Pareto ranking learning: ranking learning based on fairness of recommendation system

​Author | Wang Hao

Reviewer | Sun Shujuan

Fairness in recommendation systems is a field of artificial intelligence research that has exploded since 2017. Well-known artificial intelligence companies such as Twitter, Google, IBM and Baidu have all established artificial intelligence ethics teams or developed artificial intelligence ethics products. However, it is regrettable that artificial intelligence ethics research started late in China, and there is still a certain gap compared with foreign countries.

Ranking learning is a machine learning technology that broke out around 2010 and has been widely used in the fields of recommendation systems and information retrieval. In recent years, ranking learning has become a popular algorithm benchmark for artificial intelligence ethics research.

This article will introduce the paper Pareto Pairwise Ranking for Fairness Enhancement of Recommender Systems published at CISAT 2022 (International Conference on Computer Information Science and Application Technology) in 2022. This article mainly explains how to combine Pareto distribution and ranking learning to achieve a fair ranking learning recommendation algorithm.

##Figure 1. MovieLens dataset movie viewing rating difference probability distribution

We can draw the following conclusion based on observation (Figure 1) and statistical theory (statistical estimation of Zipf distribution): The probability distribution of the difference in ratings of different items by the same user is proportional to the difference in ratings. We modify the loss function of probability matrix decomposition and obtain the loss function formula of the new algorithm we invented, Pareto Pairwise Ranking:

Putting the observations we just made into the loss function formula, we get the following loss function formula:

We take the logarithm of L and get the following formula:

We use the stochastic gradient descent formula to solve the logarithm of the loss function and get the following formula:

The algorithm flow of Pareto ranking learning is as follows:

Figures 2 and 3 show the test results of Pareto ranking learning on the MovieLens 1 Million Dataset data set. The authors of the paper compared 10 recommendation system algorithms and found that the Pareto ranking learning algorithm performed best on the fairness index.

Figure 4 and Figure 5 show the test results of Pareto ranking learning on the LDOS-CoMoDa data set . The Pareto ranking learning algorithm still performs best on the fairness index.

The Pareto ranking learning algorithm is a rare fairness-based ranking learning recommendation system algorithm in China. The algorithm principle is simple, easy to implement and fast in operation. The author tested it on a Lenovo laptop with 16G RAM and Intel Core i5, and the execution speed was very fast. Artificial intelligence ethics research is currently a research hotspot internationally, and I hope it will attract enough attention from everyone.

About the author

Wang Hao, former head of Funplus Artificial Intelligence Laboratory, has more than 100 employees in ThoughtWorks, Douban, Sina, NetEase and other companies 11 years of R&D and management experience. He has rich technical experience in the fields of recommendation systems, chat robots, and risk control and anti-fraud. Published 30 papers in international academic conferences and journals, and won the Best Paper Award/Best Paper Report Award 3 times. 2006 ACM Regional Competition Gold Medal. Graduated from the University of Utah in the United States with a bachelor's degree and a master's degree. Part-time MBA from University of International Business and Economics.

The above is the detailed content of Pareto ranking learning: ranking learning based on fairness of recommendation system. For more information, please follow other related articles on the PHP Chinese website!