U Kang

Dongjin Choi, Jun-Gi Jang, U Kang

How can we capture the hidden properties from a tensor and a matrix data simultaneously in a fast, accurate, and scalable way? Coupled matrix-tensor factorization (CMTF) is a major tool to extract latent factors from a tensor and matrices at once. Designing an accurate and efficient CMTF method has become more crucial as the size and dimension of real-world data are growing explosively. However, existing methods for CMTF suffer from lack of accuracy, slow running time, and limited scalability. In this paper, we propose S3CMTF, a fast, accurate, and scalable CMTF method. S3CMTF achieves high speed by exploiting the sparsity of real-world tensors, and high accuracy by capturing inter-relations between factors. Also, S3CMTF accomplishes additional speed-up by lock-free parallel SGD update for multi-core shared memory systems. We present two methods, S3CMTF-naive and S3CMTF-opt. S3CMTF-naive is a basic version of S3CMTF, and S3CMTF-opt improves its speed by exploiting intermediate data. We theoretically and empirically show that S3CMTF is the fastest, outperforming existing methods. Experimental results show that S3CMTF is 11~43 times faster, and 2.1~4.1 times more accurate than existing methods. S3CMTF shows linear scalability on the number of data entries and the number of cores. In addition, we apply S3CMTF to Yelp recommendation tensor data coupled with 3 additional matrices to discover interesting properties.

Hyunsik Jeon, Jong-eun Lee, Jeongin Yun et al.

How can we recommend cold-start bundles to users? The cold-start problem in bundle recommendation is crucial because new bundles are continuously created on the Web for various marketing purposes. Despite its importance, existing methods for cold-start item recommendation are not readily applicable to bundles. They depend overly on historical information, even for less popular bundles, failing to address the primary challenge of the highly skewed distribution of bundle interactions. In this work, we propose CoHeat (Popularity-based Coalescence and Curriculum Heating), an accurate approach for cold-start bundle recommendation. CoHeat first represents users and bundles through graph-based views, capturing collaborative information effectively. To estimate the user-bundle relationship more accurately, CoHeat addresses the highly skewed distribution of bundle interactions through a popularity-based coalescence approach, which incorporates historical and affiliation information based on the bundle's popularity. Furthermore, it effectively learns latent representations by exploiting curriculum learning and contrastive learning. CoHeat demonstrates superior performance in cold-start bundle recommendation, achieving up to 193% higher nDCG@20 compared to the best competitor.

Seungcheol Park, Hojun Choi, U Kang

Given a pretrained encoder-based language model, how can we accurately compress it without retraining? Retraining-free structured pruning algorithms are crucial in pretrained language model compression due to their significantly reduced pruning cost and capability to prune large language models. However, existing retraining-free algorithms encounter severe accuracy degradation, as they fail to handle pruning errors, especially at high compression rates. In this paper, we propose K-prune (Knowledge-preserving pruning), an accurate retraining-free structured pruning algorithm for pretrained encoder-based language models. K-prune focuses on preserving the useful knowledge of the pretrained model to minimize pruning errors through a carefully designed iterative pruning process composed of knowledge measurement, knowledge-preserving mask search, and knowledge-preserving weight-tuning. As a result, K-prune shows significant accuracy improvements up to 58.02%p higher F1 score compared to existing retraining-free pruning algorithms under a high compression rate of 80% on the SQuAD benchmark without any retraining process.

Huiwen Xu, U Kang

Given a set of pre-trained models, how can we quickly and accurately find the most useful pre-trained model for a downstream task? Transferability measurement is to quantify how transferable is a pre-trained model learned on a source task to a target task. It is used for quickly ranking pre-trained models for a given task and thus becomes a crucial step for transfer learning. Existing methods measure transferability as the discrimination ability of a source model for a target data before transfer learning, which cannot accurately estimate the fine-tuning performance. Some of them restrict the application of transferability measurement in selecting the best supervised pre-trained models that have classifiers. It is important to have a general method for measuring transferability that can be applied in a variety of situations, such as selecting the best self-supervised pre-trained models that do not have classifiers, and selecting the best transferring layer for a target task. In this work, we propose TMI (TRANSFERABILITY MEASUREMENT WITH INTRA-CLASS FEATURE VARIANCE), a fast and accurate algorithm to measure transferability. We view transferability as the generalization of a pre-trained model on a target task by measuring intra-class feature variance. Intra-class variance evaluates the adaptability of the model to a new task, which measures how transferable the model is. Compared to previous studies that estimate how discriminative the models are, intra-class variance is more accurate than those as it does not require an optimal feature extractor and classifier. Extensive experiments on real-world datasets show that TMI outperforms competitors for selecting the top-5 best models, and exhibits consistently better correlation in 13 out of 17 cases.

Seungcheol Park, Jaehyeon Choi, Sojin Lee et al.

How can we compress language models without sacrificing accuracy? The number of compression algorithms for language models is rapidly growing to benefit from remarkable advances of recent language models without side effects due to the gigantic size of language models, such as increased carbon emissions and expensive maintenance fees. While numerous compression algorithms have shown remarkable progress in compressing language models, it ironically becomes challenging to capture emerging trends and identify the fundamental concepts underlying them due to the excessive number of algorithms. In this paper, we survey and summarize diverse compression algorithms including pruning, quantization, knowledge distillation, low-rank approximation, parameter sharing, and efficient architecture design. We not only summarize the overall trend of diverse compression algorithms but also select representative algorithms and provide in-depth analyses of them. We discuss the value of each category of compression algorithms, and the desired properties of low-cost compression algorithms which have a significant impact due to the emergence of large language models. Finally, we introduce promising future research topics based on our survey results.

Minjun Kim, Jaehyeon Choi, Jongkeun Lee et al.

Network quantization has proven to be a powerful approach to reduce the memory and computational demands of deep learning models for deployment on resource-constrained devices. However, traditional quantization methods often rely on access to training data, which is impractical in many real-world scenarios due to privacy, security, or regulatory constraints. Zero-shot Quantization (ZSQ) emerges as a promising solution, achieving quantization without requiring any real data. In this paper, we provide a comprehensive overview of ZSQ methods and their recent advancements. First, we provide a formal definition of the ZSQ problem and highlight the key challenges. Then, we categorize the existing ZSQ methods into classes based on data generation strategies, and analyze their motivations, core ideas, and key takeaways. Lastly, we suggest future research directions to address the remaining limitations and advance the field of ZSQ. To the best of our knowledge, this paper is the first in-depth survey on ZSQ.

Seungcheol Park, Sojin Lee, Jongjin Kim et al.

How can we accelerate large language models(LLMs) without sacrificing accuracy? The slow inference speed of LLMs hinders us to benefit from their remarkable performance in diverse applications. This is mainly because numerous sublayers are stacked together in LLMs. Sublayer pruning compresses and expedites LLMs via removing unnecessary sublayers. However, existing sublayer pruning algorithms are limited in accuracy since they naively select sublayers to prune, overlooking the different characteristics of each sublayer. In this paper, we propose SPRINT (Sublayer PRuning wIth LateNcy and Tunability Information), an accurate sublayer pruning method for LLMs. SPRINT accurately selects a target sublayer to prune by considering 1) the amount of latency reduction after pruning and 2) the tunability of sublayers. SPRINT iteratively prunes redundant sublayers and swiftly tunes the parameters of remaining sublayers. Experiments show that SPRINT achieves the best accuracy-speedup trade-off, exhibiting up to 23.88%p higher accuracy on zero-shot commonsense reasoning benchmarks compared to existing pruning algorithms.

Minjun Kim, Jaehyeon Choi, SeungJoo Lee et al.

How can we accurately classify graphs? Graph classification is a pivotal task in data mining with applications in social network analysis, web analysis, drug discovery, molecular property prediction, etc. Graph neural networks have achieved the state-of-the-art performance in graph classification, but they consistently struggle with overfitting. To mitigate overfitting, researchers have introduced various representation learning methods utilizing graph augmentation. However, existing methods rely on simplistic use of graph augmentation, which loses augmentation-induced differences and limits the expressiveness of representations. In this paper, we propose AugWard (Augmentation-Aware Training with Graph Distance and Consistency Regularization), a novel graph representation learning framework that carefully considers the diversity introduced by graph augmentation. AugWard applies augmentation-aware training to predict the graph distance between the augmented graph and its original one, aligning the representation difference directly with graph distance at both feature and structure levels. Furthermore, AugWard employs consistency regularization to encourage the classifier to handle richer representations. Experimental results show that AugWard gives the state-of-the-art performance in supervised, semi-supervised graph classification, and transfer learning.

Hyunwoo Lee, Jihyeong Jeon, Jaemin Hong et al.

How can we address distribution shifts in stock price data to improve stock price prediction accuracy? Stock price prediction has attracted attention from both academia and industry, driven by its potential to uncover complex market patterns and enhance decisionmaking. However, existing methods often fail to handle distribution shifts effectively, focusing on scaling or representation adaptation without fully addressing distributional discrepancies and shape misalignments between training and test data. We propose ReVol (Return-Volatility Normalization for Mitigating Distribution Shift in Stock Price Data), a robust method for stock price prediction that explicitly addresses the distribution shift problem. ReVol leverages three key strategies to mitigate these shifts: (1) normalizing price features to remove sample-specific characteristics, including return, volatility, and price scale, (2) employing an attention-based module to estimate these characteristics accurately, thereby reducing the influence of market anomalies, and (3) reintegrating the sample characteristics into the predictive process, restoring the traits lost during normalization. Additionally, ReVol combines geometric Brownian motion for long-term trend modeling with neural networks for short-term pattern recognition, unifying their complementary strengths. Extensive experiments on real-world datasets demonstrate that ReVol enhances the performance of the state-of-the-art backbone models in most cases, achieving an average improvement of more than 0.03 in IC and over 0.7 in SR across various settings.

Jun-Gi Jang, Jeongyoung Lee, Yong-chan Park et al.

How can we efficiently and accurately analyze an irregular tensor in a dual-way streaming setting where the sizes of two dimensions of the tensor increase over time? What types of anomalies are there in the dual-way streaming setting? An irregular tensor is a collection of matrices whose column lengths are the same while their row lengths are different. In a dual-way streaming setting, both new rows of existing matrices and new matrices arrive over time. PARAFAC2 decomposition is a crucial tool for analyzing irregular tensors. Although real-time analysis is necessary in the dual-way streaming, static PARAFAC2 decomposition methods fail to efficiently work in this setting since they perform PARAFAC2 decomposition for accumulated tensors whenever new data arrive. Existing streaming PARAFAC2 decomposition methods work in a limited setting and fail to handle new rows of matrices efficiently. In this paper, we propose Dash, an efficient and accurate PARAFAC2 decomposition method working in the dual-way streaming setting. When new data are given, Dash efficiently performs PARAFAC2 decomposition by carefully dividing the terms related to old and new data and avoiding naive computations involved with old data. Furthermore, applying a forgetting factor makes Dash follow recent movements. Extensive experiments show that Dash achieves up to 14.0x faster speed than existing PARAFAC2 decomposition methods for newly arrived data. We also provide discoveries for detecting anomalies in real-world datasets, including Subprime Mortgage Crisis and COVID-19.

Jinhong Jung, Jaemin Yoo, U Kang

Given a signed social graph, how can we learn appropriate node representations to infer the signs of missing edges? Signed social graphs have received considerable attention to model trust relationships. Learning node representations is crucial to effectively analyze graph data, and various techniques such as network embedding and graph convolutional network (GCN) have been proposed for learning signed graphs. However, traditional network embedding methods are not end-to-end for a specific task such as link sign prediction, and GCN-based methods suffer from a performance degradation problem when their depth increases. In this paper, we propose Signed Graph Diffusion Network (SGDNet), a novel graph neural network that achieves end-to-end node representation learning for link sign prediction in signed social graphs. We propose a random walk technique specially designed for signed graphs so that SGDNet effectively diffuses hidden node features. Through extensive experiments, we demonstrate that SGDNet outperforms state-of-the-art models in terms of link sign prediction accuracy.

Jaehun Jung, Jinhong Jung, U Kang

Temporal knowledge graphs (TKGs) inherently reflect the transient nature of real-world knowledge, as opposed to static knowledge graphs. Naturally, automatic TKG completion has drawn much research interests for a more realistic modeling of relational reasoning. However, most of the existing mod-els for TKG completion extend static KG embeddings that donot fully exploit TKG structure, thus lacking in 1) account-ing for temporally relevant events already residing in the lo-cal neighborhood of a query, and 2) path-based inference that facilitates multi-hop reasoning and better interpretability. In this paper, we propose T-GAP, a novel model for TKG completion that maximally utilizes both temporal information and graph structure in its encoder and decoder. T-GAP encodes query-specific substructure of TKG by focusing on the temporal displacement between each event and the query times-tamp, and performs path-based inference by propagating attention through the graph. Our empirical experiments demonstrate that T-GAP not only achieves superior performance against state-of-the-art baselines, but also competently generalizes to queries with unseen timestamps. Through extensive qualitative analyses, we also show that T-GAP enjoys from transparent interpretability, and follows human intuition in its reasoning process.

Dawon Ahn, Jun-Gi Jang, U Kang

Given a time-evolving tensor with missing entries, how can we effectively factorize it for precisely predicting the missing entries? Tensor factorization has been extensively utilized for analyzing various multi-dimensional real-world data. However, existing models for tensor factorization have disregarded the temporal property for tensor factorization while most real-world data are closely related to time. Moreover, they do not address accuracy degradation due to the sparsity of time slices. The essential problems of how to exploit the temporal property for tensor decomposition and consider the sparsity of time slices remain unresolved. In this paper, we propose TATD (Time-Aware Tensor Decomposition), a novel tensor decomposition method for real-world temporal tensors. TATD is designed to exploit temporal dependency and time-varying sparsity of real-world temporal tensors. We propose a new smoothing regularization with Gaussian kernel for modeling time dependency. Moreover, we improve the performance of TATD by considering time-varying sparsity. We design an alternating optimization scheme suitable for temporal tensor factorization with our smoothing regularization. Extensive experiments show that TATD provides the state-of-the-art accuracy for decomposing temporal tensors.

Hyun Dong Lee, Seongmin Lee, U Kang

How can we effectively regularize BERT? Although BERT proves its effectiveness in various downstream natural language processing tasks, it often overfits when there are only a small number of training instances. A promising direction to regularize BERT is based on pruning its attention heads based on a proxy score for head importance. However, heuristic-based methods are usually suboptimal since they predetermine the order by which attention heads are pruned. In order to overcome such a limitation, we propose AUBER, an effective regularization method that leverages reinforcement learning to automatically prune attention heads from BERT. Instead of depending on heuristics or rule-based policies, AUBER learns a pruning policy that determines which attention heads should or should not be pruned for regularization. Experimental results show that AUBER outperforms existing pruning methods by achieving up to 10% better accuracy. In addition, our ablation study empirically demonstrates the effectiveness of our design choices for AUBER.

Seongmin Lee, Hyunsik Jeon, U Kang

Given multiple source datasets with labels, how can we train a target model with no labeled data? Multi-source domain adaptation (MSDA) aims to train a model using multiple source datasets different from a target dataset in the absence of target data labels. MSDA is a crucial problem applicable to many practical cases where labels for the target data are unavailable due to privacy issues. Existing MSDA frameworks are limited since they align data without considering conditional distributions p(x|y) of each domain. They also miss a lot of target label information by not considering the target label at all and relying on only one feature extractor. In this paper, we propose Ensemble Multi-source Domain Adaptation with Pseudolabels (EnMDAP), a novel method for multi-source domain adaptation. EnMDAP exploits label-wise moment matching to align conditional distributions p(x|y), using pseudolabels for the unavailable target labels, and introduces ensemble learning theme by using multiple feature extractors for accurate domain adaptation. Extensive experiments show that EnMDAP provides the state-of-the-art performance for multi-source domain adaptation tasks in both of image domains and text domains.

Yong-chan Park, Jun-Gi Jang, U Kang

Given a time series vector, how can we efficiently compute a specified part of Fourier coefficients? Fast Fourier transform (FFT) is a widely used algorithm that computes the discrete Fourier transform in many machine learning applications. Despite its pervasive use, all known FFT algorithms do not provide a fine-tuning option for the user to specify one's demand, that is, the output size (the number of Fourier coefficients to be computed) is algorithmically determined by the input size. This matters because not every application using FFT requires the whole spectrum of the frequency domain, resulting in an inefficiency due to extra computation. In this paper, we propose a fast Partial Fourier Transform (PFT), a careful modification of the Cooley-Tukey algorithm that enables one to specify an arbitrary consecutive range where the coefficients should be computed. We derive the asymptotic time complexity of PFT with respect to input and output sizes, as well as its numerical accuracy. Experimental results show that our algorithm outperforms the state-of-the-art FFT algorithms, with an order of magnitude of speedup for sufficiently small output sizes without sacrificing accuracy.

Seungcheol Park, Huiwen Xu, Taehun Kim et al.

Given a set of heterogeneous source datasets with their classifiers, how can we quickly find the most useful source dataset for a specific target task? We address the problem of measuring transferability between source and target datasets, where the source and the target have different feature spaces and distributions. We propose Transmeter, a fast and accurate method to estimate the transferability of two heterogeneous multivariate datasets. We address three challenges in measuring transferability between two heterogeneous multivariate datasets: reducing time, minimizing domain gap, and extracting meaningful homogeneous representations. To overcome the above issues, we utilize a pre-trained source model, an adversarial network, and an encoder-decoder architecture. Extensive experiments on heterogeneous multivariate datasets show that Transmeter gives the most accurate transferability measurement with up to 10.3 times faster performance than its competitor. We also show that selecting the best source data with Transmeter followed by a full transfer leads to the best transfer accuracy and the fastest running time.

Jun-Gi Jang, Chun Quan, Hyun Dong Lee et al.

How can we efficiently compress Convolutional Neural Network (CNN) while retaining their accuracy on classification tasks? Depthwise Separable Convolution (DSConv), which replaces a standard convolution with a depthwise convolution and a pointwise convolution, has been used for building lightweight architectures. However, previous works based on depthwise separable convolution are limited when compressing a trained CNN model since 1) they are mostly heuristic approaches without a precise understanding of their relations to standard convolution, and 2) their accuracies do not match that of the standard convolution. In this paper, we propose FALCON, an accurate and lightweight method to compress CNN. FALCON uses GEP, our proposed mathematical formulation to approximate the standard convolution kernel, to interpret existing convolution methods based on depthwise separable convolution. By exploiting the knowledge of a trained standard model and carefully determining the order of depthwise separable convolution via GEP, FALCON achieves sufficient accuracy close to that of the trained standard model. Furthermore, this interpretation leads to developing a generalized version rank-k FALCON which performs k independent FALCON operations and sums up the result. Experiments show that FALCON 1) provides higher accuracy than existing methods based on depthwise separable convolution and tensor decomposition, and 2) reduces the number of parameters and FLOPs of standard convolution by up to a factor of 8 while ensuring similar accuracy. We also demonstrate that rank-k FALCON further improves the accuracy while sacrificing a bit of compression and computation reduction rates.

Hyunsik Jeon, Bonhun Koo, U Kang

Given a sparse rating matrix and an auxiliary matrix of users or items, how can we accurately predict missing ratings considering different data contexts of entities? Many previous studies proved that utilizing the additional information with rating data is helpful to improve the performance. However, existing methods are limited in that 1) they ignore the fact that data contexts of rating and auxiliary matrices are different, 2) they have restricted capability of expressing independence information of users or items, and 3) they assume the relation between a user and an item is linear. We propose DaConA, a neural network based method for recommendation with a rating matrix and an auxiliary matrix. DaConA is designed with the following three main ideas. First, we propose a data context adaptation layer to extract pertinent features for different data contexts. Second, DaConA represents each entity with latent interaction vector and latent independence vector. Unlike previous methods, both of the two vectors are not limited in size. Lastly, while previous matrix factorization based methods predict missing values through the inner-product of latent vectors, DaConA learns a non-linear function of them via a neural network. We show that DaConA is a generalized algorithm including the standard matrix factorization and the collective matrix factorization as special cases. Through comprehensive experiments on real-world datasets, we show that DaConA provides the state-of-the-art accuracy.

Mauro Scanagatta, Giorgio Corani, Marco Zaffalon et al.

Learning a Bayesian networks with bounded treewidth is important for reducing the complexity of the inferences. We present a novel anytime algorithm (k-MAX) method for this task, which scales up to thousands of variables. Through extensive experiments we show that it consistently yields higher-scoring structures than its competitors on complete data sets. We then consider the problem of structure learning from incomplete data sets. This can be addressed by structural EM, which however is computationally very demanding. We thus adopt the novel k-MAX algorithm in the maximization step of structural EM, obtaining an efficient computation of the expected sufficient statistics. We test the resulting structural EM method on the task of imputing missing data, comparing it against the state-of-the-art approach based on random forests. Our approach achieves the same imputation accuracy of the competitors, but in about one tenth of the time. Furthermore we show that it has worst-case complexity linear in the input size, and that it is easily parallelizable.

Haekyu Park, Hyunsik Jeon, Junghwan Kim et al.

How can we leverage social network data and observed ratings to correctly recommend proper items and provide a persuasive explanation for the recommendations? Many online services provide social networks among users, and it is crucial to utilize social information since recommendation by a friend is more likely to grab attention than the one from a random user. Also, explaining why items are recommended is very important in encouraging the users' actions such as actual purchases. Exploiting both ratings and social graph for recommendation, however, is not trivial because of the heterogeneity of the data. In this paper, we propose UniWalk, an explainable and accurate recommender system that exploits both social network and rating data. UniWalk combines both data into a unified graph, learns latent features of users and items, and recommends items to each user through the features. Importantly, it explains why items are recommended together with the recommendation results. Extensive experiments show that UniWalk provides the best explainability and achieves the state-of-the-art-accuracy.

Haekyu Park, Jinhong Jung, U Kang

Between matrix factorization or Random Walk with Restart (RWR), which method works better for recommender systems? Which method handles explicit or implicit feedback data better? Does additional information help recommendation? Recommender systems play an important role in many e-commerce services such as Amazon and Netflix to recommend new items to a user. Among various recommendation strategies, collaborative filtering has shown good performance by using rating patterns of users. Matrix factorization and random walk with restart are the most representative collaborative filtering methods. However, it is still unclear which method provides better recommendation performance despite their extensive utility. In this paper, we provide a comparative study of matrix factorization and RWR in recommender systems. We exactly formulate each correspondence of the two methods according to various tasks in recommendation. Especially, we newly devise an RWR method using global bias term which corresponds to a matrix factorization method using biases. We describe details of the two methods in various aspects of recommendation quality such as how those methods handle cold-start problem which typically happens in collaborative filtering. We extensively perform experiments over real-world datasets to evaluate the performance of each method in terms of various measures. We observe that matrix factorization performs better with explicit feedback ratings while RWR is better with implicit ones. We also observe that exploiting global popularities of items is advantageous in the performance and that side information produces positive synergy with explicit feedback but gives negative effects with implicit one.