Tree Construction Process Research Articles

A Privacy-Preserving Data Mining Through Comprehensive GNIPP Approach in Sensitive Data Sets

The quick growth of methods for analyzing data and the availability of easily available datasets have made it possible to build a thorough analytics model that can help with support decision-making. In the meantime, protecting personal privacy is crucial. A popular technique for medical evaluation and prediction, decision trees are easy to comprehend and interpret. However, the decision tree construction procedure may reveal personal information about an individual. By keeping the statistical properties intact and limiting the chance of privacy leaking within a reasonable bound, differential privacy offers a formal mathematical definition of privacy. To construct a boosting random forest that preserves privacy, we propose a Gaussian Noise Integrated Privacy Preservation (GNIPP) in this study. To address the issue of personal information breaches, we have designed a unique Gaussian distribution mechanism in GNIPP that enables the nodes with deeper depth to obtain more privacy during the decision tree construction process. We propose a comprehensive boosting technique based on the decision forest's prediction accuracy for assembling multiple decision trees into a forest. Furthermore, we propose an iterative technique to accelerate the assembly of decision trees. After all, we demonstrate through experimentation that the suggested GNIPP outperforms alternative algorithms on two real-world datasets.

PseU-KeMRF: A Novel Method for Identifying RNA Pseudouridine Sites.

Pseudouridine is a type of abundant RNA modification that is seen in many different animals and is crucial for a variety of biological functions. Accurately identifying pseudouridine sites within the RNA sequence is vital for the subsequent study of various biological mechanisms of pseudouridine. However, the use of traditional experimental methods faces certain challenges. The development of fast and convenient computational methods is necessary to accurately identify pseudouridine sites from RNA sequence information. To address this, we introduce a novel pseudouridine site prediction model called PseU-KeMRF, which can identify pseudouridine sites in three species, H. sapiens, S. cerevisiae, and M. musculus. Through comprehensive analysis, we selected four RNA coding schemes, including binary feature, position-specific trinucleotide propensity based on single strand (PSTNPss), nucleotide chemical property (NCP) and pseudo k-tuple composition (PseKNC). Then the support vector machine-recursive feature elimination (SVM-RFE) method was used for feature selection and the feature subset was optimized. Finally, the best feature subsets are input into the kernel based on multinomial random forests (KeMRF) classifier for cross-validation and independent testing. As a new classification method, compared with the traditional random forest, KeMRF not only improves the node splitting process of decision tree construction based on multinomial distribution, but also combines the easy to interpret kernel method for prediction, which makes the classification performance better. Our results indicate superior predictive performance of PseU-KeMRF over other existing models, which can prove that PseU-KeMRF is a highly competitive predictive model that can successfully identify pseudouridine sites in RNA sequences.

Application of C4.5 Algorithm Using Synthetic Minority Oversampling Technique (SMOTE) and Particle Swarm Optimization (PSO) for Diabetes Prediction

Abstract. Diabetes is the fourth or fifth leading cause of death in most developed countries and an epidemic in many developing countries. Early detection can be a preventive measure that uses a set of existing data to be processed through data mining with a classification process. Purpose: Investigate the efficacy of integrating the C4.5 algorithm with Synthetic Minority Oversampling Technique (SMOTE) and Particle Swarm Optimization (PSO) for improving the accuracy of diabetes prediction models. By employing SMOTE, the study aims to address the class imbalance issue inherent in diabetes datasets, which often contain significantly fewer instances of positive cases (diabetes) than negative cases (non-diabetes). Furthermore, by incorporating PSO, the research seeks to optimize the decision tree construction process within the C4.5 algorithm, enhancing its ability to discern complex patterns and relationships within the data. Methods/Study design/approach: This study proposes the use of the C4.5 classification algorithm by applying the synthetic minority oversampling technique (SMOTE) and particle swarm optimization (PSO) to overcome problems in the diabetes dataset, namely the Pima Indian Diabetes Database (PIDD). Result/Findings: From the research results, the accuracy obtained in applying the C4.5 algorithm without the preprocessing process is 75.97%, while the results of the SMOTE application of the C4.5 algorithm are 80%. Meanwhile, applying the C4.5 algorithm using SMOTE and PSO produces the highest accuracy, with 82.5%. This indicates an increase of 6.53% from the classification results using the C4.5 algorithm. Novelty/Originality/Value: This research contributes novelty by proposing a hybrid approach that combines the C4.5 decision tree algorithm with two advanced techniques, Synthetic Minority Oversampling Technique (SMOTE) and Particle Swarm Optimization (PSO), for the prediction of diabetes. While previous studies have explored the application of machine learning algorithms for diabetes prediction, few have examined the synergistic effects of integrating SMOTE and PSO with the C4.5 algorithm specifically.

Efficient phylogenetic tree inference for massive taxonomic datasets: harnessing the power of a server to analyze 1 million taxa.

Phylogenies play a crucial role in biological research. Unfortunately, the search for the optimal phylogenetic tree incurs significant computational costs, and most of the existing state-of-the-art tools cannot deal with extremely large datasets in reasonable times. In this work, we introduce the new VeryFastTree code (version 4.0), which is able to construct a tree on 1 server using single-precision arithmetic from a massive 1 million alignment dataset in only 36hours, which is 3 times and 3.2 times faster than its previous version and FastTree-2, respectively. This new version further boosts performance by parallelizing all tree traversal operations during the tree construction process, including subtree pruning and regrafting moves. Additionally, it introduces significant new features such as support for new and compressed file formats, enhanced compatibility across a broader range of operating systems, and the integration of disk computing functionality. The latter feature is particularly advantageous for users without access to high-end servers, as it allows them to manage very large datasets, albeit with an increase in computing time. Experimental results establish VeryFastTree as the fastest tool in the state-of-the-art for maximum likelihood phylogeny estimation. It is publicly available at https://github.com/citiususc/veryfasttree. In addition, VeryFastTree is included as a package in Bioconda, MacPorts, and all Debian-based Linux distributions.

A Decision-Focused Learning Framework for Vessel Selection Problem

Maritime transportation safety is pivotal in international trade, with port state control (PSC) inspections being crucial to vessel safety. However, port authorities need to identify substandard vessels effectively because of resource constraints and high costs. Therefore, we propose robust predictive models and optimization strategies for vessel selection, using the random forest (RF) algorithm. We first use a traditional RF model serving as a benchmark, denoted as model M0. Then, we construct model M1 by refining the RF algorithm with a batch-processing method, thereby providing a better measure of the relative relationship between the predicted deficiency counts within a batch of ships. Then, we propose model M2, incorporating a decision-focused learning (DFL) framework into the tree construction process, enhancing the decision performance of the algorithm. In addition, we propose a variant model of M2, denoted as M2-0, considering the worst-case scenario when designing the decision loss function. By conducting experiments with data from the port of Hong Kong, we demonstrate that models M1 and M2 offer superior decision-making performance compared to model M0, and model M2 outperforms model M2-0 in both decision performance and stability. We further verify the robustness of these models by testing them under various instance scales. Overall, our study enhances the PSC inspection efficiency, ultimately bolstering maritime transportation safety.

An extended smart “predict, and optimize” (SPO) framework based on similar sets for ship inspection planning

This study aims to address one critical issue in ship inspection planning optimization, where the first step is to accurately predict ship risk, and the second step is to assign scarce port inspection resources, aiming to identify as much non-compliance from the inspected ships as possible. A traditional decision tree is first developed as the benchmark. Then, to go from a good prediction to a good decision, the structure and performance of the following optimization problem are integrated in the prediction model, which we denote by integrated decision trees. Three modes are proposed to develop integrated decision trees with different combination ways and degrees. Especially, we innovatively propose the concept of “similar set” in data sets, and use the similar sets to select the hyperparameter tuple leading to the best decision optimization problem in mode 1. Then, the structure of the decision problem is considered into the decision tree construction facilitated by similar sets in mode 2. Finally, similar sets are used to integrate the performance of the following decision optimization problem directly into the decision tree construction process in mode 3. Numerical experiments show that mode 3 can achieve the best performance in the decision optimization model. Conservative estimations show that the proposed models can save at least millions to tens of millions inspection cost in Hong Kong dollars for the Hong Kong port each year, and up to 837 million inspection cost in Hong Kong dollars all over the world per year.

A modified un-realisation approach for effective data perturbation

In recent times, data has been evolving from multiple sources like social media, Facebook, Twitter, etc. in large volumes and acquiring in multiple forms. These data have multi-dimensional sensitive features from different resources entail that privacy preserving is a significant research issue. In this context, un-realisation algorithms have evolved to hide the collected data with the addition of noise to them to generate a distorted dataset while attempting privacy preservation. In this paper, a novel modified un-realisation algorithm has been proposed to generate a distorted dataset by removing duplicate elements in the dataset decreasing computational time of decision tree construction process. These techniques add noise to the original data and generate a distorted dataset by using a un-realisation algorithm. This novel approach converts the original sample datasets into different perturbed datasets by inducing the noise through set theory. It experimentally produces better results than un-realisation algorithm in terms of CPU execution time and space complexity.

372 Machine learning and pulmonary hypertension

Abstract Aims Pulmonary hypertension (IP) characterized by an average resting pulmonary pressure ≥20 mmHg can sustain various clinical conditions that differ in physiopathological, haemodynamic, and therefore therapeutic aspects. The goal of our work was to apply a machine learning algorithm that could accurately distinguish pre- and post-heart pulmonary hypertension through non-invasive methods (medical history, clinical, and echocardiographic data). Methods and results In order to achieve our goal we used the ‘decision tree’ machine learning algorithm implemented in the C5.0 package of the R development environment. The first step was the preparation of the data. The dataset of patients with IP was composed of 85 patients divided into XX precapillary IP (1) and YY postcapillary (2). Each patient is described by 11 features: some comorbidities (arterial hypertension and atrial fibrillation), BMI, right axial deviation on ECG, DLCO, and some echocardiographic measurements (e/e′, right atrial area, S wave at TDI, acceleration time on the pulmonary, inferior vena cava, diameters of the right ventricle). The dataset was divided into a data.train training subset (45 patients) and an evaluation subset (40 patients), maintaining the proportion between classes. Starting from the training dataset, the C5.0 algorithm generated the decision tree shown in Figure 1. The root node was made up of the mitral pattern e/e′, followed by the right axis deviation on the ECG and the acceleration rate on the lung that the algorithm considered the best discriminated features. The model was then validated in the validation dataset and through the Caret package and the Confusion matrix function we calculated the performance metrics of the algorithm obtaining an accuracy of 0.87, a kappa statistic of 0.742, a sensitivity of 0.913, and a specificity of 0.823. The true positive rate was 0.87 while the true negative rate was 0.87. The performance of the model was also measured using the ROC curve, obtaining an area under the curve of 0.916. Conclusions Our results show that the ‘decision tree’ algorithm starting from echocardiographic data and the ECG has a good ability to discriminate between the precapillary and postcapillary IP. In particular, the decision chain consisting of: mitral pattern and / and ratio ≤8, right axial deviation on the ECG and acceleration time on the lung ≤80 ms seems to predict the IP class with reasonable accuracy. Our results confirm that the probability of prediction and the prediction itself depend, however, on what degree of purity the partitions learned during the decision tree construction process are made up. To improve the estimation of the algorithm’s performance and thus generalize the results obtained, we believe to evaluate this approach on larger datasets also considering different machine learning algorithms. 372 Figure

Research on Online Scene Teaching Mode of Tobacco Picking Decision Tree Construction Process Integrating Deep Learning

Objectives: In view of the characteristics of online teaching during the coronavirus pandemic and the importance of practical teaching in training students’ skills in the process of graduate education, this paper proposes an online scene teaching mode that takes projects as the carrier and integrates with deep learning. In order to meet the demand for information and communication engineering professionals in the big data context, the whole teaching process is divided into four stages: Topic selection, Teaching project setting, online teaching interaction and teaching evaluation. In the teaching process of Python Data Analysis Foundations, the project “establishment process of tobacco picking decision tree based on information gain” is taken as the teaching case. Prior knowledge and references are pushed through the cloud platform before class, and The scene of tobacco picking affected by the weather is set in the online classroom to guide students to seek solutions to problems, and the results are presented with graphics to assist students to summarize, and then reset the scene to promote knowledge transfer, so as to integrate deep learning into the teaching process, and modify the corresponding stages according to the teaching evaluation results. The content of the scene is gradually increased from easy to difficult, from simple to complex, and from least to most, gradually increasing the difficulty, which enhances students’ learning interest and sense of achievement. Meanwhile, students’ initiative to participate in curriculum research further strengthens the effectiveness of the course in serving scientific research, which has a certain value of popularization and application.

Reinforcement Learning-Based Routing in Underwater Acoustic Sensor Networks

Underwater acoustic sensor networks (UASNs) play an essential role in exploring the marine world applications in rivers, ponds, and oceans. In recent years, prominent developments have addressed and resolved the numerous issues of critical underwater applications. Routing is one of the vital problems discussed widely, and various protocols are designed to solve the routing-related issues effectively. In this paper, we propose reinforcement-learning based routing using Steiner-points to minimize energy consumption and enhance the network lifetime of UASNs. Initially, the sensor nodes form a group of three nodes and calculate the Steiner point for each such group. The isolated nodes and the Steiner points in the network are vertices of the Steiner tree, which is the routing path to be constructed dynamically by the sink node agent to travel with the assistance of AUV. The sink node agent in the unknown environment initiates the Steiner tree construction process by visiting either the isolated neighbor node or neighbor sensor node’s Steiner point to collect the data from sensor nodes based on the reward function and the environmental conditions. If agent visits the isolated node, then it collects the data from that node itself. If the agent visits the Steiner point, the agent broadcasts its arrival through the beacon message and intended members of that group communicate to the agent and forwards the sensed data. The agent processes the data to remove the data’s redundancy and selects the next point to be visited and constructs the Steiner tree dynamically. The performance evaluation of the proposed scheme minimized each sensor node’s energy consumption and enhanced the network’s lifetime.

Backward Planning for a Multi-Stage Steerable Needle Lung Robot.

Lung cancer is one of the deadliest types of cancer, and early diagnosis is crucial for successful treatment. Definitively diagnosing lung cancer typically requires biopsy, but current approaches either carry a high procedural risk for the patient or are incapable of reaching many sites of clinical interest in the lung. We present a new sampling-based planning method for a steerable needle lung robot that has the potential to accurately reach targets in most regions of the lung. The robot comprises three stages: a transorally deployed bronchoscope, a sharpened piercing tube (to pierce into the lung parenchyma from the airways), and a steerable needle able to navigate to the target. Planning for the sequential deployment of all three stages under health safety concerns is a challenging task, as each stage depends on the previous one. We introduce a new backward planning approach that starts at the target and advances backwards toward the airways with the goal of finding a piercing site reachable by the bronchoscope. This new strategy enables faster performance by iteratively building a single search tree during the entire computation period, whereas previous forward approaches have relied on repeating this expensive tree construction process many times. Additionally, our method further reduces runtime by employing biased sampling and sample rejection based on geometric constraints. We evaluate this approach using simulation-based studies in anatomical lung models. We demonstrate in comparison with existing techniques that the new approach (i) is more likely to find a path to a target, (ii) is more efficient by reaching targets more than 5 times faster on average, and (iii) arrives at lower-risk paths in shorter time.

A reinforced random forest model for enhanced crop yield prediction by integrating agrarian parameters

The development in technology and science has contributed to a vast volume of data from various agrarian fields to be aggregated in the public domain. Predicting the crop yield based on climate, soil and water parameters has been a potential re- search subject. Therefore an objective arises in integrating the data with agrarian processes for crop enhancement. In this paper, a new hybrid regression-based algorithm, Reinforcement Random Forest is proposed which displays significantly enhanced performance over traditional machine learning techniques like the random forest, decision tree, gradient boosting, artificial neural network and deep Q-learning. The new strategy executes reinforcement learning at every selection of a splitting attribute amid the process of tree construction for the efficient utilization of the available samples. They analyze the variable significance measure to select the most substantial variable for node splitting process in the model development and promotes efficient utilization of training data. This integrated hybrid procedure provides significant enhancement over prevailing strategies specifically for sparse model structures. Besides performing internal cross-validation, the proposed approach requires less parameter tuning, reduces over-fitting, faster calculation and more transparent. The experimented models are evaluated with various assessment metrics like Root Mean Squared Error, Mean Squared Error, Determination Coefficient and Mean Absolute Error. The results obtained delineated that the proposed approach performs better with reduced error measures and improved accuracy of 92.2%.

A differentially private greedy decision forest classification algorithm with high utility

The rapid development of data analysis technologies and the easily accessible datasets have enabled the construction of a comprehensive analytics model, which can facilitate the decision makings involved in services. Meanwhile, the individual privacy preservation is of great necessity. Decision tree is a common method in medical prediction and diagnose, known for its simplicity of understanding and interpreting. However, the process of building a decision tree might cause individual privacy disclosure. Differential privacy provides a rigorous mathematical definition of privacy by controlling the risk of privacy leakage in a manageable range while maintaining the statistical characteristics. In this paper, we propose a Differentially Private Greedy Decision Forest with high utility (DPGDF) to build a privacy-preserving decision forest. In DPGDF, we design a novel budget allocation strategy that allows the nodes in greater depth get more privacy budgets in the decision tree construction process, which can, to some extent, mitigate the problem of excessive noises introduced to the leaf nodes. To aggregate multiple trees into a forest, we propose a selective aggregation method based on the prediction accuracy of the decision forest. In addition, we develop an iterative method to speed up the process of selective aggregation. Finally, we experimentally prove that the proposed DPGDF can achieve a better performance on two practical datasets compared with other algorithms.

Reinforced XGBoost machine learning model for sustainable intelligent agrarian applications

The development in science and technical intelligence has incited to represent an extensive amount ofdata from various fields of agriculture. Therefore an objective rises up for the examination of the available data and integrating with processes like crop enhancement, yield prediction, examination of plant infections etc. Machine learning has up surged with tremendous processing techniques to perceive new contingencies in the multi-disciplinary agrarian advancements. In this pa- per a novel hybrid regression algorithm, reinforced extreme gradient boosting is proposed which displays essentially improved execution over traditional machine learning algorithms like artificial neural networks, deep Q-Network, gradient boosting, ran- dom forest and decision tree. Extreme gradient boosting constructs new models, which are essentially, decision trees learning from the mistakes of their predecessors by optimizing the gradient descent loss function. The proposed hybrid model performs reinforcement learning at every node during the node splitting process of the decision tree construction. This leads to effective utilizationofthesamplesbyselectingtheappropriatesplitattributeforenhancedperformance. Model’sperformanceisevaluated by means of Mean Square Error, Root Mean Square Error, Mean Absolute Error, and Coefficient of Determination. To assure a fair assessment of the results, the model assessment is performed on both training and test dataset. The regression diagnostic plots from residuals and the results obtained evidently delineates the fact that proposed hybrid approach performs better with reduced error measure and improved accuracy of 94.15% over the other machine learning algorithms. Also the performance of probability density function for the proposed model delineates that, it can preserve the actual distributional characteristics of the original crop yield data more approximately when compared to the other experimented machine learning models.

Scalable Construction of Clock Trees With Useful Skew and High Timing Quality

Clock trees can be constructed based on static arrival time constraints or dynamic implied skew constraints. Dynamic implied skew constraints allow the full timing margins to be utilized. However, the dynamic skew constraints require a high run-time complexity to be evaluated. In contrast, static arrival time constraints are more restrictive but can be evaluated in constant time. Consequently, there is a tradeoff between timing margin utilization and run-time. In this paper, a scalable clock tree synthesis (CTS) framework is proposed for the construction of low-cost useful skew trees (USTs) with high timing quality. The scalability is based on combining the use of arrival time constraints with virtual minimum and maximum delay offsets, which facilitates that a pair of smaller subtrees can be joined into a larger subtree in constant time. The ability to quickly join subtrees is leveraged to perform a high degree of solution space exploration, which translates into the construction of USTs with low-cost. In particular, clock trees with various routing tree topologies, buffer tree topologies, buffer sizes, and stem wire lengths are explored. Moreover, the arrival time constraints are specified with the objective of being the least restrictive to reduce cost. Furthermore, the constraints are respecified throughout the tree construction process using a slack graph (SG) to expose additional timing margins. The high timing quality is obtained by seamlessly integrating arbitrary timing models using the SG. Finally, the proposed CTS framework is integrated with a clock tree optimization framework to demonstrate that the constructed USTs are capable of meeting timing constraints under the influence of on-chip variations.

Consistent estimation of residual variance with random forest Out-Of-Bag errors

The issue of estimating residual variance in regression models with unknown and eventually complex link-function is still an open problem. Predictions of such outcomes are usually conducted by black-box procedures without seriously restricting the link-function class. However, quantifying uncertainty by means of residual variance estimators is of primary interest in many practical applications, e.g. as a primary step towards the construction of prediction intervals. Here, we consider this issue for the random forest. Therein, the functional relationship between covariates and response variable is modeled by a weighted sum of the latter. The dependence structure is, however, involved in the weights that are constructed during the tree construction process making the model complex in mathematical analysis. Restricting to L2-consistent random forest models, we provide random forest based residual variance estimators and prove their consistency.

On the Problem of Generating a Large Number of Comparable Test Variants

The paper presents a possible way of solving the problem of creating more test variants for a large number of students divided into groups. The proposed solution may consist in introducing a parameterized automatic test generator. The principle of an automatic parameterized test generator is shown. The process of the question tree construction according to the increasing numbers of questions in the banks of the particular subjects leads to a combinatorial explosion. This often results in the excessive time of generation of the different variants of tests. To solve this problem, a heuristic method based on a pre-processing stage that precedes the construction of the searching tree is proposed. Further, the results of the experiments comparing the time of the test generation and the congruence of the test variants generated by the algorithm either using or non-using this heuristics are presented. According to these results, the use of the generator with the proposed heuristics provides a considerably shorter time of generation, and the congruence of the generated test variants is even better in most cases.

Быстродействующие реализации метода Крускала построения минимального остова графа

Constructing the minimum spanning tree of a weighted graph falls in the category of fundamental problems in the graph theory. The minimum spanning tree of a graph is generally applied for setting up communication between the nodes of discrete systems that is optimal in terms of cost (weight). Two basic methods for constructing the minimum spanning tree were known before the advent of modern computers. Reduction of computational complexity has become the main line of improving the methods. Inventions made in the field of computational structures facilitated these efforts. The previously published paper by V.S. Zubov and V.V. Kraskov gave a detailed analysis of the most sophisticated implementation of the Prim method. The competing Kruskal’s method features a high computational complexity of its first stage, at which the graph edges are sorted in ascending order of their weights. It is exactly this stage that determines the overall complexity of the method. The article analyzes the effect from constructing special data structures for the first stage of the method. As a result, the status of the Kruskal method has been changed, which becomes faster than the Prim method. Three implementations of the Kruskal method are proposed, in which an idea of partially ordering the set of edges is used. In the first of them, the weak pyramid is used. The number of edge comparison operations in the course of its construction is minimal. The average number of edges extracted from the pyramid for constructing the minimum spanning tree is L ≈ V(0.333 log2V + 0.45). The average time taken to construct the minimum spanning tree has become shorter, but the Prim method still remains more advantageous. A so-called dth pyramid was not previously used in the Kruskal method. Its special implementation contains a fewer average number of not only the main operations, but also a fewer number of the auxiliary operations. The computational complexity of the pyramid construction stage was found to be minimal at d > 4. A formula for estimating the average time taken to construct the minimum spanning tree is given, in which the first, second, and third terms correspond, respectively, to the time taken to construct the pyramid, to extract L edges from it, and to check the edges and include them into the minimum spanning tree; the formula also contains machine-dependent coefficients which are determined experimentally. This implementation is faster than the Prim method, except for low-density graphs. The system of chained lists (stacks) can be used to speed up the edge ordering process at the first stage of the new implementation of the Kruskal method. Edge lists are drawn up within a single graph description viewing cycle. Each list includes edges of equal weight. The lists are used in ascending order of edge weight, until the minimum spanning tree construction process is completed. This implementation is significantly faster than the Prim method. The new implementations of the method were tested on graphs random (pseudorandom) in terms of edge content and weights. References to the implementation codes are also given in the article.

Cost-sensitive feature selection using random forest: Selecting low-cost subsets of informative features

Feature selection aims to select a small subset of informative features that contain most of the information related to a given task. Existing feature selection methods often assume that all the features have the same cost. However, in many real world applications, different features may have different costs (e.g., different tests a patient might take in medical diagnosis). Ignoring the feature cost may produce good feature subsets in theory but they can not be used in practice. In this paper, we propose a random forest-based feature selection algorithm that incorporates the feature cost into the base decision tree construction process to produce low-cost feature subsets. In particular, when constructing a base tree, a feature is randomly selected with a probability inversely proportional to its associated cost. We evaluate the proposed method on a number of UCI datasets and apply it to a medical diagnosis problem where the real feature costs are estimated by experts. The experimental results demonstrate that our feature-cost-sensitive random forest (FCS-RF) is able to select a low-cost subset of informative features and achieves better performance than other state-of-art feature selection methods in real-world problems.

Fast Localization in Large-Scale Environments Using Supervised Indexing of Binary Features.

The essence of image-based localization lies in matching 2D key points in the query image and 3D points in the database. State-of-the-art methods mostly employ sophisticated key point detectors and feature descriptors, e.g., Difference of Gaussian (DoG) and Scale Invariant Feature Transform (SIFT), to ensure robust matching. While a high registration rate is attained, the registration speed is impeded by the expensive key point detection and the descriptor extraction. In this paper, we propose to use efficient key point detectors along with binary feature descriptors, since the extraction of such binary features is extremely fast. The naive usage of binary features, however, does not lend itself to significant speedup of localization, since existing indexing approaches, such as hierarchical clustering trees and locality sensitive hashing, are not efficient enough in indexing binary features and matching binary features turns out to be much slower than matching SIFT features. To overcome this, we propose a much more efficient indexing approach for approximate nearest neighbor search of binary features. This approach resorts to randomized trees that are constructed in a supervised training process by exploiting the label information derived from that multiple features correspond to a common 3D point. In the tree construction process, node tests are selected in a way such that trees have uniform leaf sizes and low error rates, which are two desired properties for efficient approximate nearest neighbor search. To further improve the search efficiency, a probabilistic priority search strategy is adopted. Apart from the label information, this strategy also uses non-binary pixel intensity differences available in descriptor extraction. By using the proposed indexing approach, matching binary features is no longer much slower but slightly faster than matching SIFT features. Consequently, the overall localization speed is significantly improved due to the much faster key point detection and descriptor extraction. It is empirically demonstrated that the localization speed is improved by an order of magnitude as compared with state-of-the-art methods, while comparable registration rate and localization accuracy are still maintained.