PDF Publication Title:
Text from PDF Page: 008
CIKM’18, October 2018, Turin, Italy Mark Wilhelm, Ajith Ramanathan, Alex Bonomo, et al. Figure 5: Although diversification did little to increase in- teractions directly on the homepage, it did increase the to- tal originating from the homepage (taking the related video panel into account). Error bars represent 95% confidence in- tervals. Figure 6: Evidence of a long-term “learning effect” as seen in the number of people watching videos from the homepage. The implication is a much more useful and satisfying prod- uct experience. feed. Cumulatively, it suggests that users find more videos that they enjoy compared to before. Moreover, we have been able to observe a long-term “learning effect” [17] from diversifying users’ feeds. That is, diversification results in users returning to and enjoying our service more as time goes on. We measured this effect by running two sets of long-term holdback experiments2. In the first holdback condition, users do not get DPP-diversified feeds, but that subset of the user popula- tion changes every day (these users are normally exposed to the diversified feed, except on the rare day that they end up in this holdback set). In the second holdback condition, a consistent set 2A holdback is simply an A/B experiment where users in group B do not receive the launched treatment. of users do not see DPP-diversified feeds. We can then observe whether DPP diversification results in a long-term improvement in user experience by observing the difference between the two holdbacks when compared to their respective control groups. As we can see in Figure 6, which shows the increase in number of users watching at least one video from the homepage against these two holdback groups, users who have been exposed to diversified feeds more often realize that they can find videos of interest on YouTube’s homepage. Therefore, we can say that diversified feeds lead to increased user satisfaction in the immediate term, and that this effect becomes even more pronounced over time. 6 CONCLUSIONS AND FUTURE WORK Researchers realized well over a decade ago that diversification is an important problem for recommendation systems, and for in- formation retrieval in general. Significant research efforts have invested in approaches that use a taxonomic or category-based approach, often combined with a variety of heuristics. In contrast, we propose using a method based on determinantal point processes (DPPs). Our approach performs set-wise optimization of recom- mendations. Since this approach naturally factors the problem into one of estimating item quality, and another of estimating repulsive effects between pairs of items, our stacked architecture allows us to leverage existing sophisticated investments in pointwise scoring and item analysis. In this paper, we discussed the challenges of applying DPPs in a large-scale video recommendation system. We considered several parameterizations of the DPP kernel as well as learning methods for computing the value of the kernel parameters from positive user interactions with videos. Finally, we presented live experiment results on this large-scale system, showing both an immediate short- term lift in user utility, as well as long-term effects—users looked to YouTube more often to satisfy their needs. Our work is not without limitations. First, the DPP we trained is non-personalized in that the parameters such as σ are learned from training on a large population of user data, not on a single user’s data. In the near future, we hope to develop new approaches to understand each individual user’s short-term and long-term di- versification needs. We also do not fully understand how different domains or genres might affect diversification policy. For instance, users might prefer music videos to stay within a certain boundary (no vocals, for instance), as they might be enjoyed somewhat more passively, while genres like comedy might need more diversity. Additionally, we do not have a good model that takes time into account, such as understanding weekday vs. weekend diversity pref- erences. We would like to explore the connection of diversification methods with reinforcement learning, so that we can learn a good control policy for diversification. Given the multitude of directions for future work, we feel that our current work simply “scratches the surface” of the possibilities available to improve user experiences by moving away from pointwise estimators in recommender systems. REFERENCES [1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, San- jay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Leven- berg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, MikePDF Image | Practical Diversified Recommendations on YouTube with Determinantal Point Processes
PDF Search Title:
Practical Diversified Recommendations on YouTube with Determinantal Point ProcessesOriginal File Name Searched:
cikm2018.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com (Standard Web Page)