MyMediaLite
3.10
|
Matrix factorization with explicit user and item bias, learning is performed by stochastic gradient descent More...
Public Member Functions | |
override void | AddRatings (IRatings ratings) |
Add new ratings and perform incremental training | |
BiasedMatrixFactorization () | |
Default constructor | |
virtual bool | CanPredict (int user_id, int item_id) |
Check whether a useful prediction (i.e. not using a fallback/default answer) can be made for a given user-item combination | |
Object | Clone () |
create a shallow copy of the object | |
override float | ComputeObjective () |
Compute the regularized loss | |
override void | Iterate () |
Run one iteration (= pass over the training data) | |
override void | LoadModel (string filename) |
Get the model parameters from a file | |
override float | Predict (int user_id, int item_id) |
Predict the rating of a given user for a given item | |
IList< Tuple< int, float > > | Recommend (int user_id, int n=-1, ICollection< int > ignore_items=null, ICollection< int > candidate_items=null) |
Recommend items for a given user | |
virtual System.Collections.Generic.IList < Tuple< int, float > > | Recommend (int user_id, int n=-1, System.Collections.Generic.ICollection< int > ignore_items=null, System.Collections.Generic.ICollection< int > candidate_items=null) |
override void | RemoveItem (int item_id) |
Remove all feedback by one item | |
override void | RemoveRatings (IDataSet ratings) |
Remove existing ratings and perform "incremental" training | |
override void | RemoveUser (int user_id) |
Remove all feedback by one user | |
override void | RetrainItem (int item_id) |
Updates the latent factors of an item | |
override void | RetrainUser (int user_id) |
Updates the latent factors on a user | |
override void | SaveModel (string filename) |
Save the model parameters to a file | |
IList< Tuple< int, float > > | ScoreItems (IList< Tuple< int, float >> rated_items, IList< int > candidate_items) |
Rate a list of items given a list of ratings that represent a new user | |
override string | ToString () |
Return a string representation of the recommender | |
override void | Train () |
Learn the model parameters of the recommender from the training data | |
override void | UpdateRatings (IRatings ratings) |
Update existing ratings and perform incremental training |
Protected Member Functions | |
override void | AddItem (int item_id) |
override void | AddUser (int user_id) |
double | ComputeLoss () |
Computes the value of the loss function that is currently being optimized | |
override float[] | FoldIn (IList< Tuple< int, float >> rated_items) |
Compute parameters (latent factors) for a user represented by ratings | |
override void | Iterate (IList< int > rating_indices, bool update_user, bool update_item) |
Iterate once over rating data and adjust corresponding factors (stochastic gradient descent) | |
float | Predict (int user_id, int item_id, bool bound) |
override float | Predict (float[] user_vector, int item_id) |
Predict rating for a fold-in user and an item | |
void | SetupLoss () |
Set up the common part of the error gradient of the loss function to optimize | |
override void | UpdateLearnRate () |
Updates current_learnrate after each epoch |
Protected Attributes | |
Func< double, double, float > | compute_gradient_common |
delegate to compute the common term of the error gradient | |
const int | FOLD_IN_BIAS_INDEX = 0 |
Index of the bias term in the user vector representation for fold-in | |
const int | FOLD_IN_FACTORS_START = 1 |
Start index of the user factors in the user vector representation for fold-in | |
float | global_bias |
The bias (global average) | |
double | last_loss = double.NegativeInfinity |
Loss for the last iteration, used by bold driver heuristics | |
float | max_rating |
Maximum rating value | |
float | min_rating |
Minimum rating value | |
float | rating_range_size |
size of the interval of valid ratings | |
IRatings | ratings |
rating data |
Properties | |
float | BiasLearnRate [get, set] |
Learn rate factor for the bias terms | |
float | BiasReg [get, set] |
regularization factor for the bias terms | |
bool | BoldDriver [get, set] |
Use bold driver heuristics for learning rate adaption | |
float | Decay [get, set] |
Multiplicative learn rate decay | |
bool | FrequencyRegularization [get, set] |
Regularization based on rating frequency | |
double | InitMean [get, set] |
Mean of the normal distribution used to initialize the factors | |
double | InitStdDev [get, set] |
Standard deviation of the normal distribution used to initialize the factors | |
float | LearnRate [get, set] |
Learn rate (update step size) | |
OptimizationTarget | Loss [get, set] |
The optimization target | |
int | MaxItemID [get, set] |
Maximum item ID | |
virtual float | MaxRating [get, set] |
Maximum rating value | |
int | MaxThreads [get, set] |
the maximum number of threads to use | |
int | MaxUserID [get, set] |
Maximum user ID | |
virtual float | MinRating [get, set] |
Minimum rating value | |
bool | NaiveParallelization [get, set] |
Use 'naive' parallelization strategy instead of conflict-free 'distributed' SGD | |
uint | NumFactors [get, set] |
Number of latent factors | |
uint | NumIter [get, set] |
Number of iterations over the training data | |
virtual IRatings | Ratings [get, set] |
The rating data | |
float | RegI [get, set] |
regularization constant for the item factors | |
float | RegU [get, set] |
regularization constant for the user factors | |
override float | Regularization [set] |
bool | UpdateItems [get, set] |
bool | UpdateUsers [get, set] |
Matrix factorization with explicit user and item bias, learning is performed by stochastic gradient descent
Per default optimizes for RMSE. Alternatively, you can set the Loss property to MAE or LogisticLoss. If set to log likelihood and with binary ratings, the recommender implements a simple version Menon and Elkan's LFL model, which predicts binary labels, has no advanced regularization, and uses no side information.
This recommender makes use of multi-core machines if requested. Just set MaxThreads to a large enough number (usually multiples of the number of available cores). The parallelization is based on ideas presented in the paper by Gemulla et al.
Literature:
This recommender supports incremental updates. See the paper by Rendle and Schmidt-Thieme.
|
inline |
Default constructor
|
inlinevirtualinherited |
Add new ratings and perform incremental training
ratings | the ratings |
Reimplemented from IncrementalRatingPredictor.
|
inlinevirtualinherited |
Check whether a useful prediction (i.e. not using a fallback/default answer) can be made for a given user-item combination
It is up to the recommender implementor to decide when a prediction is useful, and to document it accordingly.
user_id | the user ID |
item_id | the item ID |
Implements IRecommender.
Reimplemented in ExternalItemRecommender, ExternalRatingPredictor, BiPolarSlopeOne, SlopeOne, Constant, GlobalAverage, UserAverage, ItemAverage, and Random.
|
inlineinherited |
create a shallow copy of the object
|
inlineprotected |
Computes the value of the loss function that is currently being optimized
|
inlinevirtual |
Compute the regularized loss
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, SigmoidUserAsymmetricFactorModel, and SocialMF.
|
inlineprotectedvirtual |
Compute parameters (latent factors) for a user represented by ratings
rated_items | a list of (item ID, rating value) pairs |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidUserAsymmetricFactorModel, and SigmoidItemAsymmetricFactorModel.
|
inlinevirtual |
Run one iteration (= pass over the training data)
Reimplemented from MatrixFactorization.
|
inlineprotectedvirtual |
Iterate once over rating data and adjust corresponding factors (stochastic gradient descent)
rating_indices | a list of indices pointing to the ratings to iterate over |
update_user | true if user factors to be updated |
update_item | true if item factors to be updated |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, SigmoidUserAsymmetricFactorModel, and SocialMF.
|
inline |
Get the model parameters from a file
filename | the name of the file to read from |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, and SigmoidUserAsymmetricFactorModel.
|
inline |
Predict the rating of a given user for a given item
If the user or the item are not known to the recommender, the global average is returned. To avoid this behavior for unknown entities, use CanPredict() to check before.
user_id | the user ID |
item_id | the item ID |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, and SigmoidUserAsymmetricFactorModel.
|
inlineprotectedvirtual |
Predict rating for a fold-in user and an item
user_vector | a float vector representing the user |
item_id | the item ID |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel.
|
inherited |
Recommend items for a given user
user_id | the user ID |
n | the number of items to recommend, -1 for as many as possible |
ignore_items | collection if items that should not be returned; if null, use empty collection |
candidate_items | the candidate items to choose from; if null, use all items |
Implemented in WeightedEnsemble, and Ensemble.
|
inlinevirtual |
Remove all feedback by one item
item_id | the item ID |
Reimplemented from MatrixFactorization.
|
inlinevirtualinherited |
Remove existing ratings and perform "incremental" training
ratings | the user and item IDs of the ratings to be removed |
Reimplemented from IncrementalRatingPredictor.
|
inlinevirtual |
Remove all feedback by one user
user_id | the user ID |
Reimplemented from MatrixFactorization.
|
inlinevirtual |
Updates the latent factors of an item
item_id | the item ID |
Reimplemented from MatrixFactorization.
|
inlinevirtual |
Updates the latent factors on a user
user_id | the user ID |
Reimplemented from MatrixFactorization.
|
inline |
Save the model parameters to a file
filename | the name of the file to write to |
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, and SigmoidUserAsymmetricFactorModel.
|
inlineinherited |
Rate a list of items given a list of ratings that represent a new user
rated_items | the ratings (item IDs and rating values) representing the new user |
candidate_items | the items to be rated |
Implements IFoldInRatingPredictor.
|
inlineprotected |
Set up the common part of the error gradient of the loss function to optimize
|
inline |
Return a string representation of the recommender
The ToString() method of recommenders should list the class name and all hyperparameters, separated by space characters.
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, SigmoidUserAsymmetricFactorModel, and SocialMF.
|
inline |
Learn the model parameters of the recommender from the training data
Reimplemented from MatrixFactorization.
Reimplemented in SigmoidCombinedAsymmetricFactorModel, SigmoidItemAsymmetricFactorModel, and SigmoidUserAsymmetricFactorModel.
|
inlineprotectedvirtual |
Updates current_learnrate after each epoch
Reimplemented from MatrixFactorization.
|
inlinevirtualinherited |
Update existing ratings and perform incremental training
ratings | the ratings |
Reimplemented from IncrementalRatingPredictor.
|
protected |
delegate to compute the common term of the error gradient
|
protected |
Index of the bias term in the user vector representation for fold-in
|
protected |
Start index of the user factors in the user vector representation for fold-in
|
protectedinherited |
The bias (global average)
|
protected |
Loss for the last iteration, used by bold driver heuristics
|
protectedinherited |
Maximum rating value
|
protectedinherited |
Minimum rating value
|
protected |
size of the interval of valid ratings
|
protectedinherited |
rating data
|
getset |
Learn rate factor for the bias terms
|
getset |
regularization factor for the bias terms
|
getset |
Use bold driver heuristics for learning rate adaption
Literature:
|
getsetinherited |
Multiplicative learn rate decay
Applied after each epoch (= pass over the whole dataset)
|
getset |
Regularization based on rating frequency
Regularization proportional to the inverse of the square root of the number of ratings associated with the user or item. As described in the paper by Menon and Elkan.
|
getsetinherited |
Mean of the normal distribution used to initialize the factors
|
getsetinherited |
Standard deviation of the normal distribution used to initialize the factors
|
getsetinherited |
Learn rate (update step size)
|
getset |
The optimization target
|
getsetinherited |
Maximum item ID
|
getsetinherited |
Maximum rating value
|
getset |
the maximum number of threads to use
For parallel learning, set this number to a multiple of the number of available cores/CPUs
|
getsetinherited |
Maximum user ID
|
getsetinherited |
Minimum rating value
|
getset |
Use 'naive' parallelization strategy instead of conflict-free 'distributed' SGD
The exact sequence of updates depends on the thread scheduling. If you want reproducible results, e.g. when setting –random-seed=N, do NOT set this property.
|
getsetinherited |
Number of latent factors
|
getsetinherited |
Number of iterations over the training data
|
getset |
regularization constant for the item factors
|
getset |
regularization constant for the user factors