Learning by Collaborative and Individual-Based Recommendation Agents

Intelligent recommendation systems can be based on two basic principles: collaborative filters and individual-based agents. In this work we examine the learning function that results from these two general types of learning smart agents. There has been significant work on the predictive properties of each type, but no work has examined the patterns in their learning from feedback over repeated trials. Using simulations, we create clusters of “consumers” with heterogeneous utility functions and errorful reservation utility thresholds. The consumers go shopping with one of the designated smart agents – receive recommendations from the agents, and purchase products they like and reject ones they do not. Based on the purchase / no purchase behavior of the consumers, agents learn about the consumers and potentially improve the quality of their recommendations. We characterize learning curves by modified exponential functions with an intercept for percent of recommendations accepted at trial 0, an asymptotic rate of recommendation acceptance, and a rate at which learning moves from intercept to asymptote. We compare the learning of a baseline random recommendation agent, an individual-based logistic-regression agent, and two types of collaborative filters that rely on K-mean clustering (popular in most commercial applications) and nearest-neighbor algorithms. Compared to the collaborative filtering agents, the individual agent: 1) learns slower initially, but performs better in the long run when the environment is stable; 2) is less negatively affected by permanent changes in the consumer’s utility function; 3) is less adversely affected by error in the reservation threshold to which consumers compare a recommended product’s utility. The K-mean agent reaches a lower asymptote but approaches it faster, reflecting a surprising stickiness of target classifications after feedback from recommendations made under initial (incorrect) hypotheses.