/// TODO: this is the vanilla sgd by Tacaks 2009, I speculate that using scaling technique proposed in:
/// Towards Optimal One Pass Large Scale Learning with Averaged Stochastic Gradient Descent section 5, page 6
/// can be beneficial in term s of both speed and accuracy.
///
/// Tacaks' method doesn't calculate gradient of regularization correctly, which has non-zero elements everywhere of
/// the matrix. While Tacaks' method can only updates a single row/column, if one user has a lot of recommendation,
/// her vector will be more affected by regularization using an isolated scaling factor for both user vectors and
/// item vectors can remove this issue without inducing more update cost it even reduces it a bit by only performing
/// one addition and one multiplication.
///
/// BAD SIDE1: the scaling factor decreases fast, it has to be scaled up from time to time before dropped to zero or
/// caused roundoff error
/// BAD SIDE2: no body experiment on it before, and people generally use very small lambda
/// so it's impact on accuracy may still be unknown.
/// BAD SIDE3: don't know how to make it work for L1-regularization or
/// "pseudorank?" (sum of singular values)-regularization
protected void update(IPreference preference, double mu)
{
int userIdx = userIndex(preference.GetUserID());
int itemIdx = itemIndex(preference.GetItemID());
double[] userVector = userVectors[userIdx];
double[] itemVector = itemVectors[itemIdx];
double prediction = dot(userVector, itemVector);
double err = preference.GetValue() - prediction;
// adjust features
for (int k = FEATURE_OFFSET; k < rank; k++)
{
double userFeature = userVector[k];
double itemFeature = itemVector[k];
userVector[k] += mu * (err * itemFeature - lambda * userFeature);
itemVector[k] += mu * (err * userFeature - lambda * itemFeature);
}
// adjust user and item bias
userVector[USER_BIAS_INDEX] += biasMuRatio * mu * (err - biasLambdaRatio * lambda * userVector[USER_BIAS_INDEX]);
itemVector[ITEM_BIAS_INDEX] += biasMuRatio * mu * (err - biasLambdaRatio * lambda * itemVector[ITEM_BIAS_INDEX]);
}
public void testPreferenceShufflerWithSyntheticData()
{
setUpSyntheticData();
ParallelSGDFactorizer.PreferenceShuffler shuffler = new ParallelSGDFactorizer.PreferenceShuffler(dataModel);
shuffler.shuffle();
shuffler.stage();
FastByIDMap <FastByIDMap <bool?> > checkedLst = new FastByIDMap <FastByIDMap <bool?> >();
for (int i = 0; i < shuffler.size(); i++)
{
IPreference pref = shuffler.get(i);
float?value = dataModel.GetPreferenceValue(pref.GetUserID(), pref.GetItemID());
Assert.AreEqual(pref.GetValue(), value.Value, 0.0);
if (!checkedLst.ContainsKey(pref.GetUserID()))
{
checkedLst.Put(pref.GetUserID(), new FastByIDMap <bool?>());
}
Assert.IsNull(checkedLst.Get(pref.GetUserID()).Get(pref.GetItemID()));
checkedLst.Get(pref.GetUserID()).Put(pref.GetItemID(), true);
}
var userIDs = dataModel.GetUserIDs();
int index = 0;
while (userIDs.MoveNext())
{
long userID = userIDs.Current;
IPreferenceArray preferencesFromUser = dataModel.GetPreferencesFromUser(userID);
foreach (IPreference preference in preferencesFromUser)
{
Assert.True(checkedLst.Get(preference.GetUserID()).Get(preference.GetItemID()).Value);
index++;
}
}
Assert.AreEqual(index, shuffler.size());
}
public GenericItemPreferenceArray(IList <IPreference> prefs) : this(prefs.Count)
{
int size = prefs.Count;
long itemID = Int64.MinValue;
for (int i = 0; i < size; i++)
{
IPreference pref = prefs[i];
ids[i] = pref.GetUserID();
if (i == 0)
{
itemID = pref.GetItemID();
}
else
{
if (itemID != pref.GetItemID())
{
throw new ArgumentException("Not all item IDs are the same");
}
}
values[i] = pref.GetValue();
}
id = itemID;
}
protected override void processOneEstimate(float estimatedPreference, IPreference realPref)
{
double diff = realPref.GetValue() - estimatedPreference;
average.AddDatum(diff * diff);
}
public void Set(int i, IPreference pref) {
id = pref.GetItemID();
ids[i] = pref.GetUserID();
values[i] = pref.GetValue();
}
public void Set(int i, IPreference pref)
{
id = pref.GetItemID();
ids[i] = pref.GetUserID();
values[i] = pref.GetValue();
}
/// TODO: this is the vanilla sgd by Tacaks 2009, I speculate that using scaling technique proposed in:
/// Towards Optimal One Pass Large Scale Learning with Averaged Stochastic Gradient Descent section 5, page 6
/// can be beneficial in term s of both speed and accuracy.
///
/// Tacaks' method doesn't calculate gradient of regularization correctly, which has non-zero elements everywhere of
/// the matrix. While Tacaks' method can only updates a single row/column, if one user has a lot of recommendation,
/// her vector will be more affected by regularization using an isolated scaling factor for both user vectors and
/// item vectors can remove this issue without inducing more update cost it even reduces it a bit by only performing
/// one addition and one multiplication.
///
/// BAD SIDE1: the scaling factor decreases fast, it has to be scaled up from time to time before dropped to zero or
/// caused roundoff error
/// BAD SIDE2: no body experiment on it before, and people generally use very small lambda
/// so it's impact on accuracy may still be unknown.
/// BAD SIDE3: don't know how to make it work for L1-regularization or
/// "pseudorank?" (sum of singular values)-regularization
protected void update(IPreference preference, double mu) {
int userIdx = userIndex(preference.GetUserID());
int itemIdx = itemIndex(preference.GetItemID());
double[] userVector = userVectors[userIdx];
double[] itemVector = itemVectors[itemIdx];
double prediction = dot(userVector, itemVector);
double err = preference.GetValue() - prediction;
// adjust features
for (int k = FEATURE_OFFSET; k < rank; k++) {
double userFeature = userVector[k];
double itemFeature = itemVector[k];
userVector[k] += mu * (err * itemFeature - lambda * userFeature);
itemVector[k] += mu * (err * userFeature - lambda * itemFeature);
}
// adjust user and item bias
userVector[USER_BIAS_INDEX] += biasMuRatio * mu * (err - biasLambdaRatio * lambda * userVector[USER_BIAS_INDEX]);
itemVector[ITEM_BIAS_INDEX] += biasMuRatio * mu * (err - biasLambdaRatio * lambda * itemVector[ITEM_BIAS_INDEX]);
}
protected override void processOneEstimate(float estimatedPreference, IPreference realPref) {
average.AddDatum(Math.Abs(realPref.GetValue() - estimatedPreference));
}
protected override void processOneEstimate(float estimatedPreference, IPreference realPref)
{
average.AddDatum(Math.Abs(realPref.GetValue() - estimatedPreference));
}
protected override void processOneEstimate(float estimatedPreference, IPreference realPref) {
double diff = realPref.GetValue() - estimatedPreference;
average.AddDatum(diff * diff);
}