Cardinality estimation is a central task of cost-based database query optimization. Accurate estimates enable optimizers to identify and avoid expensive plans requiring large intermediate results. While cardinality estimation has been studied extensively in relational databases, research in the setting of graph databases has been more scarce. Furthermore, recent studies have shown that machine-learning-based methods can be utilized for cardinality estimation in both relational and graph databases. In this paper, we focus on the problem of estimating the cardinality of path patterns in graph databases, and we propose the Sequence-based Path Pattern Cardinality Estimator (SPACE). Our approach treats path patterns as sequences of node labels and edge types and assign similar cardinalities to path patterns with similar node and edge order. SPACE uses a dual approach: it encodes the sequence of nodes and edges to capture structural characteristics of the path pattern, while also incorporating a cardinality-based encoding to integrate cardinality information throughout learning. In a comprehensive experimental evaluation, we show that our method outperforms the state of the art in terms of both accuracy ( Q -error) and training time.

Cardinality estimation is an important step in cost-based database query optimization. The accuracy of the estimates directly affects the ability of an optimizer to identify the most efficient query execution plan correctly. In this paper, we study cardinality estimation of LIKE-queries, i.e., queries that use the LIKE-operator to match a pattern with wildcards against string-valued attributes. While both traditional and machine-learning-based approaches have been proposed to tackle this problem, we argue that they all suffer from drawbacks. Most importantly, many state-of-the-art approaches are not designed for patterns that contain wildcards in-between characters. Based on past research on neural language models, we introduce the LIKE-Pattern Language Model (LPLM) that uses a new language and a novel probability distribution function to capture the semantics of general LIKE-patterns. We also propose a method to generate training data for our model. We demonstrate that our method outperforms state-of-the-art approaches in terms of precision (Q-error), while offering comparable runtime performance and memory requirements.