Research on the activity of short video social media users based on machine learning

With the popularity of mobile intelligent terminals, short video social media is popular all over the world, and the continuous use of active users of short video social media is a sufficient and necessary condition for its success. The prediction of user activity plays a direct guiding role in the subsequent user loss warning. Due to the balanced distribution of positive and negative samples in the data set used in this experiment, the AUC value was decided to be used as the evaluation standard of the model after comprehensive consideration of the accuracy rate, recall rate, F1, AUC and other evaluation indicators. There are two main reasons: first, AUC value is not affected by the wide value division compared with other indicators, which can more objectively reflect the performance of the model; second, there is an R0C curve corresponding to the AUC value, and describing the R0C curve can make the experimental results in this paper more convincing. This paper first introduces the data used in the experiment when constructing the user activity prediction framework. Since the original data does not have labels, this paper defines the problem to be solved in the experiment. As the data used in the experiment is time-related, sliding window is used to divide the data. This partitioning method can not only increase the number of samples, but also forget the data before the start of the window through the window mechanism. This paper assigns labels to the raw data based on the definition of the problem and the sliding window. Due to the wide application of machine learning, the relevant research on data reliability has been paid more attention by researchers. In this paper, the abnormal data is found out by visualization and the data is cleaned, and the experimental demonstration is given. Feature engineering is the act of extracting features from raw data and transforming them into a format suitable for machine learning models. In the feature engineering, this paper mined the features in the data, and selected the features according to the importance of the features and Pearson correlation coefficient between the features. It was concluded that the feature selection method in this paper can greatly accelerate the model training speed. In this paper, five machine learning algorithms, including LightGBM, support vector machine (SVM), KN, decision tree and Naive Bayes, are used for experimental comparison under full feature conditions. Under AUC evaluation criteria, LightGBM is the optimal AUC corresponding algorithm for a single model. Model fusion is an important method to improve algorithm performance. In this paper, a model fusion algorithm based on grid search is proposed. The basic idea of the algorithm is to find the weight of each submodel by using the idea of grid search. Because the algorithm is based on the idea of grid search, it is suitable for the case where the number of submodels is small. For the five machine learning submodels used in this experiment, the algorithm calculated the weights of the five seed models, and used the weight combinations to make predictions. The AUC value after fusion reached 0.9377, which increased by 2.4% compared with the optimal AUC value of a single model.