❗❗ This repo will no longer be maintained, please visit https://github.com/milvus-io/bootcamp ❗ ❗

Graph-based Recommendation System with Milvus

This project uses graph convolutional neural networks to generate embeddings, and then uses Milvus retrieval to implement a recommendation system. It provides flask services and a front-end interface.

Prerequisite

• Milvus 0.10.2
• DGL
• MySQL

Data preparation

The data source is MovieLens million-scale dataset (ml-1m), created by GroupLens Research. Refer to ml-1m-README for more information.

1. Clone the code.

   $ git clone https://github.com/milvus-io/bootcamp.git

2. Downland and extract the MovieLens-1M dataset.

   # Make sure you are in the pinsage folder
   $ cd bootcamp/solutions/graph_based_recommend/webserver/src/pinsage
   $ wget http://files.grouplens.org/datasets/movielens/ml-1m.zip
   $ unzip ml-1m.zip

3. Processing data as a pickle file.

   # Install the requirements
   $ pip install -r ../../requirements.txt
   $ mkdir output
   $ python process_movielens1m.py ./ml-1m ./output

   You can see that two files in the output directory: data.pkl and mov_id.csv.

Run model with DGL

This model returns items embedding that are K nearest neighbors of the latest item the user has interacted. The distance between two items are measured by inner product distance of item embeddings, which are learned as outputs of PinSAGE.

$ python model.py output --num-epochs 100 --num-workers 2 --hidden-dims 256

It will generate the h_item.npy file in the output directory.

Load data

Before running the script, please modify the parameters in webserver/src/common/config.py:

Parameter	Description	Default setting
MILVUS_HOST	milvus service ip address	127.0.0.1
MILVUS_PORT	milvus service port	19530
MYSQL_HOST	postgresql service ip	127.0.0.1
MYSQL_PORT	postgresql service port	3306
MYSQL_USER	postgresql user name	root
MYSQL_PWD	postgresql password	123456
MYSQL_DB	postgresql datebase name	mysql
MILVUS_TABLE	default table name	milvus_recommend

Please modify the parameters of Milvus and MySQL based on your environment.

# Make sure you are in the src folder
$ cd ..
$ python insert_milvus.py ./pinsage/output

Run webserver

1. Download and extract the movies_poster file.

   # Make sure you are in the src folder
   $ wget https://github.com/shiyu22/user_base_recommend/raw/master/webserver/src/movies_poster.zip
   $ unzip movies_poster.zip

2. Start recommend service.

   $ python main.py
   # You are expected to see the following output.
   Using backend: pytorch
   INFO:     Started server process [2415]
   INFO:     Waiting for application startup.
   INFO:     Application startup complete.
   INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)

   You can get the API by typing http://127.0.0.1:8000/docs into your browser.

Run webclient

$ docker run -d -p 8001:80 -e API_URL=http://127.0.0.1:8000 tumao/movie-taste-demo

Note: The API_URL parameter is the webserver address.

System Usage

Type http://127.0.0.1:8001 into your mobile or client browser to get started with the movie recommendation system.

• Mark movies

  Mark your favorite or disliked movies from 16 random movies.

  

• Movie recommendations

  Recommend movies based on your marked favorites.

  

• Movie info

  Get more information with movie.