The Titanic-related data was originally compiled by the British Board of Trade as part of its investigation into the tragic sinking of the Royal Mail Ship Titanic. For many years, research on the Titanic disaster remained largely in the domains of historians and enthusiasts. Its popularity in the machine learning community surged after Kaggle released a curated version of the dataset for a data analysis competition. Since then, it has been widely adopted for data science education and research, including its use in teaching data preprocessing and analysis, as well as benchmarking the performance of different machine learning algorithms. However, there is a previously overlooked flaw in this dataset: this paper shows that the average passenger class fares computed from the Kaggle dataset differ substantially from those published by NBC Los Angeles News in June 2023. In particular, the incorrect assignment of group fares to individual passenger fares has caused systematic inflation of fare values, potentially leading to misinterpretations over the years. A methodological correction for the Fare attribute is proposed, whereby group fares are divided equally among all passengers within the same travel group. This adjustment yields a significant 15.6% improvement in Spearman's correlation between the fare and passenger class. Additionally, experimental results demonstrate that fare correction improves prediction performance in classification and regression tree. It is hoped that this correction will enhance the dataset's utility for future education and research.

This paper showcases the application of a clustering method for credit card fraud detection. Unlike supervised approaches, we micro-cluster data into small homogeneous subsegments related to fraud or non-fraud instances. We then sample each cluster. For each sampled record, the true target label is acquired. Through this approach, we significantly reduce the cost of detecting fraudulent cases. We regard clusters with fraud rates significantly higher than the estimated basic fraud rate as high-risk clusters and proceed to investigate the rest of the cases in those clusters. By targeting the high-risk clusters, we achieved a 500 -fold improvement over baseline precision for a credit card fraud detection application. In addition, by targeting only high-risk clusters, the proposed approach achieves significant cost reduction and is especially useful when fraud cases are not known beforehand.

This paper introduces a framework for identifying outliers in predictions made by regression tree models. Existing robust regression approaches tend to focus on the construction stage, which builds regression models that are less sensitive to outliers. In contrast, our approach focuses on identifying outliers during the prediction stage. The process of our proposed approach begins with building a regression tree using a training dataset. Predictions significantly deviating from the mean within each terminal node are automatically labeled as outliers. We show how the labelled data can be explored to better understand the characteristics of the outliers. We also identify the situations under which the data exploration may not work well. Further, we make use of the outlier labels and training data to construct an anomaly detector. Our results show that the proposed method can effectively detect outliers that may exist within datasets. Such outliers, when removed, result in improved data quality. Insights into its effectiveness and potential caveats are also discussed. Received: 18 July 2024 | Revised: 27 August 2024 | Accepted: 18 September 2024 Conflicts of Interest The author declares that he has no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in Kaggle at https://doi.org/10.34740/KAGGLE/DSV/9355696. Author Contribution Statement Swee Chuan Tan: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing - original draft, Writing - review & editing, Visualization, Supervision, Project administration.

It is a common knowledge that ensemble learning can be used to improve performance of predictive models, but it may not work well under certain situations, such as with stable learners. Thus, ensemble approach may come across to some practitioners as only an option when searching for an accurate model. In this paper, we highlight a situation where ensemble learning is not an option. Particularly with regression trees, we argue that ensemble learning is indispensable in enhancing the estimation performance. Regression trees, while popular, are known to produce rough regression surfaces, resulting in imprecise estimations when applied as a single individual tree model. This paper shows that ensemble learning addresses this issue by combining predictions from a diverse set of regression trees, effectively smoothing the regression surface. Experimental results across five real-world datasets show that ensemble learning consistently reduces predictive errors of regression trees, and the performance of single regression trees are consistently inferior. This underscores the efficacy of ensemble learning in smoothing prediction surfaces and reducing errors. For stable learners that do not suffer from the problem of rough prediction surfaces, such as Linear Regression and Support Vector Regression, ensemble learning offers limited to no improvements over single models. These findings emphasize the necessity of applying ensemble learning whenever regression trees are employed.

This paper presents a framework for detecting outliers found in the predictions made by regression tree models. The idea is to build a regression tree using a training dataset, and then automatically label predictions that are significantly different from the mean prediction within each terminal node, as outliers. The training inputs along with the automatically derived outlier labels are then used to construct an anomaly detector. To evaluate the framework, we send the test input data to the regression tree and anomaly detector for estimations and anomaly detections, respectively. For each test instance that has been predicted as an outlier, the corresponding value estimated by the regression tree is marked for further treatments. In our case, we simply choose to discard all estimated values that are flagged as outliers. The results show significant reductions in the Mean Absolute Errors of regression tree predictions when the approach is applied to four datasets. This method is potentially useful when a proportion of suspicious predictions can be discarded without affecting the overall quality of a system. Our analyses also provide some insights into why the approach works, as well as caveats when applying this approach.

This paper introduces a binary search algorithm for determining the optimal probability cut-point value ( C ) of binary classifiers. Cut-points are operating points on the receiver operating characteristic curve that divide positive and negative predictions. Compared to the traditional exhaustive search for optimal C value, the proposed method offers execution time efficiency ( O (log 2 ( k ))) and a small cut-point error of 1/2 n after k steps of binary search. Traditionally, the optimal C value is determined by stepping through all possible C values. This search is uninformed because there is no indication of the search direction. To address this issue, we derive the expectation of the F-Measure (aka F 1 score); and use it to guide the search process. Specifically, by comparing the F-Measure at the current cut-point with the F-Measure at expected cut-point, we can use the information to adjust C dynamically towards the optimal cut-point, resulting in optimal model performance. Our results on two classifiers trained from disease classification datasets suggest that the algorithm is robust and efficient, as compared to the traditional methods.

This paper presents a conceptual framework for developing and implementing visual analytics capabilities for operational processes. The framework has three key aspects. Firstly, the inputs, resources, and outputs of an operational process are to be defined. Secondly, the key measurements of inputs, outputs, and resources are identified. Thirdly, the time intervals for taking the identified measurements are determined. This framework, albeit simple, provides users with a methodical approach to develop a useful and effective visual analytics solution for monitoring, understanding, and diagnosing problems in operational processes. Finally, we present a case where this proposed framework has been applied in a real-world project, where visual analytics had been applied to help a call center identify bottlenecks, better manage manpower resources and improve service level performance.

Star Coordinate Plot is a simple and efficient technique for visualizing multidimensional data. Since the proposal of this method in early 2000, several researchers have attempted to address its weakness of tending to project data points toward the origin of the star coordinate space. But so far no one has provided a critical analysis of the issue in the literature. As a result, the weakness of Star Coordinate Plot is still not well understood. In this paper, we first provide an explanation of its weakness by pointing out two design constraints in the original Star Coordinate Plot. We show how these constraints result in three categories of data points that are lost in the process of translating from n - dimensional space to a two-dimensional star coordinate space. We then propose the Enhanced Star Coordinate data visualization method to address these constraints. Our experimental results show that the proposed method is superior to the original Star Coordinate Plot on several datasets used for evaluations. (c) 2018 Elsevier B.V. All rights reserved.

Association rule mining is an important data mining technique that help discover interesting attribute relationships that are useful for decision making. Most association rule mining methods use item-set manipulation approach, whereby data type must be categorical in nature. When a dataset contains numerical attributes, they will need to be discretized before rule mining. At the moment, most unsupervised data discretization methods do not account for data distributions, and users have to try different methods and discretization settings in order to improve rule mining results. In this paper, we propose using TwoStep clustering for data discretization. Unlike simple discretization methods, TwoStep automatically determines the discretization intervals by taking into account the unique data distribution property of each attribute. In our experiments, we evaluated the performance of Apriori algorithm based on four datasets, whereby each dataset was pre-processed using TwoStep and three other commonly used discretization methods. Our results show that TwoStep produced the greatest number of high-quality rules, as compared to common discretization methods.

The Star Coordinate Plot is a simple and efficient technique for visualizing multidimensional data. Since the proposal of this method in early 2000, several researchers have attempted to address its weakness of tending to project data points towards the origin of the star coordinate space. Despite such attempts, the weakness of Star Coordinate Plot is still not well understood. This paper gives the reasons for the weakness by pointing out two fundamental errors in the design of the original Star Coordinate Plot. We show how these errors result in loss of data points in the process of translating from n-dimensional space to the two-dimensional Star Coordinate space. We then propose the Enhanced Star Coordinate data visualization method to address these issues. Our experimental results show that the proposed method is superior to the original Star Coordinate Plot on several datasets used for evaluations.

This paper presents a data analytics approach for recovering test-pad information from images of printed circuit boards. The main aim is to obtain highly accurate information as input to a robotic flying probe tester. Such a tester is a mechatronic system that is able to perform a great variety of diagnostic testing on printed circuit boards without any additional circuit board documentation. In this work, a two-stage clustering process was applied on a dataset with 71040 pixel records obtained from an electronic circuit board image. In total, the method discovered 128 locations on the circuit board that are potentially the test pads. Visual inspection found that all the 120 legitimate test pads on the circuit board were retrieved. The other eight locations were not really test pads and were removed (i.e., Recall = 100%, and Precision = 93.25%). We propose this image analytics approach as an effective way to speed up the recovery of test-pad locations from printed circuit boards.

This paper studies the problem of finding similar time series of product sales in transactional data. We argue that finding such similar time series can lead to discovery of interesting and actionable business information such as previously unknown complementary products or substitutes, and hidden supply chain information. However, finding all possible pairs of n time series exhaustively results in O(n 2) time complexity. To address this issue, we propose using k-means clustering method to create small clusters of similar time series, and those clusters with very small intra-cluster variability are used to find similar time series. Finally, we demonstrate the utility of our approach to derive interesting results from real-life data.

Irrelevant attributes in real-world data sets are known to affect data mining performance as well as making a model to become opaque and hard to interpret. In supervised learning, the problem of attribute selection is commonly solved using the Supervised Wrapper Approach (SWA). This approach is well established---it searches for an attribute-subset that gives the best predictive accuracy attained by a (wrapped) predictive model. Hence, the prerequisite for SWA to work is the availability of target variables. In unsupervised learning, target variables do not exist, and the applicability of SWA to unsupervised learning problems seems counterintuitive. This paper demonstrates how SWA can be used for attribute selection in data clustering, an important task in unsupervised learning. Experimental results show that the proposed method can be used to find relevant attributes and remove irrelevant ones, resulting in smaller data size and lower dimensionality, and improved clustering performance.

Many real-world data sets exhibit imbalanced class distributions in which almost all instances are assigned to one class and far fewer instances to a smaller, yet usually interesting class. Building classification models from such imbalanced data sets is a relatively new challenge in the machine learning and data mining community because many traditional classification algorithms assume similar proportions of majority and minority classes. When the data is imbalanced, these algorithms generate models that achieve good classification accuracy for the majority class, but poor accuracy for the minority class. This paper reports our experience in applying data balancing techniques to develop a classifier for an imbalanced real-world fraud detection data set. We evaluated the models generated from seven classification algorithms with two simple data balancing techniques. Despite many ideas floating in the literature to tackle the imbalanced issue, our study shows the simplest data balancing technique is all that is required to significantly improve the accuracy in identifying the primary class of interest (i.e., the minority class) in all the seven algorithms tested. Our results also show that precision and recall are useful and effective measures for evaluating models created from artificially balanced data. Hence, we advise data mining practitioners to try simple data balancing first before exploring more sophisticated techniques to tackle the class imbalance problem.

Association rule mining is an important data mining method primarily used for market basket analysis. However, the method usually generates a large number of association rules; and it is difficult to use domain-independent objective measures to help find pragmatically important rules. To address these issues, we present a general method that succinctly summarizes rules with common consequent(s). This consequent-based approach allows user to focus on evaluating a rule set based on the practical significance of consequent(s) in an application domain, which usually outweighs the importance of objective measures such as rule confidence. We provide a case study to demonstrate how the proposed method can be used in conjunction with a heuristic procedure to find important rules generated from large real-world data, leading to discovery of important business knowledge and insights.

Scatter plot is a useful method for visualising clusters and outliers in continuous data. However, this method cannot be used directly on nominal data due to a lack of natural ordering and ‘distance’ in nominal values. One solution to this problem is to map the multi-dimensional nominal data to a numeric space, and then draw a scatter plot of the data points based on the first two principal components of the numeric space. This paper reports a study on how such plots can be generated using three types of mapping: (a) Binary Input Mapping (BImap), (b) Attribute Value Frequency Mapping (AVFmap), and (c) BImap combined with AVFmap. Results show that the combined method draws upon the complementary strengths of BImap and AVFmap, to generate meaningful scatter plots for visualising categorical outliers and achieve the highest information gain among the methods tested.

Market Basket Analysis often involves applying the de facto association rule mining method on massive sales transaction data. In this paper, we argue that association rule mining is not always the most suitable method for analysing big market-basket data. This is because the data matrix to be used for association rule mining is usually large and sparse, resulting in sluggish generation of many trivial rules with little insight. To address this problem, we summarise a real-world sales transaction data set into time series format. We then use time series clustering to discover commonly purchased items that are useful for pricing or formulating cross-selling strategies. We show that this approach uses a data set that is substantially smaller than the data to be used for association analysis. In addition, it reveals significant patterns and insights that are otherwise hard to uncover when using association analysis.