJU-DAILY/TrajEdge.

Although multi-view learning has achieved remarkable progress over the past decades, most existing methods implicitly assume that all views (or modalities) are well-aligned. In practice, however, collecting fully aligned views is challenging due to complexities and discordances in time and space, resulting in the Partially View-unaligned Problem (PVP), such as audio-video asynchrony caused by network congestion. While some methods are proposed to align the unaligned views by learning view-invariant representations, almost all of them overlook specific information across different views for complementarity, limiting performance improvement. To address these problems, we propose a robust framework, dubbed VariatIonal ConTrAstive Learning (VITAL), designed to learn both common and specific information simultaneously. To be specific, each data sample is first modeled as a Gaussian distribution in the latent space, where the mean estimates the most probable common information, while the variance indicates view-specific information. Second, by using variational inference, VITAL conducts intra- and inter-view contrastive learning to preserve common and specific semantics in the distribution representations, thereby achieving comprehensive perception. As a result, the common representation (mean) could be used to guide category-level realignment, while the specific representation (variance) complements sample semantic information, thereby boosting overall performance. Finally, considering the abundance of False Negative Pairs (FNPs) generated by unsupervised contrastive learning, we propose a robust loss function that seamlessly incorporates FNP rectification into the contrastive learning paradigm. Empirical evaluations on eight benchmark datasets reveal that VITAL outperforms ten state-of-the-art deep clustering baselines, demonstrating its efficacy in both partially and fully aligned scenarios. The Code is available at https://github.com/He-Changhao/2024-MM-VITAL.

The ubiquitous GPS sensors collect massive trajectory data from moving objects, which is useful in data mining applications. However, trajectory data is enormous in volume, and thus, directly storing and processing the raw data is expensive. Using trajectory simplification, a trajectory can be reduced to a set of continuous line segments with acceptable data loss, which is an efficient method. Although many algorithms are proposed, they still suffer from the following issues including (i) non-data driven capability as most studies rely on human-crafted rules or pre-defined parameters, (ii) bound with error measures that yield high computational cost, and (iii) focusing only on the local information preservation in trajectories, but failing in capturing the global mobility patterns for trajectory compression.To address the above issues, we propose a Seq2Seq2Seq framework, abbreviated S3, which consists of two chained Seq2Seq. With differentiable reconstruction learning, S3 enables self-supervised trajectory simplification in a lightweight manner. Besides, we deploy S3 over the graph neural architecture to capture the context-aware mobility patterns and enhance the representation paradigm of trajectories with geographical semantics, where a context-aware distance measure is designed for quality evaluation. An online extension of S3 is also developed to enable streaming trajectory simplifications. Finally, extensive experiments using two real-world datasets in both offline and online scenarios show that S3 achieves much higher efficiency (e.g., it achieves up to one order of magnitude speed-up gains) and comparable compression quality, compared with both non-learning and state-of-the-art learning-based methods.