Research On Online Video Viewing Behavior Analysis On Smart TVs And Recommender Systems

In recent years,millions of households have purchased a smart TV and con-nected it to the Internet for the purpose of watching online videos on TV screens.Compared with traditional TV channels,the online video service on smart TVs offers a greater variety of video content,and makes the video viewing experi-ence more interactive and personalized.Viewers could watch whatever they are interested in at their convenience.A key pillar to the online video service is recommender system that could automatically identify a few videos of potential interest to the user from a huge repository of videos.In order to optimize the configuration of system resources and provide better service for users,it is nec-essary to understand the characteristics and patterns of the online video viewing behavior of smart TV viewers from different perspectives,identify various factors that have an impact on the recommendation performance,and design effective recommendation algorithms.Under the support of the National Natural Science Foundation of China,we conducted a series of research about user behavior analysis and recommender system based on a large-scale log containing online video viewing records on Hisense smart TVs and other benchmark datasets.The main research contents and innovations are summarized as follows:(1)We uncover several crowd-level temporal patterns of the online video view-ing behavior on smart TVs,conform the existence of holiday effect,and analyze the temporal dynamics of different video categories,which derives a more reason-able partition of the day that improves the video recommendation performance of time-aware recommendation algorithms.Time is an important contextual factor for the viewing behavior.The amount of time spent in watching online videos on smart TVs greatly depends on our daily and weekly cycles of free and busy time.We first measure the amount of time that each household spent in watching online videos on smart TV during each hour on each day.By applying clustering techniques,we uncover several interpretable daily patterns,whose peak hours occur in different segments of the day and align well with different dayparts in broadcast programming.Based on the affinity between the viewing behavior of each household across days and d-ifferent daily patterns,we identify a few temporal habits,which could roughly characterize the differences among households and also reveal the correlations between some dayparts.Next we switch to the weekly basis,and detect several weekly patterns with different characteristics,but they are all periodic with a period of 24 hours,indicating that there seems to be a circadian rhythm at the crowd level.Further statistically analysis confirms that there exists a holiday effect in the online video viewing behavior on smart TVs.We also investigate the popularity variations of different video categories over 24 hours of the day,and observe that some categories are more popular in certain dayparts.There-upon,we derive a more reasonable partition of the day,which improves the video recommendation performance of time-aware recommendation algorithms.(2)We identify a novel profile characteristic—profile coherence——hat has an impact on the online video recommendation performance on smart TVs,design four variants of the item-based collaborative filtering(CF)method,and analyze the impact of profile coherence on the accuracy,diversity and popularity of the recommendation lists.In traditional recommender systems,an account usually represents a single user.However,a smart TV corresponds to a shared account.What is observed consists of the mixed behavior of multiple users in a household who tend to have disparate interest.Therefore,the profile coherence should influence the video recommendation performance on smart TVs.We propose to measure the profile coherence of an account by the average similarity between videos that have ever been played on a smart TV.Then we make a distinction between coherent ac-counts and incoherent accounts.We also design four variants of the item-based CF method for top-N recommendation.The main differences lie in the recom-mendation evidence selection policy and ranking score prediction policy.The experiment results confirm that incoherent accounts indeed get less accurate rec-ommendations than coherent accounts,but the performance differences in terms of diversity and popularity depend on the recommendation evidence selection policy and ranking score prediction policy adopted by different variants.(3)We propose a tag-informed item embedding method towards learning bet-ter low-dimensional representations,which improves the top-N recommendation performance and the interpretability of item embeddings.Besides,we design two metrics to quantitatively evaluate the interpretability of item embeddings from two viewpoints—individual dimensions and local neighborhoods in the latent s-pace.Recently,the item-item co-occurrence relations are exploited to learn item em-beddings,however the derived item-item co-occurrence information is biased,due to the sparsity of the user-item interaction matrix and the long tail phenomenon of item popularity.By visualizing the lengths of the learned item embeddings,we find that the flaws in the input matrices lead to bias in learned item embed-dings towards popular and/or rare items.Because tags can be readily collected from multiple data sources,are less affected by item popularity,and contain rich semantic information,we utilize the item-tag relevance information to overcome the flaws in the input matrices,alleviate the bias in learned item embeddings,and improve the interpretability of item embeddings.Specifically,we propose to jointly factorize the user-item interaction matrix,item-item co-occurrence matrix and item-tag relevance matrix by sharing the item embeddings such that differ-ent forms of information could co-operate with each other to guide the learning of item embeddings.We also design two metrics to quantitatively evaluate the interpretability of item embeddings from two perspectives—the interpretability of individual dimensions of the latent space and the semantic coherence of local neighborhoods in the latent space.The experiment results demonstrate that the item-tag relevance information is beneficial to improving the recommendation performance.The sparser the user-item interaction matrix,the larger the rela-tive improvement.Besides,the bias of the lengths of learned item embeddings towards popular and/or rare items is alleviated to some extent.Meanwhile,the interpretability of individual dimensions of the latent space and the semantic coherence of local neighborhoods in the latent space are also improved.