Authors:
Hasan Tercan
1
;
Christian Bitter
1
;
Todd Bodnar
2
;
Philipp Meisen
2
and
Tobias Meisen
1
Affiliations:
1
Chair for Technologies and Management of Digital Transformation, University of Wuppertal, Wuppertal, Germany
;
2
Breinify Inc., San Francisco, U.S.A.
Keyword(s):
Recommender System, Deep Learning, Neural Network, Embedding, Prod2vec, Word2vec.
Abstract:
Recommender systems are a central component of many online stores and product websites. An essential functionality of them is to show users new products that they do not yet know they want to buy. Since the users of the website are often unknown to the system, a product recommendation must be made using the current activities within a browser session. In this paper we address this issue in a deep learning-based product-to-product recommendation problem for a commercial website with millions of user interactions. Our proposed approach is based on a prod2vec method for product embeddings, thus recommending those products that often occur together with the target product. Following the idea of word2vec methods from the NLP domain, we train an artificial neural network on user activity data extracted from historical browser sessions. As part of several real A/B tests on the website, we prove that our approach delivers successful product recommendations and outperforms the current system
in use. In addition, the results show that the performance can be significantly improved by an appropriate selection of the training data and the time range of historical user interactions.
(More)