Citation:

If the data is used in a publication, please cite this web site:

Mathieu Guillame-Bert, http://www-prima.imag.fr/guillame-bert/?page=database

Download:

Presentation:

This page presents our computer generated symbolic time sequence dataset for temporal data mining and machine learning algorithms' evaluation. The dataset is divided into 100 parts (experiments). Each parts contain different temporal patterns with various complexity. A description of the temporal patterns of each part is available in the dataset documentation. Several parts containing the same time patterns are given in order to perform cross validation.

Additionally, the dataset contains (for every part) a list of reference perditions of events of type "a". These predictions are (with some error level) the best predictions that can be expected to do. Predictions are different from time serie data because they are expressed with time ranges instead of time points. For example, suppose a dataset containing two types of events (A and B). Suppose the only temporal pattern is that every time there is an event A at time t, there is an event B between t+10 and t+15 (uniform distribution).

The figure 1 represents graphically this pattern. The figure 2 is an example of such dataset. The figure 3 represents the predictions of events B according to this pattern.

fig1 : simple temporal pattern

fig2 : example of time serie

fig3 : example of prediction

Uniformly, there is no meaning to predict B events with a time range under 5 time units. Formally, any predictions with time range under 5 time units will have lower confidence and lower support than the kind of predictions represented on figure 3.

A Python script is provided to automatically compare user predictions with references predictions (results of machine learning techniques). The comparison's result is presented in a Html report. The figure 4 presents an example of such report. The report contains confidence, support and temporal precision of the user and reference predictions. This report also contains various analysis of the predictions such as relations between predictions confidence and several measure of the dataset complexity e.g. number of condition of the generative pattern, noise, density of events, etc..

fig4 : example of html report
(clic to enlarge)

Dataset organization:

The data set is organized in the following way:

• The "dataSetInformation.html" file presents the patterns of each part of the dataset. The content of this file is explained in the Dataset information section (see online version here).
• The "dataset" directory contains part subpart of the dataset. Names of the files are "part_[X]_[Y].event". All files with the same X contains the same temporal patterns.
• The "dataset_refPred" directory contains the references predictions of each part of the dataset. Names of the files are "part_[X]_[Y].pred". "part_[X]_[Y].pred" predictions correspond to "part_[X]_[Y].event" events.
• The "dataset_desc" directory contains descriptions of the complexity of each part of the dataset. Names of the files are "part_[X]_[Y].txt". "part_[X]_[Y].txt" descriptions correspond to "part_[X]_[Y].event" events.
• The "evaluatePrediction.py" script compares user predictions to reference predictions and generate a Html report. The script asks the user to select a directory containing its own predictions. This directory should have the same organization as the "dataset_refPred" directory. Of course, if the references prediction directory is selected as the user prediction directory, the report displays prefect results. The figure 5 shows a screen capture of script running. The more important measures given in the report are :
  • Ratio (or Gain) User's predictions' estimated confidence/Reference's predictions' estimated confidence
  • Ratio (or Gain) User's predictions' estimated support/Reference's predictions' estimated support
  • Ratio (or Gain) User's predictions' precision/Reference's predictions' precision
• The "readMe.html" file contains this page.

In order to obtain significant results, you should train and evaluate your learning algorithm on different subset of the data set. For example, you can train on "part_[X]_01.event" files and evaluate on "part_[X]_02.event" files.

fig5 : screen capture of the script
(click to enlarge)

File format :

Events files (*.event) contain line by line events. Lines follow the following schema:

[time] event.[symbol]

The regular expression is :

[+-]?([0-9]*\.[0-9]+|[0-9]+)\s+\S+

Example:

Prediction files (*.pred) contain line by line events. Lines follow the following schema:

[begin time] [end time] event.[symbol] [probability]

The regular expression is :

[-]?([0-9]*\.[0-9]+|[0-9]+)\s+[-]?([0-9]*\.[0-9]+|[0-9]+)\s+event\.\S+\s+?([0-9]*\.[0-9]+|[0-9]+)

Example:

Dataset descriptions files (*.txt) contain lists of complexity measures. Lines follow the following schema:

"[name of the measure]" [value of the measure]

Example:

Dataset information :

For each part of dataset is given several pieces of information:

Example of patterns :

Here is several example of temporal patterns:

If an event P appends at time t, therefore an event A will appends between t+0.8 and t+2.9 with 62.28% chance.

If an event I appends at time t, and an event P appended between t-15.5 and t-8.3, therefore an event A will appends between t+2.8 and t+4.5 with 62.92% chance.

If an event F appends at time t, and an event K appended at time t' with t' ∈ [ t-7.9,t-2], and an event N appended at time t'' with t'' ∈[t'-14.1,t'-5.6], therefore an event A will appends between t+1.8 and t+4.6 with 63.26% chance.

If an event K appends at time t, and there is no events D appended between t-13.6 and t-4.9, therefore an event A will appends between t+4.9 and t+7.9 with 62.27% chance.

If an event I appends at time t, and an event J appended at time t' with t' ∈ [ t-7.3,t], and there is no events E appended at time t'' with t'' ∈[t'-10.9,t'-5.4], therefore an event A will appends between t and t+1.8 with 65.15% chance.

If an event F appends at time t, and an event K appended at time t' with t' ∈ [ t-13.1,t-3.2], and an event N appended at time t'' with t'' ∈[t-7.5,t'-1.7], therefore an event A will appends between t+7.3 and t+9.4 with 64.29% chance.