Specific Instance Detection Based Multi-Instance Learning And Its Applications To Virtual Props Recommendation

In traditional supervised learning,each object has a label associated with it.How-ever,there are many weakly or ambiguous labeled data which have complicated se-mantics in the real world.Multi-instance learning(MIL)was proposed to use this kind of weak label information.In MIL,the training set consists of a number of bags with concept labels,and each bag has a number of instances unlabeled.MIL reflects the complexity of practical tasks in the real world and has been widely applied to in diverse fields such as drug activity prediction,natural scene classification,document classifi-cation and object detection.Most existing MIL approaches focus on discriminating bags instead of directly detecting the region of positive concepts.However,in various applications of MIL,finding what instances trigger the corresponding labels is still a matter of concern.For example,in game props recommendation system,the relationship between users' pur-chase activities and their in-game behavior is attractive to operators.Considering the existing MIL issues,this dissertation studies positive instance detection and its appli-cations to virtual props recommendation.The contributions can be summarized as fol-lows:1.We propose a novel positive instance detection approach based on linearly lo-calized interpolation.Inspired by the Diverse Density algorithm,we assume negative instances in positive bags have the same properties with those in negative bags and distinguish the positive instances in bags from others by testing whether they can be linearly localized interpolated by the negative instances in negative bags.We then cast the MIL into a standard supervised learning problem and use support vector machines to solve it.Empirical investigations on drug activity prediction,text and image classi-fication demonstrate the decent classification performance of the proposed method.2.We use MIL technique on a large scale virtual props recommendation system.We analyse the relationship between assumptions of MIL and the issues in virtual props recommendation system such as complicated dependency on context,long-distance in-tervention and props priority-role dependency.We then model the game props recom-mendation into a multi-instance multi-label learning task for utilizing the complicated dependencies and capturing the rank of purchase intentions.We implement the algo-rithm on Spark,and do experiments on data set from real game log records.The results show our algorithm can effectively handle tens of millions of examples and outperform the traditional collaborative filtering methods.