PROCESS CHARACTERISTICS ESTIMATION IN WEB APPLICATIONS USING K-MEANS CLUSTERING

Subject of Research. The paper presents the study of estimation problem of process characteristics for the particular case of user's activity prediction in computer online games. Various machine learning methods are considered, and the advantages of clustering-based approaches are identified. The variety of metrics for the estimation of clustering quality is studied. Method. A clustering-based approach to estimation of process characteristics was developed on the base of a hypothesis proposed during the preliminary analysis of user's activity data. Data on activity of users with the known predicted values was collected. Each user was represented as a pair of vectors: the first vector corresponded to his first days of activity, and the second one corresponded to the days with predicted performance. The vectors representing user's activity in the first days were used as training data for the K-means algorithm. A developed entropy-like loss function was used to find a value of K suitable for the problem under consideration. The clusters were matched with vectors of predicted process characteristics averaged over all users in the cluster. These matches were used as the prediction of new users' characteristics. Main Results. An approach to the determination of the suitable number of clusters is proposed, taking into account the specifics of the considered data. Numerical experiment is carried out, demonstrating the applicability of the developed method. Practical Relevance. The proposed approach application allows for the simultaneous prediction of multiple characteristics of online-game users, and, therefore, for solution of various planning and analytics problems during online-game development. For example, the method developed in the present work was used to analyze the development payback of new game elements, and to predict server load in order to increase available computational resources beforehand. The advantages of the developed method include no need for expert tagging of the training set and relatively low computational cost due to the low computational complexity of the proposed loss function used to estimate the hyperparameter K.