Establishment Of Predictive Model Of Skin Sensitive Related Ingredients Based On Machine Learning

At present,there are abundant resources for online cosmetics reviews,such as Tmall Taobao,Jingdong and Meilixiuxing,but no reports have been reported to dig deeply into consumers'subjective sensitivity.In this study,we use the web crawler method to obtain data,and further use the sentiment dictionary combined with machine learning method to mine information about different skin types of consumers,cosmetics ingredients andcosmetics online reviews.The results are as follows.1.Using Fiddler software and Requests module in Python,we crawled two Apps,Meilixiuxing and Yanjiuyuan,and obtained 100 430 consumer information including skin type,cosmetic ingredients and online reviews.2.Using the Jieba module in Python and Word2vec extended skin-sensitive related vocabulary,100 430 samples were cleaned up,and the information of skin-type-free and subjective-sensitive-related vocabulary-free online reviews was removed,and 46 000words were preliminarily screened.Furthermore,we extract the features of skin types,formula components and comments,and quantify the subjective sensitivity of online reviews by using the method of sensitive related sentiment dictionary extended by word2vec model,so as to realize the transformation from text type variables to structurednumerical types.3.Using scikit-learning module in Python,taking quantified skin type and cosmetic ingredients as input variables and quantified subjective sensibility of online reviews asoutput variables,machine learning model is built.Three models,decision tree,random forest and Gradient Boosting Decision Tree(GBDT),are constructed.GBDT model with maximum R~2and minimum RMSE are selected.The results showed that R~2and RMSE of the GBDT were 0.83 and 0.41 respectively.Compared with the reported ingredients related to skin sensitivity,the accuracy,sensitivity and specificity of the prediction model were82.4%,75%and 88.9%,respectively.It shows that the prediction model of skin sensitive related cosmetic ingredients based on GBDT is reliable.4.For the 11 differential lipids of normal and sensitive skin screened by skinlipidomics in the previous study,the above-mentioned GBDT model was used to predict the subjective sensitivity of 11 differential lipids to the skin.Lipid lower in sensitive skin,if the model predicts that the result is soothing,then the GBDT prediction result is correct,otherwise it is wrong.The same method is used to predict the lipid material corresponding to the higher content of differential lipid in sensitive skin.The prediction results of GBDT model were consistent with the results of lipidomics screening,and the correct rate was81.8%.The predicted results can be mutually verified with lipidomics results,providing new ideas for the discovery of soothing or stimulating lipid materials.5.Using Tkinter module in Python,the GUI interface of the prediction model of skin sensitive related ingredients is established,and the EXE program is further generated with py2exe module.This interface can assist cosmetic formulators to select cosmeticingredients for sensitive skin.In this study,by means of web crawler,affective dictionary and machine learning,a prediction model of skin sensitive related ingredients based on online reviews wasestablished,and potential irritation or alleviating lipid ingredients for sensitive skin were predicted,which provided a new direction and ideas for developing cosmetic ingredients for Self perceived sensitive skin population.