November 13, 2019
Lee, an assistant professor of Business Analytics in the Tepper School of Business, discusses the study in "How Do Product Attributes and Reviews Moderate the Impact of Recommender Systems Through Purchase Stages?" published in the Social Science Research Network’s electronic journal. Moreover, the research finds that the effect of recommender systems is shaped by product attributes and customer ratings.
In related research, "How Do Recommender Systems Affect Sales Diversity? A Cross-Category Investigation via Randomized Field Experiment," coauthored by Lee and Kartik Hosanagar of the University of Pennsylvania, the study breaks down the impact of the algorithms commonly used by online retailers such as Amazon, Walmart and Netflix. Most analyze consumers’ purchase histories and browsing habits to create a composite of people with similar tastes, then recommend the products those people have bought.
Lee found that recommender systems do help people explore and buy more products — for example, Netflix might recommend a movie you might not otherwise consider, so the diversity of individual sales increases.
But recommenders tend to send an aggregate of consumers to the same product over and over, which winds up hurting market share for niche products. If Netflix keeps sending viewers to popular movies, the small art-house film won’t get as much attention, so there is less diversity in overall consumption. The end result, Lee explained, might be that studios will produce more of the types of movies that are popular and invest fewer resources in the movies that aren’t benefiting from the recommendations.
A control group who had no recommender system had greater diversity, Lee noted. The bottom 80 percent of products accounted for 37 percent of sales in the control, compared to just 27 percent when a recommender system was in place.
While some products may be inherently popular, recommender systems also can create unintentional echo chambers. If a news outlet used them, the filters could significantly impact the way people consume information by driving the bulk of readers to the same stories. The same is true for consumer products.
Although Lee studied the effect of recommenders in retail, he notes that the findings are applicable to non-retail settings.
"Recommenders influence consumer purchase patterns across different stages of purchase funnel. This effect differs by stages and also across different product categories," he said.
Lee’s research also examines the effect of recommenders on different stages of the sales cycle, and describes how that effect is impacted by other factors, such as the type of product or consumer ratings.
For example, he finds that recommender systems can help boost a product that has low consumer ratings, as well as "hedonic" products — things that are unnecessary, but that people might buy for pleasure, such as wine or perfume. The boost is lower for practical items, however.
By knowing the differences in impact, Lee notes, a retailer could modify its website to address any shortcomings.
"If you have a niche product with mixed reviews, you could make the recommender signal more salient, which may bump up sales," Lee said.
Powered by Suggestion: Study Reviews Online Recommendations
Recommender systems for online retailers — the algorithms that tell you "customers who bought this item also bought" — are helpful for driving up overall sales figures but also can lead to a narrower array of product choices, according to new research by Carnegie Mellon University’s Dokyun "D.K." Lee.Lee, an assistant professor of Business Analytics in the Tepper School of Business, discusses the study in "How Do Product Attributes and Reviews Moderate the Impact of Recommender Systems Through Purchase Stages?" published in the Social Science Research Network’s electronic journal. Moreover, the research finds that the effect of recommender systems is shaped by product attributes and customer ratings.
In related research, "How Do Recommender Systems Affect Sales Diversity? A Cross-Category Investigation via Randomized Field Experiment," coauthored by Lee and Kartik Hosanagar of the University of Pennsylvania, the study breaks down the impact of the algorithms commonly used by online retailers such as Amazon, Walmart and Netflix. Most analyze consumers’ purchase histories and browsing habits to create a composite of people with similar tastes, then recommend the products those people have bought.
Lee found that recommender systems do help people explore and buy more products — for example, Netflix might recommend a movie you might not otherwise consider, so the diversity of individual sales increases.
But recommenders tend to send an aggregate of consumers to the same product over and over, which winds up hurting market share for niche products. If Netflix keeps sending viewers to popular movies, the small art-house film won’t get as much attention, so there is less diversity in overall consumption. The end result, Lee explained, might be that studios will produce more of the types of movies that are popular and invest fewer resources in the movies that aren’t benefiting from the recommendations.
A control group who had no recommender system had greater diversity, Lee noted. The bottom 80 percent of products accounted for 37 percent of sales in the control, compared to just 27 percent when a recommender system was in place.
While some products may be inherently popular, recommender systems also can create unintentional echo chambers. If a news outlet used them, the filters could significantly impact the way people consume information by driving the bulk of readers to the same stories. The same is true for consumer products.
Although Lee studied the effect of recommenders in retail, he notes that the findings are applicable to non-retail settings.
"Recommenders influence consumer purchase patterns across different stages of purchase funnel. This effect differs by stages and also across different product categories," he said.
Lee’s research also examines the effect of recommenders on different stages of the sales cycle, and describes how that effect is impacted by other factors, such as the type of product or consumer ratings.
For example, he finds that recommender systems can help boost a product that has low consumer ratings, as well as "hedonic" products — things that are unnecessary, but that people might buy for pleasure, such as wine or perfume. The boost is lower for practical items, however.
By knowing the differences in impact, Lee notes, a retailer could modify its website to address any shortcomings.
"If you have a niche product with mixed reviews, you could make the recommender signal more salient, which may bump up sales," Lee said.