Rua-Huan Tsaih

Unsupervised neural networks approach for understanding fraudulent financial reporting

Purpose – Creditor reliance on accounting‐based numbers as a persistent and traditional standard to assess a firms financial soundness and viability suggests that the integrity of financial statements is essential to credit decisions. The purpose of this paper is to provide an approach to explore fraudulent financial reporting (FFR) via growing hierarchical self‐organizing map (GHSOM), an unsupervised neural network tool, to help capital providers evaluate the integrity of financial statements, and to facilitate analysis further to reach prudent credit decisions.Design/methodology/approach – This paper develops a two‐stage approach: a classification stage that well trains the GHSOM to cluster the sample into subgroups with hierarchical relationship, and a pattern‐disclosure stage that uncovers patterns of the common FFR techniques and relevant risk indicators of each subgroup.Findings – An application is conducted and its results show that the proposed two‐stage approach can help capital providers evalua...

Topological pattern discovery and feature extraction for fraudulent financial reporting

Fraudulent financial reporting (FFR) involves conscious efforts to mislead others regarding the financial condition of a business. It usually consists of deliberate actions to deceive regulators, investors or the general public that also hinder systematic approaches from effective detection. The challenge comes from distinguishing dichotomous samples that have their major attributes falling in the same distribution. This study pioneers a novel dual GHSOM (Growing Hierarchical Self-Organizing Map) approach to discover the topological patterns of FFR, achieving effective FFR detection and feature extraction. Specifically, the proposed approach uses fraudulent samples and non-fraudulent samples to train a pair of dual GHSOMs under the same training parameters and examines the hypotheses for counterpart relationships among their subgroups taking advantage of unsupervised learning nature and growing hierarchical structures from GHSOMs. This study further presents (1) an effective classification rule to detect FFR based on the topological patterns and (2) an expert-competitive feature extraction mechanism to capture the salient characteristics of fraud behaviors. The experimental results against 762 annual financial statements from 144 public-traded companies in Taiwan (out of which 72 are fraudulent and 72 are non-fraudulent) reveal that the topological pattern of FFR follows the non-fraud-central spatial relationship, as well as shows the promise of using the topological patterns for FFR detection and feature extraction.

Exploring Fraudulent Financial Reporting with GHSOM

The issue of fraudulent financial reporting has drawn much public as well as academic attention. However, most relevant researches focus on predicting financial distress or bankruptcy. Little emphasis has been placed on exploring the financial reporting fraud itself. This study addresses the challenge of obtaining an enhanced understanding of the financial reporting fraud through the approach with the following four phases: (1) to identify a set of financial and corporate governance indicators that are significantly correlated with fraudulent financial reporting; (2) to use the Growing Hierarchical Self-Organizing Map (GHSOM) to cluster data from listed companies into fraud and non-fraud subsets; (3) to extract knowledge from the fraudulent financial reporting through observing the hierarchical relationship displayed in the trained GHSOM; and (4) to provide justification to the extracted knowledge.

The prediction approach with Growing Hierarchical Self-Organizing Map

The competitive learning nature of the Growing Hierarchical Self-Organizing Map (GHSOM), which is an unsupervised neural networks extended from Self-Organizing Map (SOM), can work as a regularity detector that is supposed to help discover statistically salient features of the sample population. With the spatial correspondent assumption, this study presents a prediction approach in which GHSOM is used to help identify the fraud counterpart of each non-fraud subgroup and vice versa. In this study, two GHSOMs-a non-fraud tree (NFT) and a fraud tree (FT) are generated via the non-fraud samples and the fraud samples, respectively. Each (fraud or non-fraud) training sample is classified into its belonging leaf nodes of NFT and FT. Then, two classification rules are tuned based upon all training samples to determine the associated discrimination boundary within each leaf node, and the rule with better classification performance is chosen as the prediction rule. With the spatial correspondent assumption, the prediction rule derived from such an integration of FT and NFT classification mechanisms should work well. This study sets up the experiment of fraudulent financial reporting (FFR), a sub-field of financial fraud detection (FFD), to justify the effectiveness of the proposed prediction approach and the result is quite acceptable.

Reasoning neural networks

The Reasoning Neural Network (RN) has a learning algorithm belonging to the weight-and-structure-change category, because it puts only one hidden node in the initial network structure, and will recruit and prune hidden nodes during the learning process. Empirical results show that learning of the RN is guaranteed to be completed, the number of required hidden nodes is reasonable, that the speed of learning is much faster than back propagation networks, and that the RN is able to develop good internal representation.

Network-traffic anomaly detection with incremental majority learning

Detecting anomaly behavior in large network traffic data has presented a great challenge in designing effective intrusion detection systems. We propose an adaptive model to learn majority patterns under a dynamic changing environment. We first propose unsupervised learning on data abstraction to extract essential features of samples. We then adopt incremental majority learning with iterative evolutions on fitting envelopes to characterize the majority of samples within moving windows. A network traffic sample is considered an anomaly if its abstract feature falls on the outside of the fitting envelope. We justify the effectiveness of the presented approach against 150000+ traffic samples from the NSL-KDD dataset in training and testing, demonstrating positive promise in detecting network attacks by identifying samples that have abnormal features.

Clustering iOS executable using self-organizing maps

We pioneer the study on applying both SOMs and GHSOMs to cluster mobile apps based on their behaviors, showing that the SOM family works well for clustering samples with more than ten thousands of attributes. The behaviors of apps are characterized by system method calls that are embedded in their executable, but may not be perceived by users. In the data preprocessing stage, we propose a novel static binary analysis to resolve and count implicit system method calls of iOS executable. Since an app can make thousands of system method calls, it is needed a large dimension of attributes to model their behaviors faithfully. On collecting 115 apps directly downloaded from Apple app store, the analysis result shows that each app sample is represented with 18000+ kinds of methods as their attributes. Theoretically, such a sample representation with more than ten thousand attributes raises a challenge to traditional clustering mechanisms. However, our experimental result shows that apps that have similar behaviors (due to having been developed from the same company or providing similar services) can be clustered together via both SOMs and GHSOMs.

A resistant learning procedure for coping with outliers

In the context of resistant learning, outliers are the observations far away from the fitting function that is deduced from a subset of the given observations and whose form is adaptable during the process. This study presents a resistant learning procedure for coping with outliers via single-hidden layer feed-forward neural network (SLFN). The smallest trimmed sum of squared residuals principle is adopted as the guidance of the proposed procedure, and key mechanisms are: an analysis mechanism that excludes any potential outliers at early stages of the process, a modeling mechanism that deduces enough hidden nodes for fitting the reference observations, an estimating mechanism that tunes the associated weights of SLFN, and a deletion diagnostics mechanism that checks to see if the resulted SLFN is stable. The lake data set is used to demonstrate the resistant-learning performance of the proposed procedure.

Challenges deploying complex technologies in a traditional organization

The National Palace Museum in Taiwan had to partner with experienced cloud providers to deliver television-quality exhibits.

Resistant learning on the envelope bulk for identifying anomalous patterns

Anomalous patterns are observations that lie far away from the fitting function deduced from the bulk of the given observations. This work addresses the research issue to effectively identify anomalous patterns in both contexts of resistant learning, where there is no assumption about the fitting function form, and of changing environments. The resistant learning means that the learning procedure is not impacted significantly by the outlying observations. In literature, there is the resistant learning with searching a near-perfect fitting function for identifying the bulk of the majority of observations. However, the learning algorithm with searching a near-perfect fitting function suffers from time inefficiency. To effectively identify anomalous patterns in both contexts of resistant learning and changing environments, this study proposes a new resistant learning algorithm with envelope module that learns to evolve a nonlinear fitting function wrapped with a constant-width envelope for containing the majority of observations and thus identifying anomalous patterns. An illustrative experiment is set up to justify the effectiveness of the envelope module and the experimental result shows the positive promise.