Determining characteristics of successful recommendations from log data
Dietmar Jannach, Malte Ludewig · 2017
Academic research in recommender systems largely focuses on the problem of predicting the relevance of (long-tail) items that the individual user presumably does not know yet. Many real-world systems however also recommend items that users have inspected in the past, items that are popular at the moment, and items currently on sale. In this work we investigate the value of including such items in recommendation lists based on an analysis of the web logs of a large online retailer. An examination of the features of successful item suggestions reveals that the chances of a recommendation leading to a purchase increase when the item is recently trending, on sale, or was recently viewed by the user. Offline simulation experiments furthermore show that considering those success factors that were identified from log data in the ranking algorithms can help to increase the prediction accuracy of recommender systems.