CM Blog > Page

“You are what apps you use!” – Transfer Learning based Personalized Content Recommendations

February 21, 2017|Data Intelligence

By Li San Hor, Technical Program Manager
Special thanks to Bo Tao, Bruce Deng, Luke Lu and Zhixian Yan for all their contributions and reviews to this blog post.

Making compelling personalized content recommendations (e.g. news, videos, images) to users has become a competent differentiator for today’s online web services. It is an essential capability to yield higher ROIs and maximize user’s engagement. “You are what apps you use!” – Based on this insight, we tapped into leveraging the user data of our applications, and built a Universal User Profile via neural network model to enable personalized content recommendations to new users of News Republic, our content application. This Transfer Learning approach has enabled us to solve the cold start problem with close to full coverage of our user base while yielding significant CTR growth in our A/B Testing.

Transfer Learning

News Republic is our top-rated news app on the App Store and Android Market. It is an app intended to spark a movement to promote global understanding, dialogue, and change. This news app is available in English, French, Spanish, Russian and Italian.

At the inception of the project, the user profile coverage for News Republic was on a slow growth trajectory, a major challenge for most of the newly launched applications. We recognized the need to scale promptly targeting more potential new users with no or limited click history. Thus, making the personalized content recommendations to them becomes the key goal – solving the typical cold start problem. In addition, we would also like to improve our personalized recommendation capturing the interests of our existing users.

Our Solution
Based on the previous publications on cross domain recommendations (Elkahky et al., 2015; Covington et al., 2016; Tang, Jie et al., 2012), we initiated the effort to leverage the great insight from our existing user profiles and transfer the learning to apply to our new users. The existing user profiles enable us to conduct aggregate analysis on user data (with user consent and no personal identifiable information). For example, we are able to correlate on what content the users are interested in based on the apps installed, etc. In our work, we constructed rich feature sets based on these user profiles information. We then apply the neural network-based classifier to train the rich feature sets to classify our data set to form Universal User Profile. New users are mapped against the Universal User Profile so that we are able to identify the similar behavior pattern and transfer the learning to enable better content recommendations to them.

Test Results
Our test results reported significant increase in CTR for new users with no or limited clicks information. In addition, after we rolled out the model to production, our user base coverage has increased to close to full coverage, a significant improvement in performance.

Next Step
Our next step is to further explore into Deep Learning to refine content analysis so that we can further bring the personalized content recommendations into the next level yielding higher performance.

Interested to Join us?
If you are interested to be part of our AI / Deep Learning / NLP initiatives, we are hiring!

Bo Tao, Bowen Zhang, Bruce Deng, Frank Dai, Jerry Wu, Luke Lu, Meng Zhang, Roy Wei, Shixing Shen, Stephen Brodsky, Tony Philip, Xu Yang, Yuewen Wang, Zhixian Yan.

Elkahky, Ali Mamdouh, Yang Song, and Xiaodong He. “A multi-view deep learning approach for cross domain user modeling in recommendation systems.” Proceedings of the 24th International Conference on World Wide Web. ACM, 2015.

Covington, Paul, Jay Adams, and Emre Sargin. “Deep Neural Networks for YouTube Recommendations.” Proceedings of the 10th ACM Conference on Recommender Systems 15 Sep. 2016: 191-198.

Tang, Jie et al. “Cross-domain collaboration recommendation.” Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining 12 Aug. 2012: 1285-1293.