Food-Recmmendation-System-Python

This is the Recommendation Engine that will be used in building the Lunchbox App, a platform for ordering food and keeping track of user expenditure and canteen sales. Regardless of whether or not this is actually implemented in all the canteens of IIT Kanpur (given the potential for frauds & cyber-attacks) I will still complete the platform.

Also, I would be open-sourcing the app so that any campus can implement a cash-less & integrated system of ordering food across their whole campus. After all, what good are IITs for if our canteens still keep track of student accounts on paper registers!

Build instructions

• git clone https://github.com/gsunit/Food-Recommendation-System-Pyhton.git
• Run the Jupyter Notebook src.ipynb

Demographic Filtering

Suggesting the items that are well-received and popular among the users. Most trending items and items with the best rating rise to the top and get shortlisted for recommendation.

import pandas as pd 
import numpy as np

# Importing db of food items across all canteens registered on the platform
df1=pd.read_csv('./db/food.csv')
df1.columns = ['food_id','title','canteen_id','price', 'num_orders', 'category', 'avg_rating', 'num_rating', 'tags']

df1

	food_id	title	canteen_id	price	num_orders	category	avg_rating	num_rating	tags
0	1	Lala Maggi	1	30	35	maggi	3.9	10	veg, spicy
1	2	Cheese Maggi	1	25	40	maggi	3.8	15	veg
2	3	Masala Maggi	1	25	10	maggi	3.0	10	veg, spicy
3	4	Veg Maggi	1	30	25	maggi	2.5	5	veg, healthy
4	5	Paneer Tikka	1	60	50	Punjabi	4.6	30	veg, healthy
5	6	Chicken Tikka	1	80	40	Punjabi	4.2	28	nonveg, healthy, spicy

Results of demographic filtering

top_rated_items[['title', 'num_rating', 'avg_rating', 'score']].head()
pop_items[['title', 'num_orders']].head()

	title	num_rating	avg_rating	score
4	Paneer Tikka	30	4.6	4.288889
5	Chicken Tikka	28	4.2	4.013953
1	Cheese Maggi	15	3.8	3.733333

	title	num_orders
4	Paneer Tikka	50
1	Cheese Maggi	40
5	Chicken Tikka	40
0	Lala Maggi	35
3	Veg Maggi	25

Content Based Filtering

A bit more personalised recommendation. We will analyse the past orders of the user and suggest back those items which are similar.

Also, since each person has a "home canteen", the user should be notified of any new items included in the menu by the vendor.

We will be use Count Vectorizer from Scikit-Learn to find similarity between items based on their title, category and tags. To bring all these properties of each item together, we create a "soup" of tags. "Soup" is a processed string correspnding to each item, formed using the constituents of tags, tile and category.

	food_id	title	canteen_id	price	num_orders	category	avg_rating	num_rating	tags	soup
0	1	Lala Maggi	1	30	35	maggi	3.9	10	veg, spicy	veg spicy lala maggi
1	2	Cheese Maggi	1	25	40	maggi	3.8	15	veg	veg cheese maggi
2	3	Masala Maggi	1	25	10	maggi	3.0	10	veg, spicy	veg spicy masala maggi

Using CountVectorizer from Scikit-Learn

# Import CountVectorizer and create the count matrix
from sklearn.feature_extraction.text import CountVectorizer
count = CountVectorizer(stop_words='english')

# df1['soup']
count_matrix = count.fit_transform(df1['soup'])

# Compute the Cosine Similarity matrix based on the count_matrix
from sklearn.metrics.pairwise import cosine_similarity
cosine_sim = cosine_similarity(count_matrix, count_matrix)

Sample Recommendation

df1.loc[get_recommendations(title="Paneer Tikka")]

	food_id	title	canteen_id	price	num_orders	category	avg_rating	num_rating	tags	soup
5	6	Chicken Tikka	1	80	40	Punjabi	4.2	28	nonveg, healthy, spicy	nonveg healthy spicy chicken tikka punjabi
3	4	Veg Maggi	1	30	25	maggi	2.5	5	veg, healthy	veg healthy maggi

After all the hard work, we finally get the recommendations

personalised_recomms(orders, df1, current_user, columns)
get_new_and_specials_recomms(new_and_specials, users, df1, current_canteen, columns)
get_top_rated_items(top_rated_items, df1, columns)
get_popular_items(pop_items, df1, columns).head(3)

	title	canteen_id	price	comment
0	Veg Maggi	1	30	based on your past orders
1	Paneer Tikka	1	60	based on your past orders
2	Chicken Tikka	1	80	based on your past orders

	title	canteen_id	price	comment
0	Cheese Maggi	1	25	new/today's special item in your home canteen

	title	canteen_id	price	comment
0	Paneer Tikka	1	60	top rated items across canteens
1	Chicken Tikka	1	80	top rated items across canteens
2	Cheese Maggi	1	25	top rated items across canteens

	title	canteen_id	price	comment
0	Paneer Tikka	1	60	most popular items across canteens
1	Cheese Maggi	1	25	most popular items across canteens
2	Chicken Tikka	1	80	most popular items across canteens

These are just simple algorithms to make personalised & general recommendations to users. We can easily use collaborative filtering or incorporate neural networks to make our prediction even better. However, these are more computationally intensive methods. Kinda overkill, IMO! Let's build that app first, then move on to other features!

Star the repository and send in your PRs if you think the engine needs any improvement or help me implement some more advanced features.