conceptdriftsimulatorexample.xml.old

著名的开源仿真软件yale

<!-- Concept drift experiment, ICML-2000 scenario A:          -->
<!-- Window of fixed size of 3 batches with SVM^light learner -->
<operator name="GlobalExperimentChain" class="Experiment">
  <parameter key="resultfile"
             value="Result.ConceptDrift.ScenarioA.FixedSize.txt"/>
  <parameter key="logfile"
             value="Log.ConceptDrift.ScenarioA.FixedSize.txt"/>
  <parameter key="logverbosity"    value="minimum"/>
  <parameter key="random_seed"     value="2001"/>
  <parameter key="temp_dir"        value="./tmp"/>
  <parameter key="keep_temp_files" value="none"/>
             <!-- 'all' or 'none' -->
  <parameter key="notification_email" 
             value="[e-mail]"/>
             <!-- use your e-mail address here -->

  <!-- =====  Read document vectors  ===== -->
  <operator name="TrecExampleSetSource" class="SparseFormatExampleSource">
    <parameter key="attribute_file" value="./data/trec/document.vectors"/>
               <!-- file with document vectors in sparse format -->
    <parameter key="label_file" value="./data/trec/document.topics"/>
               <!-- file with document labels -->
    <parameter key="dimension"    value="25410"/>
               <!-- optional because of auto-detection -->
    <parameter key="max_examples" value="2608"/>
               <!-- optional because of auto-detection -->
    <parameter key="format"       value="separate_file"/>
               <!-- document vectors and labels are provided in two separate files  -->
  </operator>

  <!-- =====  Concept drift simulation ===== -->
  <operator name="ConceptDriftSimulation" class="ConceptDriftSimulator">
    <parameter key="number_of_runs"    value="10"/>
               <!-- no. of experiment repititions for averaging -->
    <parameter key="number_of_batches" value="20"/>
               <!-- no. of time steps (batches) to be simulated -->
               <!-- in each run                                 -->
    <parameter key="number_of_streams" value="5"/>
               <!-- no. of original data streams (e.g. no. of -->
               <!-- original classes)                         -->
    <parameter key="data_stream_names"
               value="Topic1 Topic3 Topic4 Topic5 Topic6"/>
               <!-- names of original classes -->
    <parameter
      key="data_stream_relevance"
      <!-- Batch:      -->
      <!--       0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17  18  19
      -->
      value=
      <!-- Scenario A: -->
      "Topic1 : 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
       Topic3 : 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0"
      <!-- Scenario B: -->
      <!--
      "Topic1 : 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.8 0.6 0.4 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
       Topic3 : 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0"
      -->
    />
    <parameter key="learner_type" value="static_window"/>
       <!-- 'static_window': use a fixed time window ...        -->
    <parameter key="window_size"  value="3"/>
       <!--                  ... of the fixed size 3 batches    -->
       <!-- 'static':   full memory approach (use all old data) -->
       <!-- 'adaptive': use adaptive or weighted window learner -->

    <!-- Learning chain with time step model finder -->
    <operator name="TimeWinLearner"
              class="OperatorChain">
                <!-- for static windows -->
              <!-- class="BatchWindowLearner"> -->
                <!-- for adaptive window -->
              <!-- class="BatchWeightLearner"> -->
                <!-- for weighted adaptive window -->

      <operator name="Learner" class="SVMLightLearner">
        <parameter key="kernel_type" value="linear"/>
        <parameter key="additional_parameters"
                   value="-c 1000 -x 1"/>
                   <!-- set 'C' to 1000 and use xi-alpha -->
                   <!-- error estimation                 -->
      </operator>

      <!-- mySVM learner as an alternative to the SVM^light learner:
      <operator name="Learner" class="MySVMLearner">
       <parameter key="pattern"             value=""/>
                  <!-- task: classification -->
       <parameter key="type"                value="dot"/>
                  <!-- kernel: linear (dot product) -->
       <parameter key="C"                   value="1000"/>
       <parameter key="epsilon"             value="0.1"/>
       <parameter key="verbosity"           value="0"/>
       <parameter key="sparse"              value="true"/>
                  <!-- use sparse data format -->
       <parameter key="weighted_examples"   value="true"/>
                  <!-- use weighted examples -->
       <parameter key="xi_alpha_estimation" value="true"/>
                  <!-- use xi-alpha error estimation -->
      </operator>
      -->

    </operator>

    <!-- Application and evaluation chain -->
    <operator name="ConceptDriftApplierChain"
              class="OperatorChain">
      <operator name="Applier" class="ModelApplier" />
      <operator name="PerfEvaluator"
                class="PerformanceEvaluator">
        <parameter key="classification_error" value="true"/>
        <parameter key="precision"            value="true"/>
        <parameter key="recall"               value="true"/>
        <parameter key="main_criterion"
                   value="classification_error"/>
      </operator>
      <operator name="RunResultWriter" class="ResultWriter"/>
    </operator>

  </operator>  <!-- end of ConceptDriftSimulation -->

  <operator name="AverageResultWriter" class="ResultWriter"/>

</operator>  <!-- end of GlobalExperimentChain -->