edgeRec

Deep Learning(Item2vec Embedding + MLP) based Feature-Engineering & Training & Predict all in one Recommendation System that can run on small server or edge device.

Models implemented

Simple 2 layer MLP

Progress:

• Simple 2 layer MLP test on MovieLens
• Dropout and L2 regularization
• Batch Normalization

YouTube DNN

GAUC on MovieLens 20M: 0.760381

Progress:

• YouTube DNN test on MovieLens
• Dropout and L2 regularization
• Batch Normalization

Deep Interest Network

GAUC on MovieLens 20M: 0.790542

Progress:

• DIN test on MovieLens
• Euclidean Distance based attention
• Dropout and L2 regularization
• Batch Normalization

Demo

You can run the MovieLens training and predict demo by:

# download and unzip the SQLite DB file
wget https://github.com/auxten/edgeRec/files/9895974/movielens.db.zip && \
  unzip movielens.db.zip
# compile the edgeRec and put it in the current directory
GOBIN=`pwd` go install github.com/auxten/edgeRec@latest && \
  ./edgeRec

Wait for the message shown: Listening and serving HTTP on :8080.

Then test the API in another terminal:

curl --header "Content-Type: application/json" \
  --request POST \
  --data '{"userId":108,"itemIdList":[1,2,39]}' \
  http://localhost:8080/api/v1/recommend

Should get the response like this:

{"itemScoreList":[
  {"itemId":1,"score":0.7517360474797006},
  {"itemId":2,"score":0.5240565619788571},
  {"itemId":39,"score":0.38496231172036016}
]}

So, with a higher score, user #108 may prefer movie #1 over #2 and #39.

Quick Start

To create a deep learning based recommendation system, you need to follow the steps below:

if you prefer show me the code, just go to MovieLens Example

1. Implement the recommend.RecSys interface including func below:

   GetUserFeature(context.Context, int) (Tensor, error)
   GetItemFeature(context.Context, int) (Tensor, error)
   SampleGenerator(context.Context) (<-chan Sample, error)

2. Call the functions to Train and StartHttpApi

   model, _ = recommend.Train(recSys)
   recommend.StartHttpApi(model, "/api/v1/recommend", ":8080")

3. If you want better AUC with item embedding, you can implement the recommend.ItemEmbedding interface including func below:

   //ItemEmbedding is an interface used to generate item embedding with item2vec model
   //by just providing a behavior based item sequence.
   // Example: user liked items sequence, user bought items sequence, 
   //   user viewed items sequence
   type ItemEmbedding interface {
       ItemSeqGenerator() (<-chan string, error)
   }

   All you need to do is implement the functions of the gray part:

Features

• Pure Golang implementation, battery included.
• Parameter Server based Online Learning
• Training & Inference all in one binary powered by golang
• Databases support
  • MySQL support
  • SQLite support
  • Database Aggregation accelerated Feature Normalization
• Feature Engineering
  • Item2vec embedding
  • Rule based FE config
  • DeepL based Auto Feature Engineering
• Demo
  • MovieLens Demo
  • Android demo
  • iOS demo

Benchmark

Embedding

• Apple M1 Max
• Database: SQLite3
• Model: SkipGram, Optimizer: HierarchicalSoftmax
• WindowSize: 5
• Data: MovieLens 10m

read 9520886 words 12.169282375s
trained 9519544 words 17.155356791s

Search Embedding of:
   59784 "Kung Fu Panda (2008)" Action|Animation|Children|Comedy

  RANK | WORD  | SIMILARITY  | TITLE & GENRES
-------+-------+-------------+-------------
     1 | 60072 |   0.974392  | Wanted (2008) Action|Thriller
     2 | 60040 |   0.974080  | Incredible Hulk, The (2008) Action|Fantasy|Sci-Fi
     3 | 60069 |   0.973728  | WALL·E (2008) Adventure|Animation|Children|Comedy|Romance|Sci-Fi
     4 | 60074 |   0.970396  | Hancock (2008) Action|Comedy|Drama|Fantasy
     5 | 63859 |   0.969845  | Bolt (2008) Action|Adventure|Animation|Children|Comedy
     6 | 57640 |   0.969305  | Hellboy II: The Golden Army (2008) Action|Adventure|Comedy|Fantasy|Sci-Fi
     7 | 58299 |   0.967733  | Horton Hears a Who! (2008) Adventure|Animation|Children|Comedy
     8 | 59037 |   0.966410  | Speed Racer (2008) Action|Adventure|Children
     9 | 59315 |   0.964556  | Iron Man (2008) Action|Adventure|Sci-Fi
    10 | 58105 |   0.963332  | Spiderwick Chronicles, The (2008) Adventure|Children|Drama|Fantasy

Movie Recommend Performance

• Dataset: MovieLens 100k, split by 80%+20% userId randomly
• Code: example/movielens
• Training time: 28s
• AUC: 0.782

Q&A

• Q: What model do you use?

• A: Just 2 layers of neural network and item2vec embedding.

• Q: Where can I use this?

• A: Simple system with a database. With 100 lines of golang, you got a better than nothing recommendation system.

• Q: Where wouldn't I use this?

• A: Large (100+ million) dataset using SOTA models.

Thanks

To make this project work, quite a lot of code are copied and modified from the following libraries:

• Neural Network & Parameter Server:
  • gorgonia
  • goro
  • go-deep
  • pa-m/sklearn
• Feature Engineering:
  • go-featureprocessing
  • featuremill
  • wego
• FastAPI like framework:
  • go-fastapi
• Gopher logo with GIMP:
  • ashleymcnamara/gophers
• JetBrains for providing free license for this project.

Papers related

• YouTube DNN
• Deep Interest Network for Click-Through Rate Prediction
• Document Embedding with Paragraph Vectors
• EdgeRec: Recommender System on Edge in Mobile Taobao // not very identical implementation