Aliaksandr Zhastkou is the Product Owner and Head of Web Products at Yandex Delivery, where he has played a key role in shaping the service over the past three years. Starting his journey in marketing and transitioning into product leadership, Aliaksandr has been instrumental in driving the platform’s growth, helping Yandex Delivery become the largest player in express delivery and scale over 100 times.
—
What if your app could understand you, not just your clicks and taps, but your underlying needs? How about the way you communicated within the app? Machine Learning has the potential to accomplish this. Through learning with user behavior, it can craft ever engaging satisfaction that is tailored to their needs.
However, this opens room for pertinent questions. Are algorithms able to truly comprehend human emotions and nuances, or are they nothing more than pattern machines? And as we allow ML systems to power our technology and interpret the behavior of our users, how do we reduce the chance of misinterpretation, bias or simplification of automation?
Today, in this competitive digital environment, in-app communication has grown from an additional feature to one of the primary aspects of ensuring satisfaction and retention. Broad, universal fixed messages and templates used to send have lost their relevance among modern users who expect touch points created for them directly.
And now, this is the part where it all changes, the time when we change from being active listeners and turn into active speakers who talk niche ML communication, going ML communication into the future: hyper-personalised automated communication. However, with this need comes the challenge of balance between use and human-like interaction, personalisation and privacy.
Like many modern developments in technology, Yandex Go employed the power of ML and our team was among the first in Eastern Europe to adopt machine learning within large-scale in-app communication systems.
With this article, I want to open up practical insights along with relevant use-cases, important application KPIs, and machine learning practices that, I believe, can serve in building modern communications. Together, we can analyse the potential uses of machine learning in in-app communications along with the challenges faced in the domain and automated engagement marketing frameworks and learn how to create interfaces focused on the needs of customers.
Key ML Techniques for In-App Communication
1. Natural Language Processing (NLP) for Chatbots and Voice Assistants
NLP enables applications to understand, process, and generate human language. Its use in chatbots and voice assistants allows for highly contextual, real-time conversations with users. Advanced models like OpenAI’s ChatGPT have set benchmarks for fluency and comprehension.
NLP-driven chatbots excel at handling repetitive tasks like FAQs, but their true potential lies in nuanced, multi-turn conversations. For instance, “Google Assistant” leveraging NLP and contextual learning, can schedule meetings or make reservations with minimal user input. This capability reduces friction in task management and boosts overall user productivity.
2. Recommendation Systems for Personalised Messaging
In-app personalised messages are shaped by machine learning bolstered recommendation systems. They use a user’s browsing history, behavior, interactions, and preferences in real time to direct user activity with messages that are relevant and instinctively engaging, enhancing business returns and deepening trust.
Users receive tailored product recommendations, branded promotions, offers, and content driving advertising engagement, not as generic updates. This is the case with Amazon; real-time user-synchronised collaborative filtering neural network sequence model learning is further complemented by the aggregated behavior of millions turning every interaction into an insight.
Example Impact: Amazon’s recommendation engine contributes to 35% of its total sales, showcasing the revenue-driving power of personalised communication. That figure highlights a primary focus and the deep lever of trust that drives the sales system.
Without personalisation, revenue simply would not be generated. As the expectation for cross platform user experience grows, a strategy to leverage identified user behavior becomes crucial.
With personalisation becoming a common expectation on every digital interface, thorough recommendation engines are no longer optional but a tactical requirement.
3. Sentiment Analysis for Understanding User Emotions
As in-app communication advances, the need for interpreting user interactions with apps has sharpened the use of sentiment analysis as feedback mechanisms. Emotions driven by user interactions can now be detected thanks to machine learning algorithms, making communication frameworks a lot more sensitive and effective.
Sentiments based in text such as inputs derived from conversations or texts can now be scored using more complex tools like computer programs. Primarily, texts are categorised into three broad categories: positive, negative or neutral using Microsoft’s Azure Text Analytics API. This AI powered tool serves as a real-time, non-verbal telemetry, businesses use to monitor interaction sentiments and emotional engagement, prioritising important interactions to interact with them in an adequate manner.
Example Impact: Traveloka’s Twitter Sentiment Analysis: In a 2022 study by researchers from Bina Nusantara University, an analysis of 1,200 tweets related to Traveloka, a prominent travel and lifestyle application in Southeast Asia – was performed using three machine learning models: Support Vector Machine (SVM), Logistic Regression, and Naïve Bayes.
Key insights:
- Dataset Composition: 610 positive, 590 negative tweets
- Tools & Methods: With Python Scikit-learn, sentiment classification and vectorisation was performed using TF-IDF
Model Performance:
- SVM proved most accurate with 84.58% accuracy. Naïve Bayes and Logistic Regression followed at 82.91% and 82.50% respectively
- SVM also provided strong F1-score results of 0.88 and 0.82 for positive and negative sentiments respectively, showing strong capability in distinguishing feedback sentiments
Impacts:
-
- Customer Insight: Users were shown to value Traveloka for the discounts offered and ease of booking while expressing frustration with Traveloka Eats
- Strategic Value: Analysis like this allows Traveloka to adjust marketing strategy, improve services, and tackle proactive dissatisfaction before churn
Chatbots and Virtual Assistants for Customer Support
ML-powered chatbots offer consistent, 24/7 support, automating up to 80% of common queries and continuously improving through user interactions. Virtual assistants further enhance functionality by integrating with broader ecosystems.
Case Study: Duolingo’s Adaptive Chatbot
- Implementation: Duolingo’s chatbot tailors its conversational style to the user’s language proficiency. For beginners, it provides simplified responses, while for advanced users, it introduces complex vocabulary
- Analysis: This personalised approach has led to a 15% reduction in new user drop-off rates, indicating the effectiveness of context-aware communication in retaining users
- Metrics: Post-implementation, Duolingo observed a 12% increase in daily active users (DAU), reflecting enhanced user engagement
Push Notifications Tailored to User Behavior
Notifications, when tailored to specific users, can enhance engagement without being disruptive. The advent of machine learning algorithms makes it possible to evaluate user behaviors at a deeper level and more granularly than before, i.e. for specific users app usage, in-app actions, content preferences and even the timing of usage to send messages which seem timely and relevant.
Rather than relying on generic broadcasts, advanced push systems use behavioral signals to determine what to send and when. ML algorithms can detect contextual cues as well, such as when a user generally tunes into music, when an app is in use, or when a person ‘drops off’ from usage. Contextual nudges can then be re-engagement intelligently.
Push Strategy – Case Study of Spotify
Not only does Spotify leverage machine learning for recommending content, they also use it to personalise notifications about suggestions that are about to be released. Such strategically personalised notifications include highlighted playlists, new songs which are likely to be appreciated, and user-inspired listening reminders. These strategies are based on real-time user behavior like track skipping, replaying, listening times, and session frequency.
Example Impact: Spotify uses “Discover Weekly” as a showcase feature. As one of its crown jewels, it uses ML to create personalised playlists on a weekly basis. As users learn to expect the notifications, the engagement rates become higher as the engagement prompts, in the form of reminders, are customised and automated through user behavior and machine learning, maximally encouraging users to check their discovery list.
Through curated content nudging that has been powered by behavioral economics enhances daily active usage by reminding users gently of content or features they might have forgotten.
As of June 2020, through the usage of “Discover Weekly”, Spotify users streamed over 2.3 billion hours of music from July 2015. Users who actively engage with Discover Weekly tend to stream over twice as much compared to non-engaged users. This showcases the impact well-timed personalised discovery provides using push notifications.
Case Study: Spotify’s Personalised Recommendations
-
- Implementation: Spotify utilises ML to analyse user preferences, such as listening habits, skips, and replays, to suggest playlists and songs. Features like “Discover Weekly” provide users with personalised music recommendations
- Analysis: The precision of these recommendations introduces users to new music aligned with their tastes, deepening engagement with the platform.
- Metrics: Users engaging with “Discover Weekly” have streamed over 2.3 billion hours between July 2015 and June 2020, and they stream more than twice as long as non-Discover Weekly users, demonstrating increased engagement
Adaptive In-App Tutorials Based on User Interactions
Applications leverage ML to monitor user behavior, providing real-time tutorials when necessary. This approach eliminates unnecessary information and ensures relevant assistance, enhancing user satisfaction.
Case Study: Canva’s Smart Onboarding System
- Implementation: Canva’s system analyses user actions, such as missed features or prolonged inactivity, to trigger contextual tutorials. For example, if a user struggles with layering, the app suggests specific design tips
- Analysis: By addressing user-specific challenges, Canva enhances user satisfaction and accelerates the onboarding process
- Metrics: This targeted assistance has led to an 18% reduction in user drop-off during onboarding and increased adoption rates of app features
Personalised Content Delivery
ML enables applications to deliver personalised content, enhancing user engagement and satisfaction by catering to individual preferences.
Case Study: Hyper-personalisation of advertising via Yandex Crypta
In the current environment, businesses try to outperform competitors by delivering impactful hyper personalised advertising. Yandex Crypta uses the capabilities of ML which lets advertisers serve targeted advertisements with unprecedented accuracy.
Each user is represented by a set of anonymous identifiers, deepening privacy. With these representations, Crypta can reliably estimate ad relevance using various behaviors, and individual preferences, like search queries, web browsing, engagement frequency, and even device usage patterns. Without collecting personally identifying information (PII), users are mapped across devices which allows companies to extract insights for relevant advertising while respecting privacy.
Impact:
- User Engagement Desired: Advanced targeting accuracy using Yandex Crypta has improved advertising capture via in-app ads during engagements on services like Yandex Delivery by 10%
- Balanced Accuracy and Privacy: By utilising machine learning methods qualified to segment users without retaining PII, Crypta manages effective advertising and data privacy, pleasing both businesses and users
The advanced targeting options enable Yandex’s partners to optimise their ad placement and delivery to their preferred audience. Consequently, Crypta improves the effectiveness of digital marketing campaigns and enhances users’ experiences by showing them relevant ads, instead of ads for products and services of no interest to them.
The Influence of ML on Communication
ML assists in automating a number of mundane tasks freeing up teams to allocate the time spent on such tasks on strategising and innovating. Which in turn results in improved user satisfaction as teams are able to respond to user problems quicker using predictive ML insights. In addition to providing a comprehensive insight on data, machine learning communication allows businesses to gain a competitive edge in understanding the behavior of customers.
According to McKinsey’s research, some companies that have adopted modern analytics and intelligent process automation, which encompasses machine learning techniques are capable of progressing their customer journeys which could lead to positive enhancements in NPS. Such results prove how well adapted to customer relationships machine learning is and its importance to the overall competitiveness of businesses.
Challenges and Considerations
While ML offers great advantages in customer service – enhancing efficiency, reducing response times, enabling 24/7 support present several challenges. In all cases, organisations should address these obstacles in order to maintain ethical and effective deployment.
One of the primary issues is compliance with regulations. Data protection laws, such as the General Data Protection Regulation (GDPR) in the European Union and California Consumer Privacy Act (CCPA) in the United States, set strict guidelines for the collection, storage, and processing of personal data. Legally adopting these measures often involves complex business processes and incurring significant per-user operational overhead costs. Companies should deploy advanced user data privacy measures such as encryption, pseudonymisation, and identity anonymisation to limit the possibilities of user-identifiable data exposure.
Over-automation is another challenge worth considering. Fully automated systems, particularly those devoid of any context or empathy, can render uninspiring experiences for users and can be maddening when context is added. To address this issue, many companies are adopting hybrid systems that combine automated AI functions with the ability to easily transfer more complicated, sensitive, or emotional issues to human beings to resolve them.
ML bias presents an equally significant problem. AI systems without sufficient representative training data are likely to perpetuate pervasive stereotypes or discriminate against certain users. Organisations need to implement rigorous data contextualisation alongside a comprehensive auditing process that goes beyond basic intent inference. Failing that ,sustained monitoring of model performance is also crucial. Employing fairness-driven machine-learning techniques adds ethical strength to the system while increasing supervised involvement mitigates the risks associated with unbounded reliance on automation.
Understanding the potential of AI tools in customer service requires companies to adhere to three fundamental rules: removing bias in algorithms, achieving the right equilibrium between automation and human touch, and protecting customer data with robust privacy safeguards. These practices cultivate consumer trust while maintaining socio technical standards in AI-enabled customer service.
What’s Next?
ML is no longer in the research phase. It has become a fundamental component for developing scalable and adaptive systems that are human-centric. Leading product teams ML for hyper-personalised user flows as for user smart re-engagement messages for maintaining and building enduring relationships with users
At companies such as Duolingo, Spotify, Canva and Yandex, the impacts of ML integration with communication modules are evident from increased induced conversions, enhanced retention, and improved user experience: everything works synergistically.
The future of in-app ML communication with the emergence of generative AI, the next evolutionary leap of ML seems the most apparent: Cross-platform, cross-device communications will be intelligent and predictive. Today’s product leaders need to seize the moment and shift focus towards the evolution of systems designed to learn and adapt instead of messaging rules and user stagnation.
In the context of disruptive technologies, communication morphs from a singular functionality of a product into a holistic strategy for engaging users.
