Yelp Open Dataset (YOD) is a widely used dataset for Recommender Systems (RS). Multiple Knowledge Graphs (KGs) have been built for YOD, but they have various issues: the conversion processes usually do not follow state-of-the-art methodologies, fail to properly link to other KGs, do not link to existing vocabularies, ignore important data, and are generally of small size. Instead, we present the Yelp Collaborative Knowledge Graph (YCKG), where we correctly integrating taxonomies, product categories, business locations, and the Yelp social network, through common practices within the semantic web community, overcoming all these issues. As a result, the YCKG includes 150k businesses and 16.9M reviews from 1.9M distinct real users, resulting in over 244 million triples, 144 distinct predicates, for about 72 million resources, with an average in-degree and out-degree of 3.3 and 12.2, respectively. Further, we release both the data and the code used to generate the KG for inspection and further extensions. This dataset can be used to develop and test both recommendation and data-mining algorithms able to exploit rich and semantically meaningful knowledge. We publicize the code for the CKG construction on: https://github.com/MadsCorfixen/The-Yelp-Collaborative-Knowledge-Graph.

Abstract Usually, recommender systems are trained on a set of users and items and then used to recommend new user-item pairings among those seen during training. As users and items are added continuously, there is a pressing need to provide recommendations for new users and items, i.e., for users and items not seen during training. Solutions to this problem exploit techniques like meta-learning or auxiliary information encoded in knowledge graphs to learn an “inductive bias”. Yet, most existing works can either recommend for new users or new items not seen during training but not both. Further, existing methods have rarely been compared to each other. Finally, existing evaluations of these methods use a random split of training data, and thus do not consider temporal splits of ratings in training and testing. This setting ensures testing is correctly performed on user interactions that actually occur after the training period. In this paper, we propose a framework for training and testing the methods on three real world datasets, and perform a deeper analysis of each dataset to better understand the effect of emerging popularity trends. As a result, our re-evaluation of state-of-the-art methods identifies strong architectures and solutions for inductive recommendation. We find that inductive methods that perform aggregation are able to outperform non-aggregating methods in all settings; performances vary greatly across settings, pointing to new important research questions.

Inductive Recommender Systems are capable of recommending for new users and with new items thus avoiding the need to retrain after new data reaches the system. However, these methods are still trained on all the data available, requiring multiple days to train a single model, without counting hyperparameter tuning. In this work we focus on graph-based recommender systems, i.e., systems that model the data as a heterogeneous network. In other applications, graph sampling allows to study a subgraph and generalize the findings to the original graph. Thus, we investigate the applicability of sampling techniques for this task. We test on three real world datasets, with three state-of-the-art inductive methods, and using six different sampling methods. We find that its possible to maintain performance using only 50% of the training data with up to 86% percent decrease in training time; however, using less training data leads to far worse performance. Further, we find that when it comes to data for recommendations, graph sampling should also account for the temporal dimension. Therefore, we find that if higher data reduction is needed, new graph based sampling techniques should be studied and new inductive methods should be designed.

Knowledge Graphs (KGs) have been integrated in several models of recommendation to augment the informational value of an item by means of its related entities in the graph. Yet, existing datasets only provide explicit ratings on items and no information is provided about users' opinions of other (non-recommendable) entities. To overcome this limitation, we introduce a new dataset, called the MindReader dataset, providing explicit user ratings both for items and for KG entities. In this first version, the MindReader dataset provides more than 102 thousands explicit ratings collected from 1,174 real users on both items and entities from a KG in the movie domain. This dataset has been collected through an online interview application that we also release as open source. As a demonstration of the importance of this new dataset, we present a comparative study of the effect of the inclusion of ratings on non-item KG entities in a variety of state-of-the-art recommendation models. In particular, we show that most models, whether designed specifically for graph data or not, see improvements in recommendation quality when trained on explicit non-item ratings. Moreover, for some models, we show that non-item ratings can effectively replace item ratings without loss of recommendation quality. This finding, in addition to an observed greater familiarity from users towards certain descriptive entities than movies, motivates the use of KG entities for both warm and cold-start recommendations.