ezSASRec: The easiest way to utilize SASRec (Self-Attentive Sequential Recommendation) for your system
Kim Beomsoo. @SNU.
This repository contains the easiest and most efficient tool to build sequential recommendation system utilizing SASRec.
- Self-Attentive Sequential Recommendation
- A Case Study on Sampling Strategies for Evaluating Neural Sequential Item Recommendation Models
example data source: link
import pandas as pd
import pickle
from sasrec.util import filter_k_core, SASRecDataSet, load_model
from sasrec.model import SASREC
from sasrec.sampler import WarpSampler
import multiprocessingpath = 'your path'df = pd.read_csv('ratings.csv')
df = df.rename({'userId':'userID','movieId':'itemID','timestamp':'time'},axis=1)\
.sort_values(by=['userID','time'])\
.drop(['rating','time'],axis=1)\
.reset_index(drop=True)df.head()| userID | itemID | |
|---|---|---|
| 0 | 1 | 2762 |
| 1 | 1 | 54503 |
| 2 | 1 | 112552 |
| 3 | 1 | 96821 |
| 4 | 1 | 5577 |
# filter data
# every user and item will appear more than 6 times in filtered_df
filtered_df = filter_k_core(df, 7)Original: 270896 users and 45115 items
Final: 243377 users and 24068 items
# make maps (encoder)
user_set, item_set = set(filtered_df['userID'].unique()), set(filtered_df['itemID'].unique())
user_map = dict()
item_map = dict()
for u, user in enumerate(user_set):
user_map[user] = u+1
for i, item in enumerate(item_set):
item_map[item] = i+1
maps = (user_map, item_map) # Encode filtered_df
filtered_df["userID"] = filtered_df["userID"].apply(lambda x: user_map[x])
filtered_df["itemID"] = filtered_df["itemID"].apply(lambda x: item_map[x])# save data and maps
# save sasrec data
filtered_df.to_csv('sasrec_data.txt', sep="\t", header=False, index=False)
# save maps
with open('maps.pkl','wb') as f:
pickle.dump(maps, f)# load data
data = SASRecDataSet('sasrec_data.txt')
data.split() # train, val, test split
# the last interactions of each user is used for test
# the last but one will be used for validation
# others will be used for train# make model and warmsampler for batch training
max_len = 100
hidden_units = 128
batch_size = 2048
model = SASREC(
item_num=data.itemnum,
seq_max_len=max_len,
num_blocks=2,
embedding_dim=hidden_units,
attention_dim=hidden_units,
attention_num_heads=2,
dropout_rate=0.2,
conv_dims = [hidden_units, hidden_units],
l2_reg=0.00001
)
sampler = WarpSampler(data.user_train, data.usernum, data.itemnum, batch_size=batch_size, maxlen=max_len, n_workers=multiprocessing.cpu_count())# train model
model.train(
data,
sampler,
num_epochs=3,
batch_size=batch_size,
lr=0.001,
val_epoch=1,
val_target_user_n=1000,
target_item_n=-1,
auto_save=True,
path = path,
exp_name='exp_example',
)epoch 1 / 3 -----------------------------
Evaluating...
epoch: 1, test (NDCG@10: 0.04607630127474612, HR@10: 0.097)
best score model updated and saved
epoch 2 / 3 -----------------------------
Evaluating...
epoch: 2, test (NDCG@10: 0.060855185638025944, HR@10: 0.118)
best score model updated and saved
epoch 3 / 3 -----------------------------
Evaluating...
epoch: 3, test (NDCG@10: 0.07027207563856912, HR@10: 0.139)
best score model updated and saved
# load trained model
model = load_model(path,'exp_example')# get user-item score
# make inv_user_map
inv_user_map = {v: k for k, v in user_map.items()}
# sample target_user
model.sample_val_users(data, 100)
encoded_users = model.val_users
# get scores
score = model.get_user_item_score(data,
[inv_user_map[u] for u in encoded_users], # user_list containing raw(not-encoded) userID
[1,2,3], # item_list containing raw(not-encoded) itemID
user_map,
item_map,
batch_size=10
)100%|██████████| 10/10 [00:00<00:00, 29.67batch/s]
score.head()| userID | 1 | 2 | 3 | |
|---|---|---|---|---|
| 0 | 1525 | 5.596944 | 4.241653 | 3.804743 |
| 1 | 1756 | 4.535607 | 2.694459 | 0.858440 |
| 2 | 2408 | 5.883061 | 4.655960 | 4.691791 |
| 3 | 2462 | 5.084695 | 2.942075 | 2.773376 |
| 4 | 3341 | 5.532438 | 4.348150 | 4.073740 |
# get top N recommendation
reco = model.recommend_item(data,
user_map,
[inv_user_map[u] for u in encoded_users],
is_test=True,
top_n=5)100%|██████████| 100/100 [00:04<00:00, 21.10it/s]
# returned tuple contains topN recommendations for each user
reco{1525: [(456, 6.0680223),
(355, 6.033769),
(379, 5.9833336),
(591, 5.9718275),
(776, 5.8978705)],
1756: [(7088, 5.735977),
(15544, 5.5946136),
(5904, 5.500249),
(355, 5.492655),
(22149, 5.4117346)],
2408: [(456, 5.976555),
(328, 5.8824606),
(588, 5.8614006),
(264, 5.7114534),
(299, 5.649914)],
2462: [(259, 6.3445344),
(591, 6.2664876),
(295, 6.105361),
(355, 6.0698805),
(1201, 5.8477645)],
3341: [(110, 5.510764),
(1, 5.4927354),
(259, 5.4851904),
(161, 5.467624),
(208, 5.2486935)], ...}