RecSys16: Adaptive, Personalized Diversity for Visual Discovery

Arthur Lee

3 min readMay 2, 2020

Recommendation system paper challenge (10/50)

paper link

What problem do they solve?

A top-n r fashion discovery recommendation system with item diversification

What others solve this problem?

Most researchers utilized rating as numerical view. And predict the rating considering weighted (temporal effect, similar users). However, in several common scenarios, there is no direct link between the user feedback and numerical values, even though the feedback is richer than a binary “like-vs-dislike” indication.

What is their model?

SCORING ITEM RELEVANCE

Basic Model:

Linear Probit regression: map items into CTR.

where inverse of phi function is the probit function.

However, cold-item will suffer no-review and no weights to update -> exploration-exploitation dilemma.

Multi-armed bandit:

They apply Thompson sampling approach to do Multi-armed bandit by sampling each model weight from its posterior distribution conditioned on the training data. We then use the sampled weights to compute equation (1) for each item.

It makes that the items with the highest scores will be a mix of popular products with high expected values (an exploitation strategy) and underexposed products that were randomly scored high above their expected value (an exploration strategy).

SUBMODULAR DIVERSIFICATION

A = [a1, a2,… an]: denote the attribute vectors for a set of n items.

a_i : an one-hot encoded in d-dimensions defining attributes.

Let w denote a d-dimensional vector which encodes user preferences for each of the d item attributes.

PERSONALIZATION

Learning Adaptive Global Weights

How to control w (category weight)?

Learning User Specific Weights

User Modeling

User Click Signal Diffusion

To ensure a user is exposed to related categories, we diffuse our Dirichlet updates between categories.

Let M be a d-by-d matrix

where the entry Mij = (# users like i and j) / (# users like j)

In order to capture the correlation between two categories.

Data

Amazon Stream (www.amazon.com/stream).

Result

For Submodular diversifier:

Submodular is better than multinomial
adaptive global category weight is better than manual category weight by expert
personalized category weight is better than global weight