Undergraduate Graduation Design Project - AGCN

Multi-task Algorithm for item recommendation task(IR) and attribute inference task(AI)

Adaptive Graph Convolutional Networks(AGCN)

AGCN raw github: https://github.com/yimutianyang/AGCN

Install

pip install torch==1.8.0+cu102 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Usage

每个任务执行时都会分配一个唯一ID,如 20210420215755 程序运行过程中，结果会保存到 temp文件夹，程序正常结束时，结果会从temp文件夹移动到output文件夹，因此output文件夹中总会保存成功执行的结果.

Task	description	Target
AGCN	AGCN model for item recommendation and attribute inference	item recommendation and attribute inference
BPRMF	BPRMF model for item recommendation	item recommendation
FM	FM model for item recommendation	item recommendation
NGCF	NGCF model for item recommendation	item recommendation
LP	Label Propagation for attribute inference	attribute inference
Semi-GCN	Semi-GCN model for attribute inference	attribute inference

1. AGCN

Item Recommendation Task

Train

python -u main.py --task IR-Train  --dataset_name movielens1M

Test

python -u main.py --task IR-Test  --dataset_name movielens1M --train_id 20210420215755

Attribute Inference Task

python -u main.py --task AI --dataset_name movielens1M

2. BPRMF

python -u main.py --task BPRMF --dataset_name movielens1M

3. FM

python -u main.py --task FM --dataset_name movielens1M

3. NGCF

python -u main.py --task NGCF --dataset_name movielens1M

4. LP

python -u main.py --task LP --dataset_name movielens1M

5. Semi-GCN

python -u main.py --task Semi-GCN --dataset_name movielens1M

Parameters Settings

/src/conf/datasets/文件夹下保存了不同数据集的配置文件。以amazonVideoGames为例

name: amazonVideoGames
user_count: 31027
item_count: 33899

# attribute info
user_attr:
  have: false
item_attr:
  have: true
  attr_type_num: 3
  attr_dim_list: [ 14, 52, 10 ]
  attr_type_list: [ 1, 1, 0 ] # 0 means single-label attribute and 1 means multi-label attributes

# AGCN
AI:
  epoch_num: 300 # epoch times
  free_emb_dim: 32 # free embedding dim
  learning_rate: 0.0005 # learning rate
  batch_size: 5120
  gamma1: 1
  gamma2: 0.001
  lambda1: 0.001
  lambda2: 0.001
  attr_union_dim: 32
  gcn_layer: 3
  neg_item_num: 5

IR-Train:
  epoch_num: 300 # epoch times
  iter_num: 10 # iter times
  free_emb_dim: 32 # free embedding dim
  learning_rate: 0.0005 # learning rate
  batch_size: 5120
  gamma1: 1
  gamma2: 0.001
  lambda1: 0.001
  lambda2: 0.001
  attr_union_dim: 32
  gcn_layer: 3
  neg_item_num: 5

IR-Test:
  free_emb_dim: 32 # free embedding dim
  gamma: 0.001
  lambda1: 0.001
  lambda2: 0.001
  attr_union_dim: 32
  gcn_layer: 3
  test_topks: [ 5,10,15,20,25,30,35,40,45,50 ]

# baselines for attribute inference
LP:
  epoch_num: 50
  loss_threshold: 0.01
  select_count: 1000
  knn: 20

Semi-GCN:
  epoch_num: 50
  gcn_layer: 3
  attr_union_dim: 32
  layer_dim_list: [32, 32, 32, 32]
  learning_rate: 0.05

# baselines for item recommendation
NGCF:
  epoch_num: 100 # epoch times
  emb_size: 32 # free embedding dim
  learning_rate: 0.0005 # learning rate
  batch_size: 2048
  gcn_layer_num: 3
  layers: [ 32, 32, 32, 32]
  decay: 0.0001
  node_dropout: 0.1
  mess_dropout: [ 0.1, 0.1, 0.1 ]
  neg_item_num: 5
  test_topks: [ 5,10,15,20,25,30,35,40,45,50 ]
  stop_epoch: 40

BPRMF:
  epoch_num: 250 # epoch times
  learning_rate: 0.0005 # learning rate
  free_emb_dim: 32
  batch_size: 2048
  decay: 0.0001
  test_topks: [ 5,10,15,20,25,30,35,40,45,50 ]
  neg_item_num: 5
  stop_epoch: 40

FM:
  epoch_num: 300 # epoch times
  free_emb_dim: 32 # free embedding dim
  learning_rate: 0.0005 # learning rate
  batch_size: 5120
  lambda1: 0.001
  lambda2: 0.001
  attr_union_dim: 32 # attr_dim
  neg_item_num: 5
  test_topks: [ 5,10,15,20,25,30,35,40,45,50 ]
  stop_epoch: 25

Data format

Take movilens1Mdataset for example.

$U$: user set

$V$: item set

$uid$: user_id

$iid$: item_id

# attribute info
user_attr:
  have: false
item_attr:
  have: true
  attr_type_num: 3
  attr_dim_list: [ 14, 52, 10 ]
  attr_type_list: [ 1, 1, 0 ]

上述配置表明其没有user属性，有三种item属性，每个属性的标签数分别为14、52、10，前两个为多标签属性，后两个为单标签属性。

filename	format	type	description
total_U2I.npy	{uid: [iid, iid], ...}	dict<list>	total feedbacks
train_U2I.npy	{uid: [iid, iid], ...}	dict<list>	train feedbacks
val_U2I.npy	{uid: [iid, iid], ...}	dict<list>	val feedbacks
test_U2I.npy	{uid: [iid, iid], ...}	dict<list>	test feedbacks
complete_user_attr.npy	------	np.ndarray	complete user attributes matrix
missing_user_attr.npy	------	np.ndarray	missing user attributes matrix
existing_user_attr_index.npy	------	------	store user attribute infomation
complete_item_attr.npy	\|U\| * sum(item_attr_dim_list)	np.ndarray	complete item attributes matrix
missing_item_attr.npy	\|U\| * sum(item_attr_dim_list)	np.ndarray	missing item attributes matrix
existing_item_attr_index.npy	{'attr_dim_list': attr_dim_list, 'existing_index_list':[list1, list2, list3]}	dict	store item attribute infomation
user_attr_LP.npy	------	np.ndarray	inferred LP user attribute result for FM model
item_attr_LP.npy	------	np.ndarray	inferred LP item attribute result for FM model

kervias / AGCN