MIT Sloan professor’s learning algorithms mitigate impact of fraud on product rankings
CAMBRIDGE, Mass., Dec. 15, 2020 – As more people conduct holiday shopping online this year because of the pandemic, it is critical to be able to trust a platform’s rankings in order to give quality gifts. However, an ongoing challenge for platforms is fraudulent data, such as fake clicks, purchases, and reviews, which can impact product rankings. To mitigate this problem, MIT Sloan School of Management Prof. Negin (Nicki) Golrezaei and her colleagues designed new learning algorithms to combat fraud. The algorithms can determine optimal product rankings even when they do not know the identity or number of fake users.
“The sheer number of substitutable products on online shopping platforms combined with consumers’ limited attention has led to a new form of competition among products – the race for visibility. The success of a product critically depends on its ranking, which is based on user feedback,” explains Golrezaei. “As a result, click farms employ fake users to click on products to boost their popularity and mislead the platform to rank them in top positions.”
She notes that this is a widespread issue, as many platforms rely on data-driven algorithms to optimize their operational decisions. Commonly used learning algorithms for product rankings, such as ones based on upper confidence bounds, are vulnerable to fake users whose actions may mislead the algorithm to make poor decisions. Those decisions may result in the most visible positions being occupied by unpopular products, which can harm customer engagement and other metrics.
Golrezaei and her colleagues sought to design an algorithm that can efficiently learn the optimal product ranking in the presence of fake users, even when they are completely blind to the identity and number of fake users.
“While many recent works have explored how position bias can prevent platforms from accurately inferring customer preferences, this study is among the first to highlight how sellers can exploit this in their favor by employing fake users, and develop constructive solutions for preventing such a situation,” they write in their paper.
The researchers’ work presents a number of insights on how to design methods for uncertain environments to guarantee robustness in the face of manipulation. These include being more conservative in inferring key parameters and changing decisions based on limited data; employing parallelization and randomization to limit the damage caused by fake users; and augmenting a conservative approach via cross-learning.
“The key is to keep parallel copies of a learning algorithm that are different in terms of their conservatism level. Being conservative helps algorithms to not get manipulated easily by fraudulent data. Our parallel algorithms can communicate with each other to learn the right level of conservativism and make decisions about how to rank products. This is a new line of research on corrupted data. The bidirectional communication between parallel copies of the algorithm is unique,” says Golrezaei.
She adds, “Overall, our results show that online platforms can effectively combat fraudulent users without large costs by designing new learning algorithms that guarantee efficient convergence even when the number and identify of fake users is unknown. We believe this work can serve as a starting point for designing robust data-driven algorithms to tackle other operational challenges.”
Golrezaei is coauthor of “Learning product rankings robust to fake users” with Vahideh Manshadi of Yale University, Jon Schneider of Google Research, and Shreyas Sekar of the University of Toronto.
The MIT Sloan School of Management
The MIT Sloan School of Management is where smart, independent leaders come together to solve problems, create new organizations, and improve the world. Learn more at mitsloan.mit.edu.