source: branches/2288_HeuristicLab.VariableInteractionNetworks/HeuristicLab.VariableInteractionNetworks.Views/3.3/RunCollectionVariableInteractionNetworkView.cs @ 17915

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2016 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
 *
 * This file is part of HeuristicLab.
 *
 * HeuristicLab 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.
 *
 * HeuristicLab 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 HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
 */
#endregion

using System;
using System.Collections.Generic;
using System.ComponentModel;
using System.Drawing;
using System.Globalization;
using System.Linq;
using System.Text;
using System.Windows.Forms;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Core.Views;
using HeuristicLab.Data;
using HeuristicLab.MainForm;
using HeuristicLab.Optimization;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Visualization;
using Ellipse = HeuristicLab.Visualization.Ellipse;
using Rectangle = HeuristicLab.Visualization.Rectangle;

namespace HeuristicLab.VariableInteractionNetworks.Views {
  [View("Variable Interaction Network")]
  [Content(typeof(RunCollection), IsDefaultView = false)]

  public sealed partial class RunCollectionVariableInteractionNetworkView : ItemView {
    public RunCollectionVariableInteractionNetworkView() {
      InitializeComponent();
      ConfigureNodeShapes();
    }

    public new RunCollection Content {
      get { return (RunCollection)base.Content; }
      set {
        if (value != null && value != Content) {
          base.Content = value;
        }
      }
    }

    private VariableInteractionNetwork variableInteractionNetwork;

    private static void AssertSameProblemData(RunCollection runs) {
      IDataset dataset = null;
      IRegressionProblemData problemData = null;
      foreach (var run in runs) {
        var solution = (IRegressionSolution)run.Results.Values.Single(x => x is IRegressionSolution);
        var ds = solution.ProblemData.Dataset;

        if (solution.ProblemData == problemData) continue;
        if (ds == dataset) continue;
        if (problemData == null) {
          problemData = solution.ProblemData;
          continue;
        }
        if (dataset == null) {
          dataset = ds;
          continue;
        }

        if (problemData.TrainingPartition.Start != solution.ProblemData.TrainingPartition.Start || problemData.TrainingPartition.End != solution.ProblemData.TrainingPartition.End)
          throw new InvalidOperationException("The runs must share the same data.");

        if (!ds.DoubleVariables.SequenceEqual(dataset.DoubleVariables))
          throw new InvalidOperationException("The runs must share the same data.");

        foreach (var v in ds.DoubleVariables) {
          var values1 = (IList<double>)ds.GetReadOnlyDoubleValues(v);
          var values2 = (IList<double>)dataset.GetReadOnlyDoubleValues(v);

          if (values1.Count != values2.Count)
            throw new InvalidOperationException("The runs must share the same data.");

          if (!values1.SequenceEqual(values2))
            throw new InvalidOperationException("The runs must share the same data.");
        }
      }
    }

    public static RegressionEnsembleSolution CreateEnsembleSolution(IEnumerable<IRun> runs) {
      var solutions = runs.Select(x => x.Results.Values.Single(v => v is IRegressionSolution)).Cast<IRegressionSolution>();
      return new RegressionEnsembleSolution(new RegressionEnsembleModel(solutions.Select(x => x.Model)), solutions.First().ProblemData);
    }

    public static Dictionary<string, Tuple<IEnumerable<IRun>, Dictionary<string, double>>> CalculateVariableImpactsOnline(RunCollection runs, bool useBest) {
      AssertSameProblemData(runs);
      var solution = (IRegressionSolution)runs.First().Results.Values.Single(x => x is IRegressionSolution);
      var dataset = (Dataset)solution.ProblemData.Dataset;
      var originalValues = dataset.DoubleVariables.ToDictionary(x => x, x => dataset.GetReadOnlyDoubleValues(x).ToList());
      var medians = dataset.DoubleVariables.ToDictionary(x => x, x => Enumerable.Repeat(originalValues[x].Median(), originalValues[x].Count).ToList());

      var targetImpacts = new Dictionary<string, Tuple<IEnumerable<IRun>, Dictionary<string, double>>>();

      var groups = runs.GroupBy(run => {
        var sol = (IRegressionSolution)run.Results.Values.Single(x => x is IRegressionSolution);
        return Concatenate(sol.ProblemData.AllowedInputVariables) + sol.ProblemData.TargetVariable;
      });

      if (useBest) {
        // build network using only the best run for each target
        foreach (var group in groups) {
          var solutions = group.Select(run => Tuple.Create(run, (IRegressionSolution)run.Results.Values.Single(sol => sol is IRegressionSolution)));
          var best = solutions.OrderBy(x => x.Item2.TrainingRSquared).Last();
          var impacts = RegressionSolutionVariableImpactsCalculator.CalculateImpacts(best.Item2, 
            RegressionSolutionVariableImpactsCalculator.ReplacementMethodEnum.Shuffle,
            RegressionSolutionVariableImpactsCalculator.FactorReplacementMethodEnum.Best,
            RegressionSolutionVariableImpactsCalculator.DataPartitionEnum.All).ToDictionary(x => x.Item1, x => x.Item2);

          targetImpacts[best.Item2.ProblemData.TargetVariable] = Tuple.Create(new[] { best.Item1 }.AsEnumerable(), impacts);
        }
      } else {
        foreach (var group in groups) {
          // calculate average impacts
          var averageImpacts = new Dictionary<string, double>();
          solution = (IRegressionSolution)group.First().Results.Values.Single(x => x is IRegressionSolution);
          foreach (var run in group) {
            var sol = (IRegressionSolution)run.Results.Values.Single(v => v is IRegressionSolution);

            DoubleLimit estimationLimits = null;
            if (run.Parameters.ContainsKey("EstimationLimits")) {
              estimationLimits = (DoubleLimit)run.Parameters["EstimationLimits"];
            }
            var md = dataset.ToModifiable();

            var impacts = RegressionSolutionVariableImpactsCalculator.CalculateImpacts(sol,
              RegressionSolutionVariableImpactsCalculator.ReplacementMethodEnum.Shuffle,
              RegressionSolutionVariableImpactsCalculator.FactorReplacementMethodEnum.Best,
              RegressionSolutionVariableImpactsCalculator.DataPartitionEnum.All);            
            foreach (var t in impacts) {
              if (averageImpacts.ContainsKey(t.Item1))
                averageImpacts[t.Item1] += t.Item2;
              else {
                averageImpacts[t.Item1] = t.Item2;
              }
            }
          }

          var count = group.Count();
          foreach (var v in averageImpacts.Keys.ToList()) {
            averageImpacts[v] /= count;
          }

          targetImpacts[solution.ProblemData.TargetVariable] = Tuple.Create(group.AsEnumerable(), averageImpacts);
        }
      }
      return targetImpacts;
    }

    public static Dictionary<string, Tuple<IEnumerable<IRun>, Dictionary<string, double>>> CalculateVariableImpactsFromRunResults(RunCollection runs,
      string qualityResultName, bool maximization, string impactsResultName, bool useBestRunsPerTarget = false) {

      Func<IRun, double> getQuality = run => ((DoubleValue)run.Results[qualityResultName]).Value;
      var targetGroups = runs.GroupBy(x => ((IRegressionProblemData)x.Parameters["ProblemData"]).TargetVariable).ToList();
      var targetImpacts = new Dictionary<string, Tuple<IEnumerable<IRun>, Dictionary<string, double>>>();

      if (useBestRunsPerTarget) {
        foreach (var group in targetGroups) {
          var ordered = group.OrderBy(getQuality);
          var best = maximization ? ordered.Last() : ordered.First();
          var pd = (IRegressionProblemData)best.Parameters["ProblemData"];
          var target = group.Key;
          var impacts = (DoubleMatrix)best.Results[impactsResultName];
          targetImpacts[target] = Tuple.Create((IEnumerable<IRun>)new[] { best }, impacts.RowNames.Select((x, i) => new { x, i }).ToDictionary(x => x.x, x => impacts[x.i, 0]));
        }
      } else {
        foreach (var target in targetGroups) {
          var averageImpacts = CalculateAverageImpacts(new RunCollection(target), impactsResultName);
          targetImpacts[target.Key] = new Tuple<IEnumerable<IRun>, Dictionary<string, double>>(target, averageImpacts);
        }
      }
      return targetImpacts;
    }

    public static VariableInteractionNetwork CreateNetwork(Dictionary<string, Tuple<IEnumerable<IRun>, Dictionary<string, double>>> targetImpacts) {
      var nodes = new Dictionary<string, IVertex>();
      var vn = new VariableInteractionNetwork();
      foreach (var ti in targetImpacts) {
        var target = ti.Key;
        var variableImpacts = ti.Value.Item2;
        var targetRuns = ti.Value.Item1;
        IVertex targetNode;

        var variables = variableImpacts.Keys.ToList();
        if (variables.Count == 0) continue;

        if (!nodes.TryGetValue(target, out targetNode)) {
          targetNode = new VariableNetworkNode { Label = target };
          vn.AddVertex(targetNode);
          nodes[target] = targetNode;
        }

        IVertex variableNode;
        if (variables.Count > 1) {
          var variableList = new List<string>(variables) { target };
          var junctionLabel = Concatenate(variableList);
          IVertex junctionNode;
          var sb = new StringBuilder();
          if (!nodes.TryGetValue(junctionLabel, out junctionNode)) {
            var solutionsEnsemble = CreateEnsembleSolution(targetRuns);
            junctionNode = new JunctionNetworkNode { Label = solutionsEnsemble.TrainingRSquared.ToString("N3", CultureInfo.CurrentCulture), Data = solutionsEnsemble };
            vn.AddVertex(junctionNode);
            nodes[junctionLabel] = junctionNode;
            sb.AppendLine(junctionNode.Label);
          }
          IArc arc;
          foreach (var v in variables) {
            var impact = variableImpacts[v];
            if (!nodes.TryGetValue(v, out variableNode)) {
              variableNode = new VariableNetworkNode { Label = v };
              vn.AddVertex(variableNode);
              nodes[v] = variableNode;
            }
            arc = new Arc(variableNode, junctionNode) { Weight = impact, Label = impact.ToString("N3", CultureInfo.CurrentCulture) };
            sb.AppendLine(v + ": " + arc.Label);
            vn.AddArc(arc);
          }
          var jcnNode = (JunctionNetworkNode)junctionNode;
          var trainingR2 = ((IRegressionSolution)jcnNode.Data).TrainingRSquared;
          arc = new Arc(junctionNode, targetNode) { Weight = junctionNode.InArcs.Sum(x => x.Weight), Label = trainingR2.ToString("N3", CultureInfo.CurrentCulture) };
          vn.AddArc(arc);
        } else {
          foreach (var v in variables) {
            var impact = variableImpacts[v];
            if (!nodes.TryGetValue(v, out variableNode)) {
              variableNode = new VariableNetworkNode { Label = v };
              vn.AddVertex(variableNode);
              nodes[v] = variableNode;
            }
            var arc = new Arc(variableNode, targetNode) {
              Weight = impact,
              Label = impact.ToString("N3", CultureInfo.CurrentCulture)
            };
            vn.AddArc(arc);
          }
        }
      }
      return vn;
    }

    public static VariableInteractionNetwork ApplyThreshold(VariableInteractionNetwork originalNetwork, double threshold) {
      var arcs = originalNetwork.Arcs.Where(x => x.Weight >= threshold).ToList();
      if (!arcs.Any()) return originalNetwork;
      var filteredNetwork = new VariableInteractionNetwork();
      var cloner = new Cloner();
      var vertices = arcs.SelectMany(x => new[] { x.Source, x.Target }).Select(cloner.Clone).Distinct(); // arcs are not cloned
      filteredNetwork.AddVertices(vertices);
      filteredNetwork.AddArcs(arcs.Select(x => (IArc)x.Clone(cloner)));

      var unusedJunctions = filteredNetwork.Vertices.Where(x => x.InDegree == 0 && x is JunctionNetworkNode).ToList();
      filteredNetwork.RemoveVertices(unusedJunctions);
      var orphanedNodes = filteredNetwork.Vertices.Where(x => x.Degree == 0).ToList();
      filteredNetwork.RemoveVertices(orphanedNodes);
      return filteredNetwork.Vertices.Any() ? filteredNetwork : originalNetwork;
    }

    private static double CalculateAverageQuality(RunCollection runs) {
      var pd = (IRegressionProblemData)runs.First().Parameters["ProblemData"];
      var target = pd.TargetVariable;
      var inputs = pd.AllowedInputVariables;

      if (!runs.All(x => {
        var problemData = (IRegressionProblemData)x.Parameters["ProblemData"];
        return target == problemData.TargetVariable && inputs.SequenceEqual(problemData.AllowedInputVariables);
      })) {
        throw new ArgumentException("All runs must have the same target and inputs.");
      }
      return runs.Average(x => ((DoubleValue)x.Results["Best training solution quality"]).Value);
    }

    public static Dictionary<string, double> CalculateAverageImpacts(RunCollection runs, string resultName) {
      var pd = (IRegressionProblemData)runs.First().Parameters["ProblemData"];
      var target = pd.TargetVariable;
      var inputs = pd.AllowedInputVariables.ToList();

      var impacts = inputs.ToDictionary(x => x, x => 0d);

      // check if all the runs have the same target and same inputs
      if (!runs.All(x => {
        var problemData = (IRegressionProblemData)x.Parameters["ProblemData"];
        return target == problemData.TargetVariable && inputs.SequenceEqual(problemData.AllowedInputVariables);
      })) {
        throw new ArgumentException("All runs must have the same target and inputs.");
      }

      foreach (var run in runs) {
        var impactsMatrix = (DoubleMatrix)run.Results[resultName];
        int i = 0;
        foreach (var v in impactsMatrix.RowNames) {
          impacts[v] += impactsMatrix[i, 0];
          ++i;
        }
      }

      foreach (var v in inputs) {
        impacts[v] /= runs.Count;
      }

      return impacts;
    }

    private static string Concatenate(IEnumerable<string> strings) {
      var sb = new StringBuilder();
      foreach (var s in strings) {
        sb.Append(s);
      }
      return sb.ToString();
    }

    private void ConfigureNodeShapes() {
      graphChart.ClearShapes();
      var font = new Font(FontFamily.GenericSansSerif, 12);
      graphChart.AddShape(typeof(VariableNetworkNode), new LabeledPrimitive(new Ellipse(graphChart.Chart, new PointD(0, 0), new PointD(30, 30), Pens.Black, Brushes.White), "", font));
      graphChart.AddShape(typeof(JunctionNetworkNode), new LabeledPrimitive(new Rectangle(graphChart.Chart, new PointD(0, 0), new PointD(15, 15), Pens.Black, Brushes.DarkGray), "", font));
    }

    public void UpdateNetwork(VariableInteractionNetwork network) {
      if (InvokeRequired) {
        Invoke((Action<VariableInteractionNetwork>)UpdateNetwork, network);
      } else {
        graphChart.Graph = network;
      }      
    }

    #region events
    protected override void OnContentChanged() {
      base.OnContentChanged();
      var run = Content.First();
      var pd = (IRegressionProblemData)run.Parameters["ProblemData"];
      var variables = new HashSet<string>(new List<string>(pd.Dataset.DoubleVariables));
      impactResultNameComboBox.Items.Clear();
      foreach (var result in run.Results.Where(x => x.Value is DoubleMatrix)) {
        var m = (DoubleMatrix)result.Value;
        if (m.RowNames.All(x => variables.Contains(x)))
          impactResultNameComboBox.Items.Add(result.Key);
      }
      qualityResultNameComboBox.Items.Clear();
      foreach (var result in run.Results.Where(x => x.Value is DoubleValue)) {
        qualityResultNameComboBox.Items.Add(result.Key);
      }
      if (impactResultNameComboBox.Items.Count > 0) {
        impactResultNameComboBox.Text = (string)impactResultNameComboBox.Items[0];
      }
      if (qualityResultNameComboBox.Items.Count > 0) {
        qualityResultNameComboBox.Text = (string)qualityResultNameComboBox.Items[0];
      }
      if (impactResultNameComboBox.Items.Count > 0 && qualityResultNameComboBox.Items.Count > 0)
        NetworkConfigurationChanged(this, EventArgs.Empty);
    }

    private void TextBoxValidating(object sender, CancelEventArgs e) {
      double v;
      string errorMsg = "Could not parse the entered value. Please input a real number.";
      var tb = (TextBox)sender;
      if (!double.TryParse(tb.Text, out v)) {
        e.Cancel = true;
        tb.Select(0, tb.Text.Length);

        // Set the ErrorProvider error with the text to display. 
        this.errorProvider.SetError(tb, errorMsg);
        errorProvider.BlinkStyle = ErrorBlinkStyle.NeverBlink;
        errorProvider.SetIconPadding(tb, -20);
      }
    }

    private void ImpactThresholdTextBoxValidated(object sender, EventArgs e) {
      var tb = (TextBox)sender;
      errorProvider.SetError(tb, string.Empty);
      double impact;
      if (!double.TryParse(tb.Text, out impact)) {
        impact = 0.2;
      }
      var network = ApplyThreshold(variableInteractionNetwork, impact);
      graphChart.Graph = network;
    }

    private void LayoutConfigurationBoxValidated(object sender, EventArgs e) {
      var tb = (TextBox)sender;
      errorProvider.SetError(tb, string.Empty);
      LayoutConfigurationChanged(sender, e);
    }

    private void NetworkConfigurationChanged(object sender, EventArgs e) {
      var useBest = impactAggregationComboBox.SelectedIndex <= 0;
      var threshold = impactThresholdTrackBar.Value / 100.0;
      var qualityResultName = qualityResultNameComboBox.Text;
      var impactsResultName = impactResultNameComboBox.Text;
      if (string.IsNullOrEmpty(qualityResultName) || string.IsNullOrEmpty(impactsResultName))
        return;
      var maximization = maximizationCheckBox.Checked;
      var impacts = CalculateVariableImpactsFromRunResults(Content, qualityResultName, maximization, impactsResultName, useBest);
      variableInteractionNetwork = CreateNetwork(impacts);
      var network = ApplyThreshold(variableInteractionNetwork, threshold);
      graphChart.Graph = network;
    }

    private void LayoutConfigurationChanged(object sender, EventArgs e) {
      ConstrainedForceDirectedLayout.EdgeRouting routingMode;
      switch (edgeRoutingComboBox.SelectedIndex) {
        case 0:
          routingMode = ConstrainedForceDirectedLayout.EdgeRouting.None;
          break;
        case 1:
          routingMode = ConstrainedForceDirectedLayout.EdgeRouting.Polyline;
          break;
        case 2:
          routingMode = ConstrainedForceDirectedLayout.EdgeRouting.Orthogonal;
          break;
        default:
          throw new ArgumentException("Invalid edge routing mode.");
      }
      var idealEdgeLength = double.Parse(idealEdgeLengthTextBox.Text);
      if (routingMode == graphChart.RoutingMode && idealEdgeLength.IsAlmost(graphChart.DefaultEdgeLength)) return;
      graphChart.RoutingMode = routingMode;
      graphChart.PerformEdgeRouting = routingMode != ConstrainedForceDirectedLayout.EdgeRouting.None;
      graphChart.DefaultEdgeLength = idealEdgeLength;
      graphChart.Draw();
    }

    private void ControlsEnable(bool enabled) {
      qualityResultNameComboBox.Enabled
        = impactResultNameComboBox.Enabled
        = impactAggregationComboBox.Enabled
        = impactThresholdTrackBar.Enabled
        = onlineImpactCalculationButton.Enabled
        = edgeRoutingComboBox.Enabled
        = idealEdgeLengthTextBox.Enabled
        = maximizationCheckBox.Enabled = enabled;
    }

    private void onlineImpactCalculationButton_Click(object sender, EventArgs args) {
      var worker = new BackgroundWorker();
      worker.DoWork += (o, e) => {
        ControlsEnable(false);
        var impacts = CalculateVariableImpactsOnline(Content, impactAggregationComboBox.SelectedIndex == 0);
        variableInteractionNetwork = CreateNetwork(impacts);
        var threshold = impactThresholdTrackBar.Minimum + (double)impactThresholdTrackBar.Value / impactThresholdTrackBar.Maximum;
        graphChart.Graph = ApplyThreshold(variableInteractionNetwork, threshold);
      };
      worker.RunWorkerCompleted += (o, e) => ControlsEnable(true);
      worker.RunWorkerAsync();
    }

    private void relayoutGraphButton_Click(object sender, EventArgs e) {
      graphChart.Draw();
    }
    #endregion

    private void exportImpactsMatrixButton_Click(object sender, EventArgs e) {
      var graph = graphChart.Graph;
      var labels = graph.Vertices.OfType<VariableNetworkNode>().Select(x => x.Label).ToList();
      labels.Sort(); // sort variables alphabetically
      var matrix = new DoubleMatrix(labels.Count, labels.Count) { RowNames = labels, ColumnNames = labels };
      var indexes = labels.Select((x, i) => new { Label = x, Index = i }).ToDictionary(x => x.Label, x => x.Index);
      var junctions = graph.Vertices.OfType<JunctionNetworkNode>().ToList();
      foreach (var jn in junctions) {
        var target = jn.OutArcs.First().Target.Label;
        var targetIndex = indexes[target];
        foreach (var input in jn.InArcs) {
          var inputIndex = indexes[input.Source.Label];
          var inputImpact = input.Weight;
          matrix[targetIndex, inputIndex] = inputImpact;
        }
      }
      for (int i = 0; i < labels.Count; ++i) matrix[i, i] = 1;
      MainFormManager.MainForm.ShowContent(matrix);
    }

    private void impactThresholdTrackBar_ValueChanged(object sender, EventArgs e) {
      var impact = impactThresholdTrackBar.Minimum + (double)impactThresholdTrackBar.Value / impactThresholdTrackBar.Maximum;
      impactThresholdLabel.Text = impact.ToString("N3", CultureInfo.CurrentCulture);
      var network = ApplyThreshold(variableInteractionNetwork, impact);
      graphChart.Graph = network;
    }


  }
}