Intra-domain and cross-domain transfer learning for time series

2021 ◽  
Author(s):  
Erik Otović ◽  
Marko Njirjak ◽  
Dario Jozinović ◽  
Goran Mauša ◽  
Alberto Michelini ◽  
...  
Keyword(s):  
Machine Learning ◽  
Time Series ◽  
Transfer Learning ◽  
Time Series Data ◽  
The Other ◽  
Series Data ◽  
Sound Recognition ◽  
Transfer Of Knowledge ◽  
Learning Models ◽  
Machine Learning Models

<p>In this study, we compared the performance of machine learning models trained using transfer learning and those that were trained from scratch - on time series data. Four machine learning models were used for the experiment. Two models were taken from the field of seismology, and the other two are general-purpose models for working with time series data. The accuracy of selected models was systematically observed and analyzed when switching within the same domain of application (seismology), as well as between mutually different domains of application (seismology, speech, medicine, finance). In seismology, we used two databases of local earthquakes (one in counts, and the other with the instrument response removed) and a database of global earthquakes for predicting earthquake magnitude; other datasets targeted classifying spoken words (speech), predicting stock prices (finance) and classifying muscle movement from EMG signals (medicine).<br>In practice, it is very demanding and sometimes impossible to collect datasets of tagged data large enough to successfully train a machine learning model. Therefore, in our experiment, we use reduced data sets of 1,500 and 9,000 data instances to mimic such conditions. Using the same scaled-down datasets, we trained two sets of machine learning models: those that used transfer learning for training and those that were trained from scratch. We compared the performances between pairs of models in order to draw conclusions about the utility of transfer learning. In order to confirm the validity of the obtained results, we repeated the experiments several times and applied statistical tests to confirm the significance of the results. The study shows when, within the set experimental framework, the transfer of knowledge brought improvements in terms of model accuracy and in terms of model convergence rate.<br><br>Our results show that it is possible to achieve better performance and faster convergence by transferring knowledge from the domain of global earthquakes to the domain of local earthquakes; sometimes also vice versa. However, improvements in seismology can sometimes also be achieved by transferring knowledge from medical and audio domains. The results show that the transfer of knowledge between other domains brought even more significant improvements, compared to those within the field of seismology. For example, it has been shown that models in the field of sound recognition have achieved much better performance compared to classical models and that the domain of sound recognition is very compatible with knowledge from other domains. We came to similar conclusions for the domains of medicine and finance. Ultimately, the paper offers suggestions when transfer learning is useful, and the explanations offered can provide a good starting point for knowledge transfer using time series data.</p>

scholarly journals Neural Activity Classification with Machine Learning Models Trained on Interspike Interval Time-Series Data

2021 ◽  
Author(s):  
Ivan Lazarevich ◽  
Ilya Prokin ◽  
Boris Gutkin ◽  
Victor Kazantsev
Keyword(s):  
Machine Learning ◽  
Time Series ◽  
Language Processing ◽  
Neural Activity ◽  
Time Series Data ◽  
Series Data ◽  
Neural Decoding ◽  
Learning Models ◽  
Wide Range ◽  
Machine Learning Models

Modern well-performing approaches to neural decoding are based on machine learning models such as decision tree ensembles and deep neural networks. The wide range of algorithms that can be utilized to learn from neural spike trains, which are essentially time-series data, results in the need for diverse and challenging benchmarks for neural decoding, similar to the ones in the fields of computer vision and natural language processing. In this work, we propose a spike train classification benchmark, based on open-access neural activity datasets and consisting of several learning tasks such as stimulus type classification, animal’s behavioral state prediction and neuron type identification. We demonstrate that an approach based on hand-crafted time-series feature engineering establishes a strong baseline performing on par with state-of-the-art deep learning based models for neural decoding. We release the code allowing to reproduce the reported results 1.

scholarly journals Selection of Input Factors and Comparison of Machine Learning Models for Prediction of Dissolved Oxygen in Gyeongan Stream

2021 ◽  
Vol 43 (3) ◽  
pp. 206-217
Author(s):  
Min Ji Kim ◽  
Seon Jeong Byeon ◽  
Kyung Min Kim ◽  
Johng-Hwa Ahn
Keyword(s):  
Neural Network ◽  
Machine Learning ◽  
Electrical Conductivity ◽  
Time Series ◽  
Performance Evaluation ◽  
Suspended Solids ◽  
Time Series Data ◽  
Series Data ◽  
Learning Models ◽  
Machine Learning Models

Objectives : In this study, we select input factors for machine learning models to predict dissolved oxygen (DO) in Gyeongan Stream and compare results of performance evaluation indicators to find the optimal model.Methods : The water quality data from the specific points of Gyeongan Stream were collected between January 15, 1998 and December 30, 2019. The pretreatment data were divided into train and test data with the ratio of 7:3. We used random forest (RF), artificial neural network (ANN), convolutional neural network (CNN), and gated recurrent unit (GRU) among machine learning. RF and ANN were tested by both random split and time series data, while CNN and GRU conducted the experiment using only time series data. Performance evaluation indicators such as square of the correlation coefficient (R2), root mean square error (RMSE), and mean absolute error (MAE) were used to compare the optimal results for the models.Results and Discussion : Based on the RF variable importance results and references, water temperature, pH, electrical conductivity, PO<sub>4</sub>-P, NH<sub>4</sub>-N, total phosphorus, suspended solids, and NO<sub>3</sub>-N were used as input factors. Both RF and ANN performed better with time series data than random split. The model performance was good in order of RF > CNN > GRU > ANN.Conclusions : The eight input factors (water temperature, pH, electrical conductivity, PO<sub>4</sub>-P, NH<sub>4</sub>-N, total phosphorus, suspended solids, and NO<sub>3</sub>-N) were selected for machine learning models to predict DO in Gyeongan Stream. The best model for DO prediction was the RF model with time series data. Therefore, we suggest that the RF with the eight input factors could be used to predict the DO in streams.

scholarly journals Machine Learning Models with Time-Series Clinical Features to Predict Radiographic Progression in Patients with Ankylosing Spondylitis

2021 ◽  
Author(s):  
Bon San Koo ◽  
Miso Jang ◽  
Ji Seon Oh ◽  
Keewon Shin ◽  
Seunghun Lee ◽  
...  
Keyword(s):  
Machine Learning ◽  
Time Series ◽  
Ankylosing Spondylitis ◽  
Random Forest ◽  
Clinical Features ◽  
Radiographic Progression ◽  
Time Series Data ◽  
Series Data ◽  
Learning Models ◽  
Machine Learning Models

Abstract Background: Radiographic progression in patients with ankylosing spondylitis (AS) varies between individuals, and its evaluation requires a long period of time. Previous statistical studies for radiographic progression have limitations in integrating and analyzing multiple variables of various types. The purpose of this study was to establish the application of machine learning models for predicting radiographic progression in patients with AS using time-series data from electronic medical records (EMRs).Methods: EMR data, including baseline characteristics, laboratory finding, drug administration, and modified Stoke Ankylosing Spondylitis Spine Score (mSASSS), were collected from 1,123 AS patients who were followed up for 18 years at a common center at the time of first (T1), second (T2), and third (T3) visits. The radiographic progression of the (n + 1)th visit (Pn+1 = (mSASSSn+1 – mSASSSn) / (Tn+1 – Tn) ≥ 1 unit per year) was predicted using follow-up visit datasets from T1 to Tn. Three machine learning methods (logistic regression with least absolute shrinkage and selection operation, random forest, and extreme gradient boosting algorithms) with three-fold cross validation were used. Results: The random forest model using the T1 EMR dataset showed the highest performance in predicting the radioactive progression P2 among all the machine learning models tested. The mean accuracy and the area under the curves were 73.73% and 0.79, respectively. Among the variables of T1, the most important variables for predicting radiographic progression were in the order of total mSASSS, age, and alkaline phosphatase. Conclusion: Prognosis predictive models using time-series data showed reasonable performance with clinical features of the first visit dataset for predicting radiographic progression. Additional feature data such as spine radiographs or life-log data may improve the performance of these models.

scholarly journals A Spatiotemporal Prediction Framework for Air Pollution Based on Deep RNN

2017 ◽  
Vol IV-4/W2 ◽  
pp. 15-22 ◽  
Author(s):  
J. Fan ◽  
Q. Li ◽  
J. Hou ◽  
X. Feng ◽  
H. Karimian ◽  
...  
Keyword(s):  
Neural Network ◽  
Machine Learning ◽  
Time Series ◽  
Missing Values ◽  
Series Data ◽  
Gradient Boosting ◽  
Learning Models ◽  
Missing Value ◽  
Deep Recurrent Neural Network ◽  
Machine Learning Models

Time series data in practical applications always contain missing values due to sensor malfunction, network failure, outliers etc. In order to handle missing values in time series, as well as the lack of considering temporal properties in machine learning models, we propose a spatiotemporal prediction framework based on missing value processing algorithms and deep recurrent neural network (DRNN). By using missing tag and missing interval to represent time series patterns, we implement three different missing value fixing algorithms, which are further incorporated into deep neural network that consists of LSTM (Long Short-term Memory) layers and fully connected layers. Real-world air quality and meteorological datasets (Jingjinji area, China) are used for model training and testing. Deep feed forward neural networks (DFNN) and gradient boosting decision trees (GBDT) are trained as baseline models against the proposed DRNN. Performances of three missing value fixing algorithms, as well as different machine learning models are evaluated and analysed. Experiments show that the proposed DRNN framework outperforms both DFNN and GBDT, therefore validating the capacity of the proposed framework. Our results also provides useful insights for better understanding of different strategies that handle missing values.

scholarly journals A versatile computational algorithm for time-series data analysis and machine-learning models

2021 ◽  
Vol 7 (1) ◽  
Author(s):  
Taylor Chomiak ◽  
Neilen P. Rasiah ◽  
Leonardo A. Molina ◽  
Bin Hu ◽  
Jaideep S. Bains ◽  
...  
Keyword(s):  
Machine Learning ◽  
Time Series ◽  
Computational Algorithm ◽  
Time Series Data ◽  
Simulated Data ◽  
Series Data ◽  
Learning Models ◽  
Machine Learning Model ◽  
Time Series Data Analysis

AbstractHere we introduce Local Topological Recurrence Analysis (LoTRA), a simple computational approach for analyzing time-series data. Its versatility is elucidated using simulated data, Parkinsonian gait, and in vivo brain dynamics. We also show that this algorithm can be used to build a remarkably simple machine-learning model capable of outperforming deep-learning models in detecting Parkinson’s disease from a single digital handwriting test.

scholarly journals Multivariate Time Series Data Prediction Based on ATT-LSTM Network

2021 ◽  
Vol 11 (20) ◽  
pp. 9373
Author(s):  
Jie Ju ◽  
Fang-Ai Liu
Keyword(s):  
Time Series ◽  
Deep Learning ◽  
Time Series Data ◽  
Mutual Influence ◽  
Multivariate Time Series ◽  
The Other ◽  
Series Data ◽  
Data Sets ◽  
Learning Models ◽  
Prediction Problems

Deep learning models have been widely used in prediction problems in various scenarios and have shown excellent prediction effects. As a deep learning model, the long short-term memory neural network (LSTM) is potent in predicting time series data. However, with the advancement of technology, data collection has become more accessible, and multivariate time series data have emerged. Multivariate time series data are often characterized by a large amount of data, tight timeline, and many related sequences. Especially in real data sets, the change rules of many sequences will be affected by the changes of other sequences. The interacting factors data, mutation information, and other issues seriously impact the prediction accuracy of deep learning models when predicting this type of data. On the other hand, we can also extract the mutual influence information between different sequences and simultaneously use the extracted information as part of the model input to make the prediction results more accurate. Therefore, we propose an ATT-LSTM model. The network applies the attention mechanism (attention) to the LSTM to filter the mutual influence information in the data when predicting the multivariate time series data, which makes up for the poor ability of the network to process data. Weaknesses have greatly improved the accuracy of the network in predicting multivariate time series data. To evaluate the model’s accuracy, we compare the ATT-LSTM model with the other six models on two real multivariate time series data sets based on two evaluation indicators: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The experimental results show that the model has an excellent performance improvement compared with the other six models, proving the model’s effectiveness in predicting multivariate time series data.

scholarly journals Improving Current Glycated Hemoglobin Prediction in Adults: Consistency and Robustness of Machine Learning Algorithms with Electronic Health Records (Preprint)

2020 ◽  
Author(s):  
Zakhriya Alhassan ◽  
MATTHEW WATSON ◽  
David Budgen ◽  
Riyad Alshammari ◽  
Ali Alessan ◽  
...  
Keyword(s):  
Machine Learning ◽  
Deep Learning ◽  
Time Series Data ◽  
Learning Model ◽  
Series Data ◽  
Learning Models ◽  
Health Records ◽  
Conventional Machine ◽  
Deep Learning Model ◽  
Machine Learning Models

BACKGROUND Predicting the risk of glycated hemoglobin (HbA1c) elevation can help identify patients with the potential for developing serious chronic health problems such as diabetes and cardiovascular diseases. Early preventive interventions based upon advanced predictive models using electronic health records (EHR) data for such patients can ultimately help provide better health outcomes. OBJECTIVE Our study investigates the performance of predictive models to forecast HbA1c elevation levels by employing machine learning approaches using data from current and previous visits in the EHR systems for patients who had not been previously diagnosed with any type of diabetes. METHODS This study employed one statistical model and three commonly used conventional machine learning models, as well as a deep learning model, to predict patients’ current levels of HbA1c. For the deep learning model, we also integrated current visit data with historical (longitudinal) data from previous visits. Explainable machine learning methods were used to interrogate the models and have an understanding of the reasons behind the models' decisions. All models were trained and tested using a large and naturally balanced dataset from Saudi Arabia with 18,844 unique patient records. RESULTS The machine learning models achieved the best results for predicting current HbA1c elevation risk. The deep learning model outperformed the statistical and conventional machine learning models with respect to all reported measures when employing time-series data. The best performing model was the multi-layer perceptron (MLP) which achieved an accuracy of 74.52% when used with historical data. CONCLUSIONS This study shows that machine learning models can provide promising results for the task of predicting current HbA1c levels. For deep learning in particular, utilizing the patient's longitudinal time-series data improved the performance and affected the relative importance for the predictors used. The models showed robust results that were consistent with comparable studies.

scholarly journals Generation of Time-Series Working Patterns for Manufacturing High-Quality Products through Auxiliary Classifier Generative Adversarial Network

Sensors ◽  
2021 ◽  
Vol 22 (1) ◽  
pp. 29
Author(s):  
Manas Bazarbaev ◽  
Tserenpurev Chuluunsaikhan ◽  
Hyoseok Oh ◽  
Ga-Ae Ryu ◽  
Aziz Nasridinov ◽  
...  
Keyword(s):  
Time Series ◽  
Deep Learning ◽  
Product Quality ◽  
Time Series Data ◽  
The Other ◽  
Series Data ◽  
Learning Models ◽  
Generative Adversarial Network ◽  
Adversarial Network ◽  
Proposed Model

Product quality is a major concern in manufacturing. In the metal processing industry, low-quality products must be remanufactured, which requires additional labor, money, and time. Therefore, user-controllable variables for machines and raw material compositions are key factors for ensuring product quality. In this study, we propose a method for generating the time-series working patterns of the control variables for metal-melting induction furnaces and continuous casting machines, thus improving product quality by aiding machine operators. We used an auxiliary classifier generative adversarial network (AC-GAN) model to generate time-series working patterns of two processes depending on product type and additional material data. To check accuracy, the difference between the generated time-series data of the model and the ground truth data was calculated. Specifically, the proposed model results were compared with those of other deep learning models: multilayer perceptron (MLP), convolutional neural network (CNN), long short-term memory (LSTM), and gated recurrent unit (GRU). It was demonstrated that the proposed model outperformed the other deep learning models. Moreover, the proposed method generated different time-series data for different inputs, whereas the other deep learning models generated the same time-series data.

Sign in / Sign up

Export Citation Format

Share Document