High-dimensional Feature Space Research Articles

Analysing the Sensitivity of SVM Kernels on Hyperspectral Imagery for Land Use Land Cover Classification

Support Vector Machine is a non- parametric statistical learning theory-based machine learning technique that classifies the data by maximising the margin between two classes by constructing a hyperplane between them. Although SVM’s are well known for binary classification problems, for multiclass classification problems- they use certain kernels, to turn the linear boundaries into non-linear boundaries in a higher dimensional feature space. In this context, the type of kernel used in non-linear SVMs have a profound effect on final performance of SVM classifier. The current study focuses on analysing the sensitivity of various linear and non-linear SVM kernels and parameters used in them for classifying 10 different land use/cover features from remotely sensed space borne hyperspectral image. Forty five different models were tested to analyse the performance of each kernel with varying penalty values for classification of land use/land cover classes. Analysis of the classified maps was made based on the final overall accuracy, error/penalty parameter and degree of polynomial. Experimental results show that SVM with RBF kernel outperformed the other kernels with an overall accuracy of 90.63%, with an error penalty of 100 and a gamma value of 0.006 followed by polynomial kernel of degree 2.

EGFAFS: A Novel Feature Selection Algorithm Based on Explosion Gravitation Field Algorithm.

Feature selection (FS) is a vital step in data mining and machine learning, especially for analyzing the data in high-dimensional feature space. Gene expression data usually consist of a few samples characterized by high-dimensional feature space. As a result, they are not suitable to be processed by simple methods, such as the filter-based method. In this study, we propose a novel feature selection algorithm based on the Explosion Gravitation Field Algorithm, called EGFAFS. To reduce the dimensions of the feature space to acceptable dimensions, we constructed a recommended feature pool by a series of Random Forests based on the Gini index. Furthermore, by paying more attention to the features in the recommended feature pool, we can find the best subset more efficiently. To verify the performance of EGFAFS for FS, we tested EGFAFS on eight gene expression datasets compared with four heuristic-based FS methods (GA, PSO, SA, and DE) and four other FS methods (Boruta, HSICLasso, DNN-FS, and EGSG). The results show that EGFAFS has better performance for FS on gene expression data in terms of evaluation metrics, having more than the other eight FS algorithms. The genes selected by EGFAGS play an essential role in the differential co-expression network and some biological functions further demonstrate the success of EGFAFS for solving FS problems on gene expression data.

Learning for mismatch removal via graph attention networks

Recovering camera pose from two-view images is a critical problem in photogrammetry and computer vision. For complex scenarios, point correspondences that are constructed by off-the-shelf feature matcher such as SIFT, would be corrupted by heavy outliers. In this case, traditional sampling consensus- or motion/geometrical coherence-based methods would suffer a lot from ensuring their assumptions. To this end, we propose a deep technique to better extract underlying geometry information from high-dimensional feature space for two-view geometry estimation. Unlike existing deep methods that use distribution-based normalization or explicitly aggregate neighboring correspondences, we propose a graph attention operation with multi-head mechanism, termed as GANet, to latently capture fine-grain contextual/geometrical relations among these corrupted correspondences. This encourages our network to learn informative representation for ensuring high graph similarity thus focusing more on inliers and restraining outliers. On this basis, our network can more easily infer inliers that are best to recover camera pose. Moreover, we also observe that the calculation of graph similarity for each node is only supported by partial node features. In this regard, we further propose a lightweight implementation for graph attention, namely Sparse GANet, which is performed by learning a sparse attention map based on block-wise operation and Sinkhorn normalization. This sparse strategy can largely reduce the memory and computational requests while maintaining the performance. Extensive experiments of pose estimation, outlier rejection and image registration on different challenging datasets, and combinational tests with different descriptor matchers and robust estimators, demonstrate the superiority and great generalization of our method against the state-of-the-art. In particular, we achieve at least 1.5% and 0.6% mAP(%)@5° enhancement on YFCC and SUN3D data for pose estimation, respectively. And our sparse GANet can reduce the model size to only 0.28 MB and the time cost to 16 ms, which is significant superior than SuperGlue that requires 12.02 MB and 68 ms. (Source code is available at https://github.com/StaRainJ/Code-of-GANet.)

Klasifikasi Status Pembayaran Kredit Barang Elektronik dan Furniture Menggunakan Support Vector Machine

Classification is the process of finding a model or function that can describe and differentiate data into classes. One application of classification is Support Vector Machine (SVM). SVM is a learning system that uses a hypothetical space in the form of linear functions in a high-dimensional feature space, trained with a learning algorithm based on optimization theory by implementing machine learning derived from statistical learning theory. The concept of classification with SVM is to find the best hyperplane to separate the two data classes and use a support vector approach. This study uses the proportion of the distribution of training data and testing data, namely 50%:50%, 70%:30%, 90%:10% and uses the SVM algorithm Polynomial kernel function with parameters =0.01, r=0.5, d =2, and C=1. This study aims to determine the results of the classification of the credit payment status of electronic goods and furniture and the level of classification accuracy in the SVM method. The data used is the debtor data of PT. KB Finansia Multi Finance Bontang in 2020 as many as 133 data with current and non-current credit payment status and using 7 independent variables, namely age, number of dependents, length of stay, income, years of service, large credit payments, and length of credit borrowing. The results of the SVM classification show an average accuracy value of 72.25% and the best accuracy chosen is the proportion of training data distribution and testing data 90%:10%, which is 84.62%.

Comparison of Different Kernels in a Support Vector Machine to Classify Prostate Cancerous Tissues in T2-Weighted Magnetic Resonance Imaging

Background and objectivesA support vector machine (SVM) is one of the most powerful classifiers in machine learning that can be applied when a data set is introduced in two classes in a high dimensional feature space. The objective of this study is to compare different kernels of SVMs to classify prostate cancerous tissues.

A Method of Power Flow Database Generation Base on Weighted Sample Elimination Algorithm

With the rapid development of urban power grids, distributed renewable energy sources and adjustable loads have increased significantly, resulting in more complex operation conditions, increasing the difficulty of power flow calculations. The usage of artificial intelligence technology to assist in calculating power flows for large-scale urban grids has a wide range of application prospects. It is currently difficult to generate enough grid operation database with controlled distribution for artificial intelligence (AI) method research. Data is one of the important factors affecting the performance of deep learning algorithms, and the lack of research on data distribution characteristics also hinders the performance of deep learning algorithms. The distributional characteristics of data sets in high-dimensional feature spaces are difficult to represent and measure, and the algorithm design process is prone to encounter curse of dimensionality. This paper proposed a novel method for generating databases to improve the solving efficiency of data-driven power flow calculation problems. The proposed method removes samples based on the characteristics of data distribution. It constructs two databases, namely the blue noise distribution database and the variable density boundary enhanced distribution database. Compared with the classical stochastic sampling database, the proposed boundary-enhanced variable density (BEVD) database has significantly improved the judgment accuracy of power flow convergence. Finally, the China Electric Power Research Institute-36 (CEPRI-36) bus system is used to verify the effectiveness of the proposed method. The judgment accuracy was improved by 2.91%–9.5%.

DEEP-LEARNING-BASED POINT CLOUD UPSAMPLING OF NATURAL ENTITIES AND SCENES

Abstract. Limited accessibility, occlusions, or sensor placement, can generate unevenly sampled laser scanning based point clouds. Such uneven coverage and partial lack of detail can affect the computation of geometric features therein and generate a visually unpleasant site description. The application of 3-D interpolation-driven solutions has been demonstrated to generate oversmoothed results as such algorithms ignore local patterns and variations within the surface. In that respect, the introduction of deep neural networks (DNN) has the potential to learn more complicated forms, typical of the rich morphological patterns that natural landforms and entities therein tend to exhibit. While existing research has focused on the upsampling of man-made objects, little has been devoted to natural scenes and the entities therein. To address that, we propose in this paper a DNN based approach that utilizes the self-similarity of geometric details as a means to address this generally ill-posed problem. Specifically, we treat two key elements that stand at the root of point-DNN-related design, the definition and selection of neighboring points, and the interpolation at a high dimensional feature space. We show how the introduction of a graph convolutional network and an attention unit helps address these matters and demonstrate how knowledge of densely sampled regions can be learned and transferred to sparsely sampled ones through geometric learning methods.

PyRCN: A toolbox for exploration and application of Reservoir Computing Networks

Reservoir Computing Networks (RCNs) belong to a group of machine learning techniques that project the input space non-linearly into a high-dimensional feature space, where the underlying task can be solved linearly. Popular variants of RCNs are capable of solving complex tasks equivalently to widely used deep neural networks, but with a substantially simpler training paradigm based on linear regression. In this paper, we show how to uniformly describe RCNs with small and clearly defined building blocks, and we introduce the Python toolbox PyRCN (Python Reservoir Computing Networks) for optimizing, training and analyzing RCNs on arbitrarily large datasets. The tool is based on widely-used scientific packages and complies with the scikit-learn interface specification. It provides a platform for educational and exploratory analyses of RCNs, as well as a framework to apply RCNs on complex tasks including sequence processing. With a small number of building blocks, the framework allows the implementation of numerous different RCN architectures. We provide code examples on how to set up RCNs for time series prediction and for sequence classification tasks. PyRCN is around ten times faster than reference toolboxes on a benchmark task while requiring substantially less boilerplate code.

Segmentation of Infant Brain Using Nonnegative Matrix Factorization

This study develops an atlas-based automated framework for segmenting infants’ brains from magnetic resonance imaging (MRI). For the accurate segmentation of different structures of an infant’s brain at the isointense age (6–12 months), our framework integrates features of diffusion tensor imaging (DTI) (e.g., the fractional anisotropy (FA)). A brain diffusion tensor (DT) image and its region map are considered samples of a Markov–Gibbs random field (MGRF) that jointly models visual appearance, shape, and spatial homogeneity of a goal structure. The visual appearance is modeled with an empirical distribution of the probability of the DTI features, fused by their nonnegative matrix factorization (NMF) and allocation to data clusters. Projecting an initial high-dimensional feature space onto a low-dimensional space of the significant fused features with the NMF allows for better separation of the goal structure and its background. The cluster centers in the latter space are determined at the training stage by the K-means clustering. In order to adapt to large infant brain inhomogeneities and segment the brain images more accurately, appearance descriptors of both the first-order and second-order are taken into account in the fused NMF feature space. Additionally, a second-order MGRF model is used to describe the appearance based on the voxel intensities and their pairwise spatial dependencies. An adaptive shape prior that is spatially variant is constructed from a training set of co-aligned images, forming an atlas database. Moreover, the spatial homogeneity of the shape is described with a spatially uniform 3D MGRF of the second-order for region labels. In vivo experiments on nine infant datasets showed promising results in terms of the accuracy, which was computed using three metrics: the 95-percentile modified Hausdorff distance (MHD), the Dice similarity coefficient (DSC), and the absolute volume difference (AVD). Both the quantitative and visual assessments confirm that integrating the proposed NMF-fused DTI feature and intensity MGRF models of visual appearance, the adaptive shape prior, and the shape homogeneity MGRF model is promising in segmenting the infant brain DTI.

Nonlinear subspace clustering using non-convex Schatten-p norm regularization

Subspace clustering aims to seek a multi-subspace representation that is best suitable for data points taken from a high-dimensional space. Sparse representation and low-rank approximation-based methods have become one of the main melodies for subspace clustering. In the existing methods, nuclear norm is used to approximate rank minimization. However, the common deficiency still exists for nuclear norm, which always over-penalizes large singular values and results in a biased solution. In this paper, we propose a nonlinear subspace clustering model that exploits sparsity and low-rank of data in high dimensional feature space by using Schatten-p norm surrogate ([Formula: see text]) with learned low-rank kernel. By this manner, the model guarantees that the data mapped in the high-dimensional feature spaces is lower rank and self-expressive. And we show the alternating direction method of multipliers (abbreviated as ADMM) for the corresponding problem in a reproducing kernel Hilbert space. Various experiments on motion segmentation and image clustering display that the proposed model has potentiality in outperforming most of state-of-the-art models in current literature.

A Discriminative-Based Geometric Deep Learning Model for Cross Domain Recommender Systems

Recommender systems (RS) have been widely deployed in many real-world applications, but usually suffer from the long-standing user/item cold-start problem. As a promising approach, cross-domain recommendation (CDR), which has attracted a surge of interest, aims to transfer the user preferences observed in the source domain to make recommendations in the target domain. Traditional machine learning and deep learning methods are not designed to learn from complex data representations such as graphs, manifolds and 3D objects. However, current trends in data generation include these complex data representations. In addition, existing research works do not consider the complex dimensions and the locality structure of items, which however, contain more discriminative information essential for improving the performance accuracy of the recommender system. Furthermore, similar outcomes between test samples and their neighboring training data restrained in the kernel space are not fully realized from the recommended objects belonging to the same object category to capture the embedded discriminative information effectively. These challenges leave the problem of sparsity and the cold-start of items/users unsolved and hence impede the performance of the cross-domain recommender system, causing it to suggest less relevant and undistinguished items to the user. To handle these challenges, we propose a novel deep learning (DL) method, Discriminative Geometric Deep Learning (D-GDL) for cross-domain recommender systems. In the proposed D-GDL, a discriminative function based on sparse local sensitivity is introduced into the structure of the DL network. In the D-GDL, a local representation learning (i.e., a local sensitivity-based deep convolutional belief network) is introduced into the structure of the DL network to effectively capture the local geometric and visual information from the structure of the recommended 3D objects. A kernel-based method (i.e., a local sensitivity deep belief network) is also incorporated into the structure of the DL framework to map the complex structure of recommended objects into high dimensional feature space and achieve an effective recognition result. An improved kernel density estimator is created to serve as a weighing function in building a high dimensional feature space, which makes it more resistant to geometric noise and computation performance. The experiment results show that the proposed D-GDL significantly outperforms the state-of-the-art methods in both sparse and dense settings for cross-domain recommendation tasks.

TBAL: Two-stage batch-mode active learning for image classification

The success of deep learning applications relies on a large number of labeled data. Active learning aims at identifying most informative unlabeled samples for labeling so as to achieve comparable performance with as few labeled data as possible. A two-stage batch-mode active learning (TBAL) method is proposed in this work. In the first stage of TBAL, we propose a new semi-supervised learning framework that offers accurate uncertainty estimation and finds a high-dimensional feature representation for unlabeled samples. In other words, it leverages a small number of labeled data as well as a large number of unlabeled data. In the second stage of TBAL, we propose a novel technique called cluster re-balancing (CRB). It takes correlations within a batch into account and fits a query strategy in batch mode. The TBAL method can obtain informative batches of samples by considering uncertainty and diversity jointly in a high-dimensional feature space. Moreover, in order to compare the robustness of semi-supervised learning based active learning methods, a new evaluation protocol based on transfer learning is introduced. Experiments on SVHN, CIFAR-10 and CUB-200-2011 datasets show that our proposed CRB based query strategy that employs a semi-supervised model with a good balance between uncertainty and diversity outperforms various baselines as well as other state-of-the-art methods by a clear margin.

DNA Memristors and Their Application to Reservoir Computing.

This paper introduces memristors realized by molecular and DNA reactions. Molecular memristors process one input molecule, generate two output molecules, and are realized using two molecular reactions with two different rate constants. The DNA memristors are realized using five DNA strand displacement (DSD) reactions with two effective rate constants. The hysteresis behavior is preserved in the proposed memristors, and this behavior can be altered by changing the ratios of the rate constants. The state of the memristor can be computed from the concentrations of the two output molecules using bipolar fractional coding. We describe how the proposed memristors can be used to learn the spatial and temporal properties of data via the reservoir computing (RC) model. An RC system can be divided into two parts: reservoir and readout layer. The first part transfers the information from the input space to a high-dimensional spatiotemporal feature space represented by the state of reservoirs. The connectivity structure of the reservoir will remain fixed through the dynamical evaluations. The readout layer effectively maps the projected features to the target output. A dynamical memristor array is used to implement an RC system that exploits the internal dynamical processes of the memristors. The readout layer implements a matrix-vector multiplication using molecular reactions, also based on bipolar fractional coding. All molecular reactions are mapped to DSD reactions. The RC system based on the DNA reservoir and the DNA readout layer is used to solve a handwritten digit recognition task and a second-order time series prediction task. The performance of the DNA RC system is comparable to that of an electronic memristor RC system for both tasks.

Characterizing car-following behaviors of human drivers when following automated vehicles using the real-world dataset

As the market penetration rate of automated vehicles (AVs) increases, there will be a transition period when the traffic stream is composed of both AVs and human-driven vehicles (MVs) in the near future. However, the interactions between MVs and AVs, especially whether MVs will behave differently when following AVs compared to following MVs, have not been fully understood. Previous studies in this field mainly conducted traffic/numerical simulations or field experiments to investigate human drivers’ behavior changes, but these approaches all have critical drawbacks such as simplified driving environments and limited sample sizes. To fill in the knowledge gap, this study uses the high-resolution (10 Hz) Waymo Open Dataset to reveal differences in car-following behaviors between MV-following-AV and MV-following-MV cases. Driving volatility measures, time headways and time-to-collision (TTC) are adopted to quantify and compare MV-following-AV and MV-following-MV interactions. The principal component analysis (PCA) is applied on the high-dimensional feature space, followed by the hierarchical clustering on the dimension-reduced feature set to categorize MV driving styles when following AVs. The comparison results indicate that MV-following-AV events have lower driving volatility in terms of velocity and acceleration/deceleration, smaller time headways and higher TTC values. Furthermore, the clustering results reveal that human drivers when following AVs exhibit four different car-following styles: high-velocity-non-aggressive, high-velocity-aggressive, low-velocity-non-aggressive, and low-velocity-aggressive. These findings highlight the vital importance of taking into account the heterogeneity of MV-following-AV behaviors when designing mixed traffic control algorithms and can be beneficial for AV fleet operators to improve their algorithms.

Biomarker interaction selection and disease detection based on multivariate gain ratio

BackgroundDisease detection is an important aspect of biotherapy. With the development of biotechnology and computer technology, there are many methods to detect disease based on single biomarker. However, biomarker does not influence disease alone in some cases. It’s the interaction between biomarkers that determines disease status. The existing influence measure I-score is used to evaluate the importance of interaction in determining disease status, but there is a deviation about the number of variables in interaction when applying I-score. To solve the problem, we propose a new influence measure Multivariate Gain Ratio (MGR) based on Gain Ratio (GR) of single-variate, which provides us with multivariate combination called interaction.ResultsWe propose a preprocessing verification algorithm based on partial predictor variables to select an appropriate preprocessing method. In this paper, an algorithm for selecting key interactions of biomarkers and applying key interactions to construct a disease detection model is provided. MGR is more credible than I-score in the case of interaction containing small number of variables. Our method behaves better with average accuracy 93.13% than I-score of 91.73% in Breast Cancer Wisconsin (Diagnostic) Dataset. Compared to the classification results 89.80% based on all predictor variables, MGR identifies the true main biomarkers and realizes the dimension reduction. In Leukemia Dataset, the experiment results show the effectiveness of MGR with the accuracy of 97.32% compared to I-score with accuracy 89.11%. The results can be explained by the nature of MGR and I-score mentioned above because every key interaction contains a small number of variables in Leukemia Dataset.ConclusionsMGR is effective for selecting important biomarkers and biomarker interactions even in high-dimension feature space in which the interaction could contain more than two biomarkers. The prediction ability of interactions selected by MGR is better than I-score in the case of interaction containing small number of variables. MGR is generally applicable to various types of biomarker datasets including cell nuclei, gene, SNPs and protein datasets.

IPReditor-CMG: Improving a predictive RNA editor for crop mitochondrial genomes using genomic sequence features and an optimal support vector machine

In crops, RNA editing is one of the most important post-transcriptional processes in which specific cytidines (C) in virtually all mitochondrial protein-coding genes are converted to uridines (U). Despite extensive recent research in RNA editing, exploring all of the C-to-U editing events efficiently on the genomic scale remains challengeable. Developing accurate prediction methods for the detection of RNA editing sites would dramatically reduce experimental determination. Therefore, we propose a novel method, iPReditor-CMG (improved predictive RNA editor for crop mitochondrial genomes), to predict crop mitochondrial editing sites using genome sequence and an optimised support vector machine (SVM). We first selected three mitochondrial genomes with known RNA editing sites from Arabidopsis thaliana, Brassica napus and Oryza sativa, released by NCBI, as the training and test sets. The genes and their transcripts from self-sequenced tobacco mitochondrial ATPase were selected as the validation set. The iPReditor-CMG first coded the genome sequences as numerical vectors and then performed an efficient feature selection on the high-dimensional feature space, where the SVM was employed in feature selection and following modelling. The average independent prediction accuracy of intraspecific editing sites across three species was 0.85, and up to 0.91 in A. thaliana, which outperformed the reference models. For the interspecific independent prediction, the prediction accuracy between dicotyledons was 0.78 and the accuracy between dicotyledons and monocotyledons was 0.56, which implies that there might be similarity in the C-to-U editing mechanism in close relatives. Finally, the best model was identified with an independent test accuracy of 0.91 and an AUC of 0.88, which suggested that five unreported feature sequences, i.e. TGACA, ACAAC, GTAGA, CCGTT and TAACA, are closely associated with the editing phenomenon. Multiple tests supported that the iPReditor-CMG could be effectively applied to predict editing sites in crop mitochondria, which may further contribute to understanding the mechanisms of site editing and post-transcriptional events in crop mitochondria.

Learning to share by masking the non-shared for multi-domain sentiment classification

Multi-domain sentiment classification deals with the scenario where labeled data exists for multiple domains but is insufficient for training effective sentiment classifiers that work across domains. Thus, fully exploiting sentiment knowledge shared across domains is crucial for real-world applications. While many existing works try to extract domain-invariant features in high-dimensional space, such models fail to explicitly distinguish between shared and private features at the text level, which to some extent lacks interpretability. Based on the assumption that removing domain-related tokens from texts would help improve their domain invariance, we instead first transform original sentences to be domain-agnostic. To this end, we propose the BERTMasker model which explicitly masks domain-related words from texts, learns domain-invariant sentiment features from these domain-agnostic texts and uses those masked words to form domain-aware sentence representations. Empirical experiments on the benchmark multiple domain sentiment classification datasets demonstrate the effectiveness of our proposed model, which improves the accuracy on multi-domain and cross-domain settings by 1.91% and 3.31% respectively. Further analysis on masking proves that removing those domain-related and sentiment irrelevant tokens decreases texts’ domain separability, resulting in the performance degradation of a BERT-based domain classifier by over 12%.

Improved Kernel Entropy Composition Analysis method for transgenic cotton seeds recognition based on Terahertz Spectroscopy

Terahertz (THz) Spectroscopy, featured by low energy, instantaneity and spectral fingerprint, is promising in material identification. Kernel Entropy Composition Analysis (KECA), different from the commonly used THz time-domain spectroscopy analysis methods of PCA and Kernel Principal Component Analysis (KPCA), focuses on decomposing eigenvalues and eigenvectors of a kernel matrix where its original spectral data are projected into a high-dimensional feature space, selecting eigenvectors as projection vectors which contribute most to Rayleigh entropy of original data and getting a new database. However, the parameters selection of kernel function in traditional KECA method have a significant effect on the accuracy of final analysis results. Aimed at this problem, this paper has proposed a new approach called improved KECA method for transgenic cotton seeds recognition based on Terahertz Spectroscopy. The improved KECA takes as the criterion function difference between intra-class and inter-class dispersion based on angular structure for Kernel parameter optimization, and selects kernel parameter with the maximum value of the criterion function as the best option. Then, the clustering method based on angular structure distance is used for the identification of different substance species. In order to test whether the proposed method is effective or not, this paper has applied the THz time-domain spectroscopy technology to detect the three transgenic cotton seeds of Xinluzhong6, Xinqiu107 and Yingmian8 respectively. And then the absorbance spectrum data of these three cotton seeds will be subject to clustering analysis through the proposed improved KECA method.

An adaptive subspace data-driven method for nonlinear dynamic systems

Quality-related data-driven fault detection methods have been universally recognized by researchers in industry and academia. However, the detection performance of the common data-driven method will degrade if it cannot keep up with the dynamic shifts of the production process. In this paper, an adaptive kernel principal component regression (AKPCR) method is thus proposed to process the nonlinear feature and the time-varying property in the system. The approximate Hilbert space approach is provided to update the projection matrix based on the innovation information from the new data for the nonlinear high-dimensional feature space. Besides, the recursive post-decomposition technique is created from the updated projection matrix to better monitor the quality state. Compared with the fixed threshold, an adaptive one is designed in this study to detect the fault encountered in the time-varying system, which can decrease false alarms in fault detection. Finally, the Tennessee Eastman case (TEC) and the waste water treatment system (WWTS) are leveraged to illustrate the reliability of the proposed approach.

Construction and Evaluation of Robust Interpretation Models for Breast Cancer Metastasis Prediction.

Interpretability of machine learning (ML) models represents the extent to which a model's decision-making process can be understood by model developers and/or end users. Transcriptomics-based cancer prognosis models, for example, while achieving good accuracy, are usually hard to interpret, due to the high-dimensional feature space and the complexity of models. As interpretability is critical for the transparency and fairness of ML models, several algorithms have been proposed to improve the interpretability of arbitrary classifiers. However, evaluation of these algorithms often requires substantial domain knowledge. Here, we propose a breast cancer metastasis prediction model using a very small number of biologically interpretable features, and a simple yet novel model interpretation approach that can provide personalized interpretations. In addition, we contributed, to the best of our knowledge, the first method to quantitatively compare different interpretation algorithms. Experimental results show that our model not only achieved competitive prediction accuracy, but also higher inter-classifier interpretation consistency than state-of-the-art interpretation methods. Importantly, our interpretation results can improve the generalizability of the prediction models. Overall, this work provides several novel ideas to construct and evaluate interpretable ML models that can be valuable to both the cancer machine learning community and related application domains.