In this project, we will take a look at a dataset of client reviews of restaurants in Hong Kong. It contains the the reviewers’ rating to restaurants (referred to as stores in the data). Subsequently, we will build a recommendation system using matrix factorization method. In particular, we will training a model with PyTorch based on the existing reviewer’s rating to each store and break down into reviewer embeddings and store embeddings, containing latent factors for each reviewer/store. Hopefully, these embeddings are able to capture the hidden pattern/features of the reviewer/store, which can be used to predict the reviewer’s rating towards a restuarant that currently has no rating from the same reviewer.
Source: https://www.kaggle.com/datasets/bwandowando/hongkong-food-panda-restaurant-reviews
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import torch
import torch.nn as nn
import torch.nn.functional as F
# Load data
path = r'.\Data\hk_hong_kong_reviews.csv'
reviews_df = pd.read_csv(path, header = 0)
reviews_df.head(20)
| StoreId | uuid | createdAt | updatedAt | text | isAnonymous | reviewerId | replies | likeCount | isLiked | overall | restaurant_food | rider | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | z7km | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | 2023-05-28T09:09:59Z | 2023-05-28T09:09:59Z | 味道較咸, 而且經常比漏野 | False | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | [] | 1 | False | 2 | 2 | NaN |
| 1 | v2nn | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | 2023-04-30T05:02:30Z | 2023-04-30T05:02:30Z | 出品味道實在不錯，非常欣賞大廚的烹飪\n不過樓面服務禮貌實在太差了，問多一句都不耐煩，應該生... | False | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | [] | 0 | False | 1 | 1 | NaN |
| 2 | v9tr | 00026cb8-fffc-4feb-8753-1c786f71691c | 2023-07-09T13:44:31Z | 2023-07-09T13:44:31Z | Super quality and authentic. | False | hk2udxt3 | [] | 0 | False | 5 | 5 | NaN |
| 3 | h2qz | 0002b618-188d-4598-a4a3-cdc4f6557f35 | 2024-02-21T04:51:06Z | 2024-02-21T04:51:06Z | 好味！！！ | False | hkbqx1ei | [] | 0 | False | 5 | 5 | 5.0 |
| 4 | u3q3 | 0004a723-4885-4d3e-b942-b5d3f3d02326 | 2023-12-25T20:16:25Z | 2023-12-25T20:16:25Z | 腸粉好硬，不過煎餃好食 | False | hkkofhli | [] | 0 | False | 3 | 3 | NaN |
| 5 | v4vf | 00053eea-ed62-4e9d-9537-0dfa721f23b9 | 2023-06-28T05:45:13Z | 2023-06-28T05:45:13Z | 牛肉很多筋， 大部分都不能吃 吐了出來， 越來越退步 | False | hk6x2kmp | [] | 0 | False | 2 | 2 | NaN |
| 6 | h6ag | 00070043-36ca-4087-903c-dd49cbb6fd65 | 2024-01-25T01:47:06Z | 2024-01-25T01:47:06Z | 粥好綿，好味，不過啲料麻麻哋。皮蛋瘦肉粥啲皮蛋加埋唔知有冇半隻，瘦肉太鹹。 | False | hkgwy1kr | [] | 0 | False | 4 | 4 | NaN |
| 7 | w8ie | 0008130b-f90d-4ee9-912e-efbfc70bd082 | 2023-08-19T06:06:03Z | 2023-08-19T06:06:03Z | 好食 | False | hk0huoxn | [] | 0 | False | 5 | 5 | 5.0 |
| 8 | q301 | 00089793-4fb7-4535-992d-b6964ef7a680 | 2023-08-30T04:56:30Z | 2023-08-30T04:56:30Z | 真的不錯👍 | False | g0ssf7lj | [] | 0 | False | 4 | 4 | 4.0 |
| 9 | a8or | 000da2da-18db-4df0-ae5a-017609993c61 | 2023-09-21T03:59:45Z | 2023-09-21T03:59:45Z | 上次芫茜湯飯足料，今次粥好食，炸両腸脆卜卜又熱辣辣，滿意 | False | hkgqh8ew | [] | 0 | False | 5 | 5 | NaN |
| 10 | tznf | 000dab15-18ce-4d98-893f-4ec821c7afdb | 2023-08-27T14:36:26Z | 2023-08-27T14:36:26Z | yummy | False | hk81xahj | [] | 0 | False | 5 | 5 | 5.0 |
| 11 | j8pt | 000ff36f-0d50-40c6-8eb7-3ae80c7f5d8a | 2023-04-05T14:12:30Z | 2023-04-05T14:12:30Z | 明明叫咗牛扒意粉，最後變咗飯，sad | False | hkrvh5ec | [] | 0 | False | 1 | 1 | 2.0 |
| 12 | xw6y | 00109778-bc12-4fb4-b0c9-201b757a5b45 | 2023-05-15T13:35:08Z | 2023-05-15T13:35:08Z | 茄子太油，和部食物的之然味太濃，否則整體不錯 | False | hkzw04fx | [] | 0 | False | 4 | 4 | 5.0 |
| 13 | v9tj | 0013eb6f-111d-4aad-91e6-5b9bc4769e97 | 2023-05-04T12:05:53Z | 2023-05-04T12:05:53Z | 份量少 | False | hkh6799n | [] | 0 | False | 3 | 3 | 2.0 |
| 14 | r567 | 001461c5-a72d-4e70-acf1-571c07233a78 | 2023-09-17T05:39:57Z | 2023-09-17T05:39:57Z | 漏單\n打單個個仲要很無禮，好似係我錯唔關佢事咁 | False | 001461c5-a72d-4e70-acf1-571c07233a78 | [] | 0 | False | 1 | 1 | NaN |
| 15 | v3mh | 0015fb07-5339-45ca-bad2-38f030f706d8 | 2024-01-11T06:50:14Z | 2024-01-11T06:50:14Z | 包裝好, 份量夠 | False | hktze1el | [] | 0 | False | 5 | 5 | NaN |
| 16 | tbrh | 00166372-0968-4f40-b244-1649f412b667 | 2023-06-13T10:25:37Z | 2023-06-13T10:25:37Z | 我不會用40元買一盒中華沙律(大連菜）🤷🏻♀️ | False | hkjhwwe0 | [] | 1 | False | 2 | 2 | NaN |
| 17 | j8pt | 00170224-8337-42cd-aae5-c71cac584729 | 2023-09-18T12:43:04Z | 2023-09-18T12:43:04Z | 好好食，用心製作 | False | hkh9sx12 | [] | 0 | False | 5 | 5 | NaN |
| 18 | ddys | 0018c418-1cf2-4d83-bf26-47178ff64918 | 2023-05-18T14:57:02Z | 2023-05-18T14:57:02Z | 炒飯炒粉都不錯 | False | hk5qkim6 | [] | 0 | False | 4 | 4 | NaN |
| 19 | s5jy | 00194aff-9d0f-45ef-b012-c41aac008d79 | 2023-11-29T05:50:43Z | 2023-11-29T05:50:43Z | 蒜辣汁好食👍🏻 | False | x8qvv5jh | [] | 0 | False | 5 | 5 | NaN |
The csv parsing looks fine in Vs code. However, if such data is read in other platform (e.g., Databricks), some may find that the data is distorted by “\n” in the text column. Also, it is found that the text column containing line breaks are all quoted with double quototation marks (“<text>”), and there are also cases where double quoted text do not include a line break as well as cases where there are double double-quoted words (e.g., ““<word>””) within the text columns. These format may confuse the data parsing when converting into dataframe.
One option is to drop those rows using dropna(). However, in order to preserve as much data as possible, we can use regular expression to clean the data. In this case, the following code can be used to fix the above issues if encountered.
import re
from io import StringIO
with open(r'Data/hk_hong_kong_reviews.csv',"r") as txt:
t = txt.readlines()
text = ""
for line in t:
text += line
text = text.replace('""','')
pattern = '\"([^"]*|\n)*\"'
matches = re.finditer(pattern, text)
text_idx = []
for match in matches:
if text.find("\n",match.start(),match.end()) != -1:
text_idx.append((match.start(),match.end()))
last_idx = 0
cleaned_data_str = ""
for idx in text_idx:
cleaned_data_str+=text[last_idx:idx[0]]
revised_text = re.sub("\n"," ", text[idx[0]:idx[1]])
cleaned_data_str += revised_text
last_idx = idx[1]
cleaned_data_str += text[last_idx:len(text)]
with open(r'Data/hk_hong_kong_reviews(clean).csv',"w") as f_write:
f_write.write(cleaned_data_str)
data = pd.read_csv(StringIO(cleaned_data_str), header=0, quotechar='"')
Now, we can check whether the cleaned data still have the issues.
reviews_df = pd.read_csv(r'Data/hk_hong_kong_reviews(clean).csv', header=0, quotechar='"')
reviews_df.head(5)
| StoreId | uuid | createdAt | updatedAt | text | isAnonymous | reviewerId | replies | likeCount | isLiked | overall | restaurant_food | rider | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | z7km | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | 2023-05-28T09:09:59Z | 2023-05-28T09:09:59Z | 味道較咸, 而且經常比漏野 | False | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | [] | 1 | False | 2 | 2 | NaN |
| 1 | v2nn | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | 2023-04-30T05:02:30Z | 2023-04-30T05:02:30Z | 出品味道實在不錯，非常欣賞大廚的烹飪 不過樓面服務禮貌實在太差了，問多一句都不耐煩，應該生意... | False | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | [] | 0 | False | 1 | 1 | NaN |
| 2 | v9tr | 00026cb8-fffc-4feb-8753-1c786f71691c | 2023-07-09T13:44:31Z | 2023-07-09T13:44:31Z | Super quality and authentic. | False | hk2udxt3 | [] | 0 | False | 5 | 5 | NaN |
| 3 | h2qz | 0002b618-188d-4598-a4a3-cdc4f6557f35 | 2024-02-21T04:51:06Z | 2024-02-21T04:51:06Z | 好味！！！ | False | hkbqx1ei | [] | 0 | False | 5 | 5 | 5.0 |
| 4 | u3q3 | 0004a723-4885-4d3e-b942-b5d3f3d02326 | 2023-12-25T20:16:25Z | 2023-12-25T20:16:25Z | 腸粉好硬，不過煎餃好食 | False | hkkofhli | [] | 0 | False | 3 | 3 | NaN |
reviews_df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 50548 entries, 0 to 50547
Data columns (total 13 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 StoreId 50548 non-null object
1 uuid 50548 non-null object
2 createdAt 50548 non-null object
3 updatedAt 50548 non-null object
4 text 50544 non-null object
5 isAnonymous 50548 non-null bool
6 reviewerId 50548 non-null object
7 replies 50548 non-null object
8 likeCount 50548 non-null int64
9 isLiked 50548 non-null bool
10 overall 50548 non-null int64
11 restaurant_food 50548 non-null int64
12 rider 17179 non-null float64
dtypes: bool(2), float64(1), int64(3), object(7)
memory usage: 4.3+ MB
# Check for duplication
duplicated_count = reviews_df.duplicated().sum()
# Check for missing rating
missing_rating_count = len(reviews_df[reviews_df['overall'].isna()])
print(f"Number of duplicated records: {duplicated_count}")
print(f"Number of records with missing rating: {missing_rating_count}")
Number of duplicated records: 0
Number of records with missing rating: 0
Now, we will split the data into training and test set. For time-series data, we will normally split the data based on the time that the records are created. In this analysis, we will take the first 80% of data as training set and remaining 20% as test set.
# Sort by time
reviews_df = reviews_df.sort_values(by='createdAt', ascending=True)
# Check date range
min_date = reviews_df['createdAt'].min()
max_date = reviews_df['createdAt'].max()
print(f"Date ranges from {min_date} to {max_date}")
Date ranges from 2023-03-27T13:41:35Z to 2024-03-27T17:50:11Z
# Select needed columns and split data into training and test set
data = reviews_df[['reviewerId', 'StoreId', 'overall']]
data.rename(columns={'StoreId':'storeId'}, inplace=True)
print(data.columns)
# Drop duplicates, blank rating
data = data.drop_duplicates().dropna()
# Split data into train and test sets
train_size = int(len(data)*0.8)
train_data, test_data = data[0:train_size], data[train_size:]
len(train_data), len(test_data)
Index(['reviewerId', 'storeId', 'overall'], dtype='object')
C:\Users\Louis\AppData\Local\Temp\ipykernel_20616\859120231.py:3: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
data.rename(columns={'StoreId':'storeId'}, inplace=True)
(37004, 9252)
For the purpose of training the model in PyTorch, we need to encode the reviewerId and StoreId with continous Id. The encoded id will serve as the index in the reviewer/store embeddings. We can create a dictionary so that we can convert the encoded id back to original id afterward.
# Encoding reviewerId with continous Id
train_reviewerId = np.sort(np.unique(train_data.reviewerId.values))
num_reviewers = len(train_reviewerId)
reviewerid2idx = {o:i for i,o in enumerate(train_reviewerId)}
train_data['reviewerId'] = train_data['reviewerId'].map(reviewerid2idx)
test_data['reviewerId'] = test_data['reviewerId'].map(lambda x: reviewerid2idx.get(x,-1)) # -1 for users not in training
test_data = test_data[test_data['reviewerId'] >= 0].copy()
# Encoding StoreId with continous Id
train_storeId = np.sort(np.unique(train_data.storeId.values))
num_stores = len(train_storeId)
storeid2idx = {o:i for i,o in enumerate(train_storeId)}
train_data['storeId'] = train_data['storeId'].map(storeid2idx)
test_data['storeId'] = test_data['storeId'].map(lambda x: storeid2idx.get(x,-1)) # -1 for users not in training
test_data = test_data[test_data['storeId'] >= 0].copy()
print(f"Reviewer ID: {train_data['reviewerId'].unique()}")
print(f"Store ID: {train_data['storeId'].unique()}")
Reviewer ID: [ 520 26247 23528 ... 8712 5207 18748]
Store ID: [1002 4133 4657 ... 4459 1080 3226]
C:\Users\Louis\AppData\Local\Temp\ipykernel_20616\1362070608.py:5: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
train_data['reviewerId'] = train_data['reviewerId'].map(reviewerid2idx)
C:\Users\Louis\AppData\Local\Temp\ipykernel_20616\1362070608.py:6: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
test_data['reviewerId'] = test_data['reviewerId'].map(lambda x: reviewerid2idx.get(x,-1)) # -1 for users not in training
C:\Users\Louis\AppData\Local\Temp\ipykernel_20616\1362070608.py:13: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
train_data['storeId'] = train_data['storeId'].map(storeid2idx)
# Reset index
train_data.reset_index(inplace=True, drop=True)
test_data.reset_index(inplace=True, drop=True)
# Setup device-agnostic code
device = "cuda" if torch.cuda.is_available() else "cpu"
device
'cpu'
In this analysis, we will build 2 versions of model, one without bias term and one with bias term.
First, we will define the model archietecture.
class RecSys(nn.Module):
def __init__(self, num_reviewers, num_stores, emb_size=100):
super().__init__()
self.reviewers_emb = nn.Embedding(num_reviewers, emb_size)
self.stores_emb = nn.Embedding(num_stores, emb_size)
# reinitializing weights with a smaller value as it is found that this model works better if start with smaller values
self.reviewers_emb.weight.data.uniform_(0,0.05)
self.stores_emb.weight.data.uniform_(0,0.05)
def forward(self, reviewer_id, store_id):
u = self.reviewers_emb(reviewer_id) # This tells which indices of the reviewer embedding to extract
v = self.stores_emb(store_id) # This tells which indices of the store embedding to extract
return (u*v).sum(dim=1) # this is the dot product (first get element-wise multiplication then sum it)
# here we are not using data loaders because our data fits well in memory
def train_recsys(model, train:pd.DataFrame, test:pd.DataFrame, epochs=10, lr=0.01, weight_decay=0.0, device='cpu'): # Weight decay is effectively the lambda in regularization
# Training loop
for epoch in range(epochs):
model.train()
reviewers = torch.LongTensor(train.reviewerId.values).to(device)
stores = torch.LongTensor(train.storeId.values).to(device)
ratings = torch.FloatTensor(train.overall.values).to(device)
y_pred = model(reviewers, stores)
optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay)
train_loss = F.l1_loss(y_pred, ratings) # We are using L1 loss (Mean absolute errors) as this will show errors in the same unit as the data which is easier for us to conceptualize how well our model performs.
optimizer.zero_grad()
train_loss.backward()
optimizer.step()
if (epoch+1) % 10 == 0:
print(f"Epoch {epoch+1} | train loss {train_loss.item():.3f}")
# Test loop
model.eval()
with torch.inference_mode():
reviewers = torch.LongTensor(train.reviewerId.values).to(device)
stores = torch.LongTensor(train.storeId.values).to(device)
ratings = torch.FloatTensor(train.overall.values).to(device)
test_pred = model(reviewers, stores)
test_loss = F.l1_loss(test_pred, ratings)
print(f"test loss {test_loss.item():.3f}")
return train_loss, test_loss
# Initialize the model
num_reviewers = len(train_data['reviewerId'].unique())
num_stores = len(train_data['storeId'].unique())
recsys_model = RecSys(num_reviewers, num_stores, emb_size=100)
train_recsys(recsys_model, train_data, test_data, epochs=100, lr=0.01, weight_decay=1e-5)
Epoch 10 | train loss 2.022
Epoch 20 | train loss 0.518
Epoch 30 | train loss 0.402
Epoch 40 | train loss 0.395
Epoch 50 | train loss 0.392
Epoch 60 | train loss 0.390
Epoch 70 | train loss 0.387
Epoch 80 | train loss 0.385
Epoch 90 | train loss 0.383
Epoch 100 | train loss 0.381
test loss 0.361
(tensor(0.3808, grad_fn=<MeanBackward0>), tensor(0.3612))
class RecSys_bias(nn.Module):
def __init__(self, num_users, num_movies, emb_size=100):
super().__init__()
self.reviewers_emb = nn.Embedding(num_reviewers, emb_size)
self.stores_emb = nn.Embedding(num_stores, emb_size)
self.reviewers_bias = nn.Embedding(num_reviewers, 1)
self.stores_bias = nn.Embedding(num_stores, 1)
# re-initializing weights with a smaller value as it is found that this model works better if start with smaller values
self.reviewers_emb.weight.data.uniform_(0,0.05)
self.stores_emb.weight.data.uniform_(0,0.05)
self.reviewers_bias.weight.data.uniform_(-0.01,0.01)
self.stores_bias.weight.data.uniform_(-0.01,0.01)
def forward(self, user_id, movie_id):
u = self.reviewers_emb(user_id)
v = self.stores_emb(movie_id)
b_u = self.reviewers_bias(user_id).squeeze()
b_v = self.stores_bias(movie_id).squeeze()
return (u*v).sum(dim=1) + b_u +b_v # this is the dot product (first get element-wise multiplication then sum it)
# Initialize the model
num_reviewers = len(train_data['reviewerId'].unique())
num_stores = len(train_data['storeId'].unique())
recsys_bias_model = RecSys_bias(num_reviewers, num_stores, emb_size=100)
train_recsys(recsys_bias_model, train_data, test_data, epochs=100, lr=0.01, weight_decay=1e-5)
Epoch 10 | train loss 1.904
Epoch 20 | train loss 0.411
Epoch 30 | train loss 0.374
Epoch 40 | train loss 0.369
Epoch 50 | train loss 0.366
Epoch 60 | train loss 0.364
Epoch 70 | train loss 0.362
Epoch 80 | train loss 0.360
Epoch 90 | train loss 0.359
Epoch 100 | train loss 0.357
test loss 0.379
(tensor(0.3575, grad_fn=<MeanBackward0>), tensor(0.3792))
It appears that the performance of model with bias is similar to without bias. Let’s go for the model with bias. Now, let’s vary the embedding size to see identify the optimal size of model. However, the larger the embedding size, more epochs are required to converge, so let’s try 1000 epochs this time.
emb_sizes = [10, 50, 100, 150, 200]
result = pd.DataFrame(columns=['emb_size', 'last_train_loss', 'test_loss'])
for emb_size in emb_sizes:
recsys_bias_model = RecSys_bias(num_reviewers, num_stores, emb_size=emb_size)
last_train_loss, test_loss = train_recsys(recsys_bias_model, train_data, test_data, epochs=1000, lr=0.01, weight_decay=1e-5, device=device)
model_result = pd.DataFrame({'emb_size': [emb_size], 'last_train_loss': [last_train_loss.detach().numpy()], 'test_loss': [test_loss.detach().numpy()]})
result = pd.concat([result, model_result], axis=0)
result
Epoch 10 | train loss 2.883
Epoch 20 | train loss 2.357
Epoch 30 | train loss 1.830
Epoch 40 | train loss 1.278
Epoch 50 | train loss 0.762
Epoch 60 | train loss 0.362
Epoch 70 | train loss 0.318
Epoch 80 | train loss 0.306
Epoch 90 | train loss 0.302
Epoch 100 | train loss 0.300
Epoch 110 | train loss 0.298
Epoch 120 | train loss 0.297
Epoch 130 | train loss 0.296
Epoch 140 | train loss 0.296
Epoch 150 | train loss 0.295
Epoch 160 | train loss 0.295
Epoch 170 | train loss 0.294
Epoch 180 | train loss 0.294
Epoch 190 | train loss 0.293
Epoch 200 | train loss 0.293
Epoch 210 | train loss 0.293
Epoch 220 | train loss 0.292
Epoch 230 | train loss 0.292
Epoch 240 | train loss 0.291
Epoch 250 | train loss 0.291
Epoch 260 | train loss 0.291
Epoch 270 | train loss 0.291
Epoch 280 | train loss 0.290
Epoch 290 | train loss 0.290
Epoch 300 | train loss 0.290
Epoch 310 | train loss 0.290
Epoch 320 | train loss 0.290
Epoch 330 | train loss 0.290
Epoch 340 | train loss 0.290
Epoch 350 | train loss 0.290
Epoch 360 | train loss 0.290
Epoch 370 | train loss 0.290
Epoch 380 | train loss 0.290
Epoch 390 | train loss 0.290
Epoch 400 | train loss 0.289
Epoch 410 | train loss 0.289
Epoch 420 | train loss 0.289
Epoch 430 | train loss 0.289
Epoch 440 | train loss 0.289
Epoch 450 | train loss 0.289
Epoch 460 | train loss 0.289
Epoch 470 | train loss 0.289
Epoch 480 | train loss 0.289
Epoch 490 | train loss 0.289
Epoch 500 | train loss 0.289
Epoch 510 | train loss 0.289
Epoch 520 | train loss 0.289
Epoch 530 | train loss 0.289
Epoch 540 | train loss 0.289
Epoch 550 | train loss 0.289
Epoch 560 | train loss 0.289
Epoch 570 | train loss 0.289
Epoch 580 | train loss 0.289
Epoch 590 | train loss 0.289
Epoch 600 | train loss 0.289
Epoch 610 | train loss 0.289
Epoch 620 | train loss 0.289
Epoch 630 | train loss 0.289
Epoch 640 | train loss 0.289
Epoch 650 | train loss 0.289
Epoch 660 | train loss 0.289
Epoch 670 | train loss 0.288
Epoch 680 | train loss 0.288
Epoch 690 | train loss 0.288
Epoch 700 | train loss 0.288
Epoch 710 | train loss 0.288
Epoch 720 | train loss 0.288
Epoch 730 | train loss 0.288
Epoch 740 | train loss 0.288
Epoch 750 | train loss 0.288
Epoch 760 | train loss 0.288
Epoch 770 | train loss 0.288
Epoch 780 | train loss 0.288
Epoch 790 | train loss 0.288
Epoch 800 | train loss 0.288
Epoch 810 | train loss 0.288
Epoch 820 | train loss 0.288
Epoch 830 | train loss 0.288
Epoch 840 | train loss 0.288
Epoch 850 | train loss 0.288
Epoch 860 | train loss 0.288
Epoch 870 | train loss 0.288
Epoch 880 | train loss 0.288
Epoch 890 | train loss 0.288
Epoch 900 | train loss 0.288
Epoch 910 | train loss 0.288
Epoch 920 | train loss 0.288
Epoch 930 | train loss 0.287
Epoch 940 | train loss 0.287
Epoch 950 | train loss 0.287
Epoch 960 | train loss 0.287
Epoch 970 | train loss 0.287
Epoch 980 | train loss 0.287
Epoch 990 | train loss 0.287
Epoch 1000 | train loss 0.287
test loss 0.283
Epoch 10 | train loss 2.367
Epoch 20 | train loss 1.197
Epoch 30 | train loss 0.384
Epoch 40 | train loss 0.352
Epoch 50 | train loss 0.347
Epoch 60 | train loss 0.345
Epoch 70 | train loss 0.344
Epoch 80 | train loss 0.342
Epoch 90 | train loss 0.341
Epoch 100 | train loss 0.340
Epoch 110 | train loss 0.339
Epoch 120 | train loss 0.338
Epoch 130 | train loss 0.337
Epoch 140 | train loss 0.337
Epoch 150 | train loss 0.336
Epoch 160 | train loss 0.335
Epoch 170 | train loss 0.335
Epoch 180 | train loss 0.335
Epoch 190 | train loss 0.334
Epoch 200 | train loss 0.334
Epoch 210 | train loss 0.333
Epoch 220 | train loss 0.333
Epoch 230 | train loss 0.333
Epoch 240 | train loss 0.332
Epoch 250 | train loss 0.332
Epoch 260 | train loss 0.332
Epoch 270 | train loss 0.332
Epoch 280 | train loss 0.331
Epoch 290 | train loss 0.331
Epoch 300 | train loss 0.331
Epoch 310 | train loss 0.330
Epoch 320 | train loss 0.330
Epoch 330 | train loss 0.330
Epoch 340 | train loss 0.330
Epoch 350 | train loss 0.329
Epoch 360 | train loss 0.329
Epoch 370 | train loss 0.329
Epoch 380 | train loss 0.329
Epoch 390 | train loss 0.328
Epoch 400 | train loss 0.328
Epoch 410 | train loss 0.328
Epoch 420 | train loss 0.328
Epoch 430 | train loss 0.328
Epoch 440 | train loss 0.328
Epoch 450 | train loss 0.328
Epoch 460 | train loss 0.328
Epoch 470 | train loss 0.327
Epoch 480 | train loss 0.327
Epoch 490 | train loss 0.327
Epoch 500 | train loss 0.327
Epoch 510 | train loss 0.327
Epoch 520 | train loss 0.327
Epoch 530 | train loss 0.326
Epoch 540 | train loss 0.326
Epoch 550 | train loss 0.326
Epoch 560 | train loss 0.326
Epoch 570 | train loss 0.326
Epoch 580 | train loss 0.326
Epoch 590 | train loss 0.326
Epoch 600 | train loss 0.326
Epoch 610 | train loss 0.326
Epoch 620 | train loss 0.326
Epoch 630 | train loss 0.326
Epoch 640 | train loss 0.325
Epoch 650 | train loss 0.325
Epoch 660 | train loss 0.325
Epoch 670 | train loss 0.325
Epoch 680 | train loss 0.325
Epoch 690 | train loss 0.325
Epoch 700 | train loss 0.324
Epoch 710 | train loss 0.324
Epoch 720 | train loss 0.324
Epoch 730 | train loss 0.324
Epoch 740 | train loss 0.324
Epoch 750 | train loss 0.324
Epoch 760 | train loss 0.324
Epoch 770 | train loss 0.323
Epoch 780 | train loss 0.323
Epoch 790 | train loss 0.323
Epoch 800 | train loss 0.323
Epoch 810 | train loss 0.323
Epoch 820 | train loss 0.322
Epoch 830 | train loss 0.322
Epoch 840 | train loss 0.322
Epoch 850 | train loss 0.322
Epoch 860 | train loss 0.322
Epoch 870 | train loss 0.321
Epoch 880 | train loss 0.321
Epoch 890 | train loss 0.321
Epoch 900 | train loss 0.321
Epoch 910 | train loss 0.321
Epoch 920 | train loss 0.321
Epoch 930 | train loss 0.321
Epoch 940 | train loss 0.320
Epoch 950 | train loss 0.320
Epoch 960 | train loss 0.320
Epoch 970 | train loss 0.320
Epoch 980 | train loss 0.319
Epoch 990 | train loss 0.319
Epoch 1000 | train loss 0.319
test loss 0.312
Epoch 10 | train loss 1.905
Epoch 20 | train loss 0.413
Epoch 30 | train loss 0.373
Epoch 40 | train loss 0.368
Epoch 50 | train loss 0.365
Epoch 60 | train loss 0.363
Epoch 70 | train loss 0.360
Epoch 80 | train loss 0.359
Epoch 90 | train loss 0.358
Epoch 100 | train loss 0.356
Epoch 110 | train loss 0.355
Epoch 120 | train loss 0.355
Epoch 130 | train loss 0.354
Epoch 140 | train loss 0.353
Epoch 150 | train loss 0.352
Epoch 160 | train loss 0.351
Epoch 170 | train loss 0.350
Epoch 180 | train loss 0.349
Epoch 190 | train loss 0.349
Epoch 200 | train loss 0.348
Epoch 210 | train loss 0.347
Epoch 220 | train loss 0.346
Epoch 230 | train loss 0.346
Epoch 240 | train loss 0.345
Epoch 250 | train loss 0.345
Epoch 260 | train loss 0.345
Epoch 270 | train loss 0.344
Epoch 280 | train loss 0.344
Epoch 290 | train loss 0.343
Epoch 300 | train loss 0.343
Epoch 310 | train loss 0.343
Epoch 320 | train loss 0.342
Epoch 330 | train loss 0.342
Epoch 340 | train loss 0.341
Epoch 350 | train loss 0.340
Epoch 360 | train loss 0.340
Epoch 370 | train loss 0.340
Epoch 380 | train loss 0.340
Epoch 390 | train loss 0.339
Epoch 400 | train loss 0.338
Epoch 410 | train loss 0.338
Epoch 420 | train loss 0.338
Epoch 430 | train loss 0.338
Epoch 440 | train loss 0.337
Epoch 450 | train loss 0.337
Epoch 460 | train loss 0.337
Epoch 470 | train loss 0.337
Epoch 480 | train loss 0.337
Epoch 490 | train loss 0.336
Epoch 500 | train loss 0.336
Epoch 510 | train loss 0.336
Epoch 520 | train loss 0.335
Epoch 530 | train loss 0.335
Epoch 540 | train loss 0.335
Epoch 550 | train loss 0.335
Epoch 560 | train loss 0.334
Epoch 570 | train loss 0.334
Epoch 580 | train loss 0.334
Epoch 590 | train loss 0.333
Epoch 600 | train loss 0.333
Epoch 610 | train loss 0.333
Epoch 620 | train loss 0.332
Epoch 630 | train loss 0.332
Epoch 640 | train loss 0.332
Epoch 650 | train loss 0.331
Epoch 660 | train loss 0.331
Epoch 670 | train loss 0.331
Epoch 680 | train loss 0.330
Epoch 690 | train loss 0.330
Epoch 700 | train loss 0.329
Epoch 710 | train loss 0.329
Epoch 720 | train loss 0.329
Epoch 730 | train loss 0.328
Epoch 740 | train loss 0.328
Epoch 750 | train loss 0.328
Epoch 760 | train loss 0.327
Epoch 770 | train loss 0.327
Epoch 780 | train loss 0.327
Epoch 790 | train loss 0.327
Epoch 800 | train loss 0.326
Epoch 810 | train loss 0.326
Epoch 820 | train loss 0.326
Epoch 830 | train loss 0.325
Epoch 840 | train loss 0.325
Epoch 850 | train loss 0.325
Epoch 860 | train loss 0.324
Epoch 870 | train loss 0.324
Epoch 880 | train loss 0.324
Epoch 890 | train loss 0.324
Epoch 900 | train loss 0.323
Epoch 910 | train loss 0.323
Epoch 920 | train loss 0.323
Epoch 930 | train loss 0.323
Epoch 940 | train loss 0.322
Epoch 950 | train loss 0.322
Epoch 960 | train loss 0.322
Epoch 970 | train loss 0.322
Epoch 980 | train loss 0.322
Epoch 990 | train loss 0.321
Epoch 1000 | train loss 0.321
test loss 0.307
Epoch 10 | train loss 1.488
Epoch 20 | train loss 0.416
Epoch 30 | train loss 0.401
Epoch 40 | train loss 0.397
Epoch 50 | train loss 0.394
Epoch 60 | train loss 0.392
Epoch 70 | train loss 0.390
Epoch 80 | train loss 0.389
Epoch 90 | train loss 0.388
Epoch 100 | train loss 0.386
Epoch 110 | train loss 0.385
Epoch 120 | train loss 0.384
Epoch 130 | train loss 0.383
Epoch 140 | train loss 0.383
Epoch 150 | train loss 0.382
Epoch 160 | train loss 0.381
Epoch 170 | train loss 0.380
Epoch 180 | train loss 0.379
Epoch 190 | train loss 0.379
Epoch 200 | train loss 0.378
Epoch 210 | train loss 0.378
Epoch 220 | train loss 0.377
Epoch 230 | train loss 0.377
Epoch 240 | train loss 0.377
Epoch 250 | train loss 0.376
Epoch 260 | train loss 0.376
Epoch 270 | train loss 0.376
Epoch 280 | train loss 0.376
Epoch 290 | train loss 0.375
Epoch 300 | train loss 0.375
Epoch 310 | train loss 0.375
Epoch 320 | train loss 0.375
Epoch 330 | train loss 0.374
Epoch 340 | train loss 0.374
Epoch 350 | train loss 0.374
Epoch 360 | train loss 0.374
Epoch 370 | train loss 0.373
Epoch 380 | train loss 0.373
Epoch 390 | train loss 0.373
Epoch 400 | train loss 0.373
Epoch 410 | train loss 0.372
Epoch 420 | train loss 0.372
Epoch 430 | train loss 0.371
Epoch 440 | train loss 0.371
Epoch 450 | train loss 0.371
Epoch 460 | train loss 0.371
Epoch 470 | train loss 0.370
Epoch 480 | train loss 0.370
Epoch 490 | train loss 0.369
Epoch 500 | train loss 0.369
Epoch 510 | train loss 0.369
Epoch 520 | train loss 0.368
Epoch 530 | train loss 0.368
Epoch 540 | train loss 0.368
Epoch 550 | train loss 0.367
Epoch 560 | train loss 0.367
Epoch 570 | train loss 0.366
Epoch 580 | train loss 0.366
Epoch 590 | train loss 0.366
Epoch 600 | train loss 0.365
Epoch 610 | train loss 0.365
Epoch 620 | train loss 0.364
Epoch 630 | train loss 0.364
Epoch 640 | train loss 0.363
Epoch 650 | train loss 0.363
Epoch 660 | train loss 0.362
Epoch 670 | train loss 0.361
Epoch 680 | train loss 0.361
Epoch 690 | train loss 0.360
Epoch 700 | train loss 0.360
Epoch 710 | train loss 0.359
Epoch 720 | train loss 0.359
Epoch 730 | train loss 0.357
Epoch 740 | train loss 0.357
Epoch 750 | train loss 0.356
Epoch 760 | train loss 0.356
Epoch 770 | train loss 0.355
Epoch 780 | train loss 0.355
Epoch 790 | train loss 0.354
Epoch 800 | train loss 0.353
Epoch 810 | train loss 0.352
Epoch 820 | train loss 0.351
Epoch 830 | train loss 0.350
Epoch 840 | train loss 0.350
Epoch 850 | train loss 0.349
Epoch 860 | train loss 0.349
Epoch 870 | train loss 0.348
Epoch 880 | train loss 0.347
Epoch 890 | train loss 0.347
Epoch 900 | train loss 0.346
Epoch 910 | train loss 0.345
Epoch 920 | train loss 0.344
Epoch 930 | train loss 0.344
Epoch 940 | train loss 0.343
Epoch 950 | train loss 0.343
Epoch 960 | train loss 0.342
Epoch 970 | train loss 0.341
Epoch 980 | train loss 0.340
Epoch 990 | train loss 0.339
Epoch 1000 | train loss 0.338
test loss 0.310
Epoch 10 | train loss 1.188
Epoch 20 | train loss 0.415
Epoch 30 | train loss 0.406
Epoch 40 | train loss 0.402
Epoch 50 | train loss 0.399
Epoch 60 | train loss 0.398
Epoch 70 | train loss 0.396
Epoch 80 | train loss 0.395
Epoch 90 | train loss 0.393
Epoch 100 | train loss 0.392
Epoch 110 | train loss 0.390
Epoch 120 | train loss 0.389
Epoch 130 | train loss 0.388
Epoch 140 | train loss 0.387
Epoch 150 | train loss 0.386
Epoch 160 | train loss 0.385
Epoch 170 | train loss 0.384
Epoch 180 | train loss 0.383
Epoch 190 | train loss 0.382
Epoch 200 | train loss 0.380
Epoch 210 | train loss 0.379
Epoch 220 | train loss 0.378
Epoch 230 | train loss 0.376
Epoch 240 | train loss 0.375
Epoch 250 | train loss 0.373
Epoch 260 | train loss 0.372
Epoch 270 | train loss 0.371
Epoch 280 | train loss 0.370
Epoch 290 | train loss 0.369
Epoch 300 | train loss 0.368
Epoch 310 | train loss 0.367
Epoch 320 | train loss 0.366
Epoch 330 | train loss 0.364
Epoch 340 | train loss 0.363
Epoch 350 | train loss 0.361
Epoch 360 | train loss 0.360
Epoch 370 | train loss 0.360
Epoch 380 | train loss 0.359
Epoch 390 | train loss 0.358
Epoch 400 | train loss 0.357
Epoch 410 | train loss 0.356
Epoch 420 | train loss 0.355
Epoch 430 | train loss 0.355
Epoch 440 | train loss 0.354
Epoch 450 | train loss 0.353
Epoch 460 | train loss 0.353
Epoch 470 | train loss 0.352
Epoch 480 | train loss 0.352
Epoch 490 | train loss 0.350
Epoch 500 | train loss 0.349
Epoch 510 | train loss 0.349
Epoch 520 | train loss 0.348
Epoch 530 | train loss 0.347
Epoch 540 | train loss 0.346
Epoch 550 | train loss 0.345
Epoch 560 | train loss 0.344
Epoch 570 | train loss 0.344
Epoch 580 | train loss 0.343
Epoch 590 | train loss 0.342
Epoch 600 | train loss 0.341
Epoch 610 | train loss 0.340
Epoch 620 | train loss 0.340
Epoch 630 | train loss 0.339
Epoch 640 | train loss 0.338
Epoch 650 | train loss 0.338
Epoch 660 | train loss 0.337
Epoch 670 | train loss 0.336
Epoch 680 | train loss 0.335
Epoch 690 | train loss 0.335
Epoch 700 | train loss 0.334
Epoch 710 | train loss 0.333
Epoch 720 | train loss 0.333
Epoch 730 | train loss 0.332
Epoch 740 | train loss 0.331
Epoch 750 | train loss 0.330
Epoch 760 | train loss 0.329
Epoch 770 | train loss 0.329
Epoch 780 | train loss 0.328
Epoch 790 | train loss 0.328
Epoch 800 | train loss 0.328
Epoch 810 | train loss 0.327
Epoch 820 | train loss 0.327
Epoch 830 | train loss 0.326
Epoch 840 | train loss 0.326
Epoch 850 | train loss 0.325
Epoch 860 | train loss 0.325
Epoch 870 | train loss 0.324
Epoch 880 | train loss 0.324
Epoch 890 | train loss 0.323
Epoch 900 | train loss 0.323
Epoch 910 | train loss 0.323
Epoch 920 | train loss 0.322
Epoch 930 | train loss 0.322
Epoch 940 | train loss 0.321
Epoch 950 | train loss 0.321
Epoch 960 | train loss 0.320
Epoch 970 | train loss 0.320
Epoch 980 | train loss 0.319
Epoch 990 | train loss 0.319
Epoch 1000 | train loss 0.318
test loss 0.309
| emb_size | last_train_loss | test_loss | |
|---|---|---|---|
| 0 | 10 | 0.2871752 | 0.28284174 |
| 0 | 50 | 0.31930673 | 0.31212807 |
| 0 | 100 | 0.32132322 | 0.30682832 |
| 0 | 150 | 0.33827838 | 0.30968025 |
| 0 | 200 | 0.31849787 | 0.30858234 |
It appears that smaller model performs better than larger model. Also, the test loss is similar to the train loss, suggesting that there may not be overfitting. In summary, our best performing model is on average having an absolute error of 0.28 out of a rating from 0-5. This represents an accuracy rate of 94.4%.
Now, let’s use all the data to train the model with emb_size=10 and make predictions on unseen reviewer-store pair.
# Final model
final_data = data.copy()
# Encoding reviewerId with continous Id
final_reviewerId = np.sort(np.unique(final_data.reviewerId.values))
num_reviewers = len(final_reviewerId)
reviewerid2idx = {o:i for i,o in enumerate(final_reviewerId)}
final_data['reviewerId'] = final_data['reviewerId'].map(reviewerid2idx)
# Encoding StoreId with continous Id
final_storeId = np.sort(np.unique(final_data.storeId.values))
num_stores = len(final_storeId)
storeid2idx = {o:i for i,o in enumerate(final_storeId)}
final_data['storeId'] = final_data['storeId'].map(storeid2idx)
final_model = RecSys_bias(num_reviewers, num_stores, emb_size=10)
train_recsys(final_model, final_data, final_data, epochs=1500, lr=0.01, weight_decay=1e-5,device=device)
Epoch 10 | train loss 2.890
Epoch 20 | train loss 2.365
Epoch 30 | train loss 1.837
Epoch 40 | train loss 1.287
Epoch 50 | train loss 0.775
Epoch 60 | train loss 0.383
Epoch 70 | train loss 0.341
Epoch 80 | train loss 0.331
Epoch 90 | train loss 0.328
Epoch 100 | train loss 0.327
Epoch 110 | train loss 0.325
Epoch 120 | train loss 0.324
Epoch 130 | train loss 0.323
Epoch 140 | train loss 0.322
Epoch 150 | train loss 0.322
Epoch 160 | train loss 0.321
Epoch 170 | train loss 0.320
Epoch 180 | train loss 0.320
Epoch 190 | train loss 0.319
Epoch 200 | train loss 0.319
Epoch 210 | train loss 0.318
Epoch 220 | train loss 0.318
Epoch 230 | train loss 0.318
Epoch 240 | train loss 0.318
Epoch 250 | train loss 0.317
Epoch 260 | train loss 0.317
Epoch 270 | train loss 0.317
Epoch 280 | train loss 0.317
Epoch 290 | train loss 0.317
Epoch 300 | train loss 0.317
Epoch 310 | train loss 0.317
Epoch 320 | train loss 0.317
Epoch 330 | train loss 0.317
Epoch 340 | train loss 0.317
Epoch 350 | train loss 0.316
Epoch 360 | train loss 0.316
Epoch 370 | train loss 0.316
Epoch 380 | train loss 0.316
Epoch 390 | train loss 0.316
Epoch 400 | train loss 0.316
Epoch 410 | train loss 0.316
Epoch 420 | train loss 0.316
Epoch 430 | train loss 0.316
Epoch 440 | train loss 0.316
Epoch 450 | train loss 0.316
Epoch 460 | train loss 0.316
Epoch 470 | train loss 0.316
Epoch 480 | train loss 0.316
Epoch 490 | train loss 0.316
Epoch 500 | train loss 0.316
Epoch 510 | train loss 0.315
Epoch 520 | train loss 0.315
Epoch 530 | train loss 0.315
Epoch 540 | train loss 0.315
Epoch 550 | train loss 0.315
Epoch 560 | train loss 0.315
Epoch 570 | train loss 0.315
Epoch 580 | train loss 0.315
Epoch 590 | train loss 0.315
Epoch 600 | train loss 0.315
Epoch 610 | train loss 0.315
Epoch 620 | train loss 0.315
Epoch 630 | train loss 0.315
Epoch 640 | train loss 0.315
Epoch 650 | train loss 0.315
Epoch 660 | train loss 0.314
Epoch 670 | train loss 0.314
Epoch 680 | train loss 0.314
Epoch 690 | train loss 0.314
Epoch 700 | train loss 0.314
Epoch 710 | train loss 0.314
Epoch 720 | train loss 0.314
Epoch 730 | train loss 0.314
Epoch 740 | train loss 0.314
Epoch 750 | train loss 0.314
Epoch 760 | train loss 0.314
Epoch 770 | train loss 0.314
Epoch 780 | train loss 0.314
Epoch 790 | train loss 0.314
Epoch 800 | train loss 0.314
Epoch 810 | train loss 0.314
Epoch 820 | train loss 0.314
Epoch 830 | train loss 0.314
Epoch 840 | train loss 0.314
Epoch 850 | train loss 0.314
Epoch 860 | train loss 0.314
Epoch 870 | train loss 0.314
Epoch 880 | train loss 0.314
Epoch 890 | train loss 0.314
Epoch 900 | train loss 0.314
Epoch 910 | train loss 0.313
Epoch 920 | train loss 0.313
Epoch 930 | train loss 0.313
Epoch 940 | train loss 0.313
Epoch 950 | train loss 0.313
Epoch 960 | train loss 0.313
Epoch 970 | train loss 0.313
Epoch 980 | train loss 0.313
Epoch 990 | train loss 0.313
Epoch 1000 | train loss 0.313
Epoch 1010 | train loss 0.313
Epoch 1020 | train loss 0.313
Epoch 1030 | train loss 0.313
Epoch 1040 | train loss 0.313
Epoch 1050 | train loss 0.313
Epoch 1060 | train loss 0.313
Epoch 1070 | train loss 0.313
Epoch 1080 | train loss 0.313
Epoch 1090 | train loss 0.313
Epoch 1100 | train loss 0.313
Epoch 1110 | train loss 0.313
Epoch 1120 | train loss 0.313
Epoch 1130 | train loss 0.313
Epoch 1140 | train loss 0.313
Epoch 1150 | train loss 0.312
Epoch 1160 | train loss 0.312
Epoch 1170 | train loss 0.312
Epoch 1180 | train loss 0.312
Epoch 1190 | train loss 0.312
Epoch 1200 | train loss 0.312
Epoch 1210 | train loss 0.312
Epoch 1220 | train loss 0.312
Epoch 1230 | train loss 0.312
Epoch 1240 | train loss 0.312
Epoch 1250 | train loss 0.312
Epoch 1260 | train loss 0.312
Epoch 1270 | train loss 0.312
Epoch 1280 | train loss 0.312
Epoch 1290 | train loss 0.312
Epoch 1300 | train loss 0.312
Epoch 1310 | train loss 0.312
Epoch 1320 | train loss 0.312
Epoch 1330 | train loss 0.312
Epoch 1340 | train loss 0.312
Epoch 1350 | train loss 0.312
Epoch 1360 | train loss 0.312
Epoch 1370 | train loss 0.312
Epoch 1380 | train loss 0.312
Epoch 1390 | train loss 0.312
Epoch 1400 | train loss 0.312
Epoch 1410 | train loss 0.312
Epoch 1420 | train loss 0.312
Epoch 1430 | train loss 0.312
Epoch 1440 | train loss 0.312
Epoch 1450 | train loss 0.312
Epoch 1460 | train loss 0.312
Epoch 1470 | train loss 0.312
Epoch 1480 | train loss 0.312
Epoch 1490 | train loss 0.312
Epoch 1500 | train loss 0.312
test loss 0.309
(tensor(0.3116, grad_fn=<MeanBackward0>), tensor(0.3092))
Now we have the final model for our recommendation system. We can use this model to generate a list of new recommendations for each reviewer. First, we will generate a list of review-store pair that is currently not present in our dataset, then use our model to make predictions for those new pairs. Finally, we can shortlist the top n pairs with highest rating and convert the reviewerid and storeid back to original id.
# Generate table of all reviewer_store pair
import itertools
all_reviewers = [i for i in range(0,len(final_reviewerId))]
all_stores = [j for j in range(0,len(final_storeId))]
all_combinations = list(itertools.product(all_reviewers, all_stores))
all_data = pd.DataFrame(all_combinations, columns=['reviewerId', 'storeId'])
# Join all pairs with the existing data to get the real rating.
all_data = pd.merge(all_data, final_data, how='left', on=['reviewerId', 'storeId'])
all_data.rename(columns={"overall":"rating"}, inplace=True)
# Extract reviewer-store pair with no existing rating
new_pair = all_data[all_data['rating'].isna()]
# Make predictions for pairs with no existing rating
reviewers = torch.LongTensor(new_pair.reviewerId.values)
stores = torch.LongTensor(new_pair.storeId.values)
predictions = final_model(reviewers, stores).detach().numpy()
new_pair['rating'] = predictions
# compute the rank of each store to each reviewer
new_pair['rank'] = new_pair.groupby('reviewerId')['rating'].rank(method='dense', ascending=True).astype(int)
new_pair.head(10)
C:\Users\Louis\AppData\Local\Temp\ipykernel_22272\642763186.py:6: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
new_pair['rating'] = predictions
C:\Users\Louis\AppData\Local\Temp\ipykernel_22272\642763186.py:9: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
new_pair['rank'] = new_pair.groupby('reviewerId')['rating'].rank(method='dense', ascending=True).astype(int)
| reviewerId | storeId | rating | rank | |
|---|---|---|---|---|
| 0 | 0 | 0 | 2.426768 | 4734 |
| 1 | 0 | 1 | 2.427008 | 4744 |
| 2 | 0 | 2 | 0.926070 | 106 |
| 3 | 0 | 3 | 2.013690 | 1959 |
| 4 | 0 | 4 | 2.391284 | 3948 |
| 5 | 0 | 5 | 2.069200 | 2118 |
| 6 | 0 | 6 | 2.408282 | 4245 |
| 7 | 0 | 7 | 2.308590 | 3153 |
| 8 | 0 | 8 | 2.391606 | 3953 |
| 9 | 0 | 9 | 1.085932 | 447 |
# Suggest the top n stores to each reviewer
n = 3
suggestion = new_pair[new_pair['rank'] <=n]
# Convert the reviewId and storeId back to original code
idx2reviewerid = {i:o for i,o in enumerate(final_reviewerId)}
idx2storeid = {i:o for i,o in enumerate(final_storeId)}
suggestion['reviewerId_original'] = suggestion['reviewerId'].map(idx2reviewerid)
suggestion['storeId_original'] = suggestion['storeId'].map(idx2storeid)
suggestion.reset_index(drop=True, inplace=True)
suggestion
C:\Users\Louis\AppData\Local\Temp\ipykernel_22272\3678831406.py:9: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
suggestion['reviewerId_original'] = suggestion['reviewerId'].map(idx2reviewerid)
C:\Users\Louis\AppData\Local\Temp\ipykernel_22272\3678831406.py:10: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
suggestion['storeId_original'] = suggestion['storeId'].map(idx2storeid)
| reviewerId | storeId | rating | rank | reviewerId_original | storeId_original | |
|---|---|---|---|---|---|---|
| 0 | 0 | 264 | 0.411366 | 2 | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | bzra |
| 1 | 0 | 1383 | 0.374666 | 1 | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | k1vg |
| 2 | 0 | 4605 | 0.411569 | 3 | 00006b05-61f8-4f5c-9244-00ba8d97bc0e | wpvu |
| 3 | 1 | 264 | 0.118365 | 3 | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | bzra |
| 4 | 1 | 4605 | 0.099286 | 1 | 00017c85-e20e-4d9c-8d52-c224bf2b7d6f | wpvu |
| ... | ... | ... | ... | ... | ... | ... |
| 98704 | 32901 | 4605 | 0.324824 | 2 | z9vfu4jf | wpvu |
| 98705 | 32901 | 4653 | 0.327629 | 3 | z9vfu4jf | x17k |
| 98706 | 32902 | 264 | 0.215512 | 3 | z9vyx9yi | bzra |
| 98707 | 32902 | 4605 | 0.201193 | 1 | z9vyx9yi | wpvu |
| 98708 | 32902 | 4653 | 0.205706 | 2 | z9vyx9yi | x17k |
98709 rows × 6 columns
from pathlib import Path
# Create model directory path
MODEL_PATH = Path("models")
MODEL_PATH.mkdir(parents=True,
exist_ok=True)
# Create model save
MODEL_NAME = "Recommendation_system_for_food_deliveries_state_dict.pth"
MODEL_SAVE_PATH = MODEL_PATH / MODEL_NAME
# Save the model state dict
print(f"Saving model to: {MODEL_SAVE_PATH}")
torch.save(obj=final_model.state_dict(),
f=MODEL_SAVE_PATH)
Saving model to: models\Recommendation_system_for_food_deliveries_state_dict.pth
from pathlib import Path
# Create model directory path
MODEL_PATH = Path("models")
MODEL_PATH.mkdir(parents=True,
exist_ok=True)
# Create model save
MODEL_NAME = "Recommendation_system_for_food_deliveries_full_model.pth"
MODEL_SAVE_PATH = MODEL_PATH / MODEL_NAME
# Save the model state dict
print(f"Saving model to: {MODEL_SAVE_PATH}")
torch.save(obj=final_model,
f=MODEL_SAVE_PATH)
Saving model to: models\Recommendation_system_for_food_deliveries_full_model.pth
# Create a new instance
loaded_model = RecSys_bias(num_reviewers, num_stores, emb_size=10)
# Load in the save state_dict()
loaded_model.load_state_dict(torch.load(f=MODEL_SAVE_PATH))
# Send the model to the target device
loaded_model.to(device)
RecSys_bias(
(reviewers_emb): Embedding(32903, 10)
(stores_emb): Embedding(5221, 10)
(reviewers_bias): Embedding(32903, 1)
(stores_bias): Embedding(5221, 1)
)