Rerank

Title!!

Lorem Ipsum is simply dummy text of the printing and typesetting industry. Lorem Ipsum has been the industry's standard dummy text ever since the 1500s, when an unknown printer took a galley of type and scrambled it to make a type specimen book. It has survived not only five centuries, but also the leap into electronic typesetting, remaining essentially unchanged.

Saturation:

Sharpness:

Entropy:

Mainstream User

Judging by the rating history, this person seems to prefer very popular mainstream movies. Clasically, such users are treated better by standard recommendation algorithms, maintaining the same popularity level of suggested movies to watch. Not much change in reranking is expected here - the recommendation list is supposedly already rather fitting.

What our mainstream user likes

A regular SVD recommendation

SVD-Reranked recoommendation taking our user preference into account

Niche User

This person is less fond of mainstream movies and potentially would not be satisfied with a standard recommendation with higher popularity. The reranking effect is expected to be more visible in such case, lowering the popularity level of recommended movies.

What our niche user likes

A regular SVD recommendation

SVD-Reranked recommendation taking our user preference into account

Explanation

This Demo is based on the research of PHD Candidate Anastasiia Klimashevskaia @ SFI MediaFutures, UIB, who provided the recommendations and recommender system, and Developed by Research Assistant Snorre Alvsvåg @ SFI MediaFutures, UIB, who provided the full stack application.

It is aimed at demonstrating how a personalized approach can be utilized to adjust the recommendation popularity to the preference of each particular user. Two example users have been selected - one has strong affinity towards "mainstream" and highly popular movies, while the other one is gravitating more towards less popular and niche movies. Traditionally popularity bias mitigation approaches have been exposing every user to less popular items in the same manner to promote less known content. This approach, however, takes into consideration every viewer's watching history attempting to gauge their interest towards popular/unpopular movies and only adjust the recommendation accordingly. Since classic recommender systems are known for recommending mostly popular items to every user, the reranker is expected to have less influence on popularity-aligned users, while the strongest changes are predicted to be observed in the recommendation for the niche items. The algorithm will attempt to lower the general popularity of recommended items, while still retaining an acceptable relevance and accuracy. Note: The users with extremely niche preferences are unfortunately practically treated as outliers - the algorithm has a hard time of finding highly relevant niche items to safely recommend without a significant loss in accuracy.

Code found at github.com/sfimediafutures/DEMO_personalisation-reranker

Relevant paper by PHD Candidate Anastasiia Klimashevskaia

Developers note: The demo is slow to load images due to how we gather image information, and the lack of cached image url’s. As of date it is based on a chain of API calls originating from the TMDB API Service.