Open yogeswarl opened 2 years ago
@yogeswarl thank you. As we discussed, study how they measure the correctness of the reranked items.
@Rounique @yogeswarl I'm thinking of applying this paper or similar idea to team formation.
Currently, our code recommends top-k experts for a team. We can rerank the final top-k for fair team formation.
let me know of ur thought
@hosseinfani They measure the correctness of the reranked items by 3 metrics
They also measure the ranking accuracy to examine diversity-accuracy trade-off using standard Normalized Discounted cumulative Gain (NDCG). There algorithm is 2 variants of the re-ranking algorithms, Binary xQuad and smooth xQuad Binary xQuad is the original variant and the smooth xQuad takes ratio of items to the categories covered. All the 3 metrics are required to conclude their results. they measure their algorithm with a cross validation split of 80-20.
Their is more accuracy loss in the algorithm but they tend to be fair in mitigating bias.
@yogeswarl thank you. please put those papers in your reading list and put their summaries in github. thanks. Controlling Popularity Bias in Learning-to-Rank Recommendation is a good paper!
I also read this carefully and now I guess we can discuss it more for applying it to our project.
Title: Managing Popularity Bias in Recommender Systems with Personalized Re-ranking Venue: AAAI Florida Artificial Intelligence Research Society(FLAIRS’19) Year: 2019
Main problem: Collaborative filtering recommenders typically emphasize popular items over other “long-tail” items that may only be popular among small groups of users. The author’s work is concerned with the collaborative recommendations and focuses on the long-tail segment.
Input: Assuming a ranked recommendation list R has already been generated by a base recommendation algorithm for a given user U.
Output: Compute a ranking score for the next item to be added to output list S.
Motivation: The work was inspired by a novel probabilistic framework called xQuAD for web search results that explicitly account for various aspects associated with an under-specified query.
Previous Works and their Gaps
Contributions Enables the system designer to tune the system to achieve the desired trade-off between accuracy and better coverage of long-tall, less popular items.
Proposed Method Build on the xQuAD model to control popularity bias in recommendation algorithms. The modified xQuAD method produces a new re-ranked list that manages popularity bias while still being accurate
Experiments Test the proposed algorithm on two public datasets. Movielens 1M dataset and Epinions dataset with 600K+ instances. Following the data reduction procedure on Epinions, they removed users with fewer than 20 ratings, as users with longer profiles are much more likely to have long-tail items in their profiles.
code No official code was found.
Gaps of this work The work does not consider other divisions of the popularity distribution. Future work includes using this model for multi-stakeholder recommendation whee the system needs to make recommendations in the presence of different stakeholders providing the products