# Teaching Push Notifications to Learn Who's Listening > Source: https://www.recombee.com/blog/teaching-push-notifications-to-learn-who-s-listening > For the complete site index, see [llms.txt](/llms.txt). To integrate Recombee, see the developer documentation at https://docs.recombee.com/llms.txt. [Blog](https://www.recombee.com/blog) # Teaching Push Notifications to Learn Who's Listening ![](https://www.recombee.com/img/blog/authors/rodrigoalves.png) Rodrigo Alves Oct 8, 2026 ![](https://www.recombee.com/img/blog/teaching-push-notifications-to-learn-who-s-listening/main.png) Picture this: your phone buzzes. It's your streaming app, and for once the notification is the one you actually wanted: "LIVE: your team just went 1–0 up, 78th minute”. You tap it, you watch the replay, you're still buzzing about it the next morning. Now picture the other version of that same evening: five pushes for shows you'll never watch, one discount code you've already ignored twice, and, buried somewhere in the pile, the goal alert: twelve minutes late, after you'd already seen it on social media. Both of those evenings came out of the same box of tools. The difference is who got notified, about what, and when. And that difference is a recommendation problem, whether or not the team behind the notification system thinks of it that way. ## The Cold Start Nobody Talks About Most of what we write about on this blog deals with a familiar version of the cold start problem: a new user with no history, or a new item that nobody has rated yet. Outreach introduces a third, and particularly challenging, version of the same problem: the new campaign. A season kickoff, a title launch, a breaking story, or a live match all share the same constraint. The moment a new campaign needs to go out is also the moment it has zero interaction history of its own to learn from. You cannot wait a week to observe who engages before deciding who to notify. By then, the moment, and often the reason for reaching out at all, is gone. This is what we mean by _active recommendation_: not only deciding what to show someone once they have opened an app, but deciding, under a limited budget of attention and with almost no prior signal, whom to contact in the first place. The cost of getting this wrong can be measured in unsubscribes and disabled notifications. Industry data consistently shows that sending just two to five pushes per week can lead a substantial proportion of users to disable notifications altogether, while increasing that frequency further can lead some users to delete the app. Personalized outreach, by contrast, can achieve considerably higher engagement than generic broadcasts. The difference between these outcomes provides a strong motivation for studying how outreach decisions can be made more effectively. ## Our Active Recommendation Method This is the problem we studied, and it is the main topic of our paper \[1\] presented at CIKM 2025, the 34th ACM International Conference on Information and Knowledge Management. The method combines two components that are not typically used together in order to learn which users should be targeted during campaigns. First, a shallow autoencoder extracts a collaborative filtering signal: a compact and relatively fast to train representation of who tends to respond to what, learned from the interaction history available across previous campaigns. Second, we use this signal within a Thompson sampling multi armed bandit that actively decides, user by user, who should be contacted next. In doing so, it explicitly balances cases where we are relatively confident that a person will engage against cases where uncertainty remains and obtaining additional information may itself be valuable. Crucially, the bandit updates its beliefs as responses arrive during the campaign, without requiring the underlying model to be retrained between batches. This makes the approach practical and scalable for campaigns whose useful lifetime may only be a few hours. In our experiments, this combination outperformed standard retrieval baselines on metrics that are particularly relevant to outreach, including precision at the top of the ranked list. In practical terms, this asks whether the system is selecting the recipients who are most likely to engage. At the same time, the approach remains sufficiently interpretable to provide insight into why a particular recipient was or was not targeted. This is an important consideration when recommendation systems are deployed in settings where targeting decisions may need to be understood by marketing, editorial, or compliance teams. Note that push notifications, in app banners, emails, and CRM triggered SMS messages can all be viewed as instances of the same general question: given a new campaign with almost no interaction history, who should we contact, and how confident are we in that decision? The channel changes the mechanics of delivery, but the underlying outreach dynamics remain largely the same. This is also the type of pipeline we have worked on, where recommendation signals can inform the notification a user sees on their lock screen rather than only the content they encounter after opening an application. ## Why This Is Important in Media and Sports Live sports and breaking news provide particularly clear examples of this problem. A goal, a red card, a transfer, or a breaking headline can each trigger a new campaign that is created and becomes outdated within a very short period of time. The value of the information may decay within minutes or even seconds. There is therefore no opportunity to spend a week collecting interaction data for that specific event. The decision about who receives a notification has to be made using what the model already knows about each user, while incorporating new information from the users who have already responded. This is precisely the type of setting that motivates active recommendation. The system must make decisions immediately, learn as responses arrive, and update its targeting strategy without waiting for a complete model retraining cycle that may finish only after the relevant moment has passed. This is the environment our bandit based approach was designed to address. If you enjoyed this work, please cite the paper, send us your comments, or reach out to our research team directly. We are always interested in hearing where these ideas do and do not hold up in production. Our dataset is also available for researchers here. ## References \[1\] Zid, C., Alves, R., & Kordik, P. (2025). Active Recommendation for Email Outreach Dynamics. Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM 2025), pp. 5540–5544. * Recommendation Engine * Personalization ## Next Articles [![](https://www.recombee.com/img/blog/new-feature-ab-testing.png)](https://www.recombee.com/blog/new-feature-ab-testing) ### [New Feature: A/B Testing](https://www.recombee.com/blog/new-feature-ab-testing) Personalization is never finished. Every change to your recommendation strategy raises new questions. Should you boost fresh content? Change used Logic? Introduce new filters? ![](https://www.recombee.com/img/blog/authors/annapetr.png) Violeta Milarova Aug 14, 2026 * Recommendation Engine * Personalization [![](https://www.recombee.com/img/blog/mid-year-roundup-2026-so-far.png)](https://www.recombee.com/blog/mid-year-roundup-2026-so-far) ### [Mid-Year Roundup: 2026 So Far](https://www.recombee.com/blog/mid-year-roundup-2026-so-far) The digital world continues to change at an incredible pace. As technology evolves and user expectations keep rising, creating experiences that feel relevant has never been more important. ![](https://www.recombee.com/img/blog/authors/annapetr.png) Violeta Milarova Jul 20, 2026 * Recommendation Engine * Personalization [![](https://www.recombee.com/img/blog/a-2025-research-retrospective.png)](https://www.recombee.com/blog/a-2025-research-retrospective) ### [A 2025 Research Retrospective](https://www.recombee.com/blog/a-2025-research-retrospective) In 2025, I became Head of Research at Recombee. I knew it would be a challenge: and I wasn’t wrong. Recombee is a leader in recommendation-as-a-service, with thousands of clients... ![](https://www.recombee.com/img/blog/authors/rodrigoalves.png) Rodrigo Alves Jan 29, 2026 * Recommendation Engine * Personalization