Content-based filtering

Collaborative filtering vs Content-based filtering

Collaborative filtering

Recommend items to you based on ratings of uses who gave similar ratings as you

Content-based filtering

recommend items to you based on the features of use and item to find good match (requires having some features of each user)

Examples

Vector space can be different

User features - $X_u^{(j)}$ for user $j$

• Age
• Gender
• Country
• Movies watched
• Average rating per genre Movie feature - $X_m^{(i)}$ for movie $i$
• Year
• Genre/Genres
• Reviews
• Average rating

Learning to match

Predict rating of user $j$ on movie $i$ as

$$w^{(j)}\cdot x^{(i)}+b^{(j)}=V_u^{(j)}\cdot V_m^{(i)}$$

$V_u^{(j)}$ computed from $x_u^{(j)}$ $V_m^{(i)}$ computed from $x_m^{(i)}$

The size should match

Deep learning for content-based filtering

• A good way to develop a content-based filtering algorithm is to use deep learning.
• $v_u^{(i)}$ is a vector of length 32 that describes user j with features $x_u^{(j)}$
• $v_m^{(i)}$ is a vector of length 32 that describes movie i with features $x_m^{(i)}$
• To find movies similar to movie i
  • fined smallest distance: $||V_m^{(k)}-V_m^{(i)}|{|}^2$
  • With Collaborative filtering: $||X^{(k)}-X^{(i)}||$
• This can be pre-computed ahead of time

Recommending from a large catalogue

Two steps: Retrieval & Ranking

• Retrieval
  • Generate large list of plausible item candidates
    • e.g.
      • For each of the last 10 movies watched by the user, find 10 most similar movies $||V_m^{(k)}-V_m^{(i)}|{|}^2$
    • For most viewed 3 genres, find the top 10 movies
    • Top 20 movies in the country
  • Combine retrieved items into list, removing duplicates and items already watched/purchased
• Rankings
  • Take list retrieved and rank using learned model
  • Display ranked items to user

Retrieval step

• Retrieving more items results in better performance, but slower recommendations
• To analyze/optimize the trade-off, carry out offline experiments to see if retrieving additional items results in more relevant recommendations

Ethical use of recommender systems

What is the goal of the recommender system?

• Movies most likely to be rated 5 stars by user
• Products most likely to be purchased
• Ads most likely to be clicked on
• Products generating the largest profit
• Video leading to maximum watch time

Ethical considerations with recommender systems

Positive case

• Travel industry
  • Good travel experiences to more users
  • More profitable
  • Bid higher for ads
• Payday loans
  • Squeeze customers more
  • More profit
  • Bid higher for ads
• Amelioration: Do not accept ads from exploitative businesses

Problematic cases

• Maximizing user engagement has led to large social media/video sharing sites to amplify conspiracy theories and hate/toxicity
  • Amelioration: Filter out problematic content such as hate speech, fraud, scams and violent content
• Can a ranking system maximize your profit rather than user’s welfare be presented in a transparent way?
  • Amelioration: Be transparent with users

TensorFlow implementation of content-based filtering