Support multi-threadhing for ranking evaluation

Top-k recommendation for every single user is costly. It'd be
recommended to parallelize whenever possible.

src/evaluation/evaluate.jl

@@ -21,17 +21,17 @@ function evaluate(recommender::Recommender, truth_data::DataAccessor,
     validate(recommender, truth_data)
     n_users, n_items = size(truth_data.R)
 
-    accum = 0.0
+    accum = Threads.Atomic{Float64}(0.0)
 
-    for u in 1:n_users
+    Threads.@threads for u in 1:n_users
         observed_items = findall(!iszero, truth_data.R[u, :])
         if length(observed_items) == 0; continue; end
         truth = [first(t) for t in sort(collect(zip(observed_items, truth_data.R[u, observed_items])), by=t->last(t), rev=true)]
         candidates = findall(iszero, recommender.data.R[u, :]) # items that were unobserved as of building the model
         pred = [first(item_score_pair) for item_score_pair in recommend(recommender, u, topk, candidates)]
-        accum += measure(metric, truth, pred, topk)
+        Threads.atomic_add!(accum, measure(metric, truth, pred, topk))
     end
 
     # return average accuracy over the all target users
-    accum / n_users
+    accum[] / n_users
 end