Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. However, their increased computational and memory demands present significant challenges, especially when handling long sequences. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights. In response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20\% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve $2.8\times$ peak memory reduction while maintaining nearly the same quality, enabling a batch size increase from 4$\times$ (with KIVI alone) to 5$\times$ when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of ThinK. Our code has been made available at https://github.com/SalesforceAIResearch/ThinK.

Hallucinations (i.e., generating plausible but inaccurate content) and laziness (i.e. excessive refusals or defaulting to "I don't know") persist as major challenges in LLM reasoning. Current efforts to reduce hallucinations primarily focus on factual errors in knowledge-grounded tasks, often neglecting hallucinations related to faulty reasoning. Meanwhile, some approaches render LLMs overly conservative, limiting their problem-solving capabilities. To mitigate hallucination and laziness in reasoning tasks, we propose Automatic Curriculum Expert Iteration (Auto-CEI) to enhance LLM reasoning and align responses to the model’s capabilities--assertively answering within its limits and declining when tasks exceed them. In our method, Expert Iteration explores the reasoning trajectories near the LLM policy, guiding incorrect paths back on track to reduce compounding errors and improve robustness; it also promotes appropriate "I don't know" responses after sufficient reasoning attempts. The curriculum automatically adjusts rewards, incentivizing extended reasoning before acknowledging incapability, thereby pushing the limits of LLM reasoning and aligning its behaviour with these limits. We compare Auto-CEI with various SOTA baselines across logical reasoning, mathematics, and planning tasks, where Auto-CEI achieves superior alignment by effectively balancing assertiveness and conservativeness.

Achieving effective unified pretraining on large time series corpora remains an open challenge in developing time series foundation models. Existing methods, such as Moirai, introduce multiple projection layers for time series of different frequencies to account for high data heterogeneity. We identify major drawbacks to this human-imposed frequency-level model specialization. First, frequency is not a reliable indicator for grouping pretraining data. Second, time series can display varied distributions even within a short window. Frequency-level specialization overlooks the diversity at this granularity. To address these issues, this paper introduces Moirai-MoE, excluding human-defined data groupings while delegating the modeling of diverse time series patterns to the sparse mixture of experts (MoE) within Transformers. With this design, Moirai-MoE eliminates reliance on heuristics and enables automatic token-level specialization. Extensive evaluations on 39 datasets demonstrate the superiority of Moirai-MoE over state-of-the-art foundation models. This study also conducts comprehensive model analyses to explore the inner workings of time series MoE foundation models.

We introduce Reward-Guided Speculative Decoding (RSD), a novel framework aimed at improving the efficiency of inference in large language models (LLMs). RSD synergistically combines a lightweight draft model with a more powerful target model, incorporating a controlled bias to prioritize high-reward outputs, in contrast to existing speculative decoding methods that enforce strict unbiasedness. RSD employs a process reward model to evaluate intermediate decoding steps and dynamically decide whether to invoke the target model, optimizing the trade-off between computational cost and output quality. We theoretically demonstrate that a threshold-based mixture strategy achieves an optimal balance between resource utilization and performance. Extensive evaluations on challenging reasoning benchmarks, including Olympiad-level tasks, show that RSD delivers significant efficiency gains against decoding with the target model only (up to 4.4X fewer FLOPs), while achieving significant better accuracy than parallel decoding method on average (up to +3.5). These results highlight RSD as a robust and cost-effective approach for deploying LLMs in resource-intensive scenarios.

Automating GUI tasks remains challenging due to reliance on textual representations, platform-specific action spaces, and limited reasoning capabilities. We introduce Aguvis, a unified vision-based framework for autonomous GUI agents that directly operates on screen images, standardizes cross-platform interactions and incorporates structured reasoning via inner monologue. To enable this, we construct Aguvis data collection, a large-scale dataset with multimodal grounding and reasoning annotations, and develop a two-stage training pipeline that separates GUI grounding from planning and reasoning. Experiments show that Aguvis achieves state-of-the-art performance across offline and real-world online benchmarks, marking the first fully autonomous vision-based GUI agent that operates without closed-source models. We open-source all datasets, models, and training recipes at https://aguvis-project.github.io to advance future research.

Jierui Li, Hung Le, Yingbo Zhou, Caiming Xiong, Silvio Savarese, Doyen Sahoo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Graphical user interface (GUI) grounding, the ability to map natural language instructions to specific actions on graphical user interfaces, remains a critical bottleneck in computer use agent development. Current benchmarks oversimplify grounding tasks as short referring expressions, failing to capture the complexity of real-world interactions that require software commonsense, layout understanding, and fine-grained manipulation capabilities. To address these limitations, we introduce OSWorld-G, a comprehensive benchmark comprising 564 finely annotated samples across diverse task types including text matching, element recognition, layout understanding, and precise manipulation. Additionally, we synthesize and release the largest computer use grounding dataset Jedi, which contains 4 million examples through multi-perspective decoupling of tasks. Our multi-scale models trained on Jedi demonstrate its effectiveness by outperforming existing approaches on ScreenSpot-v2, ScreenSpot-Pro, and our OSWorld-G. Furthermore, we demonstrate that improved grounding with Jedi directly enhances agentic capabilities of general foundation models on complex computer tasks with state-of-the-art performance, improving from 23% to 51% on OSWorld. Through detailed ablation studies, we identify key factors contributing to grounding performance and verify that combining specialized data for different interface elements enables compositional generalization to novel interfaces. All benchmark, data, checkpoints, and code are open-sourced and available at https://osworld-grounding.github.io.

Large Language Models (LLMs) have already become quite proficient at solving simpler programming tasks like those in HumanEval or MBPP benchmarks. However, solving more complex and competitive programming tasks is still quite challenging for these models - possibly due to their tendency to generate solutions as monolithic code blocks instead of decomposing them into logical sub-tasks and sub-modules. On the other hand, experienced programmers instinctively write modularized code with abstraction for solving complex tasks, often reusing previously developed modules. To address this gap, we propose CodeChain, a novel framework for inference that elicits modularized code generation through a chain of self-revisions, each being guided by some representative sub-modules generated in previous iterations. Concretely, CodeChain first instructs the LLM to generate modularized codes through chain-of-thought prompting. Then it applies a chain of self-revisions by iterating the two steps: 1) extracting and clustering the generated sub-modules and selecting the cluster representatives as the more generic and re-usable implementations, and 2) augmenting the original chain-of-thought prompt with these selected module-implementations and instructing the LLM to re-generate new modularized solutions. We find that by naturally encouraging the LLM to reuse the previously developed and verified sub-modules, CodeChain can significantly boost both modularity as well as correctness of the generated solutions, achieving relative pass@1 improvements of 35\% on APPS and 76\% on CodeContests. It is shown to be effective on both OpenAI LLMs as well as open-sourced LLMs like WizardCoder. We also conduct comprehensive ablation studies with different methods of prompting, number of clusters, model sizes, program qualities, etc., to provide useful insights that underpin CodeChain's success.

Deep learning for time series forecasting has traditionally operated within a one-model-per-dataset framework, limiting its potential to leverage the game-changing impact of large pre-trained models. The concept of universal forecasting, emerging from pre-training on a vast collection of time series datasets, envisions a single Large Time Series Model capable of addressing diverse downstream forecasting tasks. However, constructing such a model poses unique challenges specific to time series data: (i) cross-frequency learning, (ii) accommodating an arbitrary number of variates for multivariate time series, and (iii) addressing the varying distributional properties inherent in large-scale data. To address these challenges, we present novel enhancements to the conventional time series Transformer architecture, resulting in our proposed Masked Encoder-based Universal Time Series Forecasting Transformer (Moirai). Trained on our newly introduced Large-scale Open Time Series Archive (LOTSA) featuring over 27B observations across nine domains, Moirai achieves competitive or superior performance as a zero-shot forecaster when compared to full-shot models. Code, data, and model weights can be found at https://github.com/SalesforceAIResearch/uni2ts.

Large language models (LLMs) for code are typically trained to align with natural language instructions to closely follow their intentions and requirements. However, in many practical scenarios, it becomes increasingly challenging for these models to navigate the intricate boundary between helpfulness and safety, especially against highly complex yet potentially malicious instructions. In this work, we introduce INDICT: a new framework that empowers LLMs with Internal Dialogues of Critiques for both safety and helpfulness guidance. The internal dialogue is a dual cooperative system between a safety-driven critic and a helpfulness-driven critic. Each critic provides analysis against the given task and corresponding generated response, equipped with external knowledge queried through relevant code snippets and tools like web search and code interpreter. We engage the dual critic system in both code generation stage as well as code execution stage, providing preemptive and post-hoc guidance respectively to LLMs. We evaluated INDICT on 8 diverse tasks across 8 programming languages from 5 benchmarks, using LLMs from 7B to 70B parameters. We observed that our approach can provide an advanced level of critiques of both safety and helpfulness analysis, significantly improving the quality of output codes (+10% absolute improvements in all models).

Dynamic Time Warping (DTW) has become the pragmatic choice for measuring distance between time series. However, it suffers from unavoidable quadratic time complexity when the optimal alignment matrix needs to be computed exactly. This hinders its use in deep learning architectures, where layers involving DTW computations cause severe bottlenecks. To alleviate these issues, we introduce a new metric for time series data based on the Optimal Transport (OT) framework, called Optimal Transport Warping (OTW). OTW enjoys linear time/space complexity, is differentiable and can be parallelized. OTW enjoys a moderate sensitivity to time and shape distortions, making it ideal for time series. We show the efficacy and efficiency of OTW on 1-Nearest Neighbor Classification and Hierarchical Clustering, as well as in the case of using OTW instead of DTW in Deep Learning architectures.

Deep learning has been actively applied to time series forecasting, leading to a deluge of new methods, belonging to the class of historical-value models. Yet, despite the attractive properties of time-index models, such as being able to model the continuous nature of underlying time series dynamics, little attention has been given to them. Indeed, while naive deep time-index models are far more expressive than the manually predefined function representations of classical time-index models, they are inadequate for forecasting, being unable to generalize to unseen time steps due to the lack of inductive bias. In this paper, we propose DeepTime, a meta-optimization framework to learn deep time-index models which overcome these limitations, yielding an efficient and accurate forecasting model. Extensive experiments on real world datasets in the long sequence time-series forecasting setting demonstrate that our approach achieves competitive results with state-of-the-art methods, and is highly efficient. Code is available at https://github.com/salesforce/DeepTime.

We introduce Merlion, an open-source machine learning library for time series. It features a unified interface for many commonly used models and datasets for forecasting and anomaly detection on both univariate and multivariate time series, along with standard pre/post-processing layers. It has several modules to improve ease-of-use, including a no-code visual dashboard, anomaly score calibration to improve interpetability, AutoML for hyperparameter tuning and model selection, and model ensembling. Merlion also provides an evaluation framework that simulates the live deployment of a model in production, and a distributed computing backend to run time series models at industrial scale. This library aims to provide engineers and researchers a one-stop solution to rapidly develop models for their specific time series needs and benchmark them across multiple datasets.

Food retrieval is an important task to perform analysis of food-related information, where we are interested in retrieving relevant information about the queried food item such as ingredients, cooking instructions, etc. In this paper, we investigate cross-modal retrieval between food images and cooking recipes. The goal is to learn an embedding of images and recipes in a common feature space, such that the corresponding image-recipe embeddings lie close to one another. Two major challenges in addressing this problem are 1) large intra-variance and small inter-variance across cross-modal food data; and 2) difficulties in obtaining discriminative recipe representations. To address these two problems, we propose Semantic-Consistent and Attention-based Networks (SCAN), which regularize the embeddings of the two modalities through aligning output semantic probabilities. Besides, we exploit a self-attention mechanism to improve the embedding of recipes. We evaluate the performance of the proposed method on the large-scale Recipe1M dataset, and show that we can outperform several state-of-the-art cross-modal retrieval strategies for food images and cooking recipes by a significant margin.

Video-grounded dialogues are very challenging due to (i) the complexity of videos which contain both spatial and temporal variations, and (ii) the complexity of user utterances which query different segments and/or different objects in videos over multiple dialogue turns.However, existing approaches to video-grounded dialogues often focus on superficial temporal-level visual cues, but neglect more fine-grained spatial signals from videos.To address this drawback, we propose Bi-directional Spatio-Temporal Learning (BiST), a vision-language neural framework for high-resolution queries in videos based on textual cues.Specifically, our approach not only exploits both spatial and temporal-level information, but also learns dynamic information diffusion between the two feature spaces through spatial-to-temporal and temporal-tospatial reasoning.The bidirectional strategy aims to tackle the evolving semantics of user queries in the dialogue setting.The retrieved visual cues are used as contextual information to construct relevant responses to the users.Our empirical results and comprehensive qualitative analysis show that BiST achieves competitive performance and generates reasonable responses on a large-scale AVSD benchmark.We also adapt our BiST models to the Video QA setting, and substantially outperform prior approaches on the TGIF-QA benchmark.* This work was mostly

Building an end-to-end conversational agent for multi-domain task-oriented dialogues has been an open challenge for two main reasons.First, tracking dialogue states of multiple domains is non-trivial as the dialogue agent must obtain complete states from all relevant domains, some of which might have shared slots among domains as well as unique slots specifically for one domain only.Second, the dialogue agent must also process various types of information across domains, including dialogue context, dialogue states, and database, to generate natural responses to users.Unlike the existing approaches that are often designed to train each module separately, we propose "UniConv" -a novel unified neural architecture for end-to-end conversational systems in multi-domain task-oriented dialogues, which is designed to jointly train (i) a Bi-level State Tracker which tracks dialogue states by learning signals at both slot and domain level independently, and (ii) a Joint Dialogue Act and Response Generator which incorporates information from various input components and models dialogue acts and target responses simultaneously.We conduct comprehensive experiments in dialogue state tracking, contextto-text, and end-to-end settings on the Multi-WOZ2.1 benchmark, achieving superior performance over competitive baselines.

Despite significant advances in deep learning based object detection in recent years, training effective detectors in a small data regime remains an open challenge. This is very important since labelling training data for object detection is often very expensive and time-consuming. In this paper, we investigate the problem of few-shot object detection, where a detector has access to only limited amounts of annotated data. Based on the meta-learning principle, we propose a new meta-learning framework for object detection named "Meta-RCNN", which learns the ability to perform few-shot detection via meta-learning. Specifically, Meta-RCNN learns an object detector in an episodic learning paradigm on the (meta) training data. This learning scheme helps acquire a prior which enables Meta-RCNN to do few-shot detection on novel tasks. Built on top of the popular Faster RCNN detector, in Meta-RCNN, both the Region Proposal Network (RPN) and the object classification branch are meta-learned. The meta-trained RPN learns to provide class-specific proposals, while the object classifier learns to do few-shot classification. The novel loss objectives and learning strategy of Meta-RCNN can be trained in an end-to-end manner. We demonstrate the effectiveness of Meta-RCNN in few-shot detection on three datasets (Pascal-VOC, ImageNet-LOC and MSCOCO) with promising results.

Developing Video-Grounded Dialogue Systems (VGDS), where a dialogue is conducted based on visual and audio aspects of a given video, is significantly more challenging than traditional image or text-grounded dialogue systems because (1) feature space of videos span across multiple picture frames, making it difficult to obtain semantic information; and(2) a dialogue agent must perceive and process information from different modalities (audio, video, caption, etc.) to obtain a comprehensive understanding.Most existing work is based on RNNs and sequence-to-sequence architectures, which are not very effective for capturing complex long-term dependencies (like in videos).To overcome this, we propose Multimodal Transformer Networks (MTN) to encode videos and incorporate information from different modalities.We also propose queryaware attention through an auto-encoder to extract query-aware features from non-text modalities.We develop a training procedure to simulate token-level decoding to improve the quality of generated responses during inference.We get state of the art performance on Dialogue System Technology Challenge 7 (DSTC7).Our model also generalizes to another multimodal visual-grounded dialogue task, and obtains promising performance.

Food computing is playing an increasingly important role in human daily life, and has found tremendous applications in guiding human behavior towards smart food consumption and healthy lifestyle. An important task under the food-computing umbrella is retrieval, which is particularly helpful for health related applications, where we are interested in retrieving important information about food (e.g., ingredients, nutrition, etc.). In this paper, we investigate an open research task of cross-modal retrieval between cooking recipes and food images, and propose a novel framework Adversarial Cross-Modal Embedding (ACME) to resolve the cross-modal retrieval task in food domains. Specifically, the goal is to learn a common embedding feature space between the two modalities, in which our approach consists of several novel ideas: (i) learning by using a new triplet loss scheme together with an effective sampling strategy, (ii) imposing modality alignment using an adversarial learning strategy, and (iii) imposing cross-modal translation consistency such that the embedding of one modality is able to recover some important information of corresponding instances in the other modality. ACME achieves the state-of-the-art performance on the benchmark Recipe1M dataset, validating the efficacy of the proposed technique.

An important aspect of health monitoring is effective logging of food consumption. This can help management of diet-related diseases like obesity, diabetes, and even cardiovascular diseases. Moreover, food logging can help fitness enthusiasts, and people who wanting to achieve a target weight. However, food-logging is cumbersome, and requires not only taking additional effort to note down the food item consumed regularly, but also sufficient knowledge of the food item consumed (which is difficult due to the availability of a wide variety of cuisines). With increasing reliance on smart devices, we exploit the convenience offered through the use of smart phones and propose a smart-food logging system: FoodAI, which offers state-of-the-art deep-learning based image recognition capabilities. FoodAI has been developed in Singapore and is particularly focused on food items commonly consumed in Singapore. FoodAI models were trained on a corpus of 400,000 food images from 756 different classes.

Kernel-based regression represents an important family of learning techniques for solving challenging regression tasks with non-linear patterns. Despite being studied extensively, most of the existing work suffers from two major drawbacks as follows: (i) they are often designed for solving regression tasks in a batch learning setting, making them not only computationally inefficient and but also poorly scalable in real-world applications where data arrives sequentially; and (ii) they usually assume that a fixed kernel function is given prior to the learning task, which could result in poor performance if the chosen kernel is inappropriate. To overcome these drawbacks, this work presents a novel scheme of Online Multiple Kernel Regression (OMKR), which sequentially learns the kernel-based regressor in an online and scalable fashion, and dynamically explore a pool of multiple diverse kernels to avoid suffering from a single fixed poor kernel so as to remedy the drawback of manual/heuristic kernel selection. The OMKR problem is more challenging than regular kernel-based regression tasks since we have to on-the-fly determine both the optimal kernel-based regressor for each individual kernel and the best combination of the multiple kernel regressors. We propose a family of OMKR algorithms for regression and discuss their application to time series prediction tasks including application to AR, ARMA, and ARIMA time series. We develop novel approaches to make OMKR scalable for large datasets, to counter the problems arising from an unbounded number of support vectors. We also explore the effect of kernel combination at prediction level and at the representation level. Finally, we conduct extensive experiments to evaluate the empirical performance on both real-world regression and times series prediction tasks.

Deep Neural Networks (DNNs) are typically trained by backpropagation in a batch setting, requiring the entire training data to be made available prior to the learning task. This is not scalable for many real-world scenarios where new data arrives sequentially in a stream. We aim to address an open challenge of ``Online Deep Learning" (ODL) for learning DNNs on the fly in an online setting. Unlike traditional online learning that often optimizes some convex objective function with respect to a shallow model (e.g., a linear/kernel-based hypothesis), ODL is more challenging as the optimization objective is non-convex, and regular DNN with standard backpropagation does not work well in practice for online settings. We present a new ODL framework that attempts to tackle the challenges by learning DNN models which dynamically adapt depth from a sequence of training data in an online learning setting. Specifically, we propose a novel Hedge Backpropagation (HBP) method for online updating the parameters of DNN effectively, and validate the efficacy on large data sets (both stationary and concept drifting scenarios).

One critical deficiency of traditional online kernel learning methods is their unbounded and growing number of support vectors in the online learning process, making them inefficient and non-scalable for large-scale applications. Recent studies on scalable online kernel learning have attempted to overcome this shortcoming, e.g., by imposing a constant budget on the number of support vectors. Although they attempt to bound the number of support vectors at each online learning iteration, most of them fail to bound the number of support vectors for the final output hypothesis, which is often obtained by averaging the series of hypotheses over all the iterations. In this article, we propose a novel framework for bounded online kernel methods, named “Sparse Passive-Aggressive (SPA)” learning, which is able to yield a final output kernel-based hypothesis with a bounded number of support vectors. Unlike the common budget maintenance strategy used by many existing budget online kernel learning approaches, the idea of our approach is to attain the bounded number of support vectors using an efficient stochastic sampling strategy that samples an incoming training example as a new support vector with a probability proportional to its loss suffered. We theoretically prove that SPA achieves an optimal mistake bound in expectation, and we empirically show that it outperforms various budget online kernel learning algorithms. Finally, in addition to general online kernel learning tasks, we also apply SPA to derive bounded online multiple-kernel learning algorithms, which can significantly improve the scalability of traditional Online Multiple-Kernel Classification (OMKC) algorithms while achieving satisfactory learning accuracy as compared with the existing unbounded OMKC algorithms.

Mining data streams has been an important open research problem in the era of big data analytics. This paper investigates supervised machine learning techniques for mining data streams with application to online anomaly detection. Unlike conventional data mining tasks, mining data streams for online anomaly detection has several challenges: (i) data arriving sequentially and increasing rapidly, (ii) highly class-imbalanced distributions; and (iii) complex anomaly patterns that could evolve dynamically. To tackle these challenges, we propose Cost-Sensitive Online Multiple Kernel Classification (CSOMKC) for comprehensively mining data streams and demonstrate its application to online anomaly detection. Specifically, CSOMKC learns a kernel-based cost-sensitive prediction model for imbalanced data streams in a sequential or online learning fashion, in which a pool of multiple diverse kernels is dynamically explored. The optimal kernel predictor and the multiple kernel combination are learnt together, and simultaneously class imbalance issues are addressed. We perform theoretical and extensive empirical analysis of the proposed algorithms.

Detecting temporal patterns is one of the most prevalent challenges while mining data. Often, timestamps or information about when certain instances or events occurred can provide us with critical information to recognize temporal patterns. Unfortunately, most existing techniques are not able to fully extract useful temporal information based on the time (especially at different resolutions of time). They miss out on 3 crucial factors: (i) they do not distinguish between timestamp features (which have cyclical or periodic properties) and ordinary features; (ii) they are not able to detect patterns exhibited at different resolutions of time (e.g. different patterns at the annual level, and at the monthly level); and (iii) they are not able to relate different features (e.g. multimodal features) of instances with different temporal properties (e.g. while predicting stock prices, stock fundamentals may have annual patterns, and at the same time factors like peer stock prices and global markets may exhibit daily patterns). To solve these issues, we offer a novel multiple-kernel learning view and develop Temporal Kernel Descriptors which utilize Kernel functions to comprehensively detect temporal patterns by deriving relationship of instances with the time features. We automatically learn the optimal kernel function, and hence the optimal temporal similarity between two instances. We formulate the optimization as a Multiple Kernel Learning (MKL) problem. We empirically evaluate its performance by solving the optimization using Online MKL.

On-line portfolio selection is a practical financial engineering problem, which aims to sequentially allocate capital among a set of assets in order to maximize long-term return. In recent years, a variety of machine learning algorithms have been proposed to address this challenging problem, but no comprehensive open-source toolbox has been released for various reasons. This article presents the first open-source toolbox for "On-Line Portfolio Selection" (OLPS), which implements a collection of classical and state-of-the-art strategies powered by machine learning algorithms. We hope that OLPS can facilitate the development of new learning methods and enable the performance benchmarking and comparisons of different strategies. OLPS is an open-source project released under Apache License (version 2.0), which is available at github.com/OLPS/OLPS or OLPS.stevenhoi.org.

On-line portfolio selection, a fundamental problem in computational finance, has attracted increasing interest from artificial intelligence and machine learning communities in recent years. Empirical evidence shows that stock's high and low prices are temporary and stock prices are likely to follow the mean reversion phenomenon. While existing mean reversion strategies are shown to achieve good empirical performance on many real datasets, they often make the single-period mean reversion assumption, which is not always satisfied, leading to poor performance in certain real datasets. To overcome this limitation, this article proposes a multiple-period mean reversion, or so-called “Moving Average Reversion” (MAR), and a new on-line portfolio selection strategy named “On-Line Moving Average Reversion” (OLMAR), which exploits MAR via efficient and scalable online machine learning techniques. From our empirical results on real markets, we found that OLMAR can overcome the drawbacks of existing mean reversion algorithms and achieve significantly better results, especially on the datasets where existing mean reversion algorithms failed. In addition to its superior empirical performance, OLMAR also runs extremely fast, further supporting its practical applicability to a wide range of applications. Finally, we have made all the datasets and source codes of this work publicly available at our project website: http://OLPS.stevenhoi.org/.

Kernel-based regression represents an important family of learning techniques for solving challenging regression tasks with non-linear patterns. Despite being studied extensively, most of the existing work suffers from two major drawbacks: (i) they are often designed for solving regression tasks in a batch learning setting, making them not only computationally inefficient and but also poorly scalable in real-world applications where data arrives sequentially; and (ii) they usually assume a fixed kernel function is given prior to the learning task, which could result in poor performance if the chosen kernel is inappropriate. To overcome these drawbacks, this paper presents a novel scheme of Online Multiple Kernel Regression (OMKR), which sequentially learns the kernel-based regressor in an online and scalable fashion, and dynamically explore a pool of multiple diverse kernels to avoid suffering from a single fixed poor kernel so as to remedy the drawback of manual/heuristic kernel selection. The OMKR problem is more challenging than regular kernel-based regression tasks since we have to on-the-fly determine both the optimal kernel-based regressor for each individual kernel and the best combination of the multiple kernel regressors. In this paper, we propose a family of OMKR algorithms for regression and discuss their application to time series prediction tasks. We also analyze the theoretical bounds of the proposed OMKR method and conduct extensive experiments to evaluate its empirical performance on both real-world regression and times series prediction tasks.