public void TestUserRescorer()
{
Rescorer <User> rescorer = NullRescorer <User> .Instance;
Assert.IsNotNull(rescorer);
User user = new GenericUser <String>("test", new List <Preference>());
Assert.AreEqual(1.0, rescorer.Rescore(user, 1.0));
Assert.AreEqual(1.0, rescorer.Rescore(null, 1.0));
Assert.AreEqual(0.0, rescorer.Rescore(user, 0.0));
Assert.IsTrue(Double.IsNaN(rescorer.Rescore(user, Double.NaN)));
}
public void TestItemRescorer()
{
Rescorer <Item> rescorer = NullRescorer <Item> .Instance;
Assert.IsNotNull(rescorer);
Item item = new GenericItem <String>("test");
Assert.AreEqual(1.0, rescorer.Rescore(item, 1.0));
Assert.AreEqual(1.0, rescorer.Rescore(null, 1.0));
Assert.AreEqual(0.0, rescorer.Rescore(item, 0.0));
Assert.IsTrue(Double.IsNaN(rescorer.Rescore(item, Double.NaN)));
}
public virtual void TestBasic()
{
// create a sort field and sort by it (reverse order)
Query query = new TermQuery(new Term("body", "contents"));
IndexReader r = searcher.IndexReader;
// Just first pass query
TopDocs hits = searcher.Search(query, 10);
AreEqual(3, hits.TotalHits);
AreEqual("3", r.Document(hits.ScoreDocs[0].Doc).Get("id"));
AreEqual("1", r.Document(hits.ScoreDocs[1].Doc).Get("id"));
AreEqual("2", r.Document(hits.ScoreDocs[2].Doc).Get("id"));
// Now, rescore:
Expression e = JavascriptCompiler.Compile("sqrt(_score) + ln(popularity)");
SimpleBindings bindings = new SimpleBindings();
bindings.Add(new SortField("popularity", SortField.Type_e.INT));
bindings.Add(new SortField("_score", SortField.Type_e.SCORE));
Rescorer rescorer = e.GetRescorer(bindings);
hits = rescorer.Rescore(searcher, hits, 10);
AreEqual(3, hits.TotalHits);
AreEqual("2", r.Document(hits.ScoreDocs[0].Doc).Get("id"));
AreEqual("1", r.Document(hits.ScoreDocs[1].Doc).Get("id"));
AreEqual("3", r.Document(hits.ScoreDocs[2].Doc).Get("id"));
string expl = rescorer.Explain(searcher, searcher.Explain(query, hits.ScoreDocs[0].Doc), hits.ScoreDocs[0].Doc).ToString();
// Confirm the explanation breaks out the individual
// variables:
IsTrue(expl.Contains("= variable \"popularity\""));
// Confirm the explanation includes first pass details:
IsTrue(expl.Contains("= first pass score"));
IsTrue(expl.Contains("body:contents in"));
}
public int Compare(RecommendedItem o1, RecommendedItem o2)
{
double rescored1 = rescorer.Rescore(o1.Item, o1.Value);
double rescored2 = rescorer.Rescore(o2.Item, o2.Value);
Debug.Assert(!double.IsNaN(rescored1));
Debug.Assert(!double.IsNaN(rescored2));
if (rescored1 < rescored2)
{
return(1);
}
else if (rescored1 > rescored2)
{
return(-1);
}
else
{
return(0);
}
}
public double Estimate(Item item)
{
Pair <Item, Item> pair = new Pair <Item, Item>(toItem, item);
if (rescorer.IsFiltered(pair))
{
return(Double.NaN);
}
double originalEstimate = correlation.GetItemCorrelation(toItem, item);
return(rescorer.Rescore(pair, originalEstimate));
}
public double Estimate(User user)
{
Pair <User, User> pair = new Pair <User, User>(toUser, user);
if (rescorer.IsFiltered(pair))
{
return(Double.NaN);
}
double originalEstimate = correlation.GetUserCorrelation(toUser, user);
return(rescorer.Rescore(pair, originalEstimate));
}
/**
* {@inheritDoc}
*/
public override IList <RecommendedItem> Recommend(Object userID, int howMany, Rescorer <Item> rescorer)
{
if (userID == null || rescorer == null)
{
throw new ArgumentNullException("userID or rescorer is null");
}
if (howMany < 1)
{
throw new ArgumentException("howMany must be at least 1");
}
CheckClustersBuilt();
if (log.IsDebugEnabled)
{
log.Debug("Recommending items for user ID '" + userID + '\'');
}
IList <RecommendedItem> recommended;
topRecsByUserID.TryGetValue(userID, out recommended);
if (recommended == null)
{
recommended = new List <RecommendedItem>();
return(recommended);
}
User theUser = this.DataModel.GetUser(userID);
List <RecommendedItem> rescored = new List <RecommendedItem>(recommended.Count);
// Only add items the user doesn't already have a preference for.
// And that the rescorer doesn't "reject".
foreach (RecommendedItem recommendedItem in recommended)
{
Item item = recommendedItem.Item;
if (rescorer.IsFiltered(item))
{
continue;
}
if (theUser.GetPreferenceFor(item.ID) == null &&
!double.IsNaN(rescorer.Rescore(item, recommendedItem.Value)))
{
rescored.Add(recommendedItem);
}
}
rescored.Sort(new ByRescoreComparator(rescorer));
return(rescored);
}
public double Estimate(Item item)
{
RunningAverage average = new FullRunningAverage();
foreach (Item toItem in toItems)
{
Pair <Item, Item> pair = new Pair <Item, Item>(toItem, item);
if (rescorer.IsFiltered(pair))
{
continue;
}
double estimate = correlation.GetItemCorrelation(toItem, item);
estimate = rescorer.Rescore(pair, estimate);
average.AddDatum(estimate);
}
return(average.Average);
}
public static IList <RecommendedItem> GetTopItems(int howMany,
IEnumerable <Item> allItems,
Rescorer <Item> rescorer,
Estimator <Item> estimator)
{
if (allItems == null || rescorer == null || estimator == null)
{
throw new ArgumentNullException("argument is null");
}
LinkedList <RecommendedItem> topItems = new LinkedList <RecommendedItem>();
bool full = false;
foreach (Item item in allItems)
{
if (item.IsRecommendable && !rescorer.IsFiltered(item))
{
double preference = estimator.Estimate(item);
double rescoredPref = rescorer.Rescore(item, preference);
LinkedListNode <RecommendedItem> node = topItems.Last;
if (!Double.IsNaN(rescoredPref) &&
(!full || rescoredPref > node.Value.Value))
{
// I think this is faster than Collections.binarySearch() over a LinkedList since our
// comparisons are cheap, which binarySearch() economizes at the expense of more traversals.
// We also know that the right position tends to be at the end of the list.
while (node != null && node.Previous != null)
{
node = node.Previous;
if (rescoredPref <= node.Value.Value)
{
node = node.Next;
break;
}
if (node == topItems.First)
{
break;
}
}
RecommendedItem newItem = new GenericRecommendedItem(item, rescoredPref);
if (node == null)
{
topItems.AddFirst(newItem);
}
else if (topItems.Count == 1)
{
// special handling in this case is to avoid problems
// with negative preferences. Imagine -0.3 being added
// first followed by -0.6. If we simply did AddAfter,
// those items would be out of sequence - cc
if (rescoredPref > node.Value.Value)
{
topItems.AddAfter(node, newItem);
}
else
{
topItems.AddBefore(node, newItem);
}
}
else
{
topItems.AddAfter(node, newItem);
}
if (full)
{
topItems.RemoveLast();
}
else if (topItems.Count > howMany)
{
full = true;
topItems.RemoveLast();
}
}
}
}
List <RecommendedItem> result = new List <RecommendedItem>(topItems.Count);
foreach (RecommendedItem item in topItems)
{
result.Add(item);
}
return(result);
}
public static List <User> GetTopUsers(int howMany,
IEnumerable <User> allUsers,
Rescorer <User> rescorer,
Estimator <User> estimator)
{
LinkedList <SimilarUser> topUsers = new LinkedList <SimilarUser>();
bool full = false;
foreach (User user in allUsers)
{
if (rescorer.IsFiltered(user))
{
continue;
}
double similarity = estimator.Estimate(user);
double rescoredSimilarity = rescorer.Rescore(user, similarity);
LinkedListNode <SimilarUser> node = topUsers.Last;
if (!double.IsNaN(rescoredSimilarity) &&
(!full || rescoredSimilarity > node.Value.Similarity))
{
//SimilarUser _user = new SimilarUser(user, similarity);
SimilarUser _user = new SimilarUser(user, rescoredSimilarity);
if (node == null)
{
topUsers.AddLast(_user);
}
else if (node.Previous == null) // 1 node
{
if (rescoredSimilarity > node.Value.Similarity)
{
topUsers.AddAfter(node, _user);
}
else
{
topUsers.AddBefore(node, _user);
}
}
else
{
while (node != null && node.Previous != null && (node != topUsers.First))
{
node = node.Previous;
if (rescoredSimilarity <= node.Value.Similarity)
{
topUsers.AddBefore(node, _user);
break;
}
}
}
if (full)
{
topUsers.RemoveLast();
}
else if (topUsers.Count > howMany)
{
full = true;
topUsers.RemoveLast();
}
}
}
List <User> result = new List <User>(topUsers.Count);
foreach (SimilarUser similarUser in topUsers)
{
result.Add(similarUser.User);
}
return(result);
}