Source domain dataset: Douban-Books
Target domain dataset: Douban-Music
Evaluation: all users in target dataset, ratio-based 8:1:1, full sort
Metrics: Recall, Precision, NDCG, MRR, Hit
Topk: 10, 20, 50
Properties:
field_separator: "\t"
source_domain:
dataset: DoubanBook
USER_ID_FIELD: user_id
ITEM_ID_FIELD: item_id
RATING_FIELD: rating
TIME_FIELD: timestamp
NEG_PREFIX: neg_
LABEL_FIELD: label
load_col:
inter: [user_id, item_id, rating]
user_inter_num_interval: "[5,inf)"
item_inter_num_interval: "[5,inf)"
val_interval:
rating: "[3,inf)"
drop_filter_field: True
target_domain:
dataset: DoubanMovie
USER_ID_FIELD: user_id
ITEM_ID_FIELD: item_id
RATING_FIELD: rating
TIME_FIELD: timestamp
NEG_PREFIX: neg_
LABEL_FIELD: label
load_col:
inter: [user_id, item_id, rating]
user_inter_num_interval: "[5,inf)"
item_inter_num_interval: "[5,inf)"
val_interval:
rating: "[3,inf)"
drop_filter_field: True
epochs: 500
train_batch_size: 4096
eval_batch_size: 409600
valid_metric: NDCG@10
For fairness, we restrict users' and items' embedding dimension as following. Please adjust the name of the corresponding args of different models.
embedding_size: 64| Dataset | #Users | #items | #Interactions | Sparsity |
|---|---|---|---|---|
| Douban-Book | 18085 | 33067 | 809248 | 99.86% |
| Douban-Movie | 22041 | 25802 | 2552305 | 99.55% |
Number of Overlapped User: 15434
Number of Overlapped Item: 0
| Method | Best hyper-parameters |
|---|---|
| CoNet | learning_rate=0.005 mlp_hidden_size=[64,32,16,8] reg_weight=0.01 |
| CLFM | learning_rate=0.0005 share_embedding_size=48 alpha=0.1 reg_weight=0.0001 |
| DTCDR | learning_rate=0.0005 mlp_hidden_size=[64,64] dropout_prob=0.2 alpha=0.1 base_model=NeuMF |
| DeepAPF | learning_rate=0.0005 |
| BiTGCF | learning_rate=0.0005 n_layers=2 concat_way=mean lambda_source=0.8 lambda_target=0.8 drop_rate=0.1 reg_weight=0.01 |
| CMF | learning_rate=0.0005 lambda=0.9 gamma=0.1 alpha=0.1 |
| EMCDR | learning_rate=0.001 mapping_function=non_linear mlp_hidden_size=[64] overlap_batch_size=100 reg_weight=0.01 latent_factor_model=BPR loss_type=BPR |
| NATR | learning_rate=0.001 max_inter_length=100 reg_weight=1e-5 |
| SSCDR | learning_rate=0.0005 lambda=0 margin=0.2 overlap_batch_size=1024 |
| DCDCSR | learning_rate=0.0005 mlp_hidden_size=[128] k=10 |