• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Improved Logistic Regression Algorithm Based on Kernel Density Estimation for Multi-Classification with Non-Equilibrium Samples

    2019-11-07 03:12:16YangYuZeyuXiongYueshanXiongandWeiziLi
    Computers Materials&Continua 2019年10期

    Yang Yu,Zeyu Xiong,Yueshan Xiongand Weizi Li

    Abstract:Logistic regression is often used to solve linear binary classi fication problems such as machine vision,speech recognition,and handwriting recognition.However,it usually fails to solve certain nonlinear multi-classi fication problem,such as problem with non-equilibrium samples.Many scholars have proposed some methods,such as neural network,least square support vector machine,AdaBoost meta-algorithm,etc.These methods essentially belong to machine learning categories.In this work,based on the probability theory and statistical principle,we propose an improved logistic regression algorithm based on kernel density estimation for solving nonlinear multi-classi fication.We havecomparedourapproachwithothermethodsusingnon-equilibriumsamples,theresults show that our approach guarantees sample integrity and achieves superior classi fication.

    Keywords:Logistic regression,multi-classi fication,kernel function,density estimation,non-equilibrium.

    1 Introduction

    Machine Learning has become one of the most popular fields in recent years.There are two main tasks of Machine Learning:1)classi fication,which goal is to divide instances into the appropriate categories,and 2)regression,which goal is to study relationship between samples.The most basic classi fication problem is binary classi fication.which can be solved using algorithms such as Naive Bayes(NB),support vector machine(SVM),decision tree,logistic regression,KNN,neural network,etc.More generally,multi-classi fication problems such as identifying handwritten digits 0~9,and and labeling document topics have gained much attention recently.To provide few examples,Liu et al.[Liu,Liang and Xue(2008)]proposed a multi-classi fication algorithm based on fuzzy support vector machines,whichprovidesbetterclassi ficationaccuracyandgeneralizationabilitycompared with traditional One-vs.-Rest methods.Tang et al.[Tang,Wang and Chen(2005)]proposed a new multi-classi fication algorithm based on support vector machine and binary tree structure to solve the problem of non-separable regions.

    In the existing regression algorithm,support vector machines are mostly used for multi-classi fication problem,but there are some limitations in algorithm.The logistic regression algorithm can only solve the problem of dichotomy and linear classi fication.Support vector machines typically support only small training samples and are equally dif ficult to deal with multiple classi fication problems.Naive Bayes is based on the assumption that the characteristic conditions are independent.Once the dataset does not satisfy this assumption,its classi fication accuracy will be greatly affected.

    In order to solve the problem above,towards dif ficult for implement large scale samples,not applicable to multi-classi fication and uncertainty to constraint conditions,Chen et al.[Chen,Chen,Mao et al.(2013)]proposed a model of Density-based Logistic Regression(DLR),which has a good result in practical application.Our model is based on kernel density-based logistic regression and we construct a new kernel function for multi-classi fication problems.This has three advantages:1)It makes better improvements to classi fication effect.2)It is an extension of DLR model to multi-classi fication problems.3)It shows good generalization performance on nonlinear and unbalanced data.We will describe the theoretical rationality and check classifying quality according to practical application for our new model.

    The rest of the paper is organized as the following.In Section 2,we explain background knowledge including logistic regression binary classi fication,multi-classi fication,SoftMax and DLR model.In Section 3,we introduce several solutions for multi-classi fication problems with imbalanced samples.In Section 4,we explain our approach in details.In Section 5,we compare our approach to other methods and analyze the performances.Finally,we conclude in Section 6.

    2 Logistic regression and related knowledge

    2.1 Logistic regression

    Logistic regression is based on linear regression,and a sigmoid logic function is applied,which is a logarithmic probability function.Logistic regression is represented as follows,

    In the model of sigmoid function,zvalues are distributed within the range of[0,1].When the independent variable is taken near 0,thez-value change curve is very steep,while thezvalue is relatively stable at other values.Therefore,the binary classi fication tasks can be handled well if taking 0 as the boundary.However,it is sometimes dif ficult to make the representation model approximate to the expected model.By adding a constant termbto the function,

    By substituting Eq.(2)into Eq.(1),we have

    Based on these formulae,assuming a given datasetD={xi,yi},i=1,···,N,xi∈ R D,D is the dimension of samples,andyi∈{0,1},logistic regression is described as follows:

    wherewstands for feature weight,which is a parameter to be learned.φis the characteristic transformation function.

    In LR modelφis usually de fined to be equal tox.The key step is to learn unknown parameterswandb.Ifyin Eq.(3)is regarded as posterior probability estimationp(y=1|x),Eq.(4)can be rewritten as:

    Thenwcan be obtained by the maximum likelihood estimate.With the definition ofbi=p(yi=1|xi),y=0 or 1,for a single sample,the posterior probability is,

    Then,the maximum likelihood function is represented as follows,

    For the convenience of calculation,the negative log of the maximum likelihood function is used as the objective function to be optimized,

    Since the maximizing likelihood probability is equivalent to minimizing negative likelihood probability,the last step is to minimize the Loss function.

    2.2 Density-based logistic regression

    In the DLR model,φis a function that mapsxto the eigenspace,

    whereDis the dimension of the input data,lnmeasures the contribution ofxdto the probability ofy=1,andmeasures the degree of imbalance for datasets.p(y=1)is the proportion of data in the training set,whose label isy=1.Nadaraya-Watson is usually used to estimatep(y=k|xd)wherek=0,1.

    whereDk?Dis the subset of data in class k,andK(x,y)is a Gaussian kernel function de fined as follows,

    wherehdis the bandwidth of the kernel density function.Thehdis usually set using the Silverman’s rule of thumb[Silverman and Green(1986)],

    whereNis the total number of samples andσis the standard deviation ofxd.

    Next we need to trainwthrough the learning algorithm untilwconverges.Givenbi=p(yi=1|xi),the loss function based on likelihood probability is calculated as follows,

    2.3 Extension of logistic regression to multiple classi fication

    Since the logistic classi fication is a binary classi fication model,it is necessary to extend it for multiple classi fication,common extensions include multiple binary classi fication models or SoftMax models.

    2.3.1 N-logistic model

    The N-logistic model generally adopts One-vs-Rest or One-vs.-One.When classifying a sample,we first classify the two classi fiers,then vote,and select the category with the highest score.At the same time,to prevent the same vote,we also add the probability of the class to each classi fier in the voting.The predictive accuracy of these two approaches is usually very similar,so unless there is a speci fic need for data characteristics,it is generally arbitrary to choose one approach to calculate.

    2.3.2 SoftMax model

    SoftMax regression is a generalization of logistic regression to multiple classi fication problems.Its basic form is described as follows,

    When in the test,to samplex,if there is a categoryc,for all the other categoryc *(c * /=c)meet thep(y=c|x)>p(y=c *|x),thenxbelongs to the categoryc.

    On the question of choosing N-logistic model or SoftMax model,many scholars have conducted in-depth exploration.Currently,it is accepted that it is necessary to investigate whether the various categories are mutually exclusive.If there is a mutual exclusion relationship between the categories to be classi fied,we’d better choose SoftMax classi fier.Whileifthereisnomutualexclusionbetweencategories,andthecategoriesareintersecting,it is best suited to the N-logistic classi fier.We verify this conclusion according to corresponding datasets in Section 5.

    3 Analysis of the classi fication results with unbalanced sample proportion

    In our actual classi fication tasks,there are often needs to deal with problems with unbalanced data sample proportions.For example,the ratio of positive and negative samples in a dataset is 10:1,including 100 positive classes and 10 negative classes.If using this kind of data to train a classi fier,it is very likely that the test data will be divided into positive classes.Obviously,this classi fier is invalid.

    For this kind of data,traditional logistic regression method usually fails to work.In recent years,studies on the problem of unbalanced classi fication have been very active[Ye,Wen and Lv(2009)].In this section we introduce several common approaches to solve the problem of sample imbalance classi fication.

    3.1 Obtain more samples

    For unbalanced classi fication,the first solution is to obtain more samples and expand a few samples to balance the sample proportion.However,in most cases,the sampling procedure needs speci fic conditions.Thus,it is generally dif ficult to obtain more samples under the same conditions.

    3.2 Sampling methods

    The general sampling method is mainly based on modifying the number of unbalanced samples.The research of Estabrooks et al.[Estabrooks,Jo and Japkowicz(2004)]show that the general sampling method has a better effect on solving unbalanced classi fication problems.

    3.2.1 Under-sampling method

    Under-sampling method is also called down-sampling[Gao,Ding and Han(2008)],which is to eliminate some samples from majority class samples,so that the number of samples in the whole group tends to be balanced.The commonly used method is random under-sampling downward method.The method is based onNmin,the number of minority class samples.We randomly sample from the majority class samples and eliminateNsamples,and thenNmax-N=Nmin,so the samples are balanced.

    3.2.2 Over-sampling method

    Over-sampling method is also called up-sampling,which refers to increase the number of minority class samples.The method of adding a small number of minority class samples(random over-sampling method)or re-fitting some new data in accordance with some law can be used to make the number of samples balanced.One commonly used method is Synthetic Minority Over-sampling Technique(SMOTE)[Chawla,Bowyer,Hall et al.(2002)].The method analyzes the distribution of the characteristic space of a few samples and proposes new samples.Compared to the random over-sampling method,the data added by SMOTE sampling method is completely new,which can follow the regular pattern in the original sample.The main idea of SMOTE is shown in Fig.1.

    For each samplexin a minority class,the Euclidean distance of each sample point of a minority sample is calculated,and itskneighbors are obtained.A suitable sampling ratio is set according to the sample proportion to determine the sampling rateN.For each of the minority samplex,select several samples randomly from itskneighbors.For each random nearest neighborxn,a new sample is constructed with the original sample according to the following equation,

    3.3 Modify evaluation index

    For unbalanced classi fication,using accuracy to evaluate classi fiers may biases.For example,assuming ratio of positive and negative samples in a dataset is 9:1,and all samples are labelled be positive.Although the accuracy rate is up to 90%,the classi fier is useless.

    Figure 1:The main idea of SMOTE method

    Table 1:A hybrid matrix of binary classi fication

    Therefore,accuracy can serve as a biased indicator.Davis et al.[Davis and Goadrich(2006)]proposed a new evaluation index named Precision and Recall,some factors are listed in Tab.1.

    Precisionrefers to the proportion of positive samples in all predicted positive samples,andRecallrefers to the proportion of all actual positive samples that are being correctly predicted.

    3.4 Use penalty items to modify the weights

    If samples are dif ficult to sample directly,the method of modifying sample weights can be used.It increases the weight of minority class samples and reduces the weight of the majority class samples.Because the weights of minority class samples are high,they can lead to better classi fication results.The commonly used method is to add a penalty item to the majority class samples each time when training the sample weight.In general,we use the regularization method to add a penalty parameter to a objective function,this reduces the chance of the over fitting[Goodfellow,Bengio and Courville(2017)].The regularized objective function is shown below,

    whereαis a parameter which represents the contribution of the penalty item and the objective function.The penalty can be adjusted by controllingα.Ifα=0,there is no penalty,otherwise the larger theα,the greater the penalty.

    After we chose an appropriate penalty,the training regularize the objective function.In this way,the data error and the parameter scale can be reduced,the computation efficiency can be improved.But in practice,how to select the optimal penalty item is a complicated problem,which needs more tests.

    3.5 Kernel-based methods

    Towards general classi fication problem,we can assume that the sample data can be classi fieddirectlybylinearmodel.Inotherwords,thereisahyperplanethatcanseparatethe samples and ensure that the classi fication is correct.However,in practice,there is usually no such a hyperplane to partition the original data correctly,which means that the data are not linearly separable.For such a problem,we can consider preprocessing data.Using the principle of support vector machine,data in the low-dimensional space are transformed into the high dimensional space through nonlinear transformation,so that they can be linearly separable[Zhou(2016)].Using this method,the relationship between data samples can be written as dot product.For example,the linear regression function can be rewritten as follows,

    wherex(i)is the training data.αis the coef ficient vector.Replacing the dot product with a function of the kernelk(x,x(i))=φ(x)·φ(x(i)),we can get,

    This function is nonlinear with respect tox,while it is linear with respect toφ(x).

    Kernel function can deal with nonlinear unbalanced classi fication well.It uses a convex optimization technique to address nonlinear problems in a linear manner.At the same time,this method can guarantee convergence and improve the accuracy of classi fication.And there is some simpli fication in parameter determination.In addition,it is much more efficient to use the kernel function to transform data into a transformation function[Goodfellow,Bengio and Courville(2017)].

    SVM can convert sample data into high dimensional feature space through a kernel function.According to the principle of maximum spacing of SVM,the hyperplane of the optimal classi fication can be constructed in the characteristic space of high dimension to realizetheclassi fication.Iftheintervalofclassi ficationcanbeextended,especiallybetween minority class samples and the optimal classi fication hyperplane,the generalization performance of the classi fier and the accuracy of classes with small samples can be effectively improved.This enables the correct classi fication of unbalanced data[Liu,Huang,Zhu et al.(2009)].

    4 Improved method of kernel density estimation model for multi-classi fication

    We extend the DLR model to solve the multi-classi fication problem and design an improved multi-classi fication algorithm.Assuming there areCclasses,fork=1,2,...,C,the DLR model is de fined as follows,

    wherewk=(wk1,wk2,...,wkD)is the feature weighting parameter of classk,andφk=(φk1,φk2,...,φkD)is the characteristic transformation function of classk.

    According to the Nadaraya-Watson estimator,the probability formula of classkis obtained as follows:

    Finally,we need to minimize the loss function,

    where,1yi=kis 1 if and only ifyi=k,otherwise it takes value 0.

    Now we present the process of evaluating the gradient of the Loss function with respect towk,

    We adjust the weightwkaccording to the direction of the gradient descent,until thewkconverges andwkin the model is well trained.During the testing,the same kernel function transformation is performed on the testing data.The transformedφ(x)and trainedwkare substituted into Eq.(25).Then we compare the probability of the different classes and choose the class with the largest probability as the result category.At this point,we have completed the generalization of the logistic regression to multi-classi fication based on kernel density function.

    To show the difference between kernel density estimation logistic regression and classical logistic regression,we will compare the corresponding algorithms later.

    In the DLR algorithm,the inputxis given a feature transformation to getφbefore calculating the probability in Eq.(25).And then substituteφforxas the input to the probability formula.At the same time,the probability formula is changed from the Sigmoid function to the SoftMax function.

    After conducting experiments,we have found that the differences ofφamong different labels obtained using the DLR algorithm are small.There is a large error in the final classi fication result.And the minority class samples cannot be discriminated at all.And the value of loss function is not reduced by training.Therefore,in the process of constructing the bandwidth of kernel function and preprocessing the data,we improve it by the following scheme.

    Figure 2:The process of searching for the optimal coef ficient

    First,We try to train the parameters of the kernel function by modifying the weight values on the basis of Eq.(14).We conducted 16 groups of experiments,as shown in Fig.2.In the previous experiment,since the value ofwwas too large,the characteristics of the input dataXitself were dif ficult to distinguish.Properly reducingcan limit the complexity of the model,thereby improving the generalization performance of the model.Through comparison experiments,we found that changing 1.06 in Eq.(14)to 0.02 can signi ficantly improve the accuracy of the model.According to Fig.2,we reduce the bandwidth of kernel function in Eq.(14).

    In this way,the difference ofhdhas been improved.However,it may cause the value ofybecome too large and over fl ow in subsequent calculations.Feature scaling is a crucial step in the data preprocessing process.For most machine learning and optimization algorithms,scaling the values of features to the same interval can make their performance even better.In order to accelerate loss function convergence rate,we normalizeφusing the min-max method.

    The training process of the improved model is established in Algorithm 3.

    In the next section,we will conduct a comparative test to analyze the relationship between test results and training results after using Algorithm 3.

    5 Application of improved algorithm:datasets and veri fication analysis

    In particular,we have implemented the following methods for testing.

    1)N-logistic model,One-vs-Rest methods,abbreviated as NLR.

    2)N-logistic model,One-vs-Rest methods,combined with the oversampling method,abbreviated as NLR_Sample.

    3)N-logistic model,One-vs-Rest methods,combined with the Smote method,abbreviated as NLR_Smote.

    4)SoftMax model.

    5)SoftMax model combined with Algorithm 3,abbreviated as DLR++.

    We choose three datasets for testing.The first one is the fitting datasetNumbconstructed by us.In this dataset,each data element contains 10 fl oating point values,ranging from 0 to 5.The data distribution is divided into three categories:GroupA,GroupBandGroupC.The second dataset is theIrisfrom UCI.There are four features,including calyx length,calyx width and petal width,and the eigenvalue is fl oating-point number.The target value is the classi fication result of irises,includingvirginica,versicolor,andsetosa.The third dataset is theWinefrom UCI,which uses the various parameters of theWineto predict the quality of the Wine.There are 11 characteristic values,including volatileacidity,non-volatile acid,citric acid,residual sugar,chlorine,total sulfur dioxide,free of sulfur dioxide,sulfate,concentration,PH and alcohol.There are three quality classes:1,2,or 3.

    Table 2:Accuracy(%)of different methods on three datasets

    Table 3:Time(s)for different methods on three datasets

    Table 4:The number of iterations of training Loss convergence on three datasets

    In order to keep the data more versatile,and the classi fication results more persuasive,we use k-fold cross validation and assign the dataset to the training set and testing set according to the ratio 7:3.The test results are given as follows.

    From Tab.2 to Tab.4,we can see that the DLR++algorithm shows better prediction accuracy.In the three datasets,Numbis linear,whileIrisandWineare non-linear.We can see from the results that both N-logistic and SoftMax models can solve the multi-classi fication problem well.Both oversampling and smote sampling method can be used to improve the classi fication results of the sample imbalance problem with the accuracy rate increased by 1.34%and 3.92%respectively.The improved DLR++model based on kernel density is the best among all these methods,and it has an advantage in solving nonlinear multi-classi fication problems.From Tab.2 to Tab.4,we can see that the improved DLR++model converges faster than the original logistic model,using only 1/20 of the training times.At the same time,the accuracy rate has been increased 7.04%,at the cost of a higher operation time.

    From Tab.5 to Tab.6,we can see that the improved DLR++model has a better performance on datasets of large scales and multiple categories.It offers an accuracy of 93.0%while LRoffers an accuracy of 47.0%for 10-classi fication problems.

    Table 5:Performance of DLR++on different scales of datasets

    Table 6:Performance of DLR++on different number of categories

    6 Conclusion

    In this paper,we propose an improved logistic regression model based on kernel density estimation,and it can be applied to solve nonlinear multi-classi fication problems.We have compared and tested several common algorithms for logistic regression.For the experimental results,we found that the sampling method[Gao,Ding and Han(2008);Chawla,Bowyer,Hall et al.(2002)]can improve the classi fication accuracy,but the training samples obtained are very different from the original samples,which destroys the data characteristics inherently in the original sample.However in contrast,our improved model guarantees the integrity of the samples,it has obvious advantages in classi fication accuracy,and has good generalization ability with an ideal training speed.But there is still room for optimization in training,especially in the matrix operation stage.In the future,we will reduce the size of the matrix and block calculation,expected to decline training time and improve efficiency.Combining application to document retrieval[Xiong and Wang(2018);Xiong,Shen,Wang et al.(2018)],we will also expect to check the improved method in this paper is effect to document classi fication which is interested by us.

    Acknowledgement:The authors would like to thank all anonymous reviewers for their suggestions and feedback.This work was supported by National Natural Science Foundation of China(Grant No.61379103).

    亚洲国产精品999在线| 夜夜爽天天搞| 国产精品一及| 国产亚洲精品第一综合不卡| 国产99白浆流出| 国产免费男女视频| 在线观看舔阴道视频| 久久久久免费精品人妻一区二区| 亚洲av电影不卡..在线观看| 两个人看的免费小视频| 国产亚洲精品久久久久久毛片| 天天躁狠狠躁夜夜躁狠狠躁| 欧美日韩亚洲综合一区二区三区_| av福利片在线| 一级毛片精品| 亚洲国产高清在线一区二区三| 巨乳人妻的诱惑在线观看| 亚洲精品美女久久av网站| 亚洲五月天丁香| 精品熟女少妇八av免费久了| 欧美大码av| 国产免费男女视频| 五月伊人婷婷丁香| 一个人观看的视频www高清免费观看 | 亚洲专区中文字幕在线| 床上黄色一级片| 国产精华一区二区三区| 两个人看的免费小视频| 精品一区二区三区视频在线观看免费| 久久中文看片网| 亚洲aⅴ乱码一区二区在线播放 | 国产精品一区二区三区四区久久| 最近在线观看免费完整版| 成人18禁在线播放| 一级黄色大片毛片| 少妇人妻一区二区三区视频| 黄频高清免费视频| 色av中文字幕| 国产精品免费视频内射| 1024香蕉在线观看| 午夜免费激情av| 国产成人啪精品午夜网站| 99久久99久久久精品蜜桃| 亚洲电影在线观看av| 久久精品aⅴ一区二区三区四区| 国产伦人伦偷精品视频| 99热这里只有精品一区 | 亚洲国产高清在线一区二区三| 国产主播在线观看一区二区| 1024视频免费在线观看| 正在播放国产对白刺激| 老熟妇乱子伦视频在线观看| 少妇人妻一区二区三区视频| cao死你这个sao货| 久久伊人香网站| 不卡av一区二区三区| 最好的美女福利视频网| 99国产综合亚洲精品| 非洲黑人性xxxx精品又粗又长| 国产探花在线观看一区二区| 国产精品久久久久久亚洲av鲁大| 舔av片在线| 国产麻豆成人av免费视频| 精品国内亚洲2022精品成人| 一进一出抽搐gif免费好疼| 亚洲av美国av| 欧美成人一区二区免费高清观看 | 成年版毛片免费区| 18禁国产床啪视频网站| 中亚洲国语对白在线视频| 老司机福利观看| 亚洲欧美日韩东京热| 亚洲美女黄片视频| 一夜夜www| 成年免费大片在线观看| 亚洲国产欧美网| 999久久久国产精品视频| 免费高清视频大片| 亚洲人与动物交配视频| 欧美高清成人免费视频www| 日本免费一区二区三区高清不卡| 十八禁网站免费在线| 禁无遮挡网站| 黑人欧美特级aaaaaa片| 999精品在线视频| 岛国视频午夜一区免费看| 18禁裸乳无遮挡免费网站照片| 日本在线视频免费播放| 日韩中文字幕欧美一区二区| 99在线人妻在线中文字幕| 欧美人与性动交α欧美精品济南到| 精品一区二区三区av网在线观看| 久久香蕉精品热| 美女黄网站色视频| 日韩大码丰满熟妇| 中文字幕久久专区| 欧美+亚洲+日韩+国产| 欧美一区二区国产精品久久精品 | 日本黄色视频三级网站网址| АⅤ资源中文在线天堂| 久久久国产成人精品二区| 亚洲专区字幕在线| 国产一区二区在线观看日韩 | 国产精品av视频在线免费观看| 老鸭窝网址在线观看| √禁漫天堂资源中文www| 777久久人妻少妇嫩草av网站| 亚洲精品久久成人aⅴ小说| 国产精品久久久久久人妻精品电影| 91麻豆av在线| 熟女电影av网| 国产一区二区在线av高清观看| 99精品久久久久人妻精品| 真人做人爱边吃奶动态| 亚洲乱码一区二区免费版| 亚洲精品美女久久久久99蜜臀| 男女之事视频高清在线观看| 亚洲七黄色美女视频| 1024视频免费在线观看| 老鸭窝网址在线观看| 老鸭窝网址在线观看| 很黄的视频免费| av片东京热男人的天堂| 国产片内射在线| 中文在线观看免费www的网站 | 亚洲精品国产精品久久久不卡| 搡老妇女老女人老熟妇| 母亲3免费完整高清在线观看| 成年人黄色毛片网站| 在线看三级毛片| 亚洲国产精品久久男人天堂| 香蕉久久夜色| 婷婷六月久久综合丁香| 国产一区二区三区在线臀色熟女| а√天堂www在线а√下载| 国产爱豆传媒在线观看 | 少妇裸体淫交视频免费看高清 | 午夜激情福利司机影院| 久久草成人影院| 国产精品电影一区二区三区| 欧美乱码精品一区二区三区| 在线a可以看的网站| 久久欧美精品欧美久久欧美| www日本黄色视频网| 亚洲中文av在线| 亚洲国产精品成人综合色| 日本一二三区视频观看| 欧美乱码精品一区二区三区| 欧美午夜高清在线| 可以在线观看的亚洲视频| 色综合婷婷激情| 在线观看免费视频日本深夜| 色尼玛亚洲综合影院| 久久久久性生活片| xxxwww97欧美| 精品高清国产在线一区| 精品不卡国产一区二区三区| 国产高清激情床上av| 岛国在线免费视频观看| 精品欧美一区二区三区在线| 久久精品亚洲精品国产色婷小说| 欧美又色又爽又黄视频| av有码第一页| 国内精品久久久久久久电影| 亚洲av中文字字幕乱码综合| 18禁观看日本| 变态另类丝袜制服| 天天添夜夜摸| 欧美丝袜亚洲另类 | 黄色视频不卡| 麻豆av在线久日| 国产av一区在线观看免费| 婷婷亚洲欧美| 午夜福利高清视频| 亚洲中文av在线| 黄色片一级片一级黄色片| 日日夜夜操网爽| 夜夜看夜夜爽夜夜摸| 久久99热这里只有精品18| 性色av乱码一区二区三区2| 青草久久国产| 级片在线观看| 日韩欧美在线二视频| 免费无遮挡裸体视频| 美女大奶头视频| 深夜精品福利| 亚洲五月婷婷丁香| 午夜久久久久精精品| 真人做人爱边吃奶动态| 岛国在线观看网站| 亚洲精品在线观看二区| 少妇裸体淫交视频免费看高清 | 啦啦啦观看免费观看视频高清| 成人高潮视频无遮挡免费网站| 亚洲美女黄片视频| 亚洲精品中文字幕一二三四区| 国产日本99.免费观看| 757午夜福利合集在线观看| 99re在线观看精品视频| 色精品久久人妻99蜜桃| 伦理电影免费视频| 久久久久久久久中文| 亚洲 国产 在线| 亚洲精品美女久久av网站| 成人av在线播放网站| svipshipincom国产片| 久久国产精品影院| 最好的美女福利视频网| 午夜福利在线在线| 90打野战视频偷拍视频| 国产欧美日韩一区二区精品| 亚洲欧洲精品一区二区精品久久久| 黄色毛片三级朝国网站| 欧美精品啪啪一区二区三区| 国产99白浆流出| 亚洲美女黄片视频| 国产av不卡久久| 国产伦人伦偷精品视频| 午夜精品在线福利| 法律面前人人平等表现在哪些方面| 在线观看舔阴道视频| 12—13女人毛片做爰片一| 国产欧美日韩一区二区三| 男女之事视频高清在线观看| 久久午夜亚洲精品久久| 99精品在免费线老司机午夜| 午夜福利欧美成人| av视频在线观看入口| 在线观看午夜福利视频| 丁香欧美五月| АⅤ资源中文在线天堂| 少妇裸体淫交视频免费看高清 | 国产69精品久久久久777片 | 午夜日韩欧美国产| 欧美成人午夜精品| 草草在线视频免费看| 制服丝袜大香蕉在线| 免费在线观看成人毛片| 日韩大尺度精品在线看网址| 三级男女做爰猛烈吃奶摸视频| 国产男靠女视频免费网站| 亚洲18禁久久av| 1024视频免费在线观看| 校园春色视频在线观看| 超碰成人久久| 99国产极品粉嫩在线观看| 床上黄色一级片| 麻豆国产av国片精品| 久久久久精品国产欧美久久久| 久久久久国产精品人妻aⅴ院| 一级a爱片免费观看的视频| 色播亚洲综合网| 精品久久久久久久人妻蜜臀av| 久久这里只有精品中国| 性欧美人与动物交配| 老汉色∧v一级毛片| 欧美成人性av电影在线观看| 久久久久免费精品人妻一区二区| 亚洲国产欧美人成| 亚洲欧美日韩高清在线视频| 欧美午夜高清在线| 欧洲精品卡2卡3卡4卡5卡区| 五月伊人婷婷丁香| 亚洲人成伊人成综合网2020| 麻豆成人av在线观看| 宅男免费午夜| 99riav亚洲国产免费| 亚洲欧美激情综合另类| 国产av麻豆久久久久久久| 久久久久久免费高清国产稀缺| 午夜a级毛片| 午夜亚洲福利在线播放| 精品欧美国产一区二区三| 怎么达到女性高潮| 国产99白浆流出| 黄色视频不卡| 成人av在线播放网站| 级片在线观看| 麻豆成人av在线观看| 亚洲国产精品999在线| 亚洲国产看品久久| 丝袜美腿诱惑在线| 毛片女人毛片| 天天一区二区日本电影三级| 久久人妻福利社区极品人妻图片| 麻豆一二三区av精品| 1024香蕉在线观看| 亚洲第一电影网av| 国内揄拍国产精品人妻在线| 色综合欧美亚洲国产小说| 亚洲全国av大片| 男插女下体视频免费在线播放| 岛国在线观看网站| 18禁美女被吸乳视频| 精品高清国产在线一区| 国产精品九九99| 最新美女视频免费是黄的| 亚洲国产看品久久| 亚洲欧美精品综合久久99| 天天躁夜夜躁狠狠躁躁| 久久久久九九精品影院| 欧美精品啪啪一区二区三区| 性色av乱码一区二区三区2| 精品高清国产在线一区| 国产亚洲欧美在线一区二区| 99精品在免费线老司机午夜| 老汉色av国产亚洲站长工具| 黄色a级毛片大全视频| 少妇被粗大的猛进出69影院| 久久久久国产一级毛片高清牌| 精品一区二区三区视频在线观看免费| 久久久水蜜桃国产精品网| 他把我摸到了高潮在线观看| 男人的好看免费观看在线视频 | 久久久国产欧美日韩av| 欧美zozozo另类| 黑人欧美特级aaaaaa片| 欧美av亚洲av综合av国产av| 日韩高清综合在线| 看片在线看免费视频| 国产主播在线观看一区二区| 无人区码免费观看不卡| av福利片在线观看| 日本 av在线| 黄色片一级片一级黄色片| 亚洲午夜精品一区,二区,三区| 哪里可以看免费的av片| 少妇的丰满在线观看| 国产亚洲av高清不卡| 亚洲国产欧洲综合997久久,| 好男人在线观看高清免费视频| 亚洲精品av麻豆狂野| 国产一区在线观看成人免费| 俺也久久电影网| 好看av亚洲va欧美ⅴa在| 在线观看一区二区三区| 一本一本综合久久| 欧美日韩国产亚洲二区| 久久久国产欧美日韩av| 中文字幕熟女人妻在线| 欧美一级毛片孕妇| 久久婷婷成人综合色麻豆| 制服丝袜大香蕉在线| 一本精品99久久精品77| 久久久久久九九精品二区国产 | 国产成+人综合+亚洲专区| 国产成人啪精品午夜网站| 国产精品综合久久久久久久免费| 国产精品爽爽va在线观看网站| 色av中文字幕| 精品国产超薄肉色丝袜足j| 九九热线精品视视频播放| 老司机福利观看| 白带黄色成豆腐渣| 一二三四社区在线视频社区8| 97超级碰碰碰精品色视频在线观看| 国产乱人伦免费视频| www.999成人在线观看| 免费在线观看成人毛片| 亚洲av五月六月丁香网| 国产高清视频在线观看网站| xxxwww97欧美| 国产精品免费视频内射| 变态另类丝袜制服| 国产单亲对白刺激| 女同久久另类99精品国产91| 色综合站精品国产| 亚洲av电影不卡..在线观看| 国产一级毛片七仙女欲春2| 亚洲人成77777在线视频| 午夜福利成人在线免费观看| 一区二区三区激情视频| 国产精品自产拍在线观看55亚洲| 亚洲国产精品久久男人天堂| 一进一出抽搐gif免费好疼| av视频在线观看入口| 可以在线观看毛片的网站| www日本黄色视频网| bbb黄色大片| 黄色成人免费大全| 91成年电影在线观看| 美女 人体艺术 gogo| 成人特级黄色片久久久久久久| 国内毛片毛片毛片毛片毛片| 91大片在线观看| 国产精品久久久久久亚洲av鲁大| 这个男人来自地球电影免费观看| 曰老女人黄片| 1024视频免费在线观看| 一区二区三区激情视频| 欧美成狂野欧美在线观看| 久久精品91蜜桃| 欧美黑人巨大hd| 韩国av一区二区三区四区| 欧美黑人欧美精品刺激| 免费在线观看成人毛片| 变态另类成人亚洲欧美熟女| 亚洲av日韩精品久久久久久密| 一本精品99久久精品77| 亚洲 国产 在线| 免费人成视频x8x8入口观看| 亚洲午夜理论影院| 日韩欧美免费精品| 成人高潮视频无遮挡免费网站| 国产成人精品久久二区二区91| 曰老女人黄片| 男插女下体视频免费在线播放| 不卡一级毛片| 99久久国产精品久久久| 亚洲国产精品sss在线观看| 亚洲av电影不卡..在线观看| 午夜免费观看网址| 啦啦啦韩国在线观看视频| 久久精品综合一区二区三区| 熟妇人妻久久中文字幕3abv| av超薄肉色丝袜交足视频| 亚洲人成网站在线播放欧美日韩| 久久香蕉激情| 人人妻人人澡欧美一区二区| 成人三级做爰电影| 国产私拍福利视频在线观看| 亚洲 欧美一区二区三区| av有码第一页| 一级毛片高清免费大全| 美女黄网站色视频| 亚洲无线在线观看| 别揉我奶头~嗯~啊~动态视频| 欧美日韩瑟瑟在线播放| 亚洲熟妇中文字幕五十中出| www.熟女人妻精品国产| 亚洲国产欧洲综合997久久,| netflix在线观看网站| 日本免费a在线| 白带黄色成豆腐渣| 成年人黄色毛片网站| 午夜福利高清视频| 狂野欧美激情性xxxx| 亚洲精品久久成人aⅴ小说| 最新美女视频免费是黄的| 99在线人妻在线中文字幕| 久久久久久久久中文| 午夜亚洲福利在线播放| 女同久久另类99精品国产91| 在线观看免费午夜福利视频| 精品久久久久久,| 蜜桃久久精品国产亚洲av| 丰满的人妻完整版| 免费观看精品视频网站| 听说在线观看完整版免费高清| 亚洲国产精品合色在线| 亚洲精华国产精华精| 久久这里只有精品中国| 99精品久久久久人妻精品| 日韩中文字幕欧美一区二区| 久久久水蜜桃国产精品网| 一级作爱视频免费观看| 午夜激情福利司机影院| 国内精品一区二区在线观看| 亚洲av电影在线进入| 成人18禁高潮啪啪吃奶动态图| 亚洲精品久久成人aⅴ小说| 香蕉丝袜av| 亚洲第一欧美日韩一区二区三区| 三级国产精品欧美在线观看 | 欧美乱码精品一区二区三区| 午夜福利18| a级毛片a级免费在线| 欧美最黄视频在线播放免费| av欧美777| 嫁个100分男人电影在线观看| 国产伦在线观看视频一区| av有码第一页| 亚洲av成人精品一区久久| 巨乳人妻的诱惑在线观看| 欧美zozozo另类| 女人爽到高潮嗷嗷叫在线视频| 1024香蕉在线观看| 女人被狂操c到高潮| www.自偷自拍.com| 久久精品成人免费网站| 99久久99久久久精品蜜桃| 黄色视频不卡| 成熟少妇高潮喷水视频| 亚洲成人免费电影在线观看| 免费搜索国产男女视频| 欧美日韩瑟瑟在线播放| 国产一区二区激情短视频| 国产激情久久老熟女| 在线视频色国产色| 成人手机av| 韩国av一区二区三区四区| 日韩欧美三级三区| 欧美大码av| av国产免费在线观看| 可以在线观看的亚洲视频| 午夜福利高清视频| 制服诱惑二区| 大型黄色视频在线免费观看| 熟女少妇亚洲综合色aaa.| 精华霜和精华液先用哪个| www.www免费av| 国产午夜福利久久久久久| 后天国语完整版免费观看| 久久久国产欧美日韩av| www.精华液| 日韩三级视频一区二区三区| 听说在线观看完整版免费高清| 黄色毛片三级朝国网站| 日韩欧美免费精品| 国产人伦9x9x在线观看| 国产一区二区三区在线臀色熟女| 香蕉久久夜色| 男人的好看免费观看在线视频 | 真人一进一出gif抽搐免费| ponron亚洲| 亚洲第一欧美日韩一区二区三区| av欧美777| 精品国内亚洲2022精品成人| 视频区欧美日本亚洲| 亚洲午夜理论影院| 丝袜人妻中文字幕| 999精品在线视频| 成年女人毛片免费观看观看9| 久久精品夜夜夜夜夜久久蜜豆 | 国产男靠女视频免费网站| 国产欧美日韩一区二区精品| www日本黄色视频网| 亚洲国产中文字幕在线视频| 久久精品综合一区二区三区| 国产伦在线观看视频一区| 一个人观看的视频www高清免费观看 | 欧美日韩福利视频一区二区| 好男人电影高清在线观看| 国产成人啪精品午夜网站| 精品乱码久久久久久99久播| 国产伦在线观看视频一区| 两个人视频免费观看高清| 悠悠久久av| 黄色片一级片一级黄色片| 草草在线视频免费看| 观看免费一级毛片| 一个人免费在线观看电影 | 国产真人三级小视频在线观看| √禁漫天堂资源中文www| 亚洲av日韩精品久久久久久密| av免费在线观看网站| 午夜精品在线福利| www日本黄色视频网| 国产人伦9x9x在线观看| 叶爱在线成人免费视频播放| 成人高潮视频无遮挡免费网站| 欧美人与性动交α欧美精品济南到| 欧美+亚洲+日韩+国产| www.熟女人妻精品国产| 男女床上黄色一级片免费看| 国产免费男女视频| 母亲3免费完整高清在线观看| 国产三级中文精品| 精品欧美一区二区三区在线| 19禁男女啪啪无遮挡网站| 亚洲av中文字字幕乱码综合| 一级片免费观看大全| 久久精品91蜜桃| 久久草成人影院| 久久久国产欧美日韩av| 日韩大尺度精品在线看网址| 亚洲国产欧美人成| 三级男女做爰猛烈吃奶摸视频| xxxwww97欧美| 国产精品av视频在线免费观看| 99久久99久久久精品蜜桃| 欧美在线一区亚洲| 久久欧美精品欧美久久欧美| 亚洲欧美精品综合久久99| 久久香蕉国产精品| 无限看片的www在线观看| 99在线视频只有这里精品首页| 精品人妻1区二区| 中文在线观看免费www的网站 | 一本大道久久a久久精品| 日日干狠狠操夜夜爽| 国产精品乱码一区二三区的特点| 久99久视频精品免费| 女人高潮潮喷娇喘18禁视频| 亚洲欧美日韩高清在线视频| 免费搜索国产男女视频| 久久婷婷成人综合色麻豆| 欧美中文综合在线视频| 日韩欧美三级三区| 日韩精品中文字幕看吧| 香蕉国产在线看| 欧美黑人精品巨大| 午夜精品一区二区三区免费看| 亚洲黑人精品在线| av中文乱码字幕在线| 精品日产1卡2卡| 国产高清激情床上av| 国产亚洲精品一区二区www| 久久亚洲真实| 日本五十路高清| 午夜福利成人在线免费观看| 国产精品爽爽va在线观看网站| 岛国在线观看网站| 免费在线观看亚洲国产| 国产精品一区二区三区四区久久| 国产伦一二天堂av在线观看| 好男人在线观看高清免费视频| 女人高潮潮喷娇喘18禁视频| 精品第一国产精品| 日本精品一区二区三区蜜桃| 淫妇啪啪啪对白视频| 国产午夜精品论理片| 男人舔奶头视频| 成人av在线播放网站| 欧美日韩黄片免|