Mobile Phone Game Props Recommendation Based On Multi-instance Multi-label Learning

As an emerging industry,mobile phone games play an important role in the process of industrial structure upgrading of our country,by means of transforming the high and new technology into real productivity.Sales of virtual props constitute most income of mobile phone games.Considering the majority of mobile phone game players are amateurs,personalized virtual prop recommendation can guide the players to buy more game props,so as to increase the income of mobile phone games.In this thesis,a series of innovations are proposed in research on personalized phone game prop recommendation as follows:First,formalize game prop recommendation as multi-instance multi-label learning task to deal with the problem of appendences between incidents,the correlation between props and the limit of computing resources.The experimental results show that MIML performs better than the traditional recommendation algorithms.Second,two algorithms called We-MIMLfast and Sp-MIMLfast are proposed based on prior probability assumption and sparse assumption separately to cope with the problem of concept variation.The experimental results show that these two algorithms have good performance and the ensemble of them performs better in game prop recommendation.Third,an algorithm called Co-MIMLfast is proposed based on the combination of multi-instance multi-label learning framework and co-training framework to take advantage of the game data of players with none purchasing acts.In this algorithm,none purchasing data is regarded as unlabeled data and the semi-supervised part improves the performance of recommendation.In this thesis,the game prop recommender system runs in real online mobile phone game,which is based on an ensemble of the proposed three kinds of algorithms.The operation data shows the recommender system raises the profits significantly.