Intelligent Push Notification Systems for Smartphones

1. Introduction

Push notifications from mobile phone applications allow businesses to capture the attention of the user and persuade them to use their application. It also allows users to consume time-sensitive information quickly, thereby improving the user experience as a whole. Sadly, these notifications often interrupt the user at a wrong time, hence not helping both the consumer and the sender of the notification. Furthermore, there exist many other issues with notifications being randomly pushed at a user. The issues arise when a lot of media-rich notifications are pushed at users who have certain limitations such as expensive data plans, constricted bandwidth availability or low battery levels on their mobile phones.

An adaptive system that manages the delivery of notifications to a user’s phone would ensure that the right notification is sent at the right time thereby eliminating the issues that were discussed above. I believe the following components would make up for a good adaptive notification management system – Content, Presentation and Delivery. These components will be discussed in detail in the subsequent sections of this post.

2. Contents of the notifications

2.1. Providing valuable content

The content of the notification is what draws the attention of the user. The usefulness of the notification is determined by the notification’s contents. Uddin et al. [1] propose an end-to-end framework called RichNote that deals with the delivery of rich notifications to mobile users. With RichNote they intended to provide an adaptive media-rich notification management service for the music streaming application Spotify. The evaluation for the framework was done using de-identified production traces from Spotify.

The content utility of a notification aims to gauge the likelihood of the user clicking on the notification to consume the full content that the application has to provide. This likelihood depends on the user’s personal preferences, the general popularity of the content and the relationship that the sender and receiver of the notification share. These factors form a feature space that with the right training data can be used to train a machine learning classifier that will be able to predict if a notification brings valuable content or not. [1]

2.2. Generating persuasive content

The paper by Fraser et al. [2] indicates the possibility of extracting persuasive features and using them to create notifications that have persuasive content in them. Based on Cialdini’s 6 principles of behaviour [3], six features of persuasiveness are chosen for push-notifications and they are - Scarcity, Authority, Reciprocity, Commitment and Consistency [2]. On further study, they were able to identify and rank these features in the order of importance. Using this order it is pretty straight forward to understand which features should be given higher weightage while generating persuasive content for a notification. The results of their paper clearly show that recommended persuasive notifications can be generated and that they perform significantly better than any other type of notifications (as shown in Figure 1).

2.3. Optimisation of content for better overall performance

A study on a mobile news application called “Apple Daily” was conducted by Saikia et al. [4] to analyse how users responded to mass notifications. The main parameters of this study are Notification Reaction Time, Notification Opening Rate, Volume of App Usage, Categories and Times of the Day. Notification Reaction Times revealed each user’s personal preferences, Notification Opening Rates showed that most users have low opening rates and very few users had high opening rates, Volume of App Usage showed that there are very few highly active users and that most users are less active, the Categories parameter revealed that those categories that are generally popular might not be so popular when sending it via notifications and finally Times of the Day indicated that most users accessed the application without assistance in the earlier part of the day and with the help of notifications in the latter part of the day. Based on these parameters a convex optimisation problem was formulated. The tool used in the paper to solve the convex optimization problem was CVX [5]. The optimised approach would send the maximum number of notifications wherein the number of notifications opened would be high while the reaction time of users would be less. Finally, it was observed that the optimised approach gives a better overall performance as it sacrificed the performance of non-popular categories to ensure that the overall performance was the best one [4]. This kind of optimisation of content would mean only extremely relevant and useful notifications are delivered to the user.

3. Presentation of the notification

Presentation is the amount of useful information that can be packed into a single notification. Richer the presentation, more resources are required to deliver it. As discussed by Uddin et al., [1] they relied heavily on user surveys to get an understanding as to what kind of presentations they wanted if Spotify were to give the users notifications. Combining content utility (as discussed in 2.1) with presentation levels received from user surveys, the utility of the notification is derived with varying presentation levels. According to the Spotify instance given in the paper by Uddin et al., presentation levels just means the audio sample they would let the user play in a notification. Based on two attributes (sampling rate and duration) multiple types of presentation levels can be generated, out which a few are chosen based on user surveys. These chosen few presentation levels are ordered according to their utility size (bandwidth and power-related costs) and used accordingly. The selection and delivery of rich notifications according to the RichNote framework is modelled as an optimization problem to maximise the utility of notifications while staying just under the resource budget constraints [1].

Looking at all the advantages that the RichNote framework proposes with regards to presentation utility, we can see that having an adaptive system to deliver the presentation of notifications can greatly help solve most of the issues that standard FIFO (First In First Out) or UTIL (Decreasing order of utility score) based notification management systems face.

4. Delivery of Notification

Delivering the notifications at the right time with the right techniques limits user frustration/distraction and instead increases the responsiveness towards the application.

4.1. RichNote’s round-based delivery model

The RichNote framework that was proposed by Uddin et al. [1] is excellent at the delivery of notification in mobile devices. It has a solution to almost all possible issues that might arise in the delivery of application notifications to a mobile phone. It uses a round-based model for notification delivery where each notification is added to a queue and a lot of processing is done to ensure perfect utility (content and presentation) is assigned to each notification, only after which the notifications are pushed to a mobile phone. RichNote optimises certain objectives under queue stability using Lyapunov analysis. This technique ensures that the queue is always stable (i.e. finite average length) while maximising the utility provided by the notifications.

Further in the paper, energy constraint is added as an objective thereby effectively creating a virtual queue that keeps track of how much energy is allowed to be spent in each round. At each round, the user’s energy and data budget values are updated. When a notification is pushed to the user, an appropriate amount of data and energy is deducted from their respective budgets. An important point to be noted here would be that the energy budget is variable as it depends on the current battery level of the mobile phone.

RichNote was observed to have better overall performance when compared to other notification management systems like FIFO and UTIL as they have a fixed presentation level and they do not adapt to network conditions such as when the device is connected to WiFi or cellular network. On the contrary, RichNote always adapts the presentation level to match the budgets (energy and bandwidth) and can also adapt to network conditions to provide appropriate levels of presentation in the notification. [1] Such adaptive systems prove that notification management systems can indeed be smart with how they deal with the delivery of notifications to mobile users.

4.2. Breakpoint / Interruptibility based delivery model

Okoshi et al. [6] equipped the Yahoo! Japan android app with a mobile-sensing and machine-learning interruptibility estimation logic to conduct a large-scale in-the-wild study with more than 680,000 users for three weeks. It was shown that in most cases delaying the notification delivery until an appropriate interrupt time was detected helped reduce the notification response time by 49.7%. There was a steady improvement in user engagement levels on the application throughout the entire study period thereby proving that their proposed method was a solid choice to decide when to deliver a specific notification.

In the paper published by Okoshi et al., they used activity recognition API to detect changes in the user’s activity and used these moments (breakpoints) as opportunities to deliver notifications. A simple example of how this proposed system works is depicted in Figure 2. During their study, they found that the change of activity from “STILL” to “TILTING” is the best time to deliver a notification as it probably means the mobile phone is being picked up from a desk or a user is standing up after sitting for a long while. Many such interesting findings were made between activity changes (see Figure 3) [6], which prove that breakpoint based notification delivery is a solid mechanism for an adaptive notification management system. Disturbing a person with a notification while they are engrossed in a specific task is a bad idea, whereas delivering a notification when a user is switching between tasks is shown to do a great job at keeping user engagement levels high.

4.3. Location-based delivery model

According to an article by Juniper Research, “56% of smartphone users would like to receive location-based offers on their phones when they are near a store” [7], and another article by Rishabh Software states that “42% of US consumers see potential benefits in sharing their locations with retailers via GPS” [8]. These statements indicate how important location-based notifications are for regular users.

Using exact user locations. In this research conducted by Taki et al. [9] to determine the best timings to deliver notifications, they observe that the user usually spends their nights at spot-1 and their days at spot-2. Thus, it was concluded that spot-1 was the user’s home and spot-2 was the user’s workplace or school. Further study revealed that the user’s notification responsiveness was greatly decreased at spot-2 (i.e. their workplace). These kinds of location-based notification delivery methods might prove to be useful in certain applications and must be considered during the development of an adaptive notification management system.

Grouping similar locations. The paper by Saikia and She [10] proposes a method of location-based mobile notification recommendation as they strongly believe that “sending the right notifications at the right location at the right time can boost user experience of mobile applications.” [10] The main idea of this paper is to establish a method wherein similar locations are grouped into localities, which can then be used to recommend notifications. The findings of this research paper suggest that this method could boost the mean notification opening rate by up to 62% while reducing the mean notification response time by 19%. [10] This seems to be another promising technique that should be thought about while modelling an intelligent system that adapts and recommends notifications to each of its users.

5. Conclusion

Notifications from an installed mobile application are meant to provide awareness of the content that the application provides. Due to the presence of multiple applications on the user’s mobile phone, there are too many notifications being pushed at the user. This often causes the user to get distracted and annoyed by irrelevant notifications. In addition to being a disturbance, notifications often bring with them a multitude of issues including high data bandwidth usage, unnecessarily high battery usage and contextually irrelevant information. Thus, a good adaptive notification management system is required to circumvent these issues.

Various components add up to make a good adaptive notification management system (Content, Presentation and Delivery). Each of these elements needs to be modelled separately using appropriate techniques to create meaningful notifications. The content of the notifications needs to be valuable first of all, then they need to be persuasive enough for the user to click on it and then finally based on the content, the notifications must be optimised thereby ensuring no irrelevant/ less-important content is shown to the user. An excellent presentation modelling technique was discussed in this post in section 3 which could be used to create adaptive media-rich notifications that could bring a lot of value to the end-user. Finally, three different features were shown using which an appropriate delivery model was built (Round based, breakpoint based, and location-based). If all the various techniques used in this post were considered and used to build a notification model, then it truly would be an intelligent notification management system that could benefit both the sender and the receiver.

6. References

[1] V. S. Y. Z. R. V. a. N. V. Md Yusuf Sarwar Uddin, “RichNote: Adaptive Selection and Delivery of Rich Media Notifications to Mobile Users,” IEEE 36th International Conference on Distributed Computing Systems, pp. 159-168, 2016.

[2] B. Y. a. O. C. Kieran Fraser, “Scrutable and Persuasive Push-Notifications,” Springer Nature Switzerland , pp. 67-73, 2019.

[3] R. B. Cialdini, Influence: The Psychology of Persuasion, 1987.

[4] M. C. J. S. a. S. P. Prasanta Saikia, “Effectiveness of Mobile Notification Delivery,” IEEE 18th International Conference on Mobile Data Management, pp. 21-29, 2017.

[5] S. Y. M.Grant, “Cvx: Matlab software for disciplined convex programming,” 2008. [Online]. Available: http://cvxr.com/cvx/.

[6] K. T. M. T. T. I. a. H. T. Tadashi Okoshi, “Attention and Engagement-Awareness in the Wild: A Large-Scale Study with Adaptive Notifications,” IEEE International Conference on Pervasive Computing and Communications (PerCom), 2017.

[7] J. Research, “Mobile coupons: Consumer engagement, loyalty and redemption strategies 2014-2019,” 2016. [Online]. Available: https://www.juniperresearch.com/press-release/coupons-pr1.

[8] R. Software, “Importance of A Location Based Application for iPhones in Retail Industry,” 28 October 2015. [Online]. Available: https://www.rishabhsoft.com/blog/geo-based-apps-for-iphone.

[9] Y. M. Y. A. K. Y. Kenta Taki, “Design and Implementation of Notification Information Survey System and Survey Results Toward Use-side Adaptive Notification Management,” PerPersuasion’19 - 1st International Workshop on Pervasive Persuasive System for Behavior Change, pp. 808-813, 2019.

[10] J. S. Prasanta Saikia, “Effective Mobile Notification Recommendation using Social Nature of Locations,” IEEE 15th Intl Conf on Dependable, Autonomic and Secure Computing, 15th Intl Conf on Pervasive Intelligence and Computing, 3rd Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology Congress, pp. 1265-1270, 2017.

The contents of this post were originally written for a review paper that I had to turn in as part of my assessment for the Adaptive Applications module at Trinity College Dublin.