Modeling Individual Users' Responsiveness to Maximize Recommendation Impact

Recommender systems provide personalized information based on a user's preferences. Differences in preferences among users are estimated from past records such as click logs or purchase logs. Recommender systems typically assume that users will respond to recommendations, provided that their favorite items are correctly selected. However, the responsiveness to recommendations depends on the type of users; while some users might be easily persuaded to take action, others might be more hesitant. In this paper, we propose a purchase prediction model that incorporates the differences in the responsiveness. We derived the individual users' responsiveness from a combination of purchase logs and recommendation logs. Improvement in the accuracy of purchase prediction was verified using a grocery shopping dataset. Another relatively unexplored yet important objective of recommender algorithms is to maximize recommendation impact, which is defined as the increase in purchase probability through recommendations. The impact of recommendations by our model exceeded that of a conventional model that ignores individual users' responsiveness. These results demonstrate the importance of modeling the responsiveness of individual users. In cases where recommendation logs are insufficient, the responsiveness needs to be estimated from other sources. Consequently, we investigated the correlation of the responsiveness with user attributes and item attributes. The estimates of the responsiveness from the correlated attributes outperformed the mean estimates. Furthermore, the recommendation impact of the model estimated from the correlated attributes was almost comparable to that of the model estimated from recommendation logs. These findings can help overcome the cold-start problem of inadequate recommendation logs. Our study presents a new direction in the field of personalization based on the responsiveness to recommendations.