Mixture of Online and Offline Experts for Non-Stationary Time Series
Authors: Zhilin Zhao, Longbing Cao, Yuanyu Wan
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, MOOE outperforms state-of-the-art methods for handling non-stationary time series. ... In this section, we present empirical analysis to support our proposed theory and model. ... Predictive Accuracy Real-World Non-Stationary Time Series To verify the effect of the proposed MOOE method, we perform comparison experiments following the setup (Zhao, Cai, and Zhou 2018). Specifically, we use eight real-world non-stationary time series datasets... The comparison results are reported in Table 1. |
| Researcher Affiliation | Academia | 1 School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China 2 School of Computing, Macquarie University, Sydney, Australia 3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China 4 School of Software Technology, Zhejiang University, Ningbo, China |
| Pseudocode | Yes | Algorithm 1: MOOE 1: Input: step size ν online expert w K 1 offline expert set {w1, . . . , w K 1} 2: Initialize α1 1 < α2 1 < < αK 1 according to: αk 1 = K + 1 (K + 1 k)(K + 2 k)K , k [K] 3: for t = 1,. .. ,T do 4: Receive online expert w K t 5: Assign offline expert wk t = wk, k [K 1] 6: Output weighted average: wt = PK k=1 αk t wk t 7: Receive the loss function ft( ) 8: Update expert weights: αk t = αk t e νft(wk t ) PK k =1 αk t e νft(wk t ) , k [K] 9: Send gradient ft(w K t ) to the online expert 10: end for |
| Open Source Code | Yes | 1The source code is publicly available at: https://github.com/Lawliet-zzl/MOOE. |
| Open Datasets | Yes | Regret on Synthetic and Real-World Datasets We address binary classification on non-stationary time series and compare MOOE with OGD using both synthetic and real-world datasets (ijcnn and cod-rna) from the LIBSVM repository (Chang and Lin 2011). ... Specifically, we use eight real-world non-stationary time series datasets, including Usenet (Katakis, Tsoumakas, and Vlahavas 2008), Weather (Elwell and Polikar 2011), Gas Sensor (Vergara et al. 2012), Powersupply (Dau et al. 2019), Electricity (Harries 1999) ,Covertype (Sun et al. 2018), WESAD (Schmidt et al. 2018), and Kitsune (Mirsky et al. 2018). |
| Dataset Splits | No | The paper describes a non-stationary time series scenario where data arrives continuously in online intervals which then become offline intervals of size B. It states, 'In the experiments, we adopt the maximum sample size of an interval B = 50.' While this defines how data is processed sequentially, it does not specify explicit training/validation/test splits of a total dataset in terms of percentages, absolute counts, or predefined static partitions, which are typically required for reproducing experiments in a traditional machine learning context. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory, or cloud instance types) used for running the experiments. It only discusses the experimental setup at an algorithmic level. |
| Software Dependencies | No | The paper mentions using 'Online Gradient Descent (OGD) (Zinkevich 2003) method as an instance' but does not specify any software names with version numbers for libraries, frameworks, or programming languages used for implementation. |
| Experiment Setup | Yes | In the experiments, we adopt the maximum sample size of an interval B = 50 and the maximal number of maintained experts Kmax = 25. ... The MOOE method with step sizes {ν = T , ηt = D βt, t [T]} guarantees the following regret for all 1 T B. |