Adaptive Model Selection using Meta Models and Drift Adaptation

Adaptive Model Selection using Meta Models and Drift Adaptation Abstract—Smart spaces that have deployed machine learning (ML) models for intelligent decision typically evolve over time. Model selection is a key step in deploying (ML) models, particularly in dynamic environments where data distribution shifts can impact model accuracy over time. Existing approaches often rely on evaluating many candidate models against predefined metrics, which is computationally expensive and neither suitable for real-time applications nor dynamic environments. This paper presents an adaptive model selection technique that combines meta-modeling with drift detection to improve efficiency and robustness. Meta-models are used to evaluate the suitability of candidate models under different constraints, such as predictive accuracy and computational cost, without requiring full evaluation on the target data. This reduces the overhead of model selection while preserving deployment quality. In addition, a drift detection mechanism monitors changes in the data distribution and updates the selection strategy accordingly. The proposed approach contributes to automated machine learning (AutoML) by enabling adaptive, efficient, and reliable model selection in real-time environments. Index Terms—Model selection, Drift detection, Meta Models, Model adaptation. I. INTRODUCTION Model selection is a fundamental step in deploying machine learning (ML) models, especially in real-world applications where both computational efficiency and adaptability are critical. Traditionally, selecting the most suitable model for a given dataset requires evaluating all candidate models against performance metrics such as accuracy or inference time [1]. While this exhaustive evaluation yields optimal choices, it is often impractical in real-time, resource-constrained or dynamic environments due to its computational cost. For example, in robot path planning [2], evaluating several navigation models on-the-fly to determine the best trade-off between accuracy and speed would consume significant computational resources, delaying the robot’s ability to react promptly to obstacles in a dynamic environment. Furthermore, the performance of ML models tends to degrade over time due to concept drift, where the statistical properties of the target variable change as the underlying data distribution evolves [3]. In such cases, even the most carefully selected model may become suboptimal, necessitating frequent re-evaluation or retraining. While there have been some efforts in detecting drifts and triggering re-training [4], these techniques are neither adaptive nor resource-optimal. To address these challenges, we propose an adaptive model selection technique that leverages meta-learning and drift detection. Our approach introduces the following contributions: 1) Meta-models for online model selection: we design meta-models that predict the most suitable ML model based on dataset characteristics (e.g., feature distributions, sample size, and statistical properties) and user objectives, without exhaustively evaluating all candidate models. 2) Drift-aware selection mechanisms: we introduce drift detection techniques that not only identify the presence of concept drift but also quantify its magnitude, enabling adaptive updates to the model selection strategy. 3) Multi-registry architecture: We propose a registrybased design for storing models trained under specific drift conditions. When drift is detected, the system first selects the relevant registry and then leverages metamodel predictions to identify the optimal model within it. This work contributes to the field of Automated Machine Learning (AutoML) [1] through adaptive, efficient, and robust model selection in dynamic environments with changing user objectives. It is especially suited for real-time and Edge computing applications, where rapid inference and adaptability to changing data distributions are crucial. The remainder of the paper is structured as follows. Section II reviews related work on AutoML and model selection. Section III presents our adaptive model selection technique. Section IV presents the experimental evaluation of our approach. Section V concludes the paper and outlines future work. II. RELATED WORK This section reviews existing work across three key research areas: AutoML and model selection, concept drift detection, and meta-learning approaches for model selection. AutoML and Model Selection: Existing AutoML frameworks aim to automate the process of selecting and tuning ML models for a given dataset. Auto-Sklearn [1] introduces a robust approach that combines Bayesian optimization with meta-learning to automate pipeline generation and hyperparameter selection. It leverages a historical performance database and a meta-learner to guide the search toward promising models. TPOT [5] builds on genetic programming to evolve ML pipelines, optimizing over combinations of models and pre-processing steps. Google Vizier [6] is a scalable blackbox optimization service designed for hyperparameter tuning using Gaussian Processes and transfer learning techniques. AutoMS [4] presents framework for automatic model selection in novelty detection tasks without requiring labeled validation data. The method leverages a Jackknife-based risk estimator to assess model generalization and select the most suitable detector from a pool of candidates. Gardner et al. [7] introduce a Bayesian active model selection framework that efficiently identifies the best model among a set of candidates by actively choosing informative queries. Their approach leverages Bayesian experimental design principles to minimize uncertainty about model performance while reducing the number of required evaluations. Despite their strengths, existing approaches rely heavily on full retraining and evaluation loops, which become impractical in real-time or resource-constrained scenarios. Additionally, they generally assume that data distributions remain stationary over time. Our method overcomes these limitations by bypassing the full evaluation loop at inference time through the use of offline-trained meta-models that make model recommendations instantly, based on current data characteristics and performance constraints. Concept Drift and Adaptive Learning: Real-world data often violates the independent and identically distributed (i.i.d.) assumption, resulting in a phenomenon known as concept drift, where the statistical properties of the input data change over time [3]. Gama et al. [8] provide a comprehensive taxonomy of drift types (sudden, gradual, incremental) and present frameworks such as Adaptive Sliding Window (ADWIN) [9] and Drift Detection Method (DDM) [10]), which monitor prediction error over time to trigger model updates. More recently, several approaches for dealing with concept drift have been developed, especially in Federated Learning (FL) settings. Flash [11] addresses concept drift in FL by combining client-side early stopping (to detect drift) with a server-side adaptive optimizer that adjusts learning rate when drift is detected. The method both matches the convergence guarantees of state-of-the-art adaptive optimizers and empirically shows improved performance in settings with drift, across diverse federated benchmarks. Jothimurugen et al. [12] consider the case where concept drift occurs not just over time but across different clients in a FL setup. They show that single global models suffer under such drift, and propose methods based on clustering clients and local drift detection to adapt more effectively. These adaptive strategies generally operate in an online fashion, constantly monitoring incoming data streams and retraining or reweighting models when drift is detected. While effective, such approaches are reactive and computationally intensive, particularly in high-frequency or large-scale settings. In contrast, our system proactively identifies the drift level using Maximum Mean Discrepancy (MMD) [13], a kernelbased distributional distance, and routes the input to the most appropriate pre-trained registry. This avoids frequent retraining, reduces computational burden, and provides robustness to a wide range of drift scenarios. Meta-Learning for Model Recommendation: Metalearning, or “learning to learn,” explores how past experiences across learning tasks can inform future decisions. Several meta-learning approaches for model selection have been proposed in the literature. Brazdil et al. [14] proposed early metalearning approaches where models are recommended based on meta-features such as number of features, instance count, skewness, and entropy. More recently, Triantafillou et al. [15] introduced MetaDataset, a benchmark for few-shot learning where metalearning models are trained to adapt rapidly across tasks sampled from different domains (e.g., ImageNet, Omniglot). AutoForecast [16] is a meta-learning framework that selects the best forecasting model for a new, unseen time-series dataset. It combines a “general meta-learner” capturing similarity across datasets (via meta-features) and a “time-series meta-learner” that learns how model performance evolves over successive time windows in each dataset. Hydra [17] is a meta-recommender that rapidly recommends anomaly detection models in multivariate time-series settings under real constraints However, most meta-learning research focuses on model accuracy or generalization performance, overlooking practical constraints like inference latency or memory overhead. Our approach builds on these foundations by training registry-specific meta-models that also consider user-defined constraints (e.g., target accuracy, maximum latency), enabling resource-aware and context-sensitive model selection. III. ADAPTIVE MODEL SELECTION This section presents the architecture our Adaptive Model Selector, shown in Figure 1. The Adaptive Model Selector is a technique that uses meta-models to select the best-performing ML models based on user-defined targets, such as accuracy and inference latency. The selector intelligently adapts to distributional changes in data (i.e., concept drift) by leveraging a drift detection mechanism and a multi-registry model storage strategy. We describe next each of the components of the Adaptive Model Selector. A. Model Registry The Model Registry includes a set of pre-trained ML models that can be used for a specific task. Such models typically have different properties that affect model size, inference time, and accuracy. For instance, a k-nearest neighbors and a neural network may both be stored in the model registry for defect identification in an industrial environment. This provides flexibility the Adaptive Model Selector is choosing a model to deploy. In addition, to deal with concept drift, we store models that are pre-trained to handle data drift situations. We generate kdatasets with drifted versions of features to create training variants X(k). For each training variant X(k) generated by applying controlled distributional changes to the original dataset, we train a set of ML models and store them in a registry R(k). Each registry acts as a specialized repository optimized for handling inputs statistically similar to X(k), as determined by drift detection. Therefore the model registry contains a diverse pool of models: Fig. 1: High-level overview of the adaptive model selector. R(k)={M(k) 1,M(k) 2,...,M(k) n} The stored models span a variety of learning paradigms including linear models, ensemble learners, and neural networks. Each model M(k) iis trained to minimize empirical loss (measure of how well the predictions match the training dataset): min θ∈Θ 1 N N X j=1 (yj−fθ(xj))2(1) After training, we record performance metrics for each model. These could include metrics such as: 1. Accuracy: Acc(k) i 2. Inference latency: T(k) i 3. Energy Efficiency: E(k) i 4. Re-training time: RT (k) i 5. Precision: P(k) i All models and their metadata are stored in an MLflow [18] model registry, allowing for reproducible versioning and structured retrieval based on input similarity and performance requirements. Note that in our experiments, we only make use of Acc(k) iand T(k) ias selection metrics. B. Drift Detection A core component of our Adaptive Model Selection technique is the ability to detect and quantify distributional shifts between the training data and incoming inference data. To achieve this, we employ the Maximum Mean Discrepancy (MMD) [13] metric, a non-parametric kernel-based distance metric. Our method adopts an evaluation strategy based on similarity. To measure the MMD, we perform the following steps: Data Preprocessing: For consistent comparison across datasets, we first preprocess all datasets by retaining only continuous numerical features and applying z-score normalization: Xstandardized i=Xi−µi σi (2) where µiand σiare the mean and standard deviation of feature iin the training dataset. This normalization ensures all features contribute equally to drift estimation and stabilizes kernel computations. Drift Evaluation via MMD: Given a set of dataset variants {X(1), X(2), . . . , X(K)}used to train different model registries, and an incoming inference dataset Xinference, we compute the squared MMD between Xinference and each X(k): MMD2(Xinference, X(k)) = ∥Ex∼Xuser [ϕ(x)] −Ex′∼X(k)[ϕ(x′)]∥2 H (3) where Eis the expectation and ϕ(·)refers to a particular feature in the dataset. To approximate this quantity efficiently, we use the unbiased linear-time estimator: MMD2≈1 m2X i,j xT ixj+1 n2X i,j yT iyj−2 mn X i,j xT iyj(4) where xi∈Xuser,yj∈X(k), and m, n are the respective sample sizes. Registry Selection: We define the optimal registry index k∗as: k∗= arg min kMMD2(Xuser, X(k))(5) The selected registry R(k∗)contains models trained on data statistically most similar to the user input. This approach avoids hand-tuned thresholds and allows drift detection to remain adaptive as more training variants are added. C. Meta-Model Guided Model Selection Once a model registry R(k∗)is selected via MMD-based drift detection, we apply a meta-model Ψ(k∗)trained specifically for that registry. The role of the meta-model is to identify, without training and running the available models online, the model in R(k∗)that best satisfies user-defined constraints. Inputs to the Meta-Model: To select the best performing ML model, the meta-model receives the following input: (i) Statistical summary of the input data features F(Xuser)(e.g., mean, standard deviation, maximum and minimum values). (ii) User-defined performance constraints: minimum required accuracy Accmin and maximum allowable latency Tmax. Output of the Meta-Model: The meta-model outputs the index i∗of the most suitable model M∗=M(k∗) i∗, that is chosen from the registry k∗such that: i∗= arg max iAcc(k∗) i s.t. T(k∗) i≤Tmax, Acc(k∗) i≥Amin (6) This ensures that model selection respects both prediction quality and real-time performance demands. Training the Meta-Model: Meta-models are trained using historical examples of input dataset statistics paired with the identity of the best-performing model under the user constraints. The task is treated as a multi-class classification problem, where each class corresponds to a model index in the registry. Once trained, the meta-model enables real-time selection of the most appropriate model that satisfies user objectives, without the need for costly inference-time evaluations. D. Model Selection The final model selection process begins with feature extraction from the incoming dataset. These meta-features are passed through the drift detection module to determine the drift type and magnitude. Based on this output, the most appropriate model registry is selected. Then, the meta-model corresponding to the selected registry identifies the best model within that registry for deployment. By integrating drift-awareness, registry selection, and predictive meta-learning, the Intelligent Model Selector enables efficient, dynamic, and real-time model deployment, making it suitable for both cloud and edge environments. IV. EXPERIMENTS AND RESULTS This section presents the experimental evaluation of our technique. We start first by presenting the experimental setup, then evaluate our technique when dealing with multiple drift variants. We next highlight the need for adaptive model selection when dealing with heterogeneous hardware resources. We finally test the flexibility of our approach in high-constraint scenarios. A. Dataset analysis In our experimental evaluation, we use the iV2I+ dataset [19], which is part of the AI4Mobile industrial datasets collection and provides a comprehensive multimodal record of communication and sensory measurements involving an Autonomous Guided Vehicle (AGV) interacting with infrastructure components in an industrial environment. The key features of the dataset include: 1) Communication Metrics: Includes measurements such as RSRP (Reference Signal Received Power), RSRQ (Reference Signal Received Quality), SINR (Signal to Interference plus Noise Ratio), and RSSI (Received Signal Strength Indicator), collected from LTE channels using MobileInsight and other logging tools. 2) Network Performance: Contains throughput and delay data derived from tools like TCPDump,Iperf, and Ping, capturing both uplink and downlink characteristics. 3) Sensor Data: AGV sensor inputs captured via the Robot Operating System (ROS), including: •Odometry: Position, orientation, and velocity (update rate: 10 ms, downsampled to 40 ms). •Inertial Measurements: Acceleration and gyroscopic data (10 ms). •LIDAR: 3D obstacle detection (100 ms). •Line-of-Sight (LoS) estimation: A boolean indicator computed using Fresnel ellipse modeling. The iV2I+ dataset offers a rich basis for the exploration of industrial communication reliability, sensor fusion, and the application of AI in connected mobility systems. For the communication aware path planning of robots, these features are used to estimate the ML model accuracy and inference times. We pre-process the dataset by resampling all data sources to a 1-second interval for alignment and we merge them to facilitate time-series analysis. The dataset is used to classify locations on the indoor map where thresholds on throughput, delay or signal strength may be violate a threshold value. All following experiments were conducted using Python 3.12 on two platforms: (1) a Raspberry Pi 4 Model B Rev 1.5 with 4GB RAM and a 1.5GHz quad-core ARM Cortex-A72 processor, and (2) a standard laptop with an Intel Core i712650H processor and 16GB RAM and an NVIDIA RTX 3050 Ti. The MLflow platform was used to manage model registries with different levels of drift. B. Model Performance Across Drift Variants We first test our technique across four distinct drift magnitudes: -7,-2,2, and 7. These drift magnitudes represent the median shift in the critical features values that can impact the model accuracy. For each drift magnitude, we evaluated multiple models including XGBoost, Gradient Boosting, Random Forest, Ridge, Lasso, Linear Regression, and a deep Scikitlearn MLP Regressor. Each model was evaluated in terms of accuracy and inference time on both the Raspberry Pi and the laptop. Figure 2 compares the inference time of all models across the two devices, with accuracy annotated above each bar. As shown in Figure 2, inference time increases significantly on Raspberry Pi for deep or ensemble models. However, the accuracy remains stable, demonstrating the robustness of the selected models under drift. Greedy Approach: As a comparative baseline, we design a greedy model selector that ranks all available models based on past performance and heuristically selects the one closest to the desired target (accuracy or latency). Unlike the adaptive selector, the greedy selector does not use drift information or meta-features from the incoming data. It simply selects: the most accurate model below a latency threshold, or he fastest model above an accuracy threshold. Although simple to implement, the greedy approach lacks adaptability to drift, and its performance degrades rapidly as the data distribution shifts away from the training distribution. C. Cross-Device Model Comparison To highlight the need for adaptive model selection, we now compare inference time for several ML models on different hardware resources As shown in Table I, the average inference time on the Raspberry Pi is approximately 6.87 times slower than on the laptop. This discrepancy is most pronounced for lightweight models such as Linear Regression and Ridge Regression, where the absolute differences are small, but the Pi-to-laptop ratios exceed 12x and 15x respectively. Conversely, heavier models such as XGBoost and MLP regressors also experience notable slowdowns, but their inference times remain within practical bounds for deployment on resourceconstrained devices. Fig. 2: Inference time and accuracy for different drift magnitudes spanning −7to 7. TABLE I: Average Inference Time and Pi/Laptop Ratio Across Devices. Model Avg. Laptop Time (s) Avg. Pi Time (s) Pi/Laptop Ratio XGBoost Regressor 0.00022185 0.00153750 6.93x Gradient Boosting Regressor 0.00646750 0.01431250 2.21x Random Forest Regressor 0.01923750 0.04871000 2.53x Ridge Regression 0.00002433 0.00030877 12.69x Lasso Regression 0.00418750 0.00551000 1.32x Linear Regression 0.00001623 0.00025875 15.95x Deep Scikit MLP Regressor 0.01675000 0.10781000 6.44x Overall Average 0.00670070 0.02549250 6.87x These results reinforce the motivation behind using an intelligent model selection strategy, especially in real-time or embedded scenarios. By incorporating both inference time and accuracy constraints, the meta-model ensures that only models suitable for the target device are selected. D. Scenario-Based Evaluation To validate the techniques’ effectiveness under varied deployment conditions, we test it in two real-world scenarios. Each scenario challenges the pipeline with different combinations of drift severity and constraint tightness. 1) Intermediate Drift and High Accuracy Requirement: In this scenario, the runtime dataset was generated with a drift magnitude of 4, which does not directly match any of the available training variants. The model registries were built on four dataset variants with drift magnitudes of -7,-2,2, and TABLE II: Pairwise MMD matrix for scenario 1. Drift -7 -2 2 7 -7 0 1233.93 3983.88 9641.95 -2 1232.47 0 788.56 3983.40 23980.19 195.73 0 1226.67 79632.48 3989.34 1226.68 0 7. This setup allows us to evaluate whether the pipeline could generalize to an unseen drift level and still make appropriate registry and model selections. The drift detection module, based on Maximum Mean Discrepancy (MMD), compares the inference data distribution against all registered training variants, as shown in Table II. It correctly identifies the registry trained on drift magnitude 2 (against 7) as the closest match, demonstrating the pipeline’s ability to interpolate between drift levels: 1Drift -7 MMD: 225.97 2Drift -2 MMD: 67.26 3Drift 2 MMD: 7.43 4Drift 7 MMD: 16.74 5Closest variant to inpute data: Drift 2. After selecting the appropriate registry, the accuracy-aware meta-model within that registry was invoked with the following user-defined constraints: Minimum required accuracy 99.9% and Maximum allowable inference time 0.01 seconds The meta-model successfully recommended the XGBoost Regressor, which satisfied both constraints with an accuracy of 99.9001% and an inference time of 0.00024 seconds. TABLE III: Pairwise MMD matrix for scenario 2. Drift -2 2 4 6 8 -2 0 790.433 1770.657 3148.944 4912.694 2785.421 0 196.718 788.582 1772.880 41774.203 195.73 0 195.880 789.296 63152.126 784.097 192.268 0 195.721 84926.472 1764.529 790.707 198.632 0 2) High Drift and Ultra-Low Latency Requirement: To test the robustness of the proposed system under extreme conditions, we simulated a scenario where the user input dataset exhibits a high drift magnitude of 9, which is beyond any of the trained variant registries (unseen sample drifts). The available registries were trained on data with drift magnitudes of -2,2,4,6, and 8. This setup evaluates the pipeline’s ability to handle out-of-range generalization. The user specified a highly stringent latency constraint of 0.00001 seconds and a minimum accuracy threshold of 90%. Despite the challenging drift and tight timing requirement, the pipeline correctly selected the registry trained on drift magnitude 8as the closest match, as Table III shows. The meta-model associated with that registry recommended the Linear Regression model. The selected model met both constraints, with the following performance: Accuracy 91.82% and Inference Time 5.55 µs. This example highlights the pipeline’s capacity to extrapolate effectively when the input drift exceeds the training envelope, and to enforce ultra-low-latency deployment constraints by prioritizing lightweight, efficient models. These examples validates that the pipeline can: 1) Detect the closest available training distribution via MMD, 2) Select the appropriate registry automatically, 3) Use a meta-model to enforce deployment-time constraints V. CONCLUSION AND FUTURE WORK This paper presented an adaptive model selection framework that integrates meta-learning and drift detection to enable efficient, real-time, and constraint-aware deployment of ML models in dynamic environments. By leveraging Maximum Mean Discrepancy (MMD) for drift quantification and metamodels for constraint-based model recommendation, the system achieves both adaptability and computational efficiency across diverse drift scenarios and hardware configurations. Experimental results demonstrate that the proposed technique maintains high predictive performance while significantly reducing inference overhead. Future work will focus on extending the technique to more dynamic changes such as resource and network constraints, improving meta-model generalization through continual learning, and exploring reinforcement-based approaches for proactive adaptation under unseen drift conditions. REFERENCES [1] M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter, “Efficient and robust automated machine learning,” Advances in neural information processing systems, vol. 28, 2015. [2] T. T. Mac, C. Copot, D. T. Tran, and R. De Keyser, “Heuristic approaches in robot path planning: A survey,” Robotics and Autonomous Systems, vol. 86, pp. 13–28, 2016. [3] L. Cao, “Beyond iid: Non-iid thinking, informatics, and learning,” IEEE Intelligent Systems, vol. 37, no. 4, pp. 5–17, 2022. [4] Y. Zhang, H. Jiang, H. Ren, C. Zou, and D. Dou, “Automs: automatic model selection for novelty detection with error rate control,” Advances in Neural Information Processing Systems, vol. 35, pp. 19 917–19 929, 2022. [5] R. S. Olson and J. H. Moore, “Tpot: A tree-based pipeline optimization tool for automating machine learning,” in Workshop on automatic machine learning. PMLR, 2016, pp. 66–74. [6] D. Golovin, B. Solnik, S. Moitra, G. Kochanski, J. Karro, and D. Sculley, “Google vizier: A service for black-box optimization,” in Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 2017, pp. 1487–1495. [7] J. R. Gardner, G. Malkomes, R. Garnett, K. Q. Weinberger, D. L. Barbour, and J. P. Cunningham, “Bayesian active model selection with an application to automated audiometry,” in Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, 2015, pp. 2386–2394. [8] J. Gama, I. ˇ Zliobait˙ e, A. Bifet, M. Pechenizkiy, and A. Bouchachia, “A survey on concept drift adaptation,” ACM computing surveys (CSUR), vol. 46, no. 4, pp. 1–37, 2014. [9] A. Bifet and R. Gavalda, “Kalman filters and adaptive windows for learning in data streams,” in International conference on discovery science. Springer, 2006, pp. 29–40. [10] J. Gama, P. Medas, G. Castillo, and P. Rodrigues, “Learning with drift detection,” in Brazilian symposium on artificial intelligence. Springer, 2004, pp. 286–295. [11] K. Panchal, S. Choudhary, S. Mitra, K. Mukherjee, S. Sarkhel, S. Mitra, and H. Guan, “Flash: Concept drift adaptation in federated learning,” in International Conference on Machine Learning. PMLR, 2023, pp. 26 931–26 962. [12] E. Jothimurugesan, K. Hsieh, J. Wang, G. Joshi, and P. B. Gibbons, “Federated learning under distributed concept drift,” in International Conference on Artificial Intelligence and Statistics. PMLR, 2023, pp. 5834–5853. [13] A. J. Smola, A. Gretton, and K. Borgwardt, “Maximum mean discrepancy,” in 13th international conference, ICONIP, vol. 6, 2006. [14] P. Brazdil, C. Soares, and J. Costa, “Ranking learning algorithms: Using ibl and meta-learning on accuracy and time results,” Machine Learning, vol. 50, pp. 251–277, 03 2003. [15] E. Triantafillou, T. Zhu, V. Dumoulin, P. Lamblin, U. Evci, K. Xu, R. Goroshin, C. Gelada, K. Swersky, P.-A. Manzagol et al., “Metadataset: A dataset of datasets for learning to learn from few examples,” arXiv preprint arXiv:1903.03096, 2019. [16] M. Abdallah, R. A. Rossi, K. Mahadik, S. Kim, H. Zhao, and S. Bagchi, “Evaluation-free time-series forecasting model selection via meta-learning,” ACM Transactions on Knowledge Discovery from Data, vol. 19, no. 3, pp. 1–41, 2025. [17] J. M. Navarro, A. Huet, and D. Rossi, “Meta-learning for fast model recommendation in unsupervised multivariate time series anomaly detection,” in International Conference on Automated Machine Learning. PMLR, 2023, pp. 24–1. [18] M. Zaharia, A. Chen, A. Davidson, A. Ghodsi, S. A. Hong, A. Konwinski, S. Murching, T. Nykodym, P. Ogilvie, M. Parkhe et al., “Accelerating the machine learning lifecycle with mlflow.” IEEE Data Eng. Bull., vol. 41, no. 4, pp. 39–45, 2018. [19] R. Hernang´ omez, A. Palaios, C. Watermann, D. Sch¨ aufele, P. Geuer, R. Ismayilov, M. Parvini, A. Krause, M. Kasparick, T. Neugebauer, O. D. Ramos-Cantor, H. Tchouankem, J. L. Calvo, B. Chen, G. Fettweis, and S. Sta´ nczak, “Toward an AI-Enabled Connected Industry: AGV Communication and Sensor Measurement Datasets,” IEEE Communications Magazine, vol. 62, no. 4, pp. 90–95, Apr. 2024.