Cold Start Video Recommendation Using Regression Models Based On Watch Ratio Prediction

Video recommendation system cold start problem occurs when there is little or no past interaction data of new users or new video uploaded videos, and it is difficult to predict the user accurately and personalize the video. This paper tackles the cold start case and establishes and compares three predictive models applied to predict a video watch ratio considering constant contextual circumstances; the Linear Regression, Ridge Regression, and XGBoost models. The feature engineering strategies were employed (temporal
attributes and one-hot encoding of video duration categories) to have uniform representation across the models. The experimental results demonstrate that Linear
Regression and Ridge Regression performed almost perfectly, displaying the highest possible R 2 and Adjusted R 2 values of 1.0000 and the lowest possible error values (MAE,
MSE, RMSE), the order of magnitude of the residual values in Linear Regression and Ridge Regression is 10 -1 and 10 -8, respectively. On the contrary, XGBoost achieved more
mediocre R that of 0.6338 at reduced error and residual in the range of 10 3. Further evidences of Ridge Regression as a superior predictor model include residual analysis,
recommendation ranking behavior, and Taylor Diagram evaluation, which prove that the latter has superior levels of variance, high correlation with the observed data, and
predictive stability. Even though the XGBoost embraces nonlinear association, it has more dispersions and variations when the data is set in the mentioned way. In general, the findings demonstrate that Ridge Regression offers the most balanced, accurate and reliable
solution to cold start problem in video recommendation system.We have the following important