| With the popularization of Internet information and the improvement of people’s spiritual educational level,video platforms have developed rapidly,and the number of users also continues to grow.The video platform provides users with massive video resources,which are divided into two types:paid videos and free videos.Among them,for free videos,all users can watch for free.For paid videos,the platform provides a variety of packages,such as education package,children package,etc.Each package contains corresponding types of paid videos.After purchasing a package,the user will become a member of the package and can enjoy some membership services,such as watching the videos in the package,getting early access to the latest updated episodes,etc.Moreover,membership revenue has become one of the main sources of revenue in most video platforms.Therefore,effectively predicting whether users will purchase a package in the next few days,and which package to purchase,will help the platform to push the package to the corresponding users.It can increase the revenue of the video platform,enhance user stickiness,and enrich the user’s viewing experience.So it has great research significance.At present,there are few related research to solve this problem.The solution to similar tasks is to model user behavior,extract the user’s interest,and predict the probability of a user purchasing the target item.However,in the application of these methods,there are following challenges that prevent them from predicting user purchase packages effectively.The first challenge is how to effectively learn the embedding representations of videos and packages,which can effectively capture their dynamically changing features.Most of the videos on the platform are provided to the users in the form of multiple episodes and regular updates,such as TV series,variety shows,etc.The daily view count and popularity of the videos are also constantly changing.These changing features are important for video embedding representation.In addition,due to the package containing a large number of videos,and some features of the video being dynamically changing,the package is also dynamically changing.So how to capture such dynamic features to effectively represent videos and packages is a crucial challenge to the prediction task.The second challenge is how to capture the evolution pattern of users’ purchase intentions for packages.In a video platform,the fact that a user likes to watch a certain type of video does not mean that he/she wants to buy a package that includes this type of video,because free videos might satisfy his/her interest.Moreover,with the continuous updating of videos and the frequent viewing of videos by users,the purchase intention of users gradually evolves over time.How to dig out the evolution pattern of users’purchase intention over time from the user’s behavior history is an important key point to solve the prediction task.To solve the above two challenges,this thesis carries out the following two-stage work.In the first stage,this thesis proposes a Dynamic Embedding based Fine-grained Intention Model for Membership Prediction(Dy-FIM),which is used to predict whether a user will purchase a package in the next few days.This model firstly utilizes the idea of graph embedding learning to design a dynamic embedding learning module,which can capture the dynamically changing features.It solves the challenge of representing dynamic changes.Besides,the model designs a user’s fine-grained intention extraction module to deeply mine the user’s intention to interact with the package.It can improve the prediction effect.In the second stage,this thesis proposes a Dynamic Evolution based Deep Hierarchical Intention Network for Membership Type Prediction(Dy-HIEN),to predict which package the user will purchase in the next few days.The model firstly proposes a dynamic embedding learning module to learn the video and package embedding representation.This module optimizes the video embedding learning method proposed by Dy-FIM model.It designs a Multi-relational Heterogeneous Information Network(MHIN),which can enrich the embedding and capture the dynamic features.Then,a hierarchical intention evolution module is proposed.While deeply mining the user’s intentions to purchase packages,it also models the evolution pattern of the purchase intentions over time.Thus this module can effectively solve the challenge of how to capture the evolution pattern of users’ purchase intentions for packages.The main work and contributions of this thesis can be summarized as follows:1.The Dy-FIM model is proposed to solve the prediction task of whether users will buy the package in the future.The model firstly designs a video embedding learning method.It applies the auto-encoder method,in-degree and out-degree structure matrices of network,to learn the local embedding of videos effectively.Besides,the attention mechanism is applied to learn the global embedding of videos,which can effectively capture the influence of different state of videos in the different time and the dynamically changing features.Then,the Dy-FIM model designs a fine-grained intention extraction method,which applies the attention mechanism and sequence encoding method to mine the intention to interact with the package.Finally,it is combined with the future embedding representation of in-package videos to predict the probability that the user will purchase the package.2.The Dy-HIEN model is proposed to accurately predict which type of package users will purchase.Based on the video embedding learning method,this model further proposes a dynamic embedding learning module,which designs a MHIN structure.It can extract the multiple relationships between nodes,and enrich the embedding representation.This module uses a position-aware attention mechanism to capture the dynamic features of videos and packages.Then,in order to capture the evolution pattern of users’ purchase intention to packages,Dy-HIEN model proposes a hierarchical intention evolution pattern module.After dividing the play history,this module uses time-order encoder and kernel function to mine users’real purchase intention,and deeply capture the intention evolution pattern over time.3.A large number of experiments are carried out on two real datasets,which can verify the effect of Dy-FIM model and Dy-HIEN model.This thesis selects Hisense dataset of smart TV platform and Sparkify dataset of popular music website for experiments.Experimental results show that our model can realize effective prediction of users’purchasing packages.Compared with some existing baseline methods,the results show that our model has greatly improved in several evaluation metrics.At the same time,in order to verify the effectiveness of each module in the model,sufficient ablation experiments are conducted in this thesis.The comparison and analysis of experiment results show that each module can well solve the corresponding challenges and improve the prediction effect. |
