(
To be able to analyze research questions like these, a dierent problem abstraction is required. Instead of a user-item rating matrix, the input to the recommendation problem is rather a time-ordered sequence (log) of user actions of dierent types, e.g., an item view event, a purchase, etc. Correspondingly, other computational tasks than rating prediction have to be considered, including the prediction of the next user action, the identication of trends, or the consideration of sequence constraints.
We will rst highlight the importance of what we call “sequenceaware” recommender systems in practice and categorize dierent recommendation scenarios where the main input to the problem is a time-ordered series of logged user actions. We will consider ComplexRec 2017, Como, Italy. 2017. Copyright for the individual papers remains with the authors. Copying permied for private and academic purposes. is volume is published and copyrighted by its editors. Published on CEUR-WS, Volume 1892.. dierent computational tasks in that context and then specically focus on session-based recommendation problems (where only the interactions of the current session are known) and session-aware ones (where we also know previous sessions of the current user). 2.2
In the remainder of the talk, we will focus on the specic problems mentioned in the introduction and present insights from recent research works. e case studies are based on a large data set containing logged user interactions of a major European fashion retailer.
2.2.1 Considering short- and long-term interests. In the rst case study [1], we compare the performance of dierent heuristic approaches to adapt the system’s recommendations to the estimated short-term interests of the user. e results show that while the choice of the underlying long-term model is relevant, considering short-term interests in the right way has much more impact on recommendation accuracy.
2.2.2 On the value of reminders. Since the rst study revealed that reminding users of things they have inspected (but not purchased) in the recent past can be an eective strategy, we then explore more elaborate reminding techniques than just presenting the recently viewed items in reverse chronological order [4].
2.2.3 Deriving recommendation success factors from log data. Since the available log data also contains information about which items were recommended to users and which of these items they actually inspected, we then re-construct in a systematic way which factors contributed to the success of the presented recommendations. e analysis shows that besides the consideration of the recent interests, recommendations are particularly successful when they are related to currently trending or to discounted items [2].
2.2.4 Operationalizing the insights into algorithms. Finally, we present a novel algorithmic approach to predict the next user interaction that considers all of the above-mentioned factors (short-term intents, reminders, trends and discounts) in an integrated way. We frame the recommendation problem as a classication task and our experiments show that a deep neural network leads to a beer performance than when using Random Forests or when using a weighted hybrid scoring approach. 2.3
e talk ends with a discussion of open questions and possible future directions in the eld.