Learn-to-adapt: Concept drift adaptation for hybrid multiple streams
Introduction
Social media and autonomous systems generate enormous volumes of data continually over time, i.e., data streams, such as the fields of automatic driving [1], wearable devices, weather forecasts, and e-commerce [2]. Developing efficient data analytics and learning techniques for data streams to facilitate precise decision-making in real-world applications is in high demand [3]. One of the most challenging problems in the prediction of data streams is concept drift, where the data distribution irregularly changes over time [4]. When concept drift occurs, the prediction ability of an offline-trained model will be destroyed [5] since a well-trained model with previous data cannot obtain accurate prediction results when it is applied to newly arriving data [6], [7].
Concept drift adaptation (CDA) methods aim to tackle the concept drift problem in data streams by continuously retraining outdated models [8], namely by discarding previous classifiers and training a new classifier with the newly arriving data when concept drift occurs. In this way, the classifier can always fit the newest data distribution. CDA methods have been extensively studied and developed in recent years by introducing many state-of-the-art ideas, such as automated machine learning (AutoML) based methods [9], [10]. These methods aim to design automatic operations to search for the optimal pipelines and adjust the pipelines continually when the data distribution changes over time [11]. Online incremental learning-based methods are proposed to accumulate knowledge for the evolving adaptation problem or data streaming regression [12], [13]. In addition, lifelong learning is applied to develop models that can expand their knowledge to make it suitable for new concepts [14]. For instance, the streaming decision tree [15] deals with the concept drift problem by using a proposed streaming class-conditional attribute estimation to avoid knowledge forgetting. These existing studies have shown that CDA methods can effectively deal with data streams with changing distributions [16], [17]. So far, most existing CDA methods are designed for a single-labeled stream, where the label arrives after obtaining the features for each data instance.
However, real-world applications often involve hybrid multiple streams where massive labeled and unlabeled streams arrive simultaneously [18]. For example, social media platforms like TikTok always handles multiple data streams (e.g., images or videos) from different users [19], [20], with each user generating one data stream. Real-time sentiment prediction in social media requires users’ true sentiment to train the model. However, some users are not willing to provide the ground-truth. Therefore, the prediction method needs to handle labeled and unlabeled streams at the same time. When using existing single-stream aimed CDA methods to deal with this kind of multi-stream tasks, they cannot predict the unlabeled streams because they can only handle each stream with delayed labels separately [21].
We propose a more challenging and general setting to handle such hybrid multiple streams, named concept drift adaptation for hybrid multiple streams (HMS-CDA) which contains unlabeled drifting streams. As shown in Fig. 1, online unsupervised domain adaptation (UDA) refers to the continual adaption from labeled source domains (without drifts) to the target domain [22], [12]. Existing concept drift adaptation (CDA) mainly focuses on a single drifting stream with delayed labels [5]. Labeled multiple stream adaptation (LMS-CDA) deals with labeled and drifting data streams [23]. Compared with these existing settings, HMS-CDA places fewer limitations on data streams, where streams can be stationary (non-drifting) or non-stationary (drifting), and streams can be labeled or unlabeled. Since HMS-CDA does not require fully labeled streams, this can effectively save the cost of annotation and protect the privacy of users’ information.
In this proposed setting, we assume that one drifting stream is always labeled (denoted as the labeled source stream) and other drifting streams are always unlabeled (denoted as unlabeled target streams). As shown in Fig. 2, the distribution of the labeled source stream (stream A) is biased towards any target stream (stream I or stream M), and concept drifts may occur synchronously or asynchronously between any two streams. Details of the proposed setting will be explained with our proposed framework in Section 4. Compared with the existing CDA studies, there are three main problems need to be solved in HMS-CDA: a) unlabeled adaptation and prediction, b) distribution bias between streams, c) Asynchronous drift adaptation. In other words, the goal of HMS-CDA is how to predict the unlabeled and drifting streams, or how to efficiently use the source stream with the label information to predict target streams and address the concurrent and non-concurrent drifts between two streams at the same time. A possible solution is to use unsupervised domain adaptation (UDA) methods to classify unlabeled data [24]. However, UDA assumes a fixed relationship between the labeled data and unlabeled data. The relationship means there is only domain bias between the source and target domains while the distribution of each domain is fixed instead of evolving over time. Therefore, UDA method is not designed to react to concept drift and track the evolving data, leading to prediction failure when applied in drifting streams [25].
Inspired by [26], the learning-to-learn (meta-learning) methods have made considerable advances in the few-shot learning domains, demonstrating a promising performance for fast adaptation. Thus, we cast the HMS-CDA problem into a Learn-to-Adapt framework (L2A). Specifically, the proposed framework comprises two stages as shown in Fig. 3: (1) the meta-training stage aims at building the ability of invariant representations for adapting different streams and accumulating the existing knowledge to accelerate future learning; (2) the online adaptation stage updates the old model by exploiting and adapting the knowledge of downstream concepts, preventing old knowledge from being overwritten. Compared with the existing CDA methods, L2A not only adapts the classifiers, but also integrates the feature extractor and classifiers into a feedback process.
To summarize, the contributions of our work are listed as follows:
(1) We propose a new setting concept drift adaptation for hybrid multiple streams (HMS-CDA), which is more challenging and practical in streaming learning scenarios.
(2) A learn-to-adapt framework (L2A) is designed to address the challenges of the new setting. The framework adapts knowledge from a labeled drifting stream to unlabeled target streams, but also contributes to dealing with more complex and multiple high-dimensional data streams.
(3) This study CDA via meta-learning, paves a new way to address the problem of concept drift in multiple streams while broadening the practical applications of meta-learning.
Section snippets
Related works
In this study, we focus on making adaptations for unlabeled drifting data streams. Therefore, we reviewed the most related concept drift adaptation methods in this section. In addition, we talked about some state-of-the-art research related to our proposed method, such as continual learning, meta-learning, and evolving domain adaptation.
Proposed method: L2A for HMS-CDA
In this section, we formalize the HMS-CDA setting and present the proposed Learn-to-Adapt (L2A) framework in detail. Our study assumes that the source stream and multiple target streams have some internal related representations and share the same label space. Therefore, the intuition behind this approach is to predict the target labels by transferring the knowledge from the labeled source data to the target streams independently. The source labels can be manually or automatically annotated in
Experiment
To evaluate our proposed method (L2A), we conduct comprehensive experiments to show the performance of L2A when dealing with the HMS-CDA problem. Firstly, we introduce the benchmarks and give a description of how to simulate the proposed setting in Section 4.1. Then, the detailed implementation of the framework is introduced in Section 4.2 while the baseline and ablation design are described in Section 4.3. In Section 4.4, we conduct the average classification performance of all benchmarks and
Conclusion & further study
To address the inherent problems of hybrid multiple streams, we propose a new setting, Concept Drift Adaptation for Hybrid Multiple Streams (HMS-CDA), which is more aligned to real-world streaming data mining applications. Furthermore, we introduce a Learn-to-Adapt framework (L2A) to make adaptations for drifting stream prediction. This is the first time the Concept Drift Adaptation (CDA) problem has been addressed from the perspective of deep meta-learning. As a result, extensive experiments
CRediT authorship contribution statement
En Yu: Conceptualization, Methodology, Writing – original draft. Yiliao Song: Supervision, Investigation. Guangquan Zhang: Writing – review & editing, Supervision. Jie Lu: Writting - review & editing, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the Australian Research Council (ARC) under Discovery Project DP190101733 and Laureate Fellow project FL190100149.
En Yu received the B.S. and M.S. degrees in School of Information Science and Engineering, from Shandong Normal University, China, in 2016 and 2020, respectively. He was a Visiting Student with the Language Technology Institute, School of Computer Science, Carnegie Mellon University (CMU), USA, in 2019. He is currently pursuing the Ph.D. degree at Australian Artificial Intelligence Institute, University of Technology Sydney, Australia. His research interests include concept drift adaptation,
References (40)
- et al.
Elastic gradient boosting decision tree with adaptive iterations for concept drift adaptation
Neurocomputing
(2022) - et al.
A concept drift-tolerant case-base editing technique
Artif. Intell.
(2016) - et al.
Structural property-aware multilayer network embedding for latent factor analysis
Pattern Recogn.
(2018) - H. Liu, M. Long, J. Wang, Y. Wang, Learning to adapt to evolving domains, in: NeurIPS,...
- et al.
A survey on ensemble learning for data stream classification
ACM Computing Surveys (CSUR)
(2017) - et al.
Online transfer learning for concept drifting data streams
BigMine@KDD
(2019) - et al.
Learning under concept drift: A review
IEEE Trans. Knowl. Data Eng.
(2018) - Y. Song, J. Lu, A. Liu, H. Lu, G. Zhang, A segment-based drift adaptation method for data streams, IEEE Trans. Neural...
- et al.
A framework for multistream regression with direct density ratio estimation
- et al.
Analyzing and repairing concept drift adaptation in data stream classification
Mach. Learn.
(2021)
Atl: Autonomous knowledge transfer from many streaming processes
Incremental evolving domain adaptation
IEEE Trans. Knowl. Data Eng.
A drift region-based data sample filtering method
IEEE Trans. Cybern.
Learning data streams with changing distributions and temporal dependency
IEEE Trans. Neural Networks Learn. Syst.
An adaptive framework for multistream classification
Cited by (10)
A self-adaptive ensemble for user interest drift learning
2024, NeurocomputingConcept drift detection based on decision distribution in inconsistent information system
2023, Knowledge-Based SystemsCSAL: Cost sensitive active learning for multi-source drifting stream
2023, Knowledge-Based SystemsDetecting group concept drift from multiple data streams
2023, Pattern RecognitionCitation Excerpt :However, these works were all based on a single data stream [7,8]. Nowadays, huge amounts of streaming data, known as multiple data streams [9,10], are generated from multiple sources, such as sensors and marketing activities. In these real-world situations, it is difficult to detect concept drift by monitoring each data stream over time, since we need to monitor all data streams at the same time.
Concept Drift Adaptation by Exploiting Drift Type
2024, ACM Transactions on Knowledge Discovery from DataAn Augmented Learning Approach for Multiple Data Streams Under Concept Drift
2024, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
En Yu received the B.S. and M.S. degrees in School of Information Science and Engineering, from Shandong Normal University, China, in 2016 and 2020, respectively. He was a Visiting Student with the Language Technology Institute, School of Computer Science, Carnegie Mellon University (CMU), USA, in 2019. He is currently pursuing the Ph.D. degree at Australian Artificial Intelligence Institute, University of Technology Sydney, Australia. His research interests include concept drift adaptation, autonomous data stream learning, and meta leaning.
Yiliao Song received a M.S. in probability and statistics in mathematics from the School of Mathematics and Statistics, Lanzhou University, China, in 2015. She received the Ph.D. degree in computer science from University of Technology Sydney, Australia. Her research interests include concept drift, data stream mining and real-time prediction. She is a Postdoctoral Researcher at Australian Artificial Intelligence Institute, University of Technology Sydney, Australia.
Guangquan Zhang is an Australian Research Council (ARC) QEII Fellow, Associate Professor and the Director of the Decision Systems and e-Service Intelligent (DeSI) Research Laboratory at the Australian Artificial Intelligence Institute, University of Technology Sydney, Australia. He received his Ph.D in applied mathematics from Curtin University, Australia, in 2001. From 1993 to 1997, he was a full Professor in the Department of Mathematics, Hebei University, China. His main research interests lie in the area of fuzzy multi-objective, bilevel and group decision making, fuzzy measure, and machine learning. He has published six authored monographs, five edited research books, and over 500 papers including some 300 refereed journal articles. Dr. Zhang has won ten ARC Discovery Project grants and many other research grants, supervised 35 PhD students to completion. He has served as a Guest Editor for special issues of IEEE Transactions and other international journals.
Jie Lu (F’18) is a Distinguished Professor and the Director of Australian Artificial Intelligence Institute (AAII) at the University of Technology Sydney, Australia. She is also an IFSA Fellow and Australian Laureate Fellow. She received a PhD degree from Curtin University, Australia, in 2000. Her main research expertise is in transfer learning, concept drift, decision support systems and recommender systems. She has been awarded 10+ Australian Research Council (ARC) discovery grants and led 20 industry projects. She has published over 500 papers in IEEE transactions and other journals and conferences, supervised 50 PhD students to completion. She serves as Editor-In-Chief for Knowledge-Based Systems (Elsevier) and Editor-In-Chief for International Journal on Computational Intelligence Systems (Springer). She has delivered 35 keynote speeches at international conferences. She has received the UTS Medal for research excellence (2019), the IEEE Transactions on Fuzzy Systems Outstanding Paper Award (2019) and the Australian Most Innovative Engineer Award (2019).