With the rapid development of Fintech, increasing fraudulent behaviors have posed a threat to the smooth functioning of online lending platforms, leading to substantial losses for financial institutions. Enhancing the fraud detection capability of these platforms has thus become an urgent need. In the lending network of PPDAI, a well-known lending platform, two key characteristics are observed in the user data: (1) The presence of indicative missing values in node features; (2) The presence of weak homophily within the network structure. However, in previous work, the widely used GNNs always rely on the homophily assumption, and popular GNN-based fraud detection methods overlook the utilization of missing values, which limits the performance of lending fraud detection. To address the above problems, in this paper, we propose a GNN-based model FIN-SIGN for online lending fraud detection. The proposed model leverages Missing Mask to capture indicative information from missing values, alleviates the weak homophily through structure augmentation, and finally identifies fraudulent users with a SIGN-based GNN detector. Experimental results on the PPDAI dataset DGraph-Fin demonstrates the effectiveness of FIN-SIGN.

Given a graph G = ( V , E ) on n vertices and an integer k between 1 and n − 1 , the k -token graph F k ( G ) has vertices representing the k -subsets of V , and two vertices are adjacent if their symmetric difference is the two end-vertices of an edge in E . Using the theory of Markov chains of random walks and the interchange process, it was proved that the algebraic connectivities (second smallest Laplacian eigenvalues) of G and F k ( G ) coincide, but a combinatorial/algebraic proof has been shown elusive. In this paper, we use the latter approach and prove that such equality holds for different new classes of graphs under perturbations, such as extended cycles, extended complete bipartite graphs, kite graphs, and graphs with a cut clique. Kite graphs are formed by a graph (head) with several paths (tail) rooted at the same vertex and with exciting properties. For instance, we show that the different eigenvalues of a kite graph are also eigenvalues of its perturbed graph obtained by adding edges. Moreover, as a particular case of one of our theorems, we generalize a recent result of Barik and Verma (2024) about graphs with a cut vertex of degree n − 1 . Along the way, we give conditions under which the perturbed graph G + u v , with u v ∈ E , has the same algebraic connectivity as G .

In this paper, we propose a novel test data compression technique named CacheCompress, which combines selective encoding and dynamic dictionary based encoding. Depending on the number of specified bits, a test data word is either encoded in a single code word or as a lookup in the dictionary. Explicit dictionary initialization is not required since the content of the dictionary is updated during testing. The dictionary itself only contains the most recently used patterns, thus it exhibits a behaviour similar to a cache. Experiments show that our technique achieves higher compression ratio than other recent compression schemes while the dictionary size has been dramatically reduced.