Personalized Game Recommendation Based On Implicit Feedback

With the increasing popularity of Internet applications,the problem of information overload has become more and more serious in the past decade and the cost of acquiring truly interesting content is getting higher and higher.The recommendation system is an effective tool to alleviate the problem of information overload.It models the user's preferences through the user's historical behaviors,related data of the items,and the user's profile.This helps the user to find content that meets their preferences.In the game field,accurate personalized game recommendations are important for merchants to mine potential players who may be interested in a particular game and enhance the player's experience level.The existing recommendation systems usually adopt explicit feedbacks such as ratings and reviews to implement personalized recommendation.However,the above explicit feedbacks are often difficult to be obtained in the field of game recommendation,resulting in some traditional recommendation algorithms that are very effective in other fields cannot be well applied to game recommendation.In contrast,implicit feedbacks such as click behaviors and browsing records largely exist in the game platform.This paper focuses on how to make effective use of these implicit feedbacks to achieve accurate personalized game recommendations.The main content of the paper is as follows:1)Aiming at the problem of the lack of expression of user preferences in implicit feedbacks,a pseudo-rating model based on discrete convolution formulas to characterize player's preferences for games is proposed.The model integrates the timebased implicit feedbacks,such as the times and duration of player's operations,reduces the error caused by the player's incorrect operation and the characteristics of the game itself,and avoids lagging expressions of preference without the consideration of the time factor.Experimental results show that the model can express the player's preference for the game more accurately in the current state.2)Aiming at the particularity of game recommendation field and preference model,a game recommendation algorithm suitable for pseudo-rating is proposed.On the basis of SVD++(Singular Value Decomposition++),this algorithm discards the player's rating deviation term that does not exist in the pseudo-rating and the implicit feedbacks part that already included,and adds the weight coefficient of the error in the loss function of the rating prediction to distinguish the importance of them.The weighting factor of the player's feature is added to adjust the weight among different features,making it more suitable for game recommendation scenarios based on implicit feedbacks.Experimental results show that the proposed algorithm has better recommendation precision and recall than the comparison algorithms.3)Aiming at the problem of low performance when decomposing a large-scale matrix,a sub-matrix decomposition algorithm for feature segmentation of game recommendation is proposed.The algorithm integrates the demographic-based recommendation algorithm to divide the global matrix into smaller-scale sub-matrices according to gender,region,age and other characteristics.Then the traditional recommendation algorithms can be used to achieve parallel sub-matrix decomposition to improve the game recommendation performance.The experimental results show that under a certain degree of sub-matrix partitioning,the sub-matrix decomposition algorithm can improve the computational performance and further,the precision and recall of game recommendations effectively.4)Aiming at the iterative efficiency problem of gradient descent method,a dynamic learning rate scheme with penalty for game recommendation is proposed.The scheme adjusts the learning rate based on the change of the value of the loss function dynamically,considers the effect of the successive increase and decrease times of the loss function value on the increase or decrease of the learning rate,and improves the convergence speed of the model.When the learning rate increases too quickly,a punishment mechanism is provided to prevent the iterative oscillation problem,thereby greatly reducing the training duration of the model.The experimental results show that the iterative efficiency of this learning rate scheme is significantly higher than that of the comparison schemes in game recommendation applications.