TY - JOUR

T1 - NEXT RUNNING APPLICATION PREDICTION IN OS USING TIME-SERIES FORECASTING AND XAI

AU - Ma, Ruimin

AU - Zhang, Yuyi

AU - Liu, Jing

AU - Petrosian, Ovanes

N1 - Conference code: LIII

PY - 2022

Y1 - 2022

N2 - With the proliferation of smart devices, app responsiveness is one of the most important indicators of user experience, and predicting the next app to be used by a user is crucial for management planning in the companies concerned. In this context, more complex machine learning, and even deep learning, which can achieve better forecast performance, are widely used in this field. However, the opacity of these black-box models greatly limits the trust of users and the degree of optimization of the models by developers. To address these issues, this paper first test five of the more popular forecast algorithms, including integrated models and neural networks, to select the optimal model. As an innovation, the paper also uses XAI techniques to explain the black-box models as a way to improve user trust in the results generated by the forecast models and to help developers in their work. After completing the above work, on the one hand, we found that the integrated model performed better in the time series datasets of user app usage information, especially with LightGBM, and on the other hand, we found that the results of the explaination of the forecast model using the SHAP method indicated that the target variable category characteristics and lag characteristics were important features in the forecast of the next app used by the user.

AB - With the proliferation of smart devices, app responsiveness is one of the most important indicators of user experience, and predicting the next app to be used by a user is crucial for management planning in the companies concerned. In this context, more complex machine learning, and even deep learning, which can achieve better forecast performance, are widely used in this field. However, the opacity of these black-box models greatly limits the trust of users and the degree of optimization of the models by developers. To address these issues, this paper first test five of the more popular forecast algorithms, including integrated models and neural networks, to select the optimal model. As an innovation, the paper also uses XAI techniques to explain the black-box models as a way to improve user trust in the results generated by the forecast models and to help developers in their work. After completing the above work, on the one hand, we found that the integrated model performed better in the time series datasets of user app usage information, especially with LightGBM, and on the other hand, we found that the results of the explaination of the forecast model using the SHAP method indicated that the target variable category characteristics and lag characteristics were important features in the forecast of the next app used by the user.

KW - EXPLAINABLE AI

KW - NEURAL NETWORK

KW - ENSEMBLE MODEL

KW - TIME-SERIES FORECASTING

KW - Time-series forecasting

KW - ensemble model

KW - NEURAL NETWORK

KW - explainable AI

UR - https://dspace.spbu.ru/handle/11701/37048

M3 - Conference article

VL - 9

SP - 329

EP - 338

JO - Процессы управления и устойчивость

JF - Процессы управления и устойчивость

SN - 2313-7304

IS - 1

T2 - Control Processes and Stability

Y2 - 4 April 2022 through 8 April 2022

ER -