際際滷

際際滷Share a Scribd company logo

ReTV AI4TV 2020

?
?
ReTV project
ReTV project

The document discusses using artificial intelligence to optimize the publication of video content across digital channels. It describes how attention spans for digital video are dropping and focuses have shifted to short-form content. The ReTV project aims to develop a platform that can analyze content across all channels and publish to multiple media vectors with minimal effort. The platform uses topics analysis and time series forecasting to predict the optimal time and channel for publishing certain content types and topics based on past trends. The goal is to help marketers and publishers effectively distribute their video assets.

1 of 14
Download to read offline
Optimising the Future Publication of Media Content through AI Lyndon Nixon, MODUL Technology AI4TV2020 workshop (online)
2 Video is everywhere! ★ How to get your audiences¨ attention? Business value of digital video marketing ? Consumers watch a full length video asset after being exposed to a summary (advertisement or trailer) ? Attention spans towards video on digital channels are dropping ? Many channels have focused their content distribution on short-form video (stories) Digital marketing of video contentrequires optimal selectionand summarization
3 Use of video assets in digital marketing Which of my media assets should I promote on social media in the coming week?
4 Use of video assets in digital marketing
5 Use of video assets in digital marketing Publish & track success
ReTV develops aTrans-Vector Platform (TVP) to analyse content across all channels and ＾publish to all media vectors with the effort of one￣
Topics Compass: explore trends in topics of online discourse Identifytopics of interest to the future audience
Prediction: what is the optimal publication time/channel for content according to its topic? We chose topicsin sports as there can be seen a past trend in onlinediscourserelated to sports events. We used topic mentionsin globalEnglish& German news sources for the time series training data. The training/testingdataset covered 28 months (dailyvalues = 841 data points) and we predict for the next 30 days. The first attempt lookedat different keywords(cycling,formulaone, wimbledon)and found that each keywordwill differin its behaviour and therefore predictivepower. We found in testing with ARIMA that the best functioningmodelwas SARIMAX (7,1,2) with seasonal order (0,0,0,365), below the MAE and RMSE for the prediction: CYCLING 15.3 18.5 FORMULAONE 22.2 37.0 WIMBLEDON 44.1 79.1
Prediction: what is the optimal publication time/channel for content according to its topic? We comparedSARIMAX and LSTMs for predictive accuracy measured by MAE and RMSE. LSTM Encoder Decoder with Seq2seq and Luong attention has performedbest for multi-step forecasting. However, SARIMAX outperformed LSTMs for one step forecastingwith autoregression. MAE RMSE SARIMAX 15.3 18.5 LSTM 13.7 15
Prediction: what is the best topic to choose on a future date? Our events and anniversariesAPIhighlights important events and anniversariesona specific date. TheTopicsCompass can identifytime references inWebpagesand socialmediaposts and aggregate those documents that refer to a specificdate, displayingthe top keywords extracted from the document cluster.
Prediction: what is the best date to choose for a certain topic? Left: RBB, all media, Jun 10-Dec 31 2020. Bookmark ＾Airports￣. Right: NISV, all media, Jun 10-Dec 31 2020. Bookmark ＾Events￣.
Prediction: hybrid model (best of all worlds) It appearsmulti-step time series forecastingcanhelp predict future topicaltrends when the topichas enough past data with a discernibleseasonality. Topics that emerge in recent time or lackpast seasonaltrends are more difficult: usingone step time seriesforecastingwith autoregression,the accuracy dropsas we look further into the future (10 fold validationresults in table, left, for 'cycling'usinga LSTM Encoder-Decoder modelwith TimeSeriesGenerator) Future work is to see how the other prediction features (event, future temporal references) can help improve the accuracy of prediction in longer time periods (>10 days!) Days into the future Average MAE Average RMSE 2 13.5 14.7 7 14.3 15.6 10 16.4 17.6 20 23.5 24.8 30 25.1 26.7
The ReTV StakeholderForum is your opportunity to engagewith us, be first to get updates and have the opportunity to test our tools and applications! Get news, see demos and read about our tools and case studies at https://www.retv-project.eu Sign up for the ReTV Newsletter and get an update every few months from us!
@ReTV_EU Facebook:ReTVeuwww.ReTV-Project.eu Instagram:retv_project Dr. Lyndon Nixon ReTV Project Coordinator info@retv-project.eu @ReTV_EU @ReTVeu ReTV Project retv_project

More Related Content

ReTV AI4TV 2020

  • 1. Optimising the Future Publication of Media Content through AI Lyndon Nixon, MODUL Technology AI4TV2020 workshop (online)
  • 2. 2 Video is everywhere! ★ How to get your audiences¨ attention? Business value of digital video marketing ? Consumers watch a full length video asset after being exposed to a summary (advertisement or trailer) ? Attention spans towards video on digital channels are dropping ? Many channels have focused their content distribution on short-form video (stories) Digital marketing of video contentrequires optimal selectionand summarization
  • 3. 3 Use of video assets in digital marketing Which of my media assets should I promote on social media in the coming week?
  • 4. 4 Use of video assets in digital marketing
  • 5. 5 Use of video assets in digital marketing Publish & track success
  • 6. ReTV develops aTrans-Vector Platform (TVP) to analyse content across all channels and ＾publish to all media vectors with the effort of one￣
  • 7. Topics Compass: explore trends in topics of online discourse Identifytopics of interest to the future audience
  • 8. Prediction: what is the optimal publication time/channel for content according to its topic? We chose topicsin sports as there can be seen a past trend in onlinediscourserelated to sports events. We used topic mentionsin globalEnglish& German news sources for the time series training data. The training/testingdataset covered 28 months (dailyvalues = 841 data points) and we predict for the next 30 days. The first attempt lookedat different keywords(cycling,formulaone, wimbledon)and found that each keywordwill differin its behaviour and therefore predictivepower. We found in testing with ARIMA that the best functioningmodelwas SARIMAX (7,1,2) with seasonal order (0,0,0,365), below the MAE and RMSE for the prediction: CYCLING 15.3 18.5 FORMULAONE 22.2 37.0 WIMBLEDON 44.1 79.1
  • 9. Prediction: what is the optimal publication time/channel for content according to its topic? We comparedSARIMAX and LSTMs for predictive accuracy measured by MAE and RMSE. LSTM Encoder Decoder with Seq2seq and Luong attention has performedbest for multi-step forecasting. However, SARIMAX outperformed LSTMs for one step forecastingwith autoregression. MAE RMSE SARIMAX 15.3 18.5 LSTM 13.7 15
  • 10. Prediction: what is the best topic to choose on a future date? Our events and anniversariesAPIhighlights important events and anniversariesona specific date. TheTopicsCompass can identifytime references inWebpagesand socialmediaposts and aggregate those documents that refer to a specificdate, displayingthe top keywords extracted from the document cluster.
  • 11. Prediction: what is the best date to choose for a certain topic? Left: RBB, all media, Jun 10-Dec 31 2020. Bookmark ＾Airports￣. Right: NISV, all media, Jun 10-Dec 31 2020. Bookmark ＾Events￣.
  • 12. Prediction: hybrid model (best of all worlds) It appearsmulti-step time series forecastingcanhelp predict future topicaltrends when the topichas enough past data with a discernibleseasonality. Topics that emerge in recent time or lackpast seasonaltrends are more difficult: usingone step time seriesforecastingwith autoregression,the accuracy dropsas we look further into the future (10 fold validationresults in table, left, for 'cycling'usinga LSTM Encoder-Decoder modelwith TimeSeriesGenerator) Future work is to see how the other prediction features (event, future temporal references) can help improve the accuracy of prediction in longer time periods (>10 days!) Days into the future Average MAE Average RMSE 2 13.5 14.7 7 14.3 15.6 10 16.4 17.6 20 23.5 24.8 30 25.1 26.7
  • 13. The ReTV StakeholderForum is your opportunity to engagewith us, be first to get updates and have the opportunity to test our tools and applications! Get news, see demos and read about our tools and case studies at https://www.retv-project.eu Sign up for the ReTV Newsletter and get an update every few months from us!
  • 14. @ReTV_EU Facebook:ReTVeuwww.ReTV-Project.eu Instagram:retv_project Dr. Lyndon Nixon ReTV Project Coordinator info@retv-project.eu @ReTV_EU @ReTVeu ReTV Project retv_project