Recent research has shown that single-vector embedding retrieval models face a fundamental bottleneck: the finite dimensionality of single-vector representations limits their capacity to represent arbitrary top-k relevant item combinations, even with perfect training. This inherent bottleneck substantially limits the expressive capacity of such models and reduces their ability to capture complex user-item interaction patterns. To overcome this bottleneck, we propose Adaptive Rich-kernelized Contrastive Learning (ARC), which enhances model expressiveness while maintaining the computational efficiency of single-vector retrieval. ARC replaces the fixed inner product with a learnable spherical kernel family parameterized by a truncated Gegenbauer expansion, thereby increasing the model's effective dimensionality while preserving single-vector efficiency and low-pass inductive bias for generalization. Specifically, we construct a positive-definite kernel on the unit sphere using a positive combination of Gegenbauer polynomials, and adopt a contrastive learning objective to jointly learn the polynomials weights and user–item embeddings. Through data-driven optimization, the adaptive kernel induces a more expressive representation space, enabling the model to better capture complex preferences. From a theoretical perspective, the learned kernel implicitly maps embeddings into a higher-dimensional Reproducing Kernel Hilbert Space, allowing the model to capture a broader range of top-k item combinations before reaching the geometric limit imposed by the embedding dimensionality. Consequently, the proposed ARC framework effectively alleviates the inherent representational bottleneck in traditional single-vector embedding models. Extensive experiments on four real-world datasets demonstrate the effectiveness of our method, showing consistent improvements over strong baselines and state-of-the-art models.

Collaborative filtering is a primary paradigm of modern recommender systems. A typical practice is to embed collaborative signals into a latent space and infer recommendation scores based on the similarities between user and item embeddings. Besides inter-type similarities (i.e., user-item relationships), intra-type similarities (i.e., user-user, and item-item) are also essential as they capture the intrinsic structure of users and items. However, many existing recommendation models only learn inter-type similarities using objectives like ranking loss or binary classification loss, while neglecting intra-type similarities. Consequently, the intrinsic structures of users and items are often distorted in the latent space, where users with similar historical interactions diverge more than those dissimilar. In this study, we show the importance of preserving the ordinal relations of intra-type similarities. We provide a theoretical analysis suggesting that preserving intra-type similarity rankings can enhance a model's generalizability and interpretability. In addition, we propose a regularization that enforces a constraint on the rankings of intra-type similarities, ensuring that learning inter-type similarities does not break intrinsic ordinal structures. It can be seamlessly integrated into most latent factor models and can be jointly trained with their original objectives. Extensive experiments on 4 benchmark datasets and 5 representative models show that our ordinal regularization can consistently improve recommendation performance, and enhance the intra-type similarity coherence in the latent space. The results also exhibit enhanced generalizability and interpretability of recommendations.

In many real-world dynamical systems, the effect of an input on an output is not immediate but occurs after a time-varying delay, influenced by both the input and the system's internal state. These delays are typically unobservable, making system modelling particularly challenging, as the temporal relationship between inputs and outputs is not directly accessible. While inferring such dynamic delays is crucial for accurate modelling, most existing approaches either assume fixed delays or disregard delay structure altogether. This often leads to misaligned input-output relationships, especially in nonlinear, state-dependent systems, ultimately degrading predictive performance. We propose DelayNetODE, a novel delay-aware neural architecture that infers dynamic, input-dependent delays in an unsupervised manner while modelling system behaviour in continuous time. The framework consists of three key components: (1) a graph-based encoder that estimates continuous-valued delays from local temporal features, (2) a soft attention mechanism that aligns past inputs based on inferred delays, and (3) a Neural Ordinary Differential Equation (Neural ODE) decoder that models the continuous-time evolution of the output conditioned on delay-adjusted inputs. We evaluate DeiayNetODE on both synthetic and real-world datasets with nonlinear, time-varying delays. Results show this approach consistently outperforms the state-of-the-art models in both prediction accuracy and delay estimation. By inferring delays in an unsupervised manner and aligning inputs accordingly, DelayNetODE provides a principled and explainable approach for capturing latent temporal dependencies and improving system modelling in complex dynamical settings.

In the current social media landscape, the study of influence propagation and consensus formation has gained prominence. While user interactions like retweeting are apparent, the underlying pathways of influence often remain hidden and complex. This study proposes a novel network called Latent Influence Network (LIN), which advances the analysis of influence on social media. LIN's architecture and the process of parameter selection are meticulously discussed within the comprehensive Latent Influence Detection Framework (LIDET). Based on the user's behavior label, LIN identifies the optimal network configuration, revealing more accurate influence patterns. We applied the LIDET framework to four diverse datasets, each demonstrating substantial improvements in influence pattern recognition over traditional network models. Specifically, in a case study on a COVID-19 dataset, LIN achieved a classification accuracy of 99%, significantly outperforming conventional methods. These findings underscore the utility of LIN in capturing the dynamics of influence and enhancing our understanding of opinion formation on social media.

We often compare time-dependent data elastically such that some compression or dilation along the time dimension can be ignored, for example, spatial trajectories of vehicles moving at different speeds or accelerometer data for exercises completed at variable rhythms. Traditionally this is possible via an alignment-based elastic distance measure, such as dynamic time warping (DTW). We may also control the degree of allowable warping with warping constraints. However, these elastic distance measures are not easy to use in large-scale time series classification, as they need to be evaluated pairwise and often cannot be directly converted into feature sets that we may use with arbitrary classifiers or combine with other features. In this research, we focus on the study of path signatures, a transformation with time warping invariance property, and how we may augment a time series to make its signature space representation reflect common warping constraints. We demonstrate that the comparing signatures is analogous to comparing time series with elastic distances, and that augmented signature features can serve as warping invariant or insensitive features in time series classification. Finally, we construct multiple path signatures with constraining augmentations classifier (MultiPSCA), a general-purpose minimal tuning time series classifier using augmented signatures and show that it is able to beat existing best-performing elastic time series classification algorithms without per-dataset hyperparameter tuning.

Multivariate time series classification (MTSC) has attracted significant research attention due to its diverse real-world applications. Recently, exploiting transformers for MTSC has achieved state-of the-art performance. However, existing methods focus on generic features, providing a comprehensive understanding of data, but they ignore class-specific features crucial for learning the representative characteristics of each class. This leads to poor performance in the case of imbalanced datasets or datasets with similar overall patterns but differing in minor class-specific details. In this paper, we propose a novel Shapelet Transformer (ShapeFormer), which comprises class-specific and generic transformer modules to capture both of these features. In the class-specific module, we introduce the discovery method to extract the discriminative subsequences of each class (i.e. shapelets) from the training set. We then propose a Shapelet Filter to learn the difference features between these shapelets and the input time series. We found that the difference feature for each shapelet contains important class-specific features, as it shows a significant distinction between its class and others. In the generic module, convolution filters are used to extract generic features that contain information to distinguish among all classes. For each module, we employ the transformer encoder to capture the correlation between their features. As a result, the combination of two transformer modules allows our model to exploit the power of both types of features, thereby enhancing the classification performance. Our experiments on 30 UEA MTSC datasets demonstrate that ShapeFormer has achieved the highest accuracy ranking compared to state-of-the-art methods. The code is available at https://github.com/xuanmay2701/shapeformer.

The emergence of social media as one of the main platforms for people to access news has enabled the wide dissemination of fake news, having serious impacts on society. Thus, it is really important to identify fake news with high confidence in a timely manner, which is not feasible using manual analysis. This has motivated numerous studies on automating fake news detection. Most of these approaches are supervised, which requires extensive time and labour to build a labelled dataset. Although there have been limited attempts at unsupervised fake news detection, their performance suffers due to not exploiting the knowledge from various modalities related to news records and due to the presence of various latent biases in the existing news datasets (e.g., unrealistic real and fake news distributions). To address these limitations, this work proposes an effective framework for unsupervised fake news detection, which first embeds the knowledge available in four modalities (i.e., source credibility, textual content, propagation speed, and user credibility) in news records and then proposes$(UMD)^{2}$, a novel noise-robust self-supervised learning technique, to identify the veracity of news records from the multi-modal embeddings. Also, we propose a novel technique to construct news datasets minimizing the latent biases in existing news datasets. Following the proposed approach for dataset construction, we produce a Large-scale Unlabelled News Dataset consisting 419,351 news articles related to COVID-19, acronymed asLUND-COVID. We trained the proposed unsupervised framework usingLUND-COVIDto exploit the potential of large datasets, and evaluate it using a set of existing labelled datasets. Our results show that the proposed unsupervised framework largely outperforms existing unsupervised baselines for different tasks such as multi-modal fake news detection, fake news early detection and few-shot fake news detection, while yielding notable improvements for unseen domains during training.

Many learning tasks involve multi-modal data streams, where continuous data from different modes convey a comprehensive description about objects. A major challenge in this context is how to efficiently interpret multi-modal information in complex environments. This has motivated numerous studies on learning unsupervised representations from multi-modal data streams. These studies aim to understand higher-level contextual information (e.g., a Twitter message) by jointly learning embeddings for the lower-level semantic units in different modalities (e.g., text, user, and location of a Twitter message). However, these methods directly associate each low-level semantic unit with a continuous embedding vector, which results in high memory requirements. Hence, deploying and continuously learning such models in low-memory devices (e.g., mobile devices) becomes a problem. To address this problem, we present METEOR, a novel MEmory and Time Efficient Online Representation learning technique, which: (1) learns compact representations for multi-modal data by sharing parameters within semantically meaningful groups and preserves the domain-agnostic semantics; (2) can be accelerated using parallel processes to accommodate different stream rates while capturing the temporal changes of the units; and (3) can be easily extended to capture implicit/explicit external knowledge related to multi-modal data streams. We evaluate METEOR using two types of multi-modal data streams (i.e., social media streams and shopping transaction streams) to demonstrate its ability to adapt to different domains. Our results show that METEOR preserves the quality of the representations while reducing memory usage by around 80% compared to the conventional memory-intensive embeddings.

Building spatiotemporal activity models for people's activities in urban spaces is important for understanding the ever-increasing complexity of urban dynamics. With the emergence of Geo-Tagged Social Media (GTSM) records, previous studies demonstrate the potential of GTSM records for spatiotemporal activity modeling. State-of-the-art methods for this task embed different modalities (location, time, and text) of GTSM records into a single embedding space. However, they ignore Non-GeoTagged Social Media (NGTSM) records, which generally account for the majority of posts (e.g., more than 95% in Twitter), and could represent a great source of information to alleviate the sparsity of GTSM records. Furthermore, in the current spatiotemporal embedding techniques, less focus has been given to the users, who exhibit spatially motivated behaviors. To bridge this research gap, this work proposes USTAR, a novel online learning method for User-guided SpatioTemporal Activity Representation, which (1) embeds locations, time, and text along with users into the same embedding space to capture their correlations; (2) uses a novel collaborative filtering approach to incorporate both NGTSM and GTSM records in learning; and (3) introduces a novel sampling technique to learn spatiotemporal representations in an online fashion to accommodate recent information into the embedding space, while avoiding overfitting to recent records and frequently appearing units in social media streams. Our results show that USTAR substantially improves the state-of-the-art for region retrieval and keyword retrieval and its potential to be applied to other downstream applications such as local event detection.

The supermarkets often use sales promotions to attract customers and create brand loyalty. They would often like to know if their promotions are effective for various customers, so that better timing and more suitable rate can be planned in the future. Given a transaction data set collected by an Australian national supermarket chain, in this paper we conduct a case study aimed at discovering customers' long-term purchase patterns, which may be induced by preference changes, as well as short-term purchase patterns, which may be induced by promotions. Since purchase events of individual customers may be too sparse to model, we propose to discover a number of latent purchase patterns from the data. The latent purchase patterns are modeled via a mixture of non-homogeneous Poisson processes where each Poisson intensity function is composed by long-term and short-term components. Through the case study, 1) we validate that our model can accurately estimate the occurrences of purchase events; 2) we discover easy-to-interpret long-term gradual changes and short-term periodic changes in different customer groups; 3) we identify the customers who are receptive to promotions through the correlation between behavior patterns and the promotions, which is particularly worthwhile for target marketing.