Welcome to DaisyRec-v2.0’s Documentation!

The description of all parameters is listed below.

Basic Settings

–problem_type

define a point-wise or pair-wise problem.

  • point-wise: point-wise algorithm

  • pair-wise: pair-wise algorithm

–optimization_metric

the metric to be optimized for hyper-parameter tuning via HyperOpt

  • ndcg

  • precision

  • recall

  • hr

  • map

  • mrr

–hyperopt_trail

the number of trails of HyperOpt

–hyperopt_pack

record the searching space of hyper-parameters for HyperOpt

–algo_name

the algorithm to be executed

  • mostpop

  • itemknn

  • puresvd

  • slim

  • mf

  • fm

  • neumf

  • nfm

  • ngcf

  • multi-vae

–dataset

the dataset to be evaluated

  • ml-100k

  • ml-1m

  • ml-10m

  • ml-20m

  • lastfm

  • book-x

  • amazon-cloth

  • amazon-electronic

  • amazon-book

  • amazon-music

  • epinions

  • yelp

  • citeulike

  • netflix

–prepro

the data pre-processing strategy

  • origin: adopt the raw data

  • Fcore: recursively filter users and items that have interactions no less than N, e.g., 5core

  • Ffilter: only filter users and items that have interactions no less than N once, e.g., 5filter

–val_method

training and validation data splitting strategy

  • tsbr: time-aware split-by-ratio

  • rsbr: random-aware split-by-ratio

  • tloo: time-aware leave-one-out

  • rloo: random-aware leave-one-out

–test_method

training and test data splitting strategy, which should be consistent with the settings for val_method

–val_size

ratio of validation set size in the range of (0,1), e.g., 0.1 means retaining 10% of training data as validation data

–test_size

ratio of test set size in the range of (0,1), e.g., 0.2 means retaining 20% of the whole data as test data

–topk

the length of recommendation list

–fold_num

the fold number of cross-validation

–cand_num

the number of candidate items used for ranking

–sample_method

negative sampling strategy

  • uniform: uniformly sample negative items

  • low-pop: sample popular items with low rank

  • high-pop: sample popular items with high rank

–sample_ratio

control the ratio of popularity sampling for the hybrid sampling strategy in the range of (0,1), e.g., for the hybrid sampling strategy uniform+low-pop, –sample_ratio=0.1 means 10% of the negative items are sampled via low-pop

–num_ng

the number of negative samples

–positive_threshold

the threshold for binarizing the ratings into positve samples (for exmaple if the threshold = 4, it means the items with ratings no less than 4 will be treated as positive items)

–loss_type

type of loss function

  • CL: cross-entropy loss for point-wise problem

  • SL: square error loss for point-wise problem

  • BPR: BPR loss for pair-wise problem

  • HL: hinge loss for pair-wise problem

  • TL: top-1 Loss for pair-wise problem

–gpu

the ID of GPU card

Algorithm Specific Settings

–init_method

parameter initializers

  • default: initialize parameters according to the original paper

  • normal: initialize parameters with normal distribution

  • uniform: initialize parameters with uniform distribution

  • xavier_normal: initialize parameters with xavier_normal distribution

  • xavier_uniform: initialize parameters with xavier_uniform distribution

–optimizer

optimization method for training the algorithms

  • default (optimizer in the original paper)

  • sgd

  • adam

  • adagrad

–early_stop

whether to activate the early-stop mechanism

  • true

  • false

–tune_testset

whether to directly tune on testset, and the default value is false

  • true

  • false

–factors

the dimension of latent factors (embeddings)

–reg_1

the coefficient of L1 regularization

–reg_2

the coefficient of L2 regularization

–dropout

dropout rate

–lr

learning rate

–epochs

training epochs

–batch_size

batch size for training

–num_layers

number of layers for MLP

–alpha

constant to multiply the penalty terms for SLIM

–elastic

the ElasticNet mixing parameter for SLIM in the range of (0,1)

–pop_n

the preliminary selected top-n popular candidate items to reduce the time complexity for MostPop

–maxk

the number of neighbors to take into account for ItemKNN

–node_dropout

node dropout ratio for NGCF

–mess_dropout

message dropout ratio for NGCF

–kl_reg

the coefficient of KL regularization for Multi-VAE