weka.classifiers.timeseries.core.Utils Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of timeseriesForecasting Show documentation
Show all versions of timeseriesForecasting Show documentation
Provides a time series forecasting environment for Weka. Includes a wrapper for Weka regression schemes that automates the process of creating lagged variables and date-derived periodic variables and provides the ability to do closed-loop forecasting. New evaluation routines are provided by a special evaluation module and graphing of predictions/forecasts are provided via the JFreeChart library. Includes both command-line and GUI user interfaces. Sample time series data can be found in ${WEKA_HOME}/packages/timeseriesForecasting/sample-data.
/*
* This program is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see missingReport) {
if (missingReport != null) {
if (periodicityHandler.isDateBased()) {
String timeFormat = "yyyy-MM-dd'T'HH:mm:ss";
SimpleDateFormat sdf = new SimpleDateFormat();
sdf.applyPattern(timeFormat);
Date d = new Date((long) incrTime);
String result = sdf.format(d);
// System.err.println("Creating a missing row " + result);
missingReport.add(result);
} else {
missingReport.add("" + incrTime);
}
}
}
/**
* Check to see if there are any instances (time steps) that are missing
* entirely from the data (and are not in the skip list). We make the
* assumption that the data won't contain both missing date values *and*
* entirely missing rows as this is a kind of chicken and egg situation with
* regards to detecting missing rows. E.g. if we have missing date values
* bracketing one or more missing rows then we can't do the date comparisons
* necessary for detecting whether the intermediate rows are missing.
* Similarly the missing date handling routine assumes that there aren't
* missing rows so that it can advance the previous date value by one unit in
* order to compute the value for the current missing date value.
*
* @param toInsert the instances to check
* @param timeStampAtt the name of time stamp att
* @param periodicityHandler the periodicity handler
* @param m_skipEntries the list of data "holes" that are expected (i.e. don't
* actually count as time stamp increments)
* @param missingReport will hold time stamps of any instances we insert
* @return the (potentially) modified instances
*/
public static Instances insertMissing(Instances toInsert,
Attribute timeStampAtt, TSLagMaker.PeriodicityHandler periodicityHandler,
String m_skipEntries, List missingReport) {
if (m_skipEntries != null && m_skipEntries.length() > 0) {
try {
periodicityHandler.setSkipList(m_skipEntries, "yyyy-MM-dd'T'HH:mm:ss");
} catch (Exception ex) {
ex.printStackTrace();
}
}
int timeIndex = timeStampAtt.index();
/*
* check to see if there are any instances (time steps) that are missing
* entirely from the data (and are not in the skip list). We make the
* assumption that the data won't contain both missing date values *and*
* entirely missing rows as this is a kind of chicken and egg situation with
* regards to detecting missing rows. E.g. if we have missing date values
* bracketing one or more missing rows then we can't do the date comparisons
* necessary for detecting whether the intermediate rows are missing.
* Similarly the missing date handling routine assumes that there aren't
* missing rows so that it can advance the previous date value by one unit
* in order to compute the value for the current missing date value.
*/
Instance prevInst = null;
int current = 0;
while (true) {
if (current == toInsert.numInstances()) {
break;
}
if (prevInst != null && !toInsert.instance(current).isMissing(timeIndex)) {
if (periodicityHandler.getPeriodicity() != Periodicity.MONTHLY
&& periodicityHandler.getPeriodicity() != Periodicity.QUARTERLY) {
double delta = periodicityHandler.getPeriodicity().deltaTime();
double diff = toInsert.instance(current).value(timeIndex)
- prevInst.value(timeIndex);
/*
* if the difference is more than 1 delta then insert a new row
*/
if (diff > 1.5 * delta) {
if (!periodicityHandler.isDateBased()
|| !periodicityHandler.dateInSkipList(new Date((long) prevInst
.value(timeIndex)))) {
Instance newInst = makeInstance(prevInst.value(timeIndex),
periodicityHandler, toInsert.numAttributes(), timeIndex);
Date candidate = new Date((long) newInst.value(timeIndex));
if (!periodicityHandler.dateInSkipList(candidate)
&& candidate.getTime() < toInsert.instance(current).value(
timeIndex)) {
newInst.setDataset(toInsert);
toInsert.add(current, newInst);
addToMissingReport(newInst.value(timeIndex),
periodicityHandler, missingReport);
}
}
}
} else {
Date d = new Date((long) toInsert.instance(current).value(timeIndex));
Calendar c = new GregorianCalendar();
c.setTime(d);
int currentMonth = c.get(Calendar.MONTH);
d = new Date((long) prevInst.value(timeIndex));
c.setTime(d);
int prevMonth = c.get(Calendar.MONTH);
double diff = (currentMonth + 1)
- ((prevMonth == Calendar.DECEMBER) ? 0 : prevMonth + 1);
if (periodicityHandler.getPeriodicity() == Periodicity.MONTHLY) {
/*
* if the difference is more than 1 month then insert a new row
*/
if (diff > 1.5) {
if (!periodicityHandler.isDateBased()
|| !periodicityHandler.dateInSkipList(new Date(
(long) prevInst.value(timeIndex)))) {
Instance newInst = makeInstance(prevInst.value(timeIndex),
periodicityHandler, toInsert.numAttributes(), timeIndex);
Date candidate = new Date((long) newInst.value(timeIndex));
if (!periodicityHandler.dateInSkipList(candidate)
&& candidate.getTime() < toInsert.instance(current).value(
timeIndex)) {
newInst.setDataset(toInsert);
toInsert.add(current, newInst);
addToMissingReport(newInst.value(timeIndex),
periodicityHandler, missingReport);
}
}
}
} else if (periodicityHandler.getPeriodicity() == Periodicity.QUARTERLY) {
/*
* if the difference is more than 1 quarter (3 months) then insert a
* new row
*/
if (diff > 4.5) {
if (!periodicityHandler.isDateBased()
|| !periodicityHandler.dateInSkipList(new Date(
(long) prevInst.value(timeIndex)))) {
Instance newInst = makeInstance(prevInst.value(timeIndex),
periodicityHandler, toInsert.numAttributes(), timeIndex);
Date candidate = new Date((long) newInst.value(timeIndex));
if (!periodicityHandler.dateInSkipList(candidate)
&& candidate.getTime() < toInsert.instance(current).value(
timeIndex)) {
toInsert.add(current, newInst);
addToMissingReport(newInst.value(timeIndex),
periodicityHandler, missingReport);
}
}
}
}
}
}
if (!toInsert.instance(current).isMissing(timeIndex)) {
prevInst = toInsert.instance(current);
}
current++;
}
return null;
}
/**
* Replace missing target values by interpolation. Also replaces missing date
* values (if a date time stamp has been specified and if possible).
*
* @param toReplace the instances to replace missing target values and time
* stamp values
* @param targets a list of target attributes to check for missing values
* @param timeStampName the name of the time stamp attribute (or null if there
* is no time stamp)
* @param dateOnly if true, only replace missing date values and not missing
* target values (useful for hold-out test sets)
* @param userHint user-specified hint as to the data periodicity
* @param skipEntries any skip entries (may be null)
* @param missingReport a varargs parameter that, if provided, is expected to
* be up to two lists of Integers and one list of Strings. The first
* list will be populated with the instance numbers (duplicates are
* possible) of instances that have missing targets replaced. The
* second list will be populated with the instance numbers of
* instances that have missing time stamp values replaced. The third
* list will be populated with the time stamps for any new instances
* that are inserted into the data (i.e. if we detect that there are
* "holes" in the data that aren't covered by the skip list entries.
*
* @return the instances with missing targets and (possibly) missing time
* stamp values replaced.
*/
public static Instances replaceMissing(Instances toReplace,
List targets, String timeStampName, boolean dateOnly,
Periodicity userHint, String skipEntries, Object... missingReport) {
Instances result = toReplace;
Attribute timeStampAtt = null;
TSLagMaker.PeriodicityHandler detected = null;
List missingTargetList = null;
List missingTimeStampList = null;
List missingTimeStampRows = null;
if (missingReport.length > 0) {
missingTargetList = (List) missingReport[0];
if (missingReport.length == 2) {
missingTimeStampList = (List) missingReport[1];
}
if (missingReport.length == 3) {
missingTimeStampRows = (List) missingReport[2];
}
}
if (timeStampName != null && timeStampName.length() > 0) {
timeStampAtt = toReplace.attribute(timeStampName);
// must be a non-artificial time stamp
if (timeStampAtt != null) {
detected = TSLagMaker
.determinePeriodicity(result, timeStampName, userHint);
// check insertMissing (if periodicity is not UNKNOWN)
/*
* If we do this first, then we can interpolate the missing target
* values that will be created for the rows that get inserted
*/
if (detected.getPeriodicity() != Periodicity.UNKNOWN) {
insertMissing(toReplace, timeStampAtt, detected, skipEntries,
missingTimeStampRows);
}
}
}
// do a quick check to see if we need to replace any missing values
boolean ok = true;
for (int i = 0; i < toReplace.numInstances(); i++) {
if (toReplace.instance(i).hasMissingValue()) {
// now check against targets and possibly date
if (!dateOnly) {
for (String target : targets) {
int attIndex = toReplace.attribute(target).index();
if (toReplace.instance(i).isMissing(attIndex)) {
ok = false;
break;
}
}
if (!ok) {
break; // outer loop
}
}
// check date if necessary
if (timeStampAtt != null) {
if (toReplace.instance(i).isMissing(timeStampAtt)) {
ok = false;
break;
}
}
}
}
if (ok) {
// nothing to do
return result;
}
// process the target(s) first
if (!dateOnly) {
for (String target : targets) {
if (result.attribute(target) != null) {
int attIndex = result.attribute(target).index();
double lastNonMissing = weka.core.Utils.missingValue();
// We won't handle missing target values at the start or end
// as experiments with using simple linear regression to fill
// the missing values that are created by default by the lagging
// process showed inferior performance compared to just letting
// Weka take care of it via mean/mode replacement
for (int i = 0; i < result.numInstances(); i++) {
Instance current = result.instance(i);
if (current.isMissing(attIndex)) {
if (!weka.core.Utils.isMissingValue(lastNonMissing)) {
// Search forward to the next non missing value (if any)
double futureNonMissing = weka.core.Utils.missingValue();
double x2 = 2; // number of x steps (lastNonMissing is at 0 on x
// axis)
for (int j = i + 1; j < result.numInstances(); j++) {
if (!result.instance(j).isMissing(attIndex)) {
futureNonMissing = result.instance(j).value(attIndex);
break;
}
x2++;
}
if (!weka.core.Utils.isMissingValue(futureNonMissing)) {
// Now do the linear interpolation
double offset = lastNonMissing;
double slope = (futureNonMissing - lastNonMissing) / x2;
// fill in the missing values
for (int j = i; j < i + x2; j++) {
if (result.instance(j).isMissing(attIndex)) {
double interpolated = (((j - i) + 1) * slope) + offset;
result.instance(j).setValue(attIndex, interpolated);
if (missingTargetList != null) {
missingTargetList.add(new Integer(j + 1));
}
}
}
}
} else {
// won't do anything with start/end missing values
}
} else {
lastNonMissing = current.value(attIndex);
}
}
}
}
}
// now check for missing date values (if necessary)
if (timeStampAtt != null) {
int attIndex = timeStampAtt.index(); // result.attribute(timeStampName).index();
double firstNonMissing = result.instance(0).value(attIndex);
double previousNonMissing = firstNonMissing;
int firstNonMissingIndex = -1;
boolean leadingMissingDates = weka.core.Utils
.isMissingValue(firstNonMissing);
for (int i = 0; i < result.numInstances(); i++) {
Instance current = result.instance(i);
if (current.isMissing(attIndex)) {
if (!weka.core.Utils.isMissingValue(previousNonMissing)) {
double newV = advanceSuppliedTimeValue(previousNonMissing, detected);
current.setValue(attIndex, newV);
// previousNonMissing = newV;
if (missingTimeStampList != null) {
missingTimeStampList.add(new Integer(i + 1));
}
}
} else if (firstNonMissingIndex == -1) {
firstNonMissingIndex = i;
firstNonMissing = current.value(attIndex);
}
previousNonMissing = current.value(attIndex);
}
if (leadingMissingDates) {
if (firstNonMissingIndex > 0) {
for (int i = firstNonMissingIndex - 1; i >= 0; i--) {
Instance current = result.instance(i);
double newV = decrementSuppliedTimeValue(firstNonMissing, detected);
current.setValue(attIndex, newV);
if (missingTimeStampList != null) {
missingTimeStampList.add(new Integer(i + 1));
}
firstNonMissing = newV;
}
}
}
}
return result;
}
/**
* Utility method to advance a supplied time value by one unit.
*
* @param valueToAdvance the time value to advance
* @param dateBasedPeriodicity the periodicity to use for data arithmetic
* @return the advanced value or the original value if this lag maker is not
* adjusting for trends.
*
*/
public static double advanceSuppliedTimeValue(double valueToAdvance,
TSLagMaker.PeriodicityHandler dateBasedPeriodicity) {
return advanceSuppliedTimeValue(valueToAdvance, dateBasedPeriodicity, false);
}
/**
* Utility method to decrement a supplied time value by one unit.
*
* @param valueToDecrement the time value to decrement
* @param dateBasedPeriodicity the periodicity to use for data arithmetic
* @return the advanced value or the original value if this lag maker is not
* adjusting for trends.
*
*/
public static double decrementSuppliedTimeValue(double valueToDecrement,
TSLagMaker.PeriodicityHandler dateBasedPeriodicity) {
return advanceSuppliedTimeValue(valueToDecrement, dateBasedPeriodicity,
true);
}
protected static double advanceSuppliedTimeValue(double valueToAdvance,
TSLagMaker.PeriodicityHandler dateBasedPeriodicity, boolean decrement) {
double result = valueToAdvance;
int sign = (decrement) ? -1 : 1;
// if (m_adjustForTrends) {
// result = valueToAdvance + dateBasedPeriodicity.deltaTime();//
// m_deltaTime;
if (dateBasedPeriodicity.getPeriodicity() != Periodicity.UNKNOWN) {
Date d = new Date((long) valueToAdvance);
Calendar c = new GregorianCalendar();
c.setTime(d);
do {
if (dateBasedPeriodicity.getPeriodicity() == Periodicity.YEARLY) {
c.add(Calendar.YEAR, 1 * sign);
} else if (dateBasedPeriodicity.getPeriodicity() == Periodicity.QUARTERLY) {
c.add(Calendar.MONTH, 3 * sign);
} else if (dateBasedPeriodicity.getPeriodicity() == Periodicity.MONTHLY) {
c.add(Calendar.MONTH, 1 * sign);
} else if (dateBasedPeriodicity.getPeriodicity() == Periodicity.WEEKLY) {
c.add(Calendar.WEEK_OF_YEAR, 1 * sign);
} else if (dateBasedPeriodicity.getPeriodicity() == Periodicity.DAILY) {
c.add(Calendar.DAY_OF_YEAR, 1 * sign);
} else if (dateBasedPeriodicity.getPeriodicity() == Periodicity.HOURLY) {
c.add(Calendar.HOUR_OF_DAY, 1 * sign);
}
result = c.getTimeInMillis();
} while (dateBasedPeriodicity.dateInSkipList(c.getTime()));
} else {
// just add the delta
do {
result += (dateBasedPeriodicity.deltaTime() * sign);
} while (dateBasedPeriodicity.dateInSkipList(new Date((long) result)));
}
// }
return result;
}
}