|
public static Evaluate ( IItemRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback train, ICollection |
||
| recommender | IItemRecommender | item recommender |
| test | IPosOnlyFeedback | test cases |
| train | IPosOnlyFeedback | training data |
| relevant_users | ICollection |
a collection of integers with all relevant users |
| relevant_items | ICollection |
a collection of integers with all relevant items |
| Результат | double>.Dictionary | |
/// <summary>Online evaluation for rankings of items</summary>
/// <remarks>
/// The evaluation protocol works as follows:
/// For every test user, evaluate on the test items, and then add the those test items to the training set and perform an incremental update.
/// The sequence of users is random.
/// </remarks>
/// <param name="recommender">the item recommender to be evaluated</param>
/// <param name="test">test cases</param>
/// <param name="training">training data (must be connected to the recommender's training data)</param>
/// <param name="test_users">a list of all test user IDs</param>
/// <param name="candidate_items">a list of all candidate item IDs</param>
/// <param name="candidate_item_mode">the mode used to determine the candidate items</param>
/// <returns>a dictionary containing the evaluation results (averaged by user)</returns>
static public ItemRecommendationEvaluationResults EvaluateOnline(
this IRecommender recommender,
IPosOnlyFeedback test, IPosOnlyFeedback training,
IList <int> test_users, IList <int> candidate_items,
CandidateItems candidate_item_mode)
{
var incremental_recommender = recommender as IIncrementalItemRecommender;
if (incremental_recommender == null)
{
throw new ArgumentException("recommender must be of type IIncrementalItemRecommender");
}
candidate_items = Items.Candidates(candidate_items, candidate_item_mode, test, training);
test_users.Shuffle();
var results_by_user = new Dictionary <int, ItemRecommendationEvaluationResults>();
foreach (int user_id in test_users)
{
if (candidate_items.Intersect(test.ByUser[user_id]).Count() == 0)
{
continue;
}
// prepare data
var current_test_data = new PosOnlyFeedback <SparseBooleanMatrix>();
foreach (int index in test.ByUser[user_id])
{
current_test_data.Add(user_id, test.Items[index]);
}
// evaluate user
var current_result = Items.Evaluate(recommender, current_test_data, training, current_test_data.AllUsers, candidate_items, CandidateItems.EXPLICIT);
results_by_user[user_id] = current_result;
// update recommender
var tuples = new List <Tuple <int, int> >();
foreach (int index in test.ByUser[user_id])
{
tuples.Add(Tuple.Create(user_id, test.Items[index]));
}
incremental_recommender.AddFeedback(tuples);
// TODO candidate_items should be updated properly
}
var results = new ItemRecommendationEvaluationResults();
foreach (int u in results_by_user.Keys)
{
foreach (string measure in Items.Measures)
{
results[measure] += results_by_user[u][measure];
}
}
foreach (string measure in Items.Measures)
{
results[measure] /= results_by_user.Count;
}
results["num_users"] = results_by_user.Count;
results["num_items"] = candidate_items.Count;
results["num_lists"] = results_by_user.Count;
return(results);
}
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary>
/// <param name="recommender">an item recommender</param>
/// <param name="split">a positive-only feedback dataset split</param>
/// <param name="test_users">a collection of integers with all test users</param>
/// <param name="candidate_items">a collection of integers with all candidate items</param>
/// <param name="candidate_item_mode">the mode used to determine the candidate items</param>
/// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param>
/// <param name="max_iter">the maximum number of iterations</param>
/// <param name="find_iter">the report interval</param>
/// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param>
static public void DoRatingBasedRankingIterativeCrossValidation(
this RatingPredictor recommender,
ISplit <IRatings> split,
IList <int> test_users,
IList <int> candidate_items,
CandidateItems candidate_item_mode,
RepeatedEvents repeated_events,
uint max_iter,
uint find_iter = 1,
bool show_fold_results = false)
{
if (!(recommender is IIterativeModel))
{
throw new ArgumentException("recommender must be of type IIterativeModel");
}
var split_recommenders = new RatingPredictor[split.NumberOfFolds];
var iterative_recommenders = new IIterativeModel[split.NumberOfFolds];
var fold_results = new ItemRecommendationEvaluationResults[split.NumberOfFolds];
// initial training and evaluation
Parallel.For(0, (int)split.NumberOfFolds, i =>
{
try
{
split_recommenders[i] = (RatingPredictor)recommender.Clone(); // to avoid changes in recommender
split_recommenders[i].Ratings = split.Train[i];
split_recommenders[i].Train();
iterative_recommenders[i] = (IIterativeModel)split_recommenders[i];
var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[i]);
var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[i]);
fold_results[i] = Items.Evaluate(split_recommenders[i], test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode, repeated_events);
if (show_fold_results)
{
Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, iterative_recommenders[i].NumIter);
}
}
catch (Exception e)
{
Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace);
throw;
}
});
Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), iterative_recommenders[0].NumIter);
// iterative training and evaluation
for (int it = (int)iterative_recommenders[0].NumIter + 1; it <= max_iter; it++)
{
Parallel.For(0, (int)split.NumberOfFolds, i =>
{
try
{
iterative_recommenders[i].Iterate();
if (it % find_iter == 0)
{
var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[i]);
var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[i]);
fold_results[i] = Items.Evaluate(split_recommenders[i], test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode, repeated_events);
if (show_fold_results)
{
Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, it);
}
}
}
catch (Exception e)
{
Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace);
throw;
}
});
Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), it);
}
}
/// <summary>Evaluate on the folds of a dataset split</summary>
/// <param name="recommender">an item recommender</param>
/// <param name="split">a dataset split</param>
/// <param name="candidate_items">a collection of integers with all candidate items</param>
/// <param name="candidate_item_mode">the mode used to determine the candidate items</param>
/// <param name="compute_fit">if set to true measure fit on the training data as well</param>
/// <param name="show_results">set to true to print results to STDERR</param>
/// <returns>a dictionary containing the average results over the different folds of the split</returns>
static public EvaluationResults DoRatingBasedRankingCrossValidation(
this RatingPredictor recommender,
ISplit <IRatings> split,
IList <int> candidate_items,
CandidateItems candidate_item_mode = CandidateItems.OVERLAP,
bool compute_fit = false,
bool show_results = false)
{
var avg_results = new ItemRecommendationEvaluationResults();
Parallel.For(0, (int)split.NumberOfFolds, fold =>
{
try
{
var split_recommender = (RatingPredictor)recommender.Clone(); // avoid changes in recommender
split_recommender.Ratings = split.Train[fold];
split_recommender.Train();
var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[fold]);
var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[fold]);
IList <int> test_users = test_data_posonly.AllUsers;
var fold_results = Items.Evaluate(split_recommender, test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode);
if (compute_fit)
{
fold_results["fit"] = (float)split_recommender.ComputeFit();
}
// thread-safe stats
lock (avg_results)
foreach (var key in fold_results.Keys)
{
if (avg_results.ContainsKey(key))
{
avg_results[key] += fold_results[key];
}
else
{
avg_results[key] = fold_results[key];
}
}
if (show_results)
{
Console.Error.WriteLine("fold {0} {1}", fold, fold_results);
}
}
catch (Exception e)
{
Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace);
throw;
}
});
foreach (var key in Items.Measures)
{
avg_results[key] /= split.NumberOfFolds;
}
avg_results["num_users"] /= split.NumberOfFolds;
avg_results["num_items"] /= split.NumberOfFolds;
return(avg_results);
}
/// <summary>Evaluate on the folds of a dataset split</summary>
/// <param name="recommender">an item recommender</param>
/// <param name="split">a dataset split</param>
/// <param name="test_users">a collection of integers with all test users</param>
/// <param name="candidate_items">a collection of integers with all candidate items</param>
/// <param name="candidate_item_mode">the mode used to determine the candidate items</param>
/// <param name="compute_fit">if set to true measure fit on the training data as well</param>
/// <param name="show_results">set to true to print results to STDERR</param>
/// <returns>a dictionary containing the average results over the different folds of the split</returns>
static public ItemRecommendationEvaluationResults DoCrossValidation(
this IRecommender recommender,
ISplit <IPosOnlyFeedback> split,
IList <int> test_users,
IList <int> candidate_items,
CandidateItems candidate_item_mode = CandidateItems.OVERLAP,
bool compute_fit = false,
bool show_results = false)
{
var avg_results = new ItemRecommendationEvaluationResults();
if (!(recommender is ItemRecommender))
{
throw new ArgumentException("recommender must be of type ItemRecommender");
}
Parallel.For(0, (int)split.NumberOfFolds, fold =>
{
try
{
var split_recommender = (ItemRecommender)recommender.Clone(); // avoid changes in recommender
split_recommender.Feedback = split.Train[fold];
split_recommender.Train();
var fold_results = Items.Evaluate(split_recommender, split.Test[fold], split.Train[fold], test_users, candidate_items, candidate_item_mode);
if (compute_fit)
{
fold_results["fit"] = (float)split_recommender.ComputeFit();
}
// thread-safe stats
lock (avg_results)
foreach (var key in fold_results.Keys)
{
if (avg_results.ContainsKey(key))
{
avg_results[key] += fold_results[key];
}
else
{
avg_results[key] = fold_results[key];
}
}
if (show_results)
{
Console.Error.WriteLine("fold {0} {1}", fold, fold_results);
}
}
catch (Exception e)
{
Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace);
throw;
}
});
foreach (var key in Items.Measures)
{
avg_results[key] /= split.NumberOfFolds;
}
avg_results["num_users"] /= split.NumberOfFolds;
avg_results["num_items"] /= split.NumberOfFolds;
return(avg_results);
}