The digital age has made an unprecedented volume of information available to marketers. From search histories to click-rates, e-commerce businesses can potentially know everything about users’ online profiles, and they steadfastly look for ways to gain insights that will translate to higher sales.
With this in mind, the USC Viterbi Institute for Innovation, in collaboration with the Target Corp., recently hosted the Target Data Challenge, a competition in which 20 teams used a dataset of Target customers’ transactions and product descriptions to predict future purchases.
A team of researchers from the USC Data Science Laboratory won first place in the mid-April competition, along with an award of $10,000, for coming up with an algorithm that improves customer product recommendations. The victorious team, which consisted of PhD students Chung Ming Cheung, Palash Goyal and Ajitesh Srivastava, was mentored by research associate Arash Saber Tehrani, all members of Professor Viktor Prasanna’s Data Science Lab at the USC Viterbi School of Engineering. Cheung, Goyal and Srivastava are students in the Department of Computer Science.
“I am thrilled to hear about this award,” Prasanna said. “My Data Science Lab has been looking into applications of machine learning in smart oil field and smart grid, and the Target challenge problem is a new application direction that complements our ongoing work.”
Social network analysis
Of the initial 20 teams, only four reached the semifinal round. Those teams competed on April 15 against each other at the Viterbi Hacker House, where the Data Science Lab team was selected as winner, based on three criteria: prediction results, presentation and novelty.
The team’s approach was based on formalism for social network analysis using multilevel representation, random walks and random sampling, which allowed them to discover patterns of relatedness.
“Not everything went smoothly,” Cheung said. “We tried many approaches and most did not give good results, but we were thrilled when our predictive model finally appeared to make good predictions.”
Recommender systems are nothing new. In fact, companies like Amazon and Netflix have used them for a while to increase sales. Nevertheless, these systems are still flawed, and it is not uncommon for them to make suggestions that are far from accurate.
“Current recommender systems fail to pinpoint exactly what users are looking for,” Tehrani said. “However, I do think that what we did is a small step toward reaching a perfect system.”