Challenges of Session-Aware Recommendation in E-Commerce (Keynote) Dietmar Jannach TU Dortmund, Germany dietmar.jannach@tu-dortmund.de KEYWORDS different computational tasks in that context and then specifically Recommender systems, Session-aware Recommendation 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). 1 MOTIVATION Research in the field of recommender systems is in many cases based on the matrix completion problem abstraction. While being 2.2 E-Commerce Case Studies able to assess the user’s general preferences towards individual In the remainder of the talk, we will focus on the specific problems items is important, this popular problem abstraction often cannot mentioned in the introduction and present insights from recent fully capture certain aspects that are important for the success of a research works. The case studies are based on a large data set recommender in practice, in particular in e-commerce settings [3]. containing logged user interactions of a major European fashion (1) First, in e-commerce scenarios users often visit an online retailer. shop with a very specific shopping intent, and a recom- mender system, to be successful, must be able to adapt 2.2.1 Considering short- and long-term interests. In the first case its recommendations to the particular contextual situation study [1], we compare the performance of different heuristic ap- and short-term preferences of the user. proaches to adapt the system’s recommendations to the estimated (2) Second, in some cases it might be relevant to know which short-term interests of the user. The results show that while the items the user inspected in his or her last session, and a choice of the underlying long-term model is relevant, considering recommender could use this knowledge to remind the user short-term interests in the right way has much more impact on of such items. recommendation accuracy. (3) Third, in some domains, also aside from e-commerce, con- sidering popularity trends in the user community could be 2.2.2 On the value of reminders. Since the first study revealed helpful when deciding on what to recommend to users. that reminding users of things they have inspected (but not pur- (4) Fourth, some users might be interested in certain items chased) in the recent past can be an effective strategy, we then only in case they are currently discounted. For such users, explore more elaborate reminding techniques than just presenting recommending items that are on sale could be a promising the recently viewed items in reverse chronological order [4]. strategy. 2.2.3 Deriving recommendation success factors from log data. To be able to analyze research questions like these, a different Since the available log data also contains information about which problem abstraction is required. Instead of a user-item rating matrix, items were recommended to users and which of these items they the input to the recommendation problem is rather a time-ordered actually inspected, we then re-construct in a systematic way which sequence (log) of user actions of different types, e.g., an item view factors contributed to the success of the presented recommenda- event, a purchase, etc. Correspondingly, other computational tasks tions. The analysis shows that besides the consideration of the than rating prediction have to be considered, including the pre- recent interests, recommendations are particularly successful when diction of the next user action, the identification of trends, or the they are related to currently trending or to discounted items [2]. consideration of sequence constraints. 2 CONTENTS OF THE TALK 2.2.4 Operationalizing the insights into algorithms. Finally, we present a novel algorithmic approach to predict the next user inter- 2.1 Defining Sequence-Aware action that considers all of the above-mentioned factors (short-term Recommendation intents, reminders, trends and discounts) in an integrated way. We We will first highlight the importance of what we call “sequence- frame the recommendation problem as a classification task and aware” recommender systems in practice and categorize different our experiments show that a deep neural network leads to a better recommendation scenarios where the main input to the problem performance than when using Random Forests or when using a is a time-ordered series of logged user actions. We will consider weighted hybrid scoring approach. ComplexRec 2017, Como, Italy. 2017. Copyright for the individual papers remains with the authors. Copying permitted 2.3 Outlook for private and academic purposes. This volume is published and copyrighted by its editors. Published on CEUR-WS, Volume 1892.. The talk ends with a discussion of open questions and possible future directions in the field. ComplexRec 2017, August 31, 2017, Como, Italy. Dietmar Jannach REFERENCES [3] Dietmar Jannach, Paul Resnick, Alexander Tuzhilin, and Markus Zanker. 2016. [1] Dietmar Jannach, Lukas Lerche, and Michael Jugovac. 2015. Adaptation and Recommender Systems - Beyond Matrix Completion. Commun. ACM 59, 11 Evaluation of Recommendations for Short-term Shopping Goals. In Proc. RecSys (2016), 94–102. DOI:http://dx.doi.org/10.1145/2891406 ’15. 211–218. [4] L. Lerche, D. Jannach, and M. Ludewig. 2016. On the Value of Reminders within [2] Dietmar Jannach and Malte Ludewig. 2017. Determining Characteristics of E-Commerce Recommendations. In Proc. UMAP ’16. 27–25. Successful Recommendations from Log Data – A Case Study. In Proc. SAC ’17.