Multivariate time series (MTS) forecasting involves predicting future time series data based on historical observations. Existing research primarily emphasizes the development of complex spatial-temporal models that capture spatial dependencies and temporal correlations among time series variables explicitly. However, recent advances have been impeded by challenges relating to data scarcity and model robustness. To address these issues, we propose Spatial-Temporal Masked Autoencoders (STMAE), an MTS forecasting framework that leverages masked autoencoders to enhance the performance of spatial-temporal baseline models. STMAE consists of two learning stages. In the pretraining stage, an encoder-decoder architecture is employed. The encoder processes the partially visible MTS data produced by a novel dual-masking strategy, including biased random walk-based spatial masking and patch-based temporal masking. Subsequently, the decoders aim to reconstruct the masked counterparts from both spatial and temporal perspectives. The pretraining stage establishes a challenging pretext task, compelling the encoder to learn robust spatial-temporal patterns. In the fine-tuning stage, the pretrained encoder is retained, and the original decoder from existing spatial-temporal models is appended for forecasting. Extensive experiments are conducted on multiple MTS benchmarks. The promising results demonstrate that integrating STMAE into various spatial-temporal models can largely enhance their MTS forecasting capability.
Graph Neural Networks (GNNs) have achieved state-of-the-art performance in recommender systems. Nevertheless, the process of searching and ranking from a large item corpus usually requires high latency, which limits the widespread deployment of GNNs in industry-scale applications. To address this issue, many methods compress user/item representations into the binary embedding space to reduce space requirements and accelerate inference. Also, they use the Straight-through Estimator (STE) to prevent vanishing gradients during back-propagation. However, the STE often causes the gradient mismatch problem, leading to sub-optimal results. In this work, we present the Hessian-aware Quantized GNN (HQ-GNN) as an effective solution for discrete representations of users/items that enable fast retrieval. HQ-GNN is composed of two components: a GNN encoder for learning continuous node embeddings and a quantized module for compressing full-precision embeddings into low-bit ones. Consequently, HQ-GNN benefits from both lower memory requirements and faster inference speeds compared to vanilla GNNs. To address the gradient mismatch problem in STE, we further consider the quantized errors and its second-order derivatives for better stability. The experimental results on several large-scale datasets show that HQ-GNN achieves a good balance between latency and performance.
Transformer and its variants are a powerful class of architectures for sequential recommendation, owing to their ability of capturing a user's dynamic interests from their past interactions. Despite their success, Transformer-based models often require the optimization of a large number of parameters, making them difficult to train from sparse data in sequential recommendation. To address the problem of data sparsity, previous studies have utilized self-supervised learning to enhance Transformers, such as pre-training embeddings from item attributes or contrastive data augmentations. However, these approaches encounter several training issues, including initialization sensitivity, manual data augmentations, and large batch-size memory bottlenecks. In this work, we investigate Transformers from the perspective of loss geometry, aiming to enhance the models' data efficiency and generalization in sequential recommendation. We observe that Transformers (e.g., SASRec) can converge to extremely sharp local minima if not adequately regularized. Inspired by the recent Sharpness-Aware Minimization (SAM), we propose SAMRec, which significantly improves the accuracy and robustness of sequential recommendation. SAMRec performs comparably to state-of-the-art self-supervised Transformers, such as S$^3$Rec and CL4SRec, without the need for pre-training or strong data augmentations.
Collaborative Filtering (CF) has been successfully used to help users discover the items of interest. Nevertheless, existing CF methods suffer from noisy data issue, which negatively impacts the quality of recommendation. To tackle this problem, many prior studies leverage adversarial learning to regularize the representations of users/items, which improves both generalizability and robustness. Those methods often learn adversarial perturbations and model parameters under min-max optimization framework. However, there still have two major drawbacks: 1) Existing methods lack theoretical guarantees of why adding perturbations improve the model generalizability and robustness; 2) Solving min-max optimization is time-consuming. In addition to updating the model parameters, each iteration requires additional computations to update the perturbations, making them not scalable for industry-scale datasets. In this paper, we present Sharpness-aware Collaborative Filtering (SharpCF), a simple yet effective method that conducts adversarial training without extra computational cost over the base optimizer. To achieve this goal, we first revisit the existing adversarial collaborative filtering and discuss its connection with recent Sharpness-aware Minimization. This analysis shows that adversarial training actually seeks model parameters that lie in neighborhoods around the optimal model parameters having uniformly low loss values, resulting in better generalizability. To reduce the computational overhead, SharpCF introduces a novel trajectory loss to measure the alignment between current weights and past weights. Experimental results on real-world datasets demonstrate that our SharpCF achieves superior performance with almost zero additional computational cost comparing to adversarial training.
Embedding learning transforms discrete data entities into continuous numerical representations, encoding features/properties of the entities. Despite the outstanding performance reported from different embedding learning algorithms, few efforts were devoted to structurally interpreting how features are encoded in the learned embedding space. This work proposes EmbeddingTree, a hierarchical embedding exploration algorithm that relates the semantics of entity features with the less-interpretable embedding vectors. An interactive visualization tool is also developed based on EmbeddingTree to explore high-dimensional embeddings. The tool helps users discover nuance features of data entities, perform feature denoising/injecting in embedding training, and generate embeddings for unseen entities. We demonstrate the efficacy of EmbeddingTree and our visualization tool through embeddings generated for industry-scale merchant data and the public 30Music listening/playlists dataset.
Graph Neural Networks (GNNs) have achieved impressive performance in collaborative filtering. However, GNNs tend to yield inferior performance when the distributions of training and test data are not aligned well. Also, training GNNs requires optimizing non-convex neural networks with an abundance of local and global minima, which may differ widely in their performance at test time. Thus, it is essential to choose the minima carefully. Here we propose an effective training schema, called {gSAM}, under the principle that the \textit{flatter} minima has a better generalization ability than the \textit{sharper} ones. To achieve this goal, gSAM regularizes the flatness of the weight loss landscape by forming a bi-level optimization: the outer problem conducts the standard model training while the inner problem helps the model jump out of the sharp minima. Experimental results show the superiority of our gSAM.
Pre-training on large models is prevalent and emerging with the ever-growing user-generated content in many machine learning application categories. It has been recognized that learning contextual knowledge from the datasets depicting user-content interaction plays a vital role in downstream tasks. Despite several studies attempting to learn contextual knowledge via pre-training methods, finding an optimal training objective and strategy for this type of task remains a challenging problem. In this work, we contend that there are two distinct aspects of contextual knowledge, namely the user-side and the content-side, for datasets where user-content interaction can be represented as a bipartite graph. To learn contextual knowledge, we propose a pre-training method that learns a bi-directional mapping between the spaces of the user-side and the content-side. We formulate the training goal as a contrastive learning task and propose a dual-Transformer architecture to encode the contextual knowledge. We evaluate the proposed method for the recommendation task. The empirical studies have demonstrated that the proposed method outperformed all the baselines with significant gains.
Vision transformer (ViT) expands the success of transformer models from sequential data to images. The model decomposes an image into many smaller patches and arranges them into a sequence. Multi-head self-attentions are then applied to the sequence to learn the attention between patches. Despite many successful interpretations of transformers on sequential data, little effort has been devoted to the interpretation of ViTs, and many questions remain unanswered. For example, among the numerous attention heads, which one is more important? How strong are individual patches attending to their spatial neighbors in different heads? What attention patterns have individual heads learned? In this work, we answer these questions through a visual analytics approach. Specifically, we first identify what heads are more important in ViTs by introducing multiple pruning-based metrics. Then, we profile the spatial distribution of attention strengths between patches inside individual heads, as well as the trend of attention strengths across attention layers. Third, using an autoencoder-based learning solution, we summarize all possible attention patterns that individual heads could learn. Examining the attention strengths and patterns of the important heads, we answer why they are important. Through concrete case studies with experienced deep learning experts on multiple ViTs, we validate the effectiveness of our solution that deepens the understanding of ViTs from head importance, head attention strength, and head attention pattern.
The most useful data mining primitives are distance measures. With an effective distance measure, it is possible to perform classification, clustering, anomaly detection, segmentation, etc. For single-event time series Euclidean Distance and Dynamic Time Warping distance are known to be extremely effective. However, for time series containing cyclical behaviors, the semantic meaningfulness of such comparisons is less clear. For example, on two separate days the telemetry from an athlete workout routine might be very similar. The second day may change the order in of performing push-ups and squats, adding repetitions of pull-ups, or completely omitting dumbbell curls. Any of these minor changes would defeat existing time series distance measures. Some bag-of-features methods have been proposed to address this problem, but we argue that in many cases, similarity is intimately tied to the shapes of subsequences within these longer time series. In such cases, summative features will lack discrimination ability. In this work we introduce PRCIS, which stands for Pattern Representation Comparison in Series. PRCIS is a distance measure for long time series, which exploits recent progress in our ability to summarize time series with dictionaries. We will demonstrate the utility of our ideas on diverse tasks and datasets.
Transformer-based sequential recommenders are very powerful for capturing both short-term and long-term sequential item dependencies. This is mainly attributed to their unique self-attention networks to exploit pairwise item-item interactions within the sequence. However, real-world item sequences are often noisy, which is particularly true for implicit feedback. For example, a large portion of clicks do not align well with user preferences, and many products end up with negative reviews or being returned. As such, the current user action only depends on a subset of items, not on the entire sequences. Many existing Transformer-based models use full attention distributions, which inevitably assign certain credits to irrelevant items. This may lead to sub-optimal performance if Transformers are not regularized properly.