full matrix prediction

src/macau.jl

@@ -39,7 +39,7 @@ function macau(data::RelationData;
   modes_other = map(entity -> Vector{Int64}[ find(en -> en != entity, r.entities) for r in entity.relations ],
                      data.entities)
   if full_prediction
-    yhat_full = zeros(size(r.relations[1]))
+    yhat_full = zeros(size(data.relations[1]))
   end
 
   verbose && println("Sampling")
@@ -107,6 +107,10 @@ function macau(data::RelationData;
     rel = data.relations[1]
     probe_rat = pred(rel, rel.test_vec, rel.test_F)
 
+    if full_prediction && i >= burnin ## last burnin sample is included
+      yhat_full += pred_all( data.relations[1] )
+    end
+
     if i > burnin
       if verbose && i == burnin + 1
         println("--------- Burn-in complete, averaging posterior samples ----------")
@@ -155,6 +159,9 @@ function macau(data::RelationData;
   result["RMSE"]        = rmse_avg
   result["accuracy"]    = err_avg
   result["ROC"]         = roc_avg
+  if full_prediction
+    result["predictions_full"] = yhat_full / (burnin + 1)
+  end
   if numTest(data.relations[1]) > 0
     rel = data.relations[1]
     result["predictions"] = copy(rel.test_vec)

src/sampling.jl

@@ -1,4 +1,6 @@
-export pred
+using Iterators
+
+export pred, pred_all
 
 function pred(r::Relation, probe_vec::DataFrame, F)
   if ! hasFeatures(r)
@@ -39,7 +41,7 @@ function pred_all(r::Relation)
   if hasFeatures(r)
     error("Prediction of all elements is not possible when Relation has features.")
   end
-  udot(r) + r.model.mean_value
+  udot_all(r) + r.model.mean_value
 end
 
 function makeClamped(x, clamp::Vector{Float64})

test/runtests.jl

@@ -57,6 +57,18 @@ assignToTest!(rd.relations[1], 2)
 result = macau(rd, burnin = 10, psamples = 10, verbose = false)
 @test size(result["predictions"],1) == 2
 
+# testing pred_all
+Yhat = pred_all(rd.relations[1])
+@test size(Yhat) == (15, 10)
+@test_approx_eq Yhat[2,3] (rd.entities[1].model.sample[2,:] * rd.entities[2].model.sample[3,:]')[1] + rd.relations[1].model.mean_value
+
+# predict all
+result1 = macau(rd, burnin = 10, psamples = 10, verbose = false, full_prediction = true)
+@test size(result1["predictions_full"]) == (15, 10)
+x1 = result1["predictions"][1, 1:2]
+y1 = result1["predictions"][:pred][1]
+@test_approx_eq result1["predictions_full"][x1[1], x1[2]] y1
+
 # custom function on latent variables
 f1(a) = length(a)
 result2 = macau(rd, burnin = 5, psamples = 6, verbose = false, f = f1)

test/tensor.jl

@@ -17,3 +17,11 @@ end
 rd = RelationData(df)
 assignToTest!(rd.relations[1], 10)
 result = macau(rd, burnin=50, psamples=10, num_latent=2, verbose=false)
+
+## checking prediction for whole matrix for tensors
+Yhat = pred_all(rd.relations[1])
+@test size(Yhat) == size(X)
+yprod   = vec(rd.entities[1].model.sample[4,:])
+yprod .*= vec(rd.entities[2].model.sample[2,:])
+yprod .*= vec(rd.entities[3].model.sample[1,:])
+@test_approx_eq Yhat[4,2,1] sum(yprod)+rd.relations[1].model.mean_value