Design And Implementation Of Video Recommendation System Based On Deep Watching Interest Network

With the development of 4G,5G and big data technology,various types of video applications have emerged.These applications have gradually become one of the indispensable applications for people.Recommendation systems are also playing an increasingly important role in the context of information overload.Therefore,we aim to design and implement an effective video recommendation system for a video application,and optimize the recommendation algorithm to improve user click-through rates and conversion rates.Essentially,the recommendation algorithm aims to model the user’s preferences,so as to select the item that the user may be interested in and recommend it to the user.However,most of the current mainstream models for click-through rate directly use the user’s click behavior as a reflection of the user’s interest.This approach is not accurate.Other feedback information needs to be considered to assist the learning of the user’s interest.In the video recommendation scenario,the user’s watching duration of the video contains more information.Therefore,we want to use users’watching duration to assist the training of the interest model.In order to reduce the impact of the original video duration on the watching duration,we use watching depth instead of watching duration.Our model uses user’s interest vector to fit the user’s watching depth.Estimation error of watching depth participates in model optimization as an additional loss,and regularizes some parameters of interest extraction.We adopt a deep neural network for the model structure,and use attention mechanism to mine the correlation between the user’s historical behaviors and between historical behaviors and the target video.Then we verify the effectiveness of the model by comparing with other mainstream algorithms.The results show that our model has a certain improvement in GAUC.In addition,we have designed and implemented a recommendation system to provide recommendation services for a video application.The system includes five modules,which are data acquisition module,data storage module,recommendation calculation module,request processing module and recommendation management module.In order to alleviate the cold-start problem and data sparse problem of the recommendation system,improve the coverage and diversity of the recommendation list,and optimize the user experience,we use hybrid recommendation method and apply multiple different types of recommendation models in the recommendation calculation module.At the same time,considering that there are multiple optimization indicators in the recommendation system,we design a fusion formula to calculate the unified recommendation score.After the development of the recommended system is completed,we test this system to verify that the functions and performance of the system meet expectations.