Временной ряд (библиотека примеров)

Материал из MachineLearning.

Текущая версия

Time series is a sequence of equally spaced data measurements. This page lists a number of examples of time series for testing forecasting algorithms.

File structure

Data is stored in comma-separated .csv format, with decimals searated by periods. The first column contains timestamps. The second column stores the forecasted time series, other columns may store complementary time series. Each dataset is followed with a tsNameReadme.txt file, with specifies:

• data source (or a problem that was solved),
• timestamps format,
• interpretation of the data, column-wise,
• type of data scale for each column,
• periodicity, if present,
• other information.

Examples

Synthetic time series (in ts format - see. TSForecastingInterfaces)

• Constant Constant
• Sine Sine
• Two sines 2Sines
• Triangles Saw
• Trapezoid Trapezium

Highly periodic

• Electricity consumption EnergyConsumption
• Machinery
• Sounds
• Music LedZeppelin

Noisy periodic time series

• Electricity prices
• Prices for consumables and commodities
• Retail sales RetailSalesItems
• Sugar prices SugarPrice
• Bread prices WhiteBreadPrices
• Drink consumption
• Weather: tempreture, humidity, wind GermanWeather
• Passenger (and freight) transportation

Complex periodicity

• ECG ECG
• Pulse wave
• MEG
• Reflected time series

Aperiodic

• Flu propagation FluUSA
• Migration

High noise

• Stock exchange Cisco
• Market indices DowJonesIndustrialAverage
• Option prices

Event-driven

• Earthquakes ArkansasEarthquakes.csv
• Financial Bubbles FinancialBubbles.csv
• Records

Accelerometry

• [1]

OPPORTUNITY Activity Recognition Data Set for Human Activity Recognition from Wearable, Object, and Ambient Sensors is a dataset devised to benchmark human activity recognition algorithms (classification, etc.).

• http://hasc.jp/hc2010/HASC2010corpus/hasc2010corpus-en.html (2010) 540 subjects, 6 activities (stay, walk, jog, skip, stair up, stair down). Includes segmented data (only one activity type, 20 seconds) and sequence data.

• http://www-scf.usc.edu/~mizhang/datasets.html (2012) 14 subjects, 12 activities (walk forward, walk left, walk right, go upstairs, go downstairs, run forward, jump up and down, sit and fidget, stand, sleep, elevator up, and elevator down). The data is captured by the MotionNode inertial sensing device which integrates an 3-axis accelerometer (+-6g) and an 3-axis gyroscope (+-500dps), sampled at 100 Hz.

• http://www.cis.fordham.edu/wisdm/dataset.php (2010-2011) 36 subjects, 6 activities (stay, walk, jog, sit, stair up, stair down).

• http://www.opportunity-project.eu/challengedatasetdownload (2011), 4 subjects. An annotated dataset of complex, interleaved and hierarchical activities, with a particularly large number of atomic activities (around 30’000), collected in a rich sensor environment. The full setup including both ambient and on-body sensors comprises 72 sensors of 10 modalities, integrated in the environment and on the body.

• http://smartlab.ws/component/content/article?id=60 (2013) 30 subjects, 6 activities, fixed set of features from

• http://www.ife.ee.ethz.ch/research/groups/Dataset/skoda_mini_checkpoint/SkodaMiniCP.zip Skoda Mini Checkpoint. The dataset contains acceleration meaurements (calibrated and raw) of 10 manipulative gestures performed in a car maintenance scenario (1 subject, 70 instances per activity).
  • Some of the datasets, available from http://www.ife.ee.ethz.ch/research/groups/Dataset are described here.

• https://cloud5.cs.fau.de/owncloud/public.php?service=files&t=9a07b48c7950d1b61d8fb8b0382ff6c7 12 subjects, 4 swimming styles (butterfly, backstroke, breaststroke and freestyle), two states (swimming/resting), one type of events (turns). See Classification of Kinematic Swimming Data with Emphasis on Resource Consumption

Other

• http://llmpp.nih.gov/lymphoma/: classification of DLBCL (Diffuse large B-cell lymphoma) patients via gene expression (pdf). Raw data for all Lymphochip microarrays are available here. For each microarray, two scan files were generated, one for each fluorescence emission wavelength corresponding to the fluorophor used in the reverse transcription labeling reaction.
• http://www.cse.ust.hk/~qyang/ICDMDMC07/: indoor location and transferlearning. The task is to predict the location of each collection of received signal strength (RSS) values in an indoor environment, received from the WiFi Access Points (APs).
  1. The training data a set of (RSS values, Location Label) pairs, where the location labels are discrete (non-sequential), and a collection of partially labelled user traces, which corresponds to a sequence of RSS values collected as a user continuously walks around a building.
  2. The training data and test data are collected at different time periods. Some test data objects are associated with location labels to use as benchmarks.
  • Similar dataset http://www.cse.ust.hk/~derekhh/ActivityRecognition/dataset/hkust.rar set was used in the papers High-level Goal Recognition in a Wireless LAN and Multiple-Goal Recognition from Low-Level Signals for "inferring high-level user-behavior patterns from low-level sensory data through location-based plan recognition".
• CAIDA collects several different types Internet-related of data at geographically and topologically diverse locations.

Kaggle competitions

• https://www.kaggle.com/c/seizure-prediction: predict seizures in intracranial EEG recordings. Intracranial EEG was recorded from dogs with naturally occurring epilepsy using an ambulatory monitoring system. EEG was sampled from 16 electrodes at 400 Hz, and recorded voltages were referenced to the group average. These are long duration recordings, spanning multiple months up to a year and recording up to a hundred seizures in some dogs. Preictal training and testing data segments are provided covering one hour prior to seizure with a five minute seizure horizon.
• https://www.kaggle.com/c/belkin-energy-disaggregation-competition/data: SmartHouse energy consumption prediction. Electromagnetic Interference (EMI) is measured using a special sensor built at the Ubicomp Lab to identify what appliance is being used and how much energy it is consuming. The data is available from 4 homes (H1-H4) consisting of both training datasets and testing datasets. The training set includes information about which appliance was turned ON or OFF and at what timestamps.
• https://www.kaggle.com/c/predicting-parkinson-s-disease-progression-with-smartphone-data: measure the symptoms of Parkinson’s disease with a smartphone. The data was collected from 9 PD patients, at varying stages of the disease, and 7 healthy controls over a period wthin 4 months. The data includes the following streams: audio, accelerometry (3D, for each of the 3 axes: mean, absolute central moment, standard deviation, maximum deviation, power spectral density across four separate bands), GPS (latitude, longitude, altitude), compass (for each of the 3 axes: mean, absolute central moment, standard deviation, maximum deviation).
• https://www.kaggle.com/c/accelerometer-biometric-competition/data: recognize users of mobile devices from accelerometer data. The dataset contains approximately 60 million unique samples of accelerometer data collected from 387 different devices. These are split into equal sets for training and test. Samples in the training set are labeled with the unique device from which the data was collected. The test set is demarcated into 90k sequences of consecutive samples from one device.
• https://www.kaggle.com/c/connectomics: network structure reconstruction. Test data includes time series of neural activities obtained from fluorescence signals and (x, y) coordinates of the neurons. (Neurons are arranged on a flat surface simulating a neural culture). For training data, the network connectivity is also provided. The task is to reconstruct their connectivity from activity data.
• https://www.kaggle.com/c/grasp-and-lift-eeg-detection: identify hand motions from EEG recordings. Dataset contains EEG time series for 12 subjects in total, 10 series of trials for each subject (8 series in training set and 2 series in test set), and approximately 30 trials within each series. The task is to detect each of six events: HandStart, FirstDigitTouch, BothStartLoadPhase, LiftOff, Replace or BothReleased.

Databases

Medicine:

• http://www.physionet.org/ contains collections of recorded physiologic signals (accelerometry, ECG, EEG, EHG, EMG, blood pressure, hart rate, auditory brainstem response, etc.)
• http://www.ebi.ac.uk/arrayexpress/experiments/browse.html is a database of genomic data. Data can be searched by a number of parameters, such as molecule (DNA, RNA, amplicon, metabolite, protein) or experimntal technology (array, high-throughput sequencing, mass spectrometry)
• https://www.ieeg.org/ includes a large database of scientific data and tools to analyze epilepsy datasets.
• https://sleepdata.org/datasets offers six public datasets of sleep research data collected in children and adults across the U.S.

Cross-disciplinary data repositories, data collections and data search engines (from datacentral.com):

• AWS public data sets
• https://datahub.io/ - data management platform from the Open Knowledge Foundation, based on the CKAN data management system

Sensors

• Smart City Includes Vehicle Traffic, Pollution and Weather data.
• At the bottom of the page there are links to the Live data set, which contains sensor data from the meeting room, including presence of people in the room with temperature, humidity, oxygen and carbon dioxide values.

Spatial-time series

• Copernicus: European's eye on Earth

See also

• Временной ряд

The dataset is used in:

• «исследование свойств алгоритмов прогноза»,
• «выбор прогностических моделей».