Coding agents powered by large language models (LLMs) have gained traction for automating code generation through iterative problem-solving with minimal human involvement. Despite the emergence of various frameworks, e.g., LangChain, AutoML, and AIDE, ML scientists still struggle to effectively review and adjust the agents' coding process. The current approach of manually inspecting individual outputs is inefficient, making it difficult to track code evolution, compare coding iterations, and identify improvement opportunities. To address this challenge, we introduce a visual analytics system designed to enhance the examination of coding agent behaviors. Focusing on the AIDE framework, our system supports comparative analysis across three levels: (1) Code-Level Analysis, which reveals how the agent debugs and refines its code over iterations; (2) Process-Level Analysis, which contrasts different solution-seeking processes explored by the agent; and (3) LLM-Level Analysis, which highlights variations in coding behavior across different LLMs. By integrating these perspectives, our system enables ML scientists to gain a structured understanding of agent behaviors, facilitating more effective debugging and prompt engineering. Through case studies using coding agents to tackle popular Kaggle competitions, we demonstrate how our system provides valuable insights into the iterative coding process.

Chia-Yuan Chang, Zhimeng Jiang, Vineeth Rakesh, Menghai Pan, Chin-Chia Michael Yeh, Guanchu Wang, Mingzhi Hu, Zhichao Xu, Yan Zheng, Mahashweta Das, Na Zou. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

State space models (SSMs) have emerged as an efficient alternative to transformer-based models, offering linear complexity that scales better than transformers. One of the latest advances in SSMs, Mamba, introduces a selective scan mechanism that assigns trainable weights to input tokens, effectively mimicking the attention mechanism. Mamba has also been successfully extended to the vision domain by decomposing 2D images into smaller patches and arranging them as 1D sequences. However, it remains unclear how these patches interact with (or attend to) each other in relation to their original 2D spatial location. Additionally, the order used to arrange the patches into a sequence also significantly impacts their attention distribution. To better understand the attention between patches and explore the attention patterns, we introduce a visual analytics tool specifically designed for vision-based Mamba models. This tool enables a deeper understanding of how attention is distributed across patches in different Mamba blocks and how it evolves throughout a Mamba model. Using the tool, we also investigate the impact of different patch-ordering strategies on the learned attention, offering further insights into the model’s behavior.

Traffic forecasting, crucial for urban planning, requires accurate predictions of spatial-temporal traffic patterns across urban areas. Existing research mainly focuses on designing complex spatial-temporal models to capture these dependencies. However, this field faces challenges related to data scarcity and model stability, which results in limited performance improvement. To address these issues, we propose Spatial-Temporal Masked AutoEncoders (STMAE), a plug-and-play framework designed to enhance existing spatial-temporal models on traffic prediction. STMAE operates in two stages. In the pretraining stage, an encoder processes partially visible traffic data produced by a dual-masking strategy, including biased random walk-based spatial masking and patch-based temporal masking. Subsequently, two decoders aim to reconstruct the masked counterparts from both spatial and temporal perspectives. The fine-tuning stage retains the pretrained encoder and integrates it with decoders from existing backbones to improve traffic forecasting accuracy. Our results on traffic benchmarks show that STMAE can largely enhance the forecasting capabilities of various spatial-temporal models.

The Matrix Profile (MP), a versatile tool for time series data mining, has been shown effective in time series anomaly detection (TSAD). This paper delves into the problem of anomaly detection in multidimensional time series, a common occurrence in real-world applications. For instance, in a manufacturing factory, multiple sensors installed across the site collect time-varying data for analysis. The Matrix Profile, named for its role in profiling the matrix storing pairwise distance between subsequences of univariate time series, becomes complex in multidimensional scenarios. If the input univariate time series has$n$subsequences, the pairwise distance matrix is a$n\times n$matrix. In a multidimensional time series with$d$dimensions, the pairwise distance information must be stored in a$n\times n\times d$tensor. In this paper, we first analyze different strategies for condensing this tensor into a profile vector. We then investigate the potential of extending the MP to efficiently find k-nearest neighbors for anomaly detection. Finally, we benchmark the multidimensional MP against 19 baseline methods on 119 multidimensional TSAD datasets. The experiments covers three learning setups: unsupervised, supervised, and semi-supervised. MP is the only method that consistently delivers high performance across all setups.

Spatial-temporal forecasting systems play a crucial role in addressing numerous real-world challenges. In this paper, we investigate the potential of addressing spatial-temporal forecasting problems using general time series forecasting models, i.e., models that do not leverage the spatial relationships among the nodes. We propose a all-Multi-Layer Perceptron (all-MLP) time series forecasting architecture called RPMixer. The all-MLP architecture was chosen due to its recent success in time series forecasting benchmarks. Furthermore, our method capitalizes on the ensemble-like behavior of deep neural networks, where each individual block within the network behaves like a base learner in an ensemble model, particularly when identity mapping residual connections are incorporated. By integrating random projection layers into our model, we increase the diversity among the blocks' outputs, thereby improving the overall performance of the network. Extensive experiments conducted on the largest spatial-temporal forecasting benchmark datasets demonstrate that the proposed method outperforms 14 alternative methods.

Prompts, which are pieces of text injected into the inputs of NLP models, can significantly enhance these models’ performance. Despite their effectiveness, a straightforward method for exploring a large number of possible prompts and assessing their quality in down-stream tasks is lacking. This paper introduces PromptLandscape, a visualization approach to exploring prompts in the zero-shot image classification context. Our focus is on the CLIP model [26], which has been pre-trained on a large number of image-text pairs and exhibits exceptional performance in zero-shot image classifications. However, its performance is also considerably influenced by the text labels. Prompts were introduced to refine the text labels and improve the classification performance. Nevertheless, the prompts are often proposed through trial-and-error and it is challenging to find the best prompts for different image datasets. PromptLandscape presents the large number of possible prompts as points in a 3D space to guide their exploration and enables users to decode any points back as texts for better interpretation. Through concrete case studies, we demonstrate the effectiveness of PromptLandscape in (1) finding the best prompts for image classifications and (2) interpreting the prompts to reveal the crucial features of the studied image datasets.

In recent years there has been significant progress in time series anomaly detection. However, after detecting an (perhaps tentative) anomaly, can we explain it? Such explanations would be useful to triage anomalies. For example, in an oil refinery, should we respond to an anomaly by dispatching a hydraulic engineer, or an intern to replace the battery on a sensor? There have been some parallel efforts to explain anomalies, however many proposed techniques produce explanations that are indirect, and often seem more complex than the anomaly they seek to explain. Our review of the literature/checklists/user-manuals used by frontline practitioners in various domains reveals an interesting near-universal commonality. Most practitioners discuss, explain and report anomalies in the following format: The anomaly would be like normal data A, if not for the corruption B. The reader will appreciate that is a type of counterfactual explanation. In this work we introduce a domain agnostic counterfactual explanation technique to produce explanations for time series anomalies. As we will show, our method can produce both visual and text-based explanations that are objectively correct, intuitive and in many circumstances, directly actionable.

Applications of neural networks like MLPs and ResNets in temporal data mining has led to improvements on the problem of time series classification. Recently, a new class of networks called Temporal Convolution Networks (TCNs) have been proposed for various time series tasks. Instead of time invariant convolutions they use temporally causal convolutions, this makes them more constrained than ResNets but surprisingly good at generalization. This raises an important question: How does a network with causal convolution solve these tasks when compared to a network with acausal convolutions? As the first attempt at answering these questions, we analyze different architectures through a lens of representational subspace similarity. We demonstrate that the evolution of input representations in the layers of TCNs is markedly different from ResNets and MLPs. We find that acausal networks are prone to form groupings of similar layers and TCNs on the other hand learn representations that are much more diverse throughout the network. Next, we study the convergence properties of internal layers across different architecture families and discover that the behaviour of layers inside Acausal network is more homogeneous when compared to TCNs. Our extensive empirical studies offer new insights into internal mechanisms of convolution networks in the domain of time series analysis and may assist practitioners gaining deeper understanding of each network.

Graph Transformers have garnered significant attention for learning graph-structured data, thanks to their superb ability to capture long-range dependencies among nodes. However, the quadratic space and time complexity hinders the scalability of Graph Transformers, particularly for large-scale recommendation. Here we propose an efficient Masked Graph Transformer, named MGFormer, capable of capturing all-pair interactions among nodes with a linear complexity. To achieve this, we treat all user/item nodes as independent tokens, enhance them with positional embeddings, and feed them into a kernelized attention module. Additionally, we incorporate learnable relative degree information to appropriately reweigh the attentions. Experimental results show the superior performance of our MGFormer, even with a single attention layer.

Word vector embeddings have been shown to contain and amplify biases in the data they are extracted from. Consequently, many techniques have been proposed to identify, mitigate, and attenuate these biases in word representations. In this article, we utilize interactive visualization to increase the interpretability and accessibility of a collection of state-of-the-art debiasing techniques. To aid this, we present the Visualization of Embedding Representations for deBiasing (VERB) system, an open-source web-based visualization tool that helps users gain a technical understanding and visual intuition of the inner workings of debiasing techniques, with a focus on their geometric properties. In particular, VERB offers easy-to-follow examples that explore the effects of these debiasing techniques on the geometry of high-dimensional word vectors. To help understand how various debiasing techniques change the underlying geometry, VERB decomposes each technique into interpretable sequences of primitive transformations and highlights their effect on the word vectors using dimensionality reduction and interactive visual exploration. VERB is designed to target natural language processing (NLP) practitioners who are designing decision-making systems on top of word embeddings and researchers working with the fairness and ethics of machine learning systems in NLP. It can also serve as a visual medium for education, which helps an NLP novice understand and mitigate biases in word embeddings.

Recent advancements in pre-trained language-image models have ushered in a new era of visual comprehension. Leveraging the power of these models, this article tackles two issues within the realm of visual analytics: (1) the efficient exploration of large-scale image datasets and identification of data biases within them; (2) the evaluation of image captions and steering of their generation process. On the one hand, by visually examining the captions generated from language-image models for an image dataset, we gain deeper insights into the visual contents, unearthing data biases that may be entrenched within the dataset. On the other hand, by depicting the association between visual features and textual captions, we expose the weaknesses of pre-trained language-image models in their captioning capability and propose an interactive interface to steer caption generation. The two parts have been coalesced into a coordinated visual analytics system, fostering the mutual enrichment of visual and textual contents. We validate the effectiveness of the system with domain practitioners through concrete case studies with large-scale image datasets.

Publishing and sharing data is crucial for the data mining community, allowing collaboration and driving open innovation. However, many researchers cannot release their data due to privacy regulations or fear of leaking confidential business information. To alleviate such issues, we propose the Time Series Synthesis Using the Matrix Profile (TSSUMP) method, where synthesized time series can be released in lieu of the original data. The TSSUMP method synthesizes time series by preserving similarity join information (i.e., Matrix Profile) while reducing the correlation between the synthesized and the original time series. As a result, neither the values for the individual time steps nor the local patterns (or shapes) from the original data can be recovered, yet the resulting data can be used for downstream tasks that data analysts are interested in. We concentrate on similarity joins because they are one of the most widely applied time series data mining routines across different data mining tasks. We test our method on a case study of ECG and gender masking prediction. In this case study, the gender information is not only removed from the synthesized time series, but the synthesized time series also preserves enough information from the original time series. As a result, unmodified data mining tools can obtain near-identical performance on the synthesized time series as on the original time series.

Time series data is ubiquitous across various domains such as finance, healthcare, and manufacturing, but their properties can vary significantly depending on the domain they originate from. The ability to perform Content-based Time Series Retrieval (CTSR) is crucial for identifying unknown time series examples. However, existing CTSR works typically focus on retrieving time series from a single domain database, which can be inadequate if the user does not know the source of the query time series. This limitation motivates us to investigate the CTSR problem in a scenario where the database contains time series from multiple domains. To facilitate this investigation, we introduce a CTSR benchmark dataset that comprises time series data from a variety of domains, such as motion, power demand, and traffic. This dataset is sourced from a publicly available time series classification dataset archive, making it easily accessible to researchers in the field. We compare several popular methods for modeling and retrieving time series data using this benchmark dataset. Additionally, we propose a novel distance learning model that outperforms the existing methods. Overall, our study highlights the importance of addressing the CTSR problem across multiple domains and provides a useful benchmark dataset for future research.

Time series classification is a widely studied problem in the field of time series data mining. Previous research has predominantly focused on scenarios where relevant or foreground subsequences have already been extracted, with each subsequence corresponding to a single label. However, real-world time series data often contain foreground subsequences that are intertwined with background subsequences. Successfully classifying these relevant subsequences requires not only distinguishing between different classes but also accurately identifying the foreground subsequences amidst the background. To address this challenge, we propose a novel subsequence classification method that represents each subsequence as an ego-network, providing crucial nearest neighbor information to the model. The ego-networks of all subsequences collectively form a time series subsequence graph, and we introduce an algorithm to efficiently construct this graph. Furthermore, we have demonstrated the significance of enforcing temporal consistency in the prediction of adjacent subsequences for the subsequence classification problem. To evaluate the effectiveness of our approach, we conducted experiments using 128 univariate and 30 multivariate time series datasets. The experimental results demonstrate the superior performance of our method compared to alternative approaches. Specifically, our method outperforms the baseline on 104 out of 158 datasets.

Time series discords are a useful primitive for time series anomaly detection, and the matrix profile is capable of capturing discord effectively. There exist many research efforts to improve the scalability of discord discovery with respect to the length of time series. However, there is surprisingly little work focused on reducing the time complexity of matrix profile computation associated with dimensionality of a multidimensional time series. In this work, we propose a sketch for discord mining among multi-dimensional time series. After an initial pre-processing of the sketch as fast as reading the data, the discord mining has runtime independent of the dimensionality of the original data. On several real world examples from water treatment and transportation, the proposed algorithm improves the throughput by at least an order of magnitude (50X) and only has minimal impact on the quality of the approximated solution. Additionally, the proposed method can handle the dynamic addition or deletion of dimensions inconsequential overhead. This allows a data analyst to consider “what-if” scenarios in real time while exploring the data.

Spatial-temporal graph neural networks (STGNNs) are promising in solving real-world spatial-temporal forecasting problems. Recognizing the inherent sequential relationship of spatial-temporal data, it is natural to explore the integration of boosting training mechanism to further enhance the performance of STGNNs. However, few studies have touched this research area. To bridge this gap, in this work, we propose spatial-temporal graph boosting networks, namely STGBN, which to the best of our knowledge is the first attempt to leverage gradient boosting for enhancing STGNNs. STGBN follows the general training procedure of conventional gradient boosting, but incorporates two distinctive designs to improve its efficiency in training on spatial-temporal graphs. Specifically, we design an incremental learning strategy that progressively includes spatial-temporal data into training. Additionally, we enforce an identical architecture for the base learner in all boosting iterations with each base learner inheriting from the one in the previous iteration. These designs facilitate rapid convergence of the base learner and expedite the overall training process. The base learner in STGBN is designed as a Transformer sandwich, which consists of two temporal Transformers on the top and bottom and a spatial Transformer in the middle. Structuring them in such a way helps the model capture long-range temporal dynamics, global spatial dependencies, and deep spatial-temporal interactions. We perform extensive spatial-temporal forecasting experiments on four spatial-temporal graph benchmarks. Promising results demonstrate the outstanding performance of STGBN against a wide range of state-of-the-art baseline models.

A Content-based Time Series Retrieval (CTSR) system is an information retrieval system for users to interact with time series emerged from multiple domains, such as finance, healthcare, and manufacturing. For example, users seeking to learn more about the source of a time series can submit the time series as a query to the CTSR system and retrieve a list of relevant time series with associated metadata. By analyzing the retrieved metadata, users can gather more information about the source of the time series. Because the CTSR system is required to work with time series data from diverse domains, it needs a high-capacity model to effectively measure the similarity between different time series. On top of that, the model within the CTSR system has to compute the similarity scores in an efficient manner as the users interact with the system in real-time. In this paper, we propose an effective and efficient CTSR model that outperforms alternative models, while still providing reasonable inference runtimes. To demonstrate the capability of the proposed method in solving business problems, we compare it against alternative models using our in-house transaction data. Our findings reveal that the proposed model is the most suitable solution compared to others for our transaction data problem.

A foundation model is a machine learning model trained on a large and diverse set of data, typically using self-supervised learning-based pre-training techniques, that can be adapted to various downstream tasks. However, current research on time series pre-training has predominantly focused on models trained exclusively on data from a single domain. As a result, these models possess domain-specific knowledge that may not be easily transferable to time series from other domains. In this paper, we aim to develop an effective time series foundation model by leveraging unlabeled samples from multiple domains. To achieve this, we repurposed the publicly available UCR Archive and evaluated four existing self-supervised learning-based pre-training methods, along with a novel method, on the datasets. We tested these methods using four popular neural network architectures for time series to understand how the pre-training methods interact with different network designs. Our experimental results show that pre-training improves downstream classification tasks by enhancing the convergence of the fine-tuning process. Furthermore, we found that the proposed pre-training method, when combined with the Transformer, outperforms the alternatives. The proposed method outperforms or achieves equal performance compared to the second best method in ~93% of downstream tasks.

Sequential tabular data is one of the most commonly used data types in real-world applications. Different from conventional tabular data, where rows in a table are independent, sequential tabular data contains rich contextual and sequential information, where some fields aredynamically changing over time and others arestatic. Existing transformer-based approaches analyzing sequential tabular data overlook the differences between dynamic and static fields by replicating and filling static fields into each record, and ignore temporal information between rows, which leads to three major disadvantages: (1) computational overhead, (2) artificially simplified data for masked language modeling pre-training task that may yield less meaningful representations, and (3) disregarding the temporal behavioral patterns implied by time intervals. In this work, we propose FATA-Trans, a model with two field transformers for modeling sequential tabular data, where each processes static and dynamic field information separately. FATA-Trans isfield - andtime -aware for sequential tabular data. Thefield -type embedding in the method enables FATA-Trans to capture differences between static and dynamic fields. Thetime -aware position embedding exploits both order and time interval information between rows, which helps the model detect underlying temporal behavior in a sequence. Our experiments on three benchmark datasets demonstrate that the learned representations from FATA-Trans consistently outperform state-of-the-art solutions in the downstream tasks. We also present visualization studies to highlight the insights captured by the learned representations, enhancing our understanding of the underlying data. Our codes are available at https://github.com/zdy93/FATA-Trans.

Multitask learning (MTL) aims to develop a unified model that can handle a set of closely related tasks simultaneously. By optimizing the model across multiple tasks, MTL generally surpasses its non-MTL counterparts in terms of generalizability. Although MTL has been extensively researched in various domains such as computer vision, natural language processing, and recommendation systems, its application to time series data has received limited attention. In this paper, we investigate the application of MTL to the time series classification (TSC) problem. However, when we integrate the state-of-the-art 1D convolution-based TSC model with MTL, the performance of the TSC model actually deteriorates. By comparing the 1D convolution-based models with the Dynamic Time Warping (DTW) distance function, it appears that the underwhelming results stem from the limited expressive power of the 1D convolutional layers. To overcome this challenge, we propose a novel design for a 2D convolution-based model that enhances the model's expressiveness. Leveraging this advantage, our proposed method outperforms competing approaches on both the UCR Archive and an industrial transaction TSC dataset.

Transformer-based models are powerful for modeling temporal dynamics of user preference in sequential recommendation. Most of the variants adopt the Softmax transformation in the self-attention layers to generate dense attention probabilities. However, real-world item sequences are often noisy, containing a mixture of true-positive and false-positive interactions. Such dense attentions inevitably assign probability mass to noisy or irrelevant items, leading to sub-optimal performance and poor explainability. Here we propose a Probabilistic Masked Attention Network (PMAN) to identify the sparse pattern of attentions, which is more desirable for pruning noisy items in sequential recommendation. Specifically, we employ a probabilistic mask to achieve sparse attentions under a constrained optimization framework. As such, PMAN allows to select which information is critical to be retained or dropped in a data-driven fashion. Experimental studies on real-world benchmark datasets show that PMAN is able to improve the performance of Transformers significantly.

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.

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.

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.

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.

Graph Neural Networks (GNNs) have achieved impressive performance in collaborative filtering. However, recent studies show that GNNs tend to yield inferior performance when the distributions of training and test data are not aligned well. Moreover, training GNNs often 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 develop an optimization strategy that can choose the minima carefully, which can yield strong generalization performance on unseen data. Here we propose an effective training schema, called gSAM, under the principle that theflatter minima has a better generalization ability than thesharper 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.

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.

With the fast-growing number of classification models being produced every day, numerous model interpretation and comparison solutions have also been introduced. For example, LIME [1] and SHAP [2] can interpret what input features contribute more to a classifier's output predictions. Different numerical metrics (e.g., accuracy) can be used to easily compare two classifiers. However, few works can interpret the contribution of a data feature to a classifier in comparison with its contribution to another classifier. This comparative interpretation can help to disclose the fundamental difference between two classifiers, select classifiers in different feature conditions, and better ensemble two classifiers. To accomplish it, we propose a learning-from-disagreement (LFD) framework to visually compare two classification models. Specifically, LFD identifies data instances with disagreed predictions from two compared classifiers and trains a discriminator to learn from the disagreed instances. As the two classifiers' training features may not be available, we train the discriminator through a set of meta-features proposed based on certain hypotheses of the classifiers to probe their behaviors. Interpreting the trained discriminator with the SHAP values of different meta-features, we provide actionable insights into the compared classifiers. Also, we introduce multiple metrics to profile the importance of meta-features from different perspectives. With these metrics, one can easily identify meta-features with the most complementary behaviors in two classifiers, and use them to better ensemble the classifiers. We focus on binary classification models in the financial services and advertising industry to demonstrate the efficacy of our proposed framework and visualizations.

Node classification for graph-structured data aims to classify nodes whose labels are unknown. While studies on static graphs are prevalent, few studies have focused on dynamic graph node classification. Node classification on dynamic graphs is challenging for two reasons. First, the model needs to capture both structural and temporal information, particularly on dynamic graphs with a long history and require large receptive fields. Second, model scalability becomes a significant concern as the size of the dynamic graph increases. To address these problems, we propose the Time Augmented Dynamic Graph Neural Network (TADGNN) framework. TADGNN consists of two modules: 1) a time augmentation module that captures the temporal evolution of nodes across time structurally, creating a time-augmented spatio-temporal graph, and 2) an information propagation module that learns the dynamic representations for each node across time using the constructed time-augmented graph. We perform node classification experiments on four dynamic graph benchmarks. Experimental results demonstrate that TADGNN framework outperforms several static and dynamic state-of-the-art (SOTA) GNN models while demonstrating superior scalability. We also conduct theoretical and empirical analyses to validate the efficiency of the proposed method.

Using a newly developed model to upgrade a legacy model is a common practice in machine learning applications. After the upgrade, it is expected that the new model should outperform the legacy model in the regions of interest. However, it is observed that the new model often makes incorrect decisions on some instances where the legacy model still performs well. For a binary classification model (e.g., click-through-rate/CTR prediction model), such undesirable behavior could even occur in the low false positive region of the receiver operating characteristic (ROC) curve. Finding the reasons behind this phenomenon can help business partners in an organization gain confidence in adopting the new model and help modelers to improve the new model in future releases. In this paper, we present the "Learning from Disagreement" framework to understand and improve the performance of a predictive model. Under the setting of a binary classification task, this proposed approach focuses on instances that lead to contradictory decisions between a pair of models at a given operating point. We perform feature importance analysis exclusively on these instances, gain insights into the pair of models without even knowing their inner operations, and offer actionable feedback for model improvement. We demonstrate the usefulness of this framework on two real-world event detection datasets.

Diagnosing a performance bug triggered in production cloud environments is notoriously challenging. Extracting performance bug signatures can help cloud operators quickly pinpoint the problem and avoid repeating manual efforts for diagnosing similar performance bugs. In this paper, we present PerfSig, a multi-modality performance bug signature extraction tool which can identify principal anomaly patterns and root cause functions for performance bugs. PerfSig performs fine-grained anomaly detection over various machine data such as system metrics, system logs, and function call traces. We then conduct causal analysis across different machine data using information theory method to pinpoint the root cause function of a performance bug. PerfSig generates bug signatures as the combination of the identified anomaly patterns and root cause functions. We have implemented a prototype of PerfSig and conducted evaluation using 20 real world performance bugs in six commonly used cloud systems. Our experimental results show that PerfSig captures various kinds of fine-grained anomaly patterns from different machine data and successfully identifies the root cause functions through multi-modality causal analysis for 19 out of 20 tested performance bugs.

Graph neural networks (GNNs) are deep learning models designed specifically for graph data, and they typically rely on node features as the input to the first layer. When applying such a type of network on the graph without node features, one can extract simple graph-based node features (e.g., number of degrees) or learn the input node representations (i.e., embeddings) when training the network. While the latter approach, which trains node embeddings, more likely leads to better performance, the number of parameters associated with the embeddings grows linearly with the number of nodes. It is therefore impractical to train the input node embeddings together with GNNs within graphics processing unit (GPU) memory in an end-to-end fashion when dealing with industrial-scale graph data. Inspired by the embedding compression methods developed for natural language processing (NLP) tasks, we develop a node embedding compression method where each node is compactly represented with a bit vector instead of a floating-point vector. The parameters utilized in the compression method can be trained together with GNNs. We show that the proposed node embedding compression method achieves superior performance compared to the alternatives.

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.

There has been an explosion of interest in designing various Knowledge Graph Neural Networks (KGNNs), which achieve state-of-the-art performance and provide great explainability for recommendation. The promising performance is mainly resulting from their capability of capturing high-order proximity messages over the knowledge graphs. However, training KGNNs at scale is challenging due to the high memory usage. In the forward pass, the automatic differentiation engines (e.g., TensorFlow/PyTorch) generally need to cache all intermediate activation maps in order to compute gradients in the backward pass, which leads to a large GPU memory footprint. Existing work solves this problem by utilizing multi-GPU distributed frameworks. Nonetheless, this poses a practical challenge when seeking to deploy KGNNs in memory-constrained environments, especially for industry-scale graphs.

The matrix profile is an effective data mining tool that provides similarity join functionality for time series data. Since the introduction of the matrix profile five years ago, multiple efforts have been made to speed up the computation with approximate joins; however, the majority of these efforts only focus on self-joins. In this work, we show that it is possible to efficiently perform approximate inter-time series similarity joins with error bounded guarantees by creating a compact “dictionary” representation of time series. Using the dictionary representation instead of the original time series, we are able to improve the throughput of an anomaly mining system by at least 20X, with essentially no decrease in accuracy. As a side effect, the dictionaries also summarize the time series in a semantically meaningful way and can provide intuitive and actionable insights.

Graph Neural Networks (GNNs) have emerged as powerful tools for collaborative filtering. A key challenge of recommendations is to distill long-range collaborative signals from user-item graphs. Typically, GNNs generate embeddings of users/items by propagating and aggregating the messages between local neighbors. Thus, the ability of GNNs to capture long-range dependencies heavily depends on their depths. However, simply training deep GNNs has several bottleneck effects, e.g., over-fitting & over-smoothing, which may lead to unexpected results if GNNs are not well regularized.

Deep reinforcement learning (DRL) targets to train an autonomous agent to interact with a pre-defined environment and strives to achieve specific goals through deep neural networks (DNN). Recurrent neural network (RNN) based DRL has demonstrated superior performance, as RNNs can effectively capture the temporal evolution of the environment and respond with proper agent actions. However, apart from the outstanding performance, little is known about how RNNs understand the environment internally and what has been memorized over time. Revealing these details is extremely important for deep learning experts to understand and improve DRLs, which in contrast, is also challenging due to the complicated data transformations inside these models. In this article, we propose Deep Reinforcement Learning Interactive Visual Explorer (DRLIVE), a visual analytics system to effectively explore, interpret, and diagnose RNN-based DRLs. Having focused on DRL agents trained for different Atari games, DRLIVE accomplishes three tasks: game episode exploration, RNN hidden/cell state examination, and interactive model perturbation. Using the system, one can flexibly explore a DRL agent through interactive visualizations, discover interpretable RNN cells by prioritizing RNN hidden/cell states with a set of metrics, and further diagnose the DRL model by interactively perturbing its inputs. Through concrete studies with multiple deep learning experts, we validated the efficacy of DRLIVE.

Predicting metrics associated with entities' transnational behavior within payment processing networks is essential for system monitoring. Multivariate time series, aggregated from the past transaction history, can provide valuable insights for such prediction. The general multivariate time series prediction problem has been well studied and applied across several domains, including manufacturing, medical, and entomology. However, new domain-related challenges associated with the data such as concept drift and multi-modality have surfaced in addition to the real-time requirements of handling the payment transaction data at scale. In this work, we study the problem of multivariate time series prediction for estimating transaction metrics associated with entities in the payment transaction database. We propose a model with five unique components to estimate the transaction metrics from multi-modality data. Four of these components capture interaction, temporal, scale, and shape perspectives, and the fifth component fuses these perspectives together. We also propose a hybrid offline/online training scheme to address concept drift in the data and fulfill the real-time requirements. Combining the estimation model with a graphical user interface, the prototype transaction metric estimation system has demonstrated its potential benefit as a tool for improving a payment processing company's system monitoring capability.

Embeddings are one of the fundamental building blocks for data analysis tasks. Embeddings are already essential tools for large language models and image analysis, and their use is being extended to many other research domains. The generation of these distributed representations is often a data-and computation-expensive process; yet the holistic analysis and adjustment of them after they have been created is still a developing area. In this paper, we first propose a very general quantitatively measure for the presence of features in the embedding data based on if it can be learned. We then devise a method to remove or alleviate undesired features in the embedding while retaining the essential structure of the data. We use a Domain Adversarial Network (DAN) to generate a non-affine transformation, but we add constraints to ensure the essential structure of the embedding is preserved. Our empirical results demonstrate that the proposed algorithm significantly outperforms the state-of-art unsupervised algorithm on several data sets, including novel applications from the industry.

Graph Convolutional Networks (GCNs) are powerful for collaborative filtering. The key component of GCNs is to explore neighborhood aggregation mechanisms to extract high-level representations of users and items. However, real-world user-item graphs are often incomplete and noisy. Aggregating misleading neighborhood information may lead to sub-optimal performance if GCNs are not regularized properly. Also, the real-world user-item graphs are often sparse and low rank. These two intrinsic graph properties are widely used in shallow matrix completion models, but far less studied in graph neural models. Here we propose Structured Graph Convolutional Networks (SGCNs) to enhance the performance of GCNs by exploiting graph structural properties of sparsity and low rank. To achieve sparsity, we attach each layer of a GCN with a trainable stochastic binary mask to prune noisy and insignificant edges, resulting in a clean and sparsified graph. To preserve its low-rank property, the nuclear norm regularization is applied. We jointly learn the parameters of stochastic binary masks and original GCNs by solving a stochastic binary optimization problem. An unbiased gradient estimator is further proposed to better backpropagate the gradients of binary variables. Experimental results demonstrate that SGCNs achieve better performance compared with the state-of-the-art GCNs.

Digital payment volume has proliferated in recent years with the rapid growth of small businesses and online shops. When processing these digital transactions, recognizing each merchant’s real identity (i.e., business type) is vital to ensure the integrity of payment processing systems. Conventionally, this problem is formulated as a time series classification problem solely using the merchant transaction history. However, with the large scale of the data, and changing behaviors of merchants and consumers over time, it is extremely challenging to achieve satisfying performance from off-the-shelf classification methods. In this work, we approach this problem from a multi-modal learning perspective, where we use not only the merchant time series data but also the information of merchant-merchant relationship (i.e., affinity) to verify the self-reported business type (i.e., merchant category) of a given merchant. Specifically, we design two individual encoders, where one is responsible for encoding temporal information and the other is responsible for affinity information, and a mechanism to fuse the outputs of the two encoders to accomplish the identification task. Our experiments on real-world credit card transaction data between 71,668 merchants and 433,772,755 customers have demonstrated the effectiveness and efficiency of the proposed model.

Since its introduction several years ago, the Matrix Profile has received significant attention for two reasons. First, it is a very general representation, allowing for the discovery of time series motifs, discords, chains, joins, shapelets, segmentations etc. Secondly, it can be computed very efficiently, allowing for fast exact computation and ultra-fast approximate computation. For analysts that use the Matrix Profile frequently, its incremental computability means that they can perform ad-hoc analytics at any time, with almost no delay time. However, they can only issue global queries. That is, queries that consider all the data from time zero to the current time. This is a significant limitation, as they may be interested in localized questions about a contiguous subset of the data. For example, "do we have any unusual motifs that correspond with that unusually cool summer two years ago". Such ad-hoc queries would require recomputing the Matrix Profile for the time period in question. This is not an untenable computation, but it could not be done in interactive time. In this work we introduce a novel indexing framework that allows queries about arbitrary ranges to be answered in quasilinear time, allowing such queries to be interactive for the first time.

In manufacturing, a golden batch is an idealized realization of the perfect process to produce the desired item, typically represented as a multidimensional time series of pressures, temperatures, flow-rates and so forth. The golden batch is sometimes produced from first-principle models, but it is typically created by recording a batch produced by the most experienced engineers on carefully cleaned and calibrated machines. In most cases, the golden batch is only used in post-mortem analysis of a product with an unexpectedly inferior quality, as plant managers attempt to understand where and when the last production attempt went wrong. In this work, we make two contributions to golden batch processing. We introduce an online algorithm that allows practitioners to understand if the process is currently deviating from the golden batch in real-time, allowing engineers to intervene and potentially save the batch. This may be done, for example, by cooling a boiler that is running unexpectedly hot. In addition, we show that our ideas can greatly expand the purview of golden batch monitoring beyond industrial manufacturing. In particular, we show that golden batch monitoring can be used for anomaly detection, attention focusing, and personalized training/skill assessment in a host of novel domains.

Unsupervised semantic segmentation in the time series domain is a much studied problem due to its potential to detect unexpected regularities and regimes in poorly understood data. However, the current techniques have several shortcomings, which have limited the adoption of time series semantic segmentation beyond academic settings for four primary reasons. First, most methods require setting/learning many parameters and thus may have problems generalizing to novel situations. Second, most methods implicitly assume that all the data is segmentable and have difficulty when that assumption is unwarranted. Thirdly, many algorithms are only defined for the single dimensional case, despite the ubiquity of multi-dimensional data. Finally, most research efforts have been confined to the batch case, but online segmentation is clearly more useful and actionable. To address these issues, we present a multi-dimensional algorithm, which is domain agnostic, has only one, easily-determined parameter, and can handle data streaming at a high rate. In this context, we test the algorithm on the largest and most diverse collection of time series datasets ever considered for this task and demonstrate the algorithm's superiority over current solutions.

The article Domain agnostic online semantic segmentation for multi-dimensional time series, written by Shaghayegh Gharghabi, Chin-Chia Michael Yeh, Yifei Ding, Wei Ding, Paul Hibbing, Samuel LaMunion, Andrew Kaplan, Scott E. Crouter, Eamonn Keogh was originally published electronically on the publisher’s internet portal (currently SpringerLink) on 25 September 2018 without open access.

Most algorithms for music data mining and retrieval analyze the similarity between feature sets extracted from the raw audio. A conventional approach to assess similarities within or between recordings is to create similarity matrices. However, this method requires quadratic space for each comparison and typically requires costly post-processing of the matrix. We have recently proposed SiMPle, a powerful representation based on subsequence similarity join, which is applicable in several music analysis tasks. In this paper, we propose SiMPle-Fast a highly efficient method for exact computation of SiMPle that is up to one order of magnitude faster than SiMPle. Furthermore, we demonstrate the utility of SiMPle-Fast in cover music recognition and thumbnailing tasks and show that our method is significantly faster and more accurate than the state-of-the-art.

Time series motif discovery is an important primitive for time series analytics, and is used in domains as diverse as neuroscience, music and sports analytics. In recent years, algorithmic advances (coupled with hardware improvements) have greatly expanded the purview of motif discovery. Nevertheless, we argue that there is an insatiable need for further scalability. This is because more than most types of analytics, motif discovery benefits from interactivity. The two state-of-the-art algorithms to find motifs are STOMP, which requires O(n2) time, and STAMP, which, despite being an O(logn) factor slower, is the preferred solution for most applications, as it is a fast converging anytime algorithm. In favorable scenarios STAMP needs only to be run to a small fraction of completion to provide a very accurate approximation of the top-k motifs. In this work we introduce SCRIMP++, an O(n2) time algorithm that is also an anytime algorithm, combining the best features of STOMP and STAMP. As we shall show, SCRIMP++ maintains all the desirable properties of the original algorithms, but converges much faster, in almost all scenarios producing the correct output after spending a tiny fraction of the full computation time. We argue that for many end-users, this allows motif discovery to be performed in interactive sessions. Moreover, this interactivity can be game changing in terms of the analytics that can be performed.

Poultry farms are an important contributor to the human food chain. Worldwide, humankind keeps an enormous number of domesticated birds (e.g. chickens) for their eggs and their meat, providing rich sources of low-fat protein. However, around the world, there have been growing concerns about the quality of life for the livestock in poultry farms; and increasingly vocal demands for improved standards of animal welfare. Recent advances in sensing technologies and machine learning allow the possibility of automatically assessing the health of some individual birds, and employing the lessons learned to improve the welfare for all birds. This task superficially appears to be easy, given the dramatic progress in recent years in classifying human behaviors, and given that human behaviors are presumably more complex. However, as we shall demonstrate, classifying chicken behaviors poses several unique challenges, chief among which is creating a generalizable "dictionary" of behaviors from sparse and noisy data. In this work we introduce a novel time series dictionary learning algorithm that can robustly learn from weakly labeled data sources.

Unsupervised semantic segmentation in the time series domain is a much-studied problem due to its potential to detect unexpected regularities and regimes in poorly understood data. However, the current techniques have several shortcomings, which have limited the adoption of time series semantic segmentation beyond academic settings for three primary reasons. First, most methods require setting/learning many parameters and thus may have problems generalizing to novel situations. Second, most methods implicitly assume that all the data is segmentable, and have difficulty when that assumption is unwarranted. Finally, most research efforts have been confined to the batch case, but online segmentation is clearly more useful and actionable. To address these issues, we present an algorithm which is domain agnostic, has only one easily determined parameter, and can handle data streaming at a high rate. In this context, we test our algorithm on the largest and most diverse collection of time series datasets ever considered, and demonstrate our algorithm's superiority over current solutions. Furthermore, we are the first to show that semantic segmentation may be possible at superhuman performance levels.

Time series motifs are approximately repeating patterns in real-valued time series data. They are useful for exploratory data mining and are often used as inputs for various time series clustering, classification, segmentation, rule discovery, and visualization algorithms. Since the introduction of the first motif discovery algorithm for univariate time series in 2002, multiple efforts have been made to generalize motifs to the multidimensional case. In this work, we show that these efforts, which typically attempt to find motifs on all dimensions, will not produce meaningful motifs except in the most contrived situations. We explain this finding and introduce mSTAMP, an algorithm that allows meaningful discovery of multidimensional motifs. Beyond producing objectively and subjectively meaningful results, our algorithm has a host of additional advantages, including being much faster, requiring fewer parameters and supporting streaming data. We demonstrate the utility of our mSTAMP-based motif discovery framework on domains as diverse as audio processing, industry, and sports analytics.

In academic settings over the last decade, there has been significant progress in time series classification. However, much of this work makes assumptions that are simply unrealistic for deployed industrial applications. Examples of these unrealistic assumptions include the following: assuming that data subsequences have a single fixed-length, are precisely extracted from the data, and are correctly labeled according to their membership in a set of equal-size classes. In real-world industrial settings, these patterns can be of different lengths, the class annotations may only belong to a general region of the data, may contain errors, and finally, the class distribution is typically highly skewed. Can we learn from such weakly labeled data? In this work, we introduce SDTS, a scalable algorithm that can learn in such challenging settings. We demonstrate the utility of our ideas by learning from diverse datasets with millions of datapoints. As we shall demonstrate, our domain-agnostic parameter-free algorithm can be competitive with domain-specific algorithms used in neuroscience and entomology, even when those algorithms have been tuned by domain experts to incorporate domain knowledge.

Multidimensional Scaling (MDS) is one of the most versatile tools used for exploratory data mining. It allows a first glimpse of possible structure in the data, which can inform the choice of analyses used. Its uses are multiple. It can give the user an idea as to the cluster ability or linear separability of the data. It can help spot outliers, or can hint at the intrinsic dimensionality of the data. Moreover, it can sometimes reveal unexpected latent dimensions in the data. With all these uses, MDS is increasingly used in areas as diverse as marketing, medicine, genetics, music and linguistics. One of the strengths of MDS is that it is essentially agnostic to data type, as we can use any distance measure to create the distance matrix, which is the only required input to the MDS algorithm. In spite of this generality, we make the following claim. MDS is not (well) defined for an increasingly important data type, time series subsequences. In this work we explain why this is the case, and we propose a scalable solution. We demonstrate the utility of our ideas on several diverse real-world datasets. At the core of our approach is a novel Minimum Description Length (MDL) subsequence extraction algorithm. Beyond MDS visualization, this subsequence extraction subroutine may be a useful tool in its own right.

The all-pairs-similarity-search (or similarity join) problem has been extensively studied for text and a handful of other datatypes. However, surprisingly little progress has been made on similarity joins for time series subsequences. The lack of progress probably stems from the daunting nature of the problem. For even modest sized datasets the obvious nested-loop algorithm can take months, and the typical speed-up techniques in this domain (i.e., indexing, lower-bounding, triangular-inequality pruning and early abandoning) at best produce one or two orders of magnitude speedup. In this work we introduce a novel scalable algorithm for time series subsequence all-pairs-similarity-search. For exceptionally large datasets, the algorithm can be trivially cast as an anytime algorithm and produce high-quality approximate solutions in reasonable time. The exact similarity join algorithm computes the answer to the time series motif and time series discord problem as a side-effect, and our algorithm incidentally provides the fastest known algorithm for both these extensively-studied problems. We demonstrate the utility of our ideas for two time series data mining problems, including motif discovery and novelty discovery.

Time series motifs have been in the literature for about fifteen years, but have only recently begun to receive significant attention in the research community. This is perhaps due to the growing realization that they implicitly offer solutions to a host of time series problems, including rule discovery, anomaly detection, density estimation, semantic segmentation, etc. Recent work has improved the scalability to the point where exact motifs can be computed on datasets with up to a million data points in tenable time. However, in some domains, for example seismology, there is an insatiable need to address even larger datasets. In this work we show that a combination of a novel algorithm and a high-performance GPU allows us to significantly improve the scalability of motif discovery. We demonstrate the scalability of our ideas by finding the full set of exact motifs on a dataset with one hundred million subsequences, by far the largest dataset ever mined for time series motifs. Furthermore, we demonstrate that our algorithm can produce actionable insights in seismology and other domains.

Representing music information using audio codewords has led to state-of-the-art performance on various music classifcation benchmarks. Comparing to conventional audio descriptors, audio words offer greater fexibility in capturing the nuance of music signals, in that each codeword can be viewed as a quantization of the music universe and that the quantization goes finer as the size of the dictionary (i.e., audio codebook) increases. In practice, however, the high computational cost of codeword assignment might discourage the use of a large dictionary. This paper presents two modifications of a LASSO screening technique developed in the compressive sensing field to speed up the codeword assignment process. The first modification exploits the repetitive nature of music signals, whereas the second one relaxes a screening constraint that is specific to reconstruction but not for classifcation. Our experiments show that the proposed method enables the use of a dictionary of 10,000 codewords with runtime close to the case of using a dictionary of 1,000 codewords. Moreover, using the larger dictionary significantly improves the mean average precision (MAP) from 0.219 to 0.246 for tagging thousands of tracks with 147 possible genre tags.

Bagging is one the most classic ensemble learning techniques in the machine learning literature. The idea is to generate multiple subsets of the training data via bootstrapping (random sampling with replacement), and then aggregate the output of the models trained from each subset via voting or averaging. As music is a temporal signal, we propose and study two bagging methods in this paper: the inter-song instance bagging that bootstraps song-level features, and the intra-song instance bagging that draws bootstrapping samples directly from short-time features for each training song. In particular, we focus on the latter method, as it better exploits the temporal information of music signals. The bagging methods result in surprisingly effective models for music auto-tagging: incorporating the idea to a simple linear support vector machine (SVM) based system yields accuracies that are comparable or even superior to state-of-the-art, possibly more sophisticated methods for three different datasets. As the bagging method is a meta algorithm, it holds the promise of improving other MIR systems.

This paper presents the AWtoolbox, an open-source software designed for extracting the audio word (AW) representation of audio signals. The toolbox comes with a graphical user interface that helps a user design custom AW extraction pipelines and various algorithms for feature encoding, dictionary learning, result rectification, pooling, normalization and others. This paper also reports a benchmark comparing eight AW representations computed by the toolbox against state-of-the-art low-level and mid-level timbre, rhythmic and tonal descriptors of music and sound. The evaluation result shows that sparse coding (SC) based AW representation leads to very competitive performances across the three tested sound and music classification tasks. AWtoolbox is available for download at http://mac.citi.sinica.edu.tw/awtoolbox.

There has been an increasing attention on learning feature representations from the complex, high-dimensional audio data applied in various music information retrieval (MIR) problems. Unsupervised feature learning techniques, such as sparse coding and deep belief networks have been utilized to represent music information as a term-document structure comprising of elementary audio codewords. Despite the widespread use of such bag-of-frames (BoF) model, few attempts have been made to systematically compare different component settings. Moreover, whether techniques developed in the text retrieval community are applicable to audio codewords is poorly understood. To further our understanding of the BoF model, we present in this paper a comprehensive evaluation that compares a large number of BoF variants on three different MIR tasks, by considering different ways of low-level feature representation, codebook construction, codeword assignment, segment-level and song-level feature pooling, tf-idf term weighting, power normalization, and dimension reduction. Our evaluations lead to the following findings: 1) modeling music information by two levels of abstraction improves the result for difficult tasks such as predominant instrument recognition, 2) tf-idf weighting and power normalization improve system performance in general, 3) topic modeling methods such as latent Dirichlet allocation does not work for audio codewords.

This paper concerns the development of a music dictionary-based model for summarizing local feature descriptors computed over time. Comparing to a holistic representation, this text-like, bag-of-frames representation better captures the rich and time-varying information of music. However, the dictionary used in classical bag-of-frames model only captures frame-level elements of the music; thus, there exists a semantic gap between the dictionary element and commonly seen music description. In order to reduce the gap, a new feature representation called dual-layer bag-of-frames is proposed in this paper. It models the music with a two layer structure, where the first-layer dictionary captures the frame-level characteristics, and the second-layer dictionary captures the segment-level semantics. This hierarchical structure resembles the alphabet-word-document structure of text. Our result demonstrates that the proposed dual-layer bag-of-frames feature achieves state-of-the-art accuracy of music genre classification. The classification accuracy for the GTZAN benchmark reaches 86.7% with dictionary trained from GTZAN, and 83.6% with dictionary trained from another data set USPOP.

This paper concerns the development of a music codebook for summarizing local feature descriptors computed over time. Comparing to a holistic representation, this text-like representation better captures the rich and time-varying information of music. We systematically compare a number of existing codebook generation techniques and also propose a new one that incorporates labeled data in the dictionary learning process. Several aspects of the encoding system such as local feature extraction and codeword encoding are also analyzed. Our result demonstrates the superiority of sparsity-enforced dictionary learning over conventional VQ-based or exemplar-based methods. With the new supervised dictionary learning algorithm and the optimal settings inferred from the performance study, we achieve state-of-the-art accuracy of music genre classification using just the log-power spectrogram as the local feature descriptor. The classification accuracies for benchmark datasets GTZAN and IS-MIR2004Genre are 84.7% and 90.8%, respectively.

Thanks to the development of music audio analysis, state-of-the-art techniques can now detect musical attributes such as timbre, rhythm, and pitch with certain level of reliability and effectiveness. An emerging body of research has begun to model the high-level perceptual properties of music listening, including the mood and the preferable listening context of a music piece. Towards this goal, we propose a novel text-like feature representation that encodes the rich and time-varying information of music using a composite of features extracted from the song lyrics and audio signals. In particular, we investigate dictionary learning algorithms to optimize the generation of local feature descriptors and also probabilistic topic models to group semantically relevant text and audio words. This text-like representation leads to significant improvement in automatic mood classification over conventional audio features.