A case study in a recommender system based on purchase data

Abstract

Collaborative filtering has been extensively studied in the context of ratings prediction. However, industrial recommender systems often aim at predicting a few items of immediate interest to the user, typically products that (s)he is likely to buy in the near future. In a collaborative filtering setting, the prediction may be based on the user's purchase history rather than rating information, which may be unreliable or unavailable. In this paper, we present an experimental evaluation of various collaborative filtering algorithms on a real-world dataset of purchase history from customers in a store of a French home improvement and building supplies chain. These experiments are part of the development of a prototype recommender system for salespeople in the store. We show how different settings for training and applying the models, as well as the introduction of domain knowledge may dramatically influence both the absolute and the relative performances of the different algorithms. To the best of our knowledge, the influence of these parameters on the quality of the predictions of recommender systems has rarely been reported in the literature.

Towards ‘Human/System Synergistic Development’: How Emergent System Characteristics Change Software Development

“…2) We find that the performance of simpler but faster algorithms such as association rules can often be superior to more elaborate techniques such as matrix factorization (MF). This confirms the results of other recent works on purchase data, such as [3]. 3) Our experiments show that training on all available data is inferior to training on recent purchases for most algorithms.…”

“…Just as in our study, they experimented on a fashion and a book store. A more recent case study by Pradel et al [3] experimentally evaluated various CF algorithms on a dataset of a French home improvement and building supplies chain. The simple bigram rules recommender was found to yield the best overall accuracy.…”

“…In [6], the reduced training set contains a fraction of randomly sampled events, whereas in [3] it contains the event history of the last two weeks of the training data. In our paper, we adopt a more flexible setting where we evaluate algorithms in various history lengths (see Section III-B).…”

“…That was already apparent in 2006, when Amazon 2 claimed that 35% of its sales were triggered by recommendations [1]. The same year, Netflix 3 announced its famous million-dollar prize for improving its recommendation engine. This contest has spurred active research within the field, yielding a variety of recommendation algorithms over the past decade.…”

“…The connection between research contributions and industrial solutions has rarely been in focus of RS literature. As Pradel et al [3] noted, "casestudies are necessary to better understand the specificities of purchase datasets and the factors that impact recommender systems for retailers". Our work is such a case study.…”