Research On User's Preference Prediction Model Based On Short Video Content Understanding

In recent years,with the rapid emergence of social networking media,the amount of information on the internet has grown exponentially.The recommender systems are now emerging as important selection supporting gear in the online advertising market.Recommender systems analyze and filter complicated information methodically to assist customers in finding thrilling items.They also minimize the issue of information overflow on websites as well as on mobile applications.Likewise,it increases loyalty by developing a value-added relationship between the websites and their users.An enormous amount of research work has been dedicated to video recommendation techniques for web sites.In this paper,we have looked into an effective recommendation method for short video applications.Short video applications have received applaud from customers across the world.Also,it has always been an excellent goal for the artificial intelligence system of a corporation to get a better understanding of video content and subsequently suggest the preferences of users representing their choices and insights.In the last few years,short video applications have become part of our daily lives,and a plethora of short videos are being watched by customers every second.Applications of short videos such as Tik Tok,has achieved global popularity because of the pervasiveness of social media and easy accessibility of recording devices like cell phones.Similarly,with the arrival of the information age,the use of past users behaviour from multi-modal resources has a vital role in the video recommendation systems.In order to keep users amused and engaged by videos every time while they open the Tik Tok,the recommending strategy must be established based on user's preferences.This fascinating system includes and reflects the user's behaviours and the information of video contents,consisting of the subject,the quality of editing,background music,and the author.Numerous attempts have been made to address the issue of video recommendation systems.Majority of the recommender applications are being developed for particular websites.But in this paper,we have shed light on the user's prediction based on short videos understanding.We have developed a universal mechanism for quick video recommendations by forecasting the chances of watching the whole video.To analyze user behaviour,we have extracted multi-modal video features,consisting of visual capabilities,text features and audio capabilities.In addition to consumer interactive conduct information,such as clicks,likes,and observe from the dataset,we have modeled the user's preferences using a video and user interaction behaviour information,and then predict the consumer's click behaviour on any other dataset.We have made numerous efforts in model structure and feature engineering.We have utilized deep learning-based framework and gradient boosting technique to solve this problem effectively.Deep Learning is a sub-field within Machine Learning that has shown promising results in prediction problems over the last few years.Gradient boosting is also a machine learning technique that produces a prediction model in the form of an ensemble of weak prediction models.We've developed two neural network-based models and one gradient boosting decision tree model,and also conducted an experiment on ensemble learning.We have used the ensemble method to combine three baseline models in order to produce one optimal predictive model.In ensemble learning,average models are great and straightforward.We have shown our results from experiments conducted on a real dataset extracted from biendata company.Our experimental results consist of two parts,the first one is the experimental results consists of the models that we trained,and second is the final score got by the fusion model.We have use Area Under Cover(AUC)as our evaluation metric.Our results have clearly shown that there is room for improvement in the AUC whenever we use hybrid model.