Recommender Systems III

General

Where we focus

• User-Item Recommendation
• Item-Item Recommendation
  • Content based recommendation
  • Collaborative Filtering

Recall: Notes on Item-Item Recommendations

• recommendation based on last item: “If you are interested in this, you probably also like …”
• doesn’t need a user model
• Item model needed → How to model similarities between items?

Recall: Collaborative Filter

…

…

….

Content based Recommendation

• we have more information about the film than just a number, we know the content of the film → how to find similar items based on item properties and/or meta data?

Possible solutions:

• Compute distance between movies for example (e.g. calculate sequences) → possible with DL, but too expensive in most cases
• use meta data → Categorization, actors/directors, year, …
• use simpler measure → e.g. analyze movie poster

Recall: Benefits of DL

• capacity of the learned mapping → tensor input/output, like pictures, videos, tables, …

• End-to-End Learning → learn decision space and function in one optimization problem

Different Approaches of using DL in Recommender Systems

Autoencoder

• instead of a SVD a non-linear operation to compress data
• consists of two parts:
  • encoder → compresses data in a latent space
  • Decoder →
• we measure the distances in the compressed space
• we can add additional evaluations to our loss function based on previous ratings

Basic CNN Approach

• Use CNNs for item (and user) content feature extraction
• Examples:
  • Movie similarity by image similarity of posters
  • Clothing and fashion items, e.g. I like this kind heel, suggest me other shoes that have a similar heel
• can directly combined with AE or SVD

RNN based Recommender Systems

• Using RNNs for text feature extraction:

1. User features: reviews and comments by a user → if somebody often insults, a negative rating doesn’t value much → if somebody only writes positive things, a negative rating has more impact
2. Item features: text description and reviews of item

• s1, s2, … are words/sequences of the text

Hybrid Recommender Systems

Combined Measurements:

• Film poster
• Text (Meta data)
• User Ratings

Code Exercises

(Links to Github)

Autoencoder_with_Keras_week8.ipynb

collaborative_filtering_with_Keras_week8.ipynb