Search

Model averaging has received much attention in the past two decades, which integrates available information by averaging over potential models. Although various model averaging methods have been developed, there are few literatures on the theoretical properties of model averaging from the perspective of stability, and the majority of these methods constrain model weights to a simplex. The aim of this paper is to introduce stability from statistical learning theory into model averaging. Thus, we define the stability, asymptotic empirical risk minimizer, generalization, and consistency of model averaging and study the relationship among them. Our results indicate that stability can ensure that model averaging has good generalization performance and consistency under reasonable conditions, where consistency means model averaging estimator can asymptotically minimize the mean squared prediction error. We also propose a L2-penalty model averaging method without limiting model weights and prove that it has stability and consistency. In order to reduce the impact of tuning parameter selection, we use 10-fold cross-validation to select a candidate set of tuning parameters and perform a weighted average of the estimators of model weights based on estimation errors. The Monte Carlo simulation and an illustrative application demonstrate the usefulness of the proposed method., Comment: 47 pages, 14 figures, 1 table. This article was submitted to Journal of Machine Learning Research in July 2023

This paper offers a new approach to address the model uncertainty in (potentially) divergent-dimensional single-index models (SIMs). We propose a model-averaging estimator based on cross-validation, which allows the dimension of covariates and the number of candidate models to increase with the sample size. We show that when all candidate models are misspecified, our model-averaging estimator is asymptotically optimal in the sense that its squared loss is asymptotically identical to that of the infeasible best possible averaging estimator. In a different situation where correct models are available in the model set, the proposed weighting scheme assigns all weights to the correct models in the asymptotic sense. We also extend our method to average regularized estimators and propose pre-screening methods to deal with cases with high-dimensional covariates. We illustrate the merits of our method via simulations and two empirical applications.

Support vector machine (SVM) is a well-known statistical technique for classification problems in machine learning and other fields. An important question for SVM is the selection of covariates (or features) for the model. Many studies have considered model selection methods. As is well-known, selecting one winning model over others can entail considerable instability in predictive performance due to model selection uncertainties. This paper advocates model averaging as an alternative approach, where estimates obtained from different models are combined in a weighted average. We propose a model weighting scheme and provide the theoretical underpinning for the proposed method. In particular, we prove that our proposed method yields a model average estimator that achieves the smallest hinge risk among all feasible combinations asymptotically. To remedy the computational burden due to a large number of feasible models, we propose a screening step to eliminate the uninformative features before combining the models. Results from real data applications and a simulation study show that the proposed method generally yields more accurate estimates than existing methods.

The fish target detection algorithm lacks a good quality data set, and the algorithm achieves real-time detection with lower power consumption on embedded devices, and it is difficult to balance the calculation speed and identification ability. To this end, this paper collected and annotated a data set named "Aquarium Fish" of 84 fishes containing 10042 images, and based on this data set, proposed a multi-scale input fast fish target detection network (BTP-yoloV3) and its optimization method. The experiment uses Depthwise convolution to redesign the backbone of the yoloV4 network, which reduces the amount of calculation by 94.1%, and the test accuracy is 92.34%. Then, the training model is enhanced with MixUp, CutMix, and mosaic to increase the test accuracy by 1.27%; Finally, use the mish, swish, and ELU activation functions to increase the test accuracy by 0.76%. As a result, the accuracy of testing the network with 2000 fish images reached 94.37%, and the computational complexity of the network BFLOPS was only 5.47. Comparing the YoloV3~4, MobileNetV2-yoloV3, and YoloV3-tiny networks of migration learning on this data set. The results show that BTP-Yolov3 has smaller model parameters, faster calculation speed, and lower energy consumption during operation while ensuring the calculation accuracy. It provides a certain reference value for the practical application of neural network.

Model averaging is an alternative to model selection for dealing with model uncertainty, which is widely used and very valuable. However, most of the existing model averaging methods are proposed based on the least squares loss function, which could be very sensitive to the presence of outliers in the data. In this paper, we propose an outlier-robust model averaging approach by Mallows-type criterion. The key idea is to develop weight choice criteria by minimising an estimator of the expected prediction error for the function being convex with an unique minimum, and twice differentiable in expectation, rather than the expected squared error. The robust loss functions, such as least absolute deviation and Huber's function, reduce the effects of large residuals and poor samples. Simulation study and real data analysis are conducted to demonstrate the finite-sample performance of our estimators and compare them with other model selection and averaging methods.

Model averaging considers the model uncertainty and is an alternative to model selection. In this paper, we propose a frequentist model averaging estimator for composite quantile regressions. In recent years, research on these topics has been added as a separate method, but no study has investigated them in combination. We apply a delete-one cross-validation method to estimate the model weights, and prove that the jackknife model averaging estimator is asymptotically optimal in terms of minimizing out-of-sample composite final prediction error. Simulations are conducted to demonstrate the good finite sample properties of our estimator and compare it with commonly used model selection and averaging methods. The proposed method is applied to the analysis of the stock returns data and the wage data and performs well.

Smoothed AIC (S-AIC) and Smoothed BIC (S-BIC) are very widely used in model averaging and are very easily to implement. Especially, the optimal model averaging method MMA and JMA have only been well developed in linear models. Only by modifying, they can be applied to other models. But S-AIC and S-BIC can be used in all situations where AIC and BIC can be calculated. In this paper, we study the asymptotic behavior of two commonly used model averaging estimators, the S-AIC and S-BIC estimators, under the standard asymptotic with general fixed parameter setup. In addition, the resulting coverage probability in Buckland et al. (1997) is not studied accurately, but it is claimed that it will be close to the intended. Our derivation make it possible to study accurately. Besides, we also prove that the confidence interval construction method in Hjort and Claeskens (2003) still works in linear regression with normal distribution error. Both the simulation and applied example support our theory conclusion.

The Horvitz-Thompson (HT) estimator is widely used in survey sampling. However, the variance of the HT estimator becomes large when the inclusion probabilities are highly heterogeneous. To overcome this shortcoming, in this paper, a hard-threshold method is used for the first-order inclusion probabilities, that is, we carefully choose a threshold value, then replace the inclusion probabilities smaller than the threshold by the threshold. By this shrinkage strategy, we propose a new estimator called improved Horvitz-Thompson (IHT) estimator to estimate the population total. The IHT estimator increases the estimation accuracy although it brings bias which is relatively small. We derive the IHT estimator's MSE and its unbiased estimator, and theoretically compare the IHT estimator with the HT estimator. We also apply our idea to construct the improved ratio estimator. We numerically analyze simulated and real data sets to illustrate that the proposed estimators are more efficient and robust than the classical estimators.

Small area estimation under linear mixed models often assumes that the small area effect is random effect in almost all previous studies. However, in this paper a new approach is proposed explaining small area effect as the unknown function of an area-indicative variable, which is a kind of semiparametric models. The nonparametric part for area effect is represented by using penalized splines, from which the estimation and inference are done in the linear mixed model framework. The mean-squared error of empirical estimators is shown, and the testing for small area effect also considered. Additionally, some numerical simulations are demonstrated to express the good performance of this method., Comment: It has been submitted