Recommenders take data, analyze it, and on the basis of these results offer recommendations to end users about the decisions to make: which restaurant would be the most appropriate considering their preferences and location, what concert to attend, and so on. They are indeed a main part of digital assistants; these software agents are an important piece of the digital transformation our society is experiencing. This means that improvements on recommenders is an active field of work.
This paper presents a recommendation model. As the title suggests, its main significant attributes are: 1) being general and 2) being designed for heterogeneous networks. The recommendation problem is treated as a graph problem in which heterogeneous entities are represented as different types of nodes, and interactions between entities are modeled as different types of edges. This method, called HeteRS, extends on previous works on recommenders: it extends multivariate Markov chain (MMC) models by introducing an optimization framework to automatically learn the influence weights of each type of relation between nodes. The thesis is that these weights vary depending on the recommendation item that is of interest for a user in each query, for example, time versus geographical proximity.
The paper starts by describing the problem treated --recommendations in heterogeneous networks-- and contextualizing the new approach, HeteRS, with respect to previous solutions for recommenders. The contributions are clearly stated from the first section of the paper. In the following section, related work on recommendation is reviewed and the limitations that motivated this research are described. Next, the proposal is presented. Finally, in order to validate it and compare it with previous approaches, HeteRS and baseline methods are applied to three event databases, and the results obtained with all methods are compared. The datasets are public, which is good when considering the reproducibility of experiments by other researchers.
This is a scientific paper, suitable for researchers working with the problem of recommenders. For this community, it is definitely of interest. Contributions are clearly stated and the validation is correctly documented. The experiments can be reproduced or used as a starting point for further research, which is a valuable feature in research.