We describe the selection, implementation and online evaluation of two e-commerce recommender systems developed with our partner company, Prediggo. The first one is based on the novel method of Bayesian Variable-order Markov Modeling (BVMM). The second, SSAGD, is a novel variant of the Matrix-Factorization technique (MF), which is considered state-of-the-art in the recommender literature.We discuss the offline tests we carried out to select the best MF variant, and present the results of two A/B tests performed on live ecommerce websites after the deployment of the new algorithms. Comparing the new recommenders and Prediggo’s proprietary algorithm of Ontology Filtering, we show that the BVMM significantly outperforms the two others in terms of CTR and prediction speed, and leads to a strong increase in recommendation-mediated sales. Although MF exhibits reasonably good accuracy, the BVMM is still significantly more accurate and avoids the high memory requirements of MF. This scalability is essential for its application in online businesses.

Recommender systems typically determine the items they should recommend by learning models of user-preferences. Most often, those preferences are modeled as static and independent of context. In real life however, users consider items in sequence: TV series are watched episode by episode and accessories are chosen after the main appliance. Unfortunately, since sequences are more complex to model, they are often not taken into account. We developed an efficient sequence-modeling approach based on Bayesian Variable-order Markov Models and combined it with an existing content-based system, the Ontology Filtering. We tested this approach through live evaluations on two e-commerce sites. It dramatically increased performance, more than doubling the CTR and strongly increasing recommendation-mediated sales. These tests also confirm that the technique works efficiently and reliably in a production setting.

Because of the abundance of news on the web, news recommendation is an important problem. We compare three approaches for personalized news recommendation: collaborative filtering at the level of news items, content-based system recommending items with similar topics, and a hybrid technique. We observe that recommending items according to the topic profile of the current browsing session seems to give poor results. Although news articles change frequently and thus data about their popularity is sparse, collaborative filtering applied to individual articles provides the best results.

Ontologies are being successfully used to overcome semanticheterogeneity, and are becoming fundamental elements of the SemanticWeb. Recently, it has also been shown that ontologies can be used tobuild more accurate and more personalized recommendation systems byinferencing missing user's preferences. However, these systemsassume the existence of ontologies, without considering theirconstruction. With product catalogs changing continuously, newtechniques are required in order to build these ontologies in realtime, and autonomously from any expert intervention.This paper focuses on this problem and show that it is possible tolearn ontologies autonomously by using clustering algorithms. Results on the MovieLens and Jester data sets show that recommendersystem with learnt ontologies significantly outperform the classical recommendation approach.