Page 1 | Page 2 |  Page 3 | Page 4

Click model 2 example

Here we implement the same model as in click model 1 but with shared variables. There are a number of reasons why one might want to use shared variables including memory problems, parallelisation, and more control over the schedule which might be necessary if there are convergence problems. Infer.NET provides a SharedVariable class and a Model class which ensure that the correct messages get marshalled between the different models. This model is available as Model2 in the example code. It mirrors the Model1 code except for the following:

  • SharedVariable objects are created in place of Variable objects for all variables that we want to infer; these are initialised with the priors.
  • Model code must be changed to refer to the instance of the SharedVariable for the current chunk.
  • The data is divided into identically sized chunks.
  • We explicitly loop over chunks, and do inference on each chunk. We need to loop over all chunks several times, checking marginals between each pass to test for convergence.
  • For each chunk, we use SharedVariable and Model class methods to obtain the variables for each submodel, and to perform inference on these variables, respectively.

Let's look at each of these in turn.

Creating the shared variables

There is a concept called Model that takes care of tedious plumbing needed for sharing information between models. The SharedVariable class is a convenient wrapper class used to specify the variables that are shared between the models.  First, we create an instance of the model. Simultaneously, we also specify the variables that are shared across all the models. For more detailed description check out this link on shared variables.

Model model = new Model(numChunks);
SharedVariable<double> scoreMean = SharedVariable<double>.Random(priorScoreMean).Named("scoreMean");
SharedVariable<double> scorePrec = SharedVariable<double>.Random(priorScorePrec).Named("scorePrec");
SharedVariable<double> judgePrec = SharedVariable<double>.Random(priorJudgePrec).Named("judgePrec");
SharedVariable<double> clickPrec = SharedVariable<double>.Random(priorClickPrec).Named("clickPrec");
SharedVariable<double>[] thresholds = new SharedVariable<double>[numThresholds];
for (int t = 0; t < numThresholds; t++) {
  thresholds[t] = SharedVariable<double>.Random(priorThresholds[t]).Named("threshold" + t);

Changes in the model code

Changes are needed in the model code to take care of the fact that scoreMean etc. are no longer Variables. These changes just require referring to the GetCopyFor method of the SharedVariable instance:

scores[r] = Variable<double>.GaussianFromMeanAndPrecision(
    scoreMean.GetCopyFor(model), scorePrec.GetCopyFor(model)).ForEach(r);
scoresJ[r] = Variable<double>.GaussianFromMeanAndPrecision(
    scores[r], judgePrec.GetCopyFor(model));
scoresC[r] = Variable<double>.GaussianFromMeanAndPrecision(
    scores[r], clickPrec.GetCopyFor(model));

    scoresJ[r], thresholds[i].GetCopyFor(model), thresholds[i + 1].GetCopyFor(model));

Dividing the data into chunks

A method is provided for this example (i.e. not part of Infer.NET) which divides the data into chunks and returns an array of arrays of arrays of Gaussians. The first index of this triply-indexed array is the chunk index, the second is the label index, and the third is the query/document index:

Gaussian[][][] allObs = getClickObservations(numLabels,chunkSize,labels,clicks,exams);
int numChunks = allObs.Length;

When a given chunk, index c, is processed, the numberOfObservations and observationDistrib parameters must be set according to the corresonding data:

for (int i = 0; i < numLabels; i++)
    numberOfObservations[i].ObservedValue = allObs[c][i].Length;
    observationDistribs[i].ObservedValue = allObs[c][i];

Looping over chunks, and convergence

We do several passes over the chunks.  We can infer all the variables jointly by using InferShared method of Model class. After each full pass the marginals of key shared variables are tested against the marginals of the previous pass in order to determine convergence:

for (int pass = 0; pass < maxPasses; pass++)
    prevMargScoreMean = marginals.marginalScoreMean;
    prevMargJudgePrec = marginals.marginalJudgePrec;
    prevMargClickPrec = marginals.marginalClickPrec;
    for (int c = 0; c < numChunks; c++)
        // set values

        // perform inference

// Retrieve marginals

        marginals.marginalScoreMean = scoreMean.Marginal<Gaussian>();
        marginals.marginalScorePrec = scorePrec.Marginal(Gamma);
        marginals.marginalJudgePrec = judgePrec.Marginal(Gamma);
        marginals.marginalClickPrec = clickPrec.Marginal(Gamma);

        // Test for convergence
if (marginals.marginalScoreMean.MaxDiff(prevMargScoreMean) < convergenceThresh &&
            marginals.marginalJudgePrec.MaxDiff(prevMargJudgePrec) < convergenceThresh &&
            marginals.marginalClickPrec.MaxDiff(prevMargClickPrec) < convergenceThresh)

Running Model 2

We can run model 2 using the same prediction model as for model 1. We present the data in chunks of 200; the inference converges after 4 passes, and gets identical results (up to some small numeric tolerance) with model 1.

Page 1 | Page 2 |  Page 3 | Page 4

©2009-2015 Microsoft Corporation. All rights reserved.  Terms of Use | Trademarks | Privacy Statement