FOLLOW THE SLIDES HERE

https://slides.com/kirillkasjanov/recommender-systems

E-Commerce Recommender Systems

TWO-TOWER DEEP NEURAL NETWORK APPLICATION

FOR CANDIDATE RETRIEVAL

INTRODUCTION

  • Recommender Systems are one of the main application of ML/DL in B2C Businesses

 

  • Learning from user experience and interactions is the foundation

 

  • Generate constantly quality recommendations is difficult and often impossible

 

  • Big enterprise (Amazon, Netflix, Google etc.) invest and develop but still cannot perfect them

Challenges in Model Design

  • No "Ground Truth"
  • Changing Environment
  • Manage adverse Effects

Recommend items that are related to what the customer likes.

Collaborative Filtering

Recommend items also based on customer similarity.

Content-based Filtering

Recommends items based on multidimensional similarity of both customers and products

Matrix Factorization

Main Approaches

How it works?

X
X
X
X

Rock

Jazz

Pop

X
X
X

Metallica

Taylor Swift

Miles Davis

X

Content-Based Filtering

Collaborative Filtering

Metallica

Taylor Swift

Miles Davis

Energetic

Calm

Emotional lyrics

No Lyrics

Madonna

+1

-1

+1

-1

How it works?

0.9 0.3 -0.8 0.5
0.2 0.8 -0.8 0.3
-0.4 -0.3 0.5 -0.3
-0.7 0.1 0.3 -0.3

Metallica

Taylor Swift

Miles Davis

Madonna

0.9 0.2 -0.7 0.5
0.2 0.9 -0.9 0.3

Energetic - Calm

Emotional Lyrics - Instrumental

1.0 -0.1
0.0 0.9
-0.4 -0.3
-0.9 0.3
x
x
x
x

Metallica

Taylor Swift

Miles Davis

Madonna

0.9 0.2 -0.7 0.5
0.2 0.9 -0.9 0.3

Energetic - Calm

Emotional Lyrics - Instrumental

1.0 -0.1
0.0 0.9
-0.4 -0.3
-0.9 0.3

Generate Recommendations with Matrix Factorization

Two-Tower DNN

Extending MF to Deep Learning 

Candidate Tower

Query Tower

Xinyang et al. 2019
 

0.4

0.3

0.5

0.8

0.6

(IN-BATCH) SOFTMAX

LOSS

Hidden Layers

EMBEDDINGS

Customer Query

Retrieval

Embedding Computation

Embedding Space Lookup

Request Recommendations

Retrieving non-linear candidate computations as recommendations

Portfolio Project

Recommender System for Beauty Care Products

Context and Data

  • Beauty Care Products 6.7k
  • Customer:  250k
  • Timestamped Full Year Data

Customer:

  • Anonymous ID
  • Timestamps
  • Interactions

Product:

  • Brand, Series
  • Category, Gender
  • Price, Vol
  • Ingredients, Effects, Scent, Hairtype, Finish, Skintype

+ Images

Embedding Size = 32

1 Hidden Layer as Output

Embedding Size = 32

1 Hidden Layer as Output

Embedding Size = 32

1 Hidden Layer as Output

Adding Hidden Layers

(non-linear)

Larger dense layers non-linear activations impair learning...

Adding Hidden Layers

(non-linear)

Customer Actions: Categorical Encoding

Product Category: Categorical Encoding

Product Type: Categorial Encoding

Product Gender: Categorical Encoding

Special Product Features: One-Hot Encoding

 

TRANSFORM TO

EMBEDDING

 

Adding Hidden Layers

+ Add More Embeddings

Generating Recommendations

Go to Demo Notebook
  • Approach does not require to have prior knowledge of the domain to be applied

 

  • Encoded all variables in embeddings (exc. numerical) show to yield best results

 

  • Linearity can be prevailing over non-linear activations. Try both setups to see what works best.

Takeaways and Insights

  • More data on customer to enrich the query tower

 

  • Explore more methodology like Rating vs. Retrieval and DCN

 

  • Experiment more with CNN for product images to identify more features

Future Work

THANK YOU!

Recommender Systems

By Kirill Kasjanov