My Content

Challenge

My Content is a start-up company that wants to help people read more by recommending books that they will love .

To overcome this challenge, we need to :

• represent books and users by meaningful features
• understand users book preferences
• recommend the right books to the right users

Goals

In this project, we are going to :

• define the meaningful features of books and users
• compare different recommendation systems
• integrate the best recommendation system with a mobile application

Recommender Systems assumptions

• similar users are likely to have similar content interrests
• similar contents are likely to interest users similarly
• user's content preferences may be :
  • explicit : based on actual feedback from users (reaction, rating, comment, ...)
  • implicit : based on the user's behavior (clicks, views, time spent, purchase, ...)

State of the art

There are three main categories of recommender systems :

• Content-Based Filtering : based solely on content features and the user's preferences
  • find content similar to what the user likes
• Collaborative Filtering : based solely on the user's preferences and the user profiles
  • find content that similar users like
• Hybrid recommender systems : combination of content and collaborative filtering

Exploratory data analysis (EDA)

For this MVP, we used the Globo.com dataset :

• 364047 articles described by :
  • 3 metadata fields (category, creation time and word count)
  • 250 dimensions embedding of the article's content
• 2988181 clicks (from Oct. 1st to Oct. 17th , 2017)
  • on 46033 news articles
  • by 322897 users
  • described by 10 fields (session, device and location)

User preferences modeling

Without an explicit rating of the articles by the users, we must model an implicit rating : $Rating (user, article) = \frac{\#Clicks(user, article)}{\#Clicks(user)}$

As for a user's preferences, we tested 3 models :

• last clicked article
• average of user's clicked articles over last session
• average of all clicked articles

Models

• Content-based Filtering : Cosine Similarity between user preferences and articles
• Collaborative Filtering
  • Surprise library
    • BaslineOnly : just estmate user and articles biases
    • SVD : a simple matrix factorisation technique
  • Implicit library : ALS a more performant MF technique
• Hybrid model
  • LightFM library : uses MF technique on latent features

Protocol

To evaluate our models, we left out the last click of each user.

We train our models on the remaining clicks.
We then predict recommendations (ranking of all articles) for the next article the user will click.

Our model score is the average predicted rank of the actual article the user has clicked last.

Results

Curent MVP System Architecture

Demo

• Start Azure CosmosDB
• Start Azure Function
• Launch App in emulator

Target Production System Architecture

Next steps

• Test the recommender system on real data
• Fine-tune the model's hyper parameters
• Implement the taret production system architecture
• Improve the Mobile App (UX, security, ...)