fbpx Skip to main content
Spread the love


This domain, which encompasses consulting firms, loyalty program enablers, software providers, and performance marketing platforms, plays a pivotal role in the strategies for customer acquisition and retention. To truly excel, these companies must leverage state-of-the-art machine learning techniques to maximize the value they deliver to the brands they partner with.

The adoption of recommendation systems within the loyalty ecosystem can be categorized into three distinct segments. Firstly, there are the non-adopters, who have yet to integrate any form of recommendation system into their operations. Then, there are the followers, who typically rely on mainstream, often ready-to-use SaaS solutions. Lastly, the early adopters stand out, as they endeavor to develop cutting-edge models, tailoring the latest technological advancements to their unique data and use-cases.

In loyalty, early adopters stand out, as they endeavor to develop cutting-edge models, tailoring the latest technological advancements to their unique data and use-cases

This article will focus on a case study of an early-adopter performance marketing platform that transitioned from traditional machine learning models to developing interactive machine learning systems, specifically tailored to address their unique recommendation challenges. While this particular use-case may not be universally applicable, it serves as a compelling illustration of the necessity to customize technology, especially machine learning approaches, to each specific scenario.

Crafted to cater to a broad audience, including those without a technical background, this article aims to demystify the complexities of these models. By maintaining a high-level overview of the technologies involved, we strive to make the content accessible and engaging for everyone, ensuring a comprehensive understanding of the transformative impact of interactive machine learning in the loyalty industry.

The main challenges we solve

Our platform is designed to allow users to earn rewards through a variety of activities. These activities range from ‘Play to Earn’ games, participation in online testing and paid surveys, to earning cashback on purchases. The primary objective of our recommendation system is twofold: firstly, to accurately recommend the most suitable offers from these four categories to each user, and secondly, to enhance the engagement and conversion rates by selecting the most effective visual for each product.

  • Objective #1 : recommend rewards
  • Objective #2 : recommend visuals 

The effectiveness of our recommendation system is crucial in achieving long-term revenue generation. This involves not only identifying the right category of offer for each user but also entails selecting the most appealing visual from a set of available options. Our initial experiments, like naive A/B testing, have demonstrated significant differences in Click-Through Rates (CTRs) between different visuals up to 10%. This insight underscores the need for a recommendation system that can discern and suggest the most impactful visual for each product, thereby optimizing user engagement and response.

Item visual impact: A/B testing has demonstrated up to 10% differences in Click-Through Rates (CTRs) between different visuals

An additional, yet critical, constraint we face is the need for rapid adaptability. The recommendation system must be capable of quickly learning and adapting to any changes in incentive rates. For instance, if there’s a change in the cashback percentage on an offer, the system should be able to reflect this change in the ranking of its recommendations within a short timeframe, preferably in hours and at most within a few days. This agility is essential to maintain the relevance and effectiveness of our recommendations in a dynamic market environment where timely response to change is a key driver of success.

By tackling these challenges, our recommendation system aims not just to meet the immediate needs of our users, but to do so in a way that ensures sustained engagement and revenue generation, marking a significant advancement in the intersection of user experience and business strategy.

Limitations of the traditional recommendation systems

What do we call “traditional recommenders” ?

Traditional recommender systems, often central in the loyalty industry, primarily rely on scoring methods to suggest items to users. Imagine wanting to recommend a product to a user; these systems score each item based on the likelihood of that specific user appreciating it. This process hinges on analyzing the interaction data of various users with different items, typically employing deep learning algorithms. The most prevalent of these methods is known as ‘collaborative filtering,’ a technique extensively utilized in e-commerce sectors.

The essence of collaborative filtering lies in its dependence on historical interaction data. To accurately score items for a user, these systems require a substantial amount of past user-item interaction data, usually spanning several months. While it’s possible to periodically retrain the model to adapt to shifting user behaviors, setting too short a historical window for training compromises the model’s accuracy. In essence, the trade-off is between freshness of recommendations and their relevance or accuracy.

Once the model is trained, it can deliver personalized recommendations each time users visit the associated webpage. This mechanism, while effective to a degree, has its limitations, particularly in its reliance on historical data and less flexibility in real-time adaptability.

For those interested in a more in-depth exploration of these methods, our previously published article “Recommendation Systems for Reward and Loyalty Platforms: 2024 Guide” offers a comprehensive analysis.

The two reasons why traditional recommenders will not work

1. Traditional recommenders are static

Traditional recommendation systems are fundamentally limited due to their static nature. As illustrated above, these models rely on a predetermined time window of historical data to “learn” user preferences among various items. The essence of their operation is to use as extensive a historical window as possible to accurately gauge these preferences.

However, this approach encounters significant limitations when user behavior changes. Any shift in preferences is only recognized and reflected in the recommendations after a duration equivalent to the predefined time window. This lag in response time can be critical in dynamic markets or where the reward is changed for a specific product.

Moreover, traditional recommendation algorithms prioritize the accuracy of prediction over the adaptability of these predictions. While accuracy is undoubtedly important, this prioritization often comes at the expense of swiftly adjusting to new user behaviors or reward changes. In scenarios where adaptability is key, these systems fall short.

Consider, for instance, a situation where the reward value for an item is altered. This change is intended to incentivize users towards a specific item, potentially leading to an increased click-through rate (CTR) and average revenue for that item. The logical course of action would be to promote this item more prominently on a webpage. However, due to their inherent design, traditional systems like collaborative filtering will only adjust the item’s ranking after the model has “learned” of the changes affecting user preferences. This delayed response can mean missed opportunities in rapidly changing market conditions.

This graph would effectively demonstrate how the ranking of a product changes over time in response to a modification in the reward value, highlighting the lag in traditional recommendation systems’ ability to adapt to such changes.

2. Traditional recommenders can’t recommend visuals

A significant limitation emerges when we consider the integration of both products and their associated visuals. These systems are fundamentally engineered to process numerical interaction data such as clicks, views, and revenue metrics. However, the challenge arises when we wish to recommend not just a product, but also a specific visual representation of that product, from a limited selection of three or four options. To accommodate this, each visual would need to be treated as a distinct product, effectively multiplying the product catalog by three or four times.

A significant limitation emerges when we consider the integration of both products and their associated visuals

This approach inadvertently fragments the web traffic across these multiple visuals, diluting the interaction data for each one. As a result, the amount of data available to train the recommendation algorithm for each individual product visual is reduced to a mere third or quarter of what it would otherwise be. In an environment where the volume of data is directly related to the accuracy of the recommendations, this division poses a considerable setback. The accuracy of the system is inevitably compromised, as it now operates on a substantially smaller dataset for each product-visual combination.

Moreover, this issue is compounded by the inherent static nature of traditional recommendation systems, as highlighted in the previous section. While these systems excel in processing and responding to numerical data, they lack the dynamic capability to effectively adapt to the nuanced and varied visual preferences of users. This shortcoming, when combined with the aforementioned dilution of data, underscores the need for a more sophisticated approach that can seamlessly integrate the complexities of recommending both products and their corresponding visuals.

A new approach with interactive AI recommendation

Reframing the Challenge: Optimization through Interactive AI

In the preceding sections, we discussed the limitations inherent in traditional recommendation systems, particularly their inadequacy in accurately recommending the most suitable items—be it tasks, cashback offers, or play-to-earn games—while adhering to our constraints of accuracy, adaptability, and multi-modality (which includes visual item recommendation). To effectively address these challenges, it’s imperative to shift our perspective: we should not view this as a mere recommendation problem, but rather as an optimization challenge.

It is imperative to shift our perspective: we should not view this as a mere recommendation problem, but rather as a revenue optimization problem

This paradigm shift leads us to focus on optimizing the revenue of the platform. The critical question we face is determining what items to display on the screen, in each position, for every individual user, with the goal of maximizing the revenue generated by these items. Traditional recommendation approaches fall short in this regard, as they often rely on static algorithms that fail to adapt to the dynamic preferences of users and the ever-changing inventory of items.

Enter Interactive AI, or more specifically, Reinforcement Learning techniques, which offer a robust solution to this conundrum. Unlike conventional systems, Interactive AI thrives on learning from user interactions in real-time, constantly adjusting its recommendations based on evolving user behavior and preferences. This approach allows for a more nuanced understanding of what items will most likely drive revenue when displayed to particular users at specific times. It’s a sophisticated balancing act—Interactive AI doesn’t just recommend items; it predicts and adapts to what users are likely to find engaging and valuable, thereby optimizing both user satisfaction and platform revenue.

In summary, by redefining our challenge not as one of mere recommendation but as one of optimization, and by harnessing the power of Interactive AI, we position ourselves to significantly enhance the effectiveness and efficiency of our recommendation system, leading to increased revenue and improved user satisfaction. This innovative approach is a game-changer, marking a significant leap from traditional methodologies.

Interactive AI Explained

Interactive AI, particularly when implemented through contextual bandit algorithms, marks a significant leap in how items are suggested on web pages. This advanced technology can be understood as a more sophisticated form of A/B testing. 

Unlike traditional methods that often rely on historical data spanning extensive time windows, interactive AI algorithms learn in real-time, adapting their suggestions based on immediate user interactions. This means they are incredibly efficient in understanding and reacting to user preferences, often requiring much less data volume for training compared to static recommendation systems. This efficiency is crucial in the fast-paced online environment, where user trends and interests can shift rapidly. The iterative nature of these algorithms allows for a dynamic distribution of web traffic across different visuals or content, ensuring that the most relevant and engaging material is presented to each user.

The iterative nature of these algorithms allows for a dynamic distribution of web traffic across different visuals or content, ensuring that the most relevant and engaging material is presented to each user

Interactive AI, particularly through contextual bandit algorithms, not only offers real-time adaptation to user preferences but also optimizes a crucial balance between exploration and exploitation. This balance is key in these algorithms. While exploitation leverages existing knowledge to maximize immediate performance, exploration involves trying new options to discover potentially better strategies that are not yet known. This tradeoff is essential for ensuring that while the system efficiently utilizes known data to provide relevant suggestions, it also continuously seeks new data or strategies to enhance its recommendation accuracy over time. This dynamic approach allows interactive AI to remain adaptable and effective, even in the ever-changing landscape of user trends and interests.

This not only enhances the user experience but also drives greater loyalty and engagement, as recommendations are continuously refined and personalized. The use of interactive AI in recommendations represents a paradigm shift in the loyalty industry, offering an agile, data-driven approach that keeps pace with the evolving digital landscape.

Model Comparison

Traditional recommendation systems Interactive recommendation systems
Machine Learning Model Collaborative filtering Contextual bandit algorithms
Fast learning No Yes
Optimize data exploration-exploitation tradeoff No Yes
Recommend visuals No Yes
Learns long-term behavior Yes No


Traditional recommender systems can be developed with built-in cloud services such as Amazon Personalize, which offers a straightforward method for integrating machine learning-based personalization into various customer interaction points. This service simplifies the creation of user-specific content recommendations, streamlining the user experience without requiring in-depth machine learning knowledge.

For dynamic and interactive AI, Amazon SageMaker RL serves as an excellent starting point. SageMaker RL takes advantage of continual learning with contextual bandits, optimizing the balance between exploiting known user preferences and exploring potential new interests. It provides a sophisticated infrastructure for not just making recommendations but also for evolving the recommendation model in real-time based on ongoing user interaction data. This continual adaptation fosters a personalized experience that keeps pace with changes in user behavior, which is particularly beneficial in the fast-moving loyalty industry.


In summary, the adoption of Interactive AI in the loyalty industry marks a significant advancement beyond traditional recommendation systems. These new AI-driven models offer real-time adaptability and a deep understanding of user preferences, including the effective integration of product visuals. This transition not only overcomes the limitations of static algorithms but also significantly enhances user engagement and revenue potential.

The case study of an early-adopter performance marketing platform illustrates the transformative impact of Interactive AI. Businesses in the loyalty sector are now recognizing that adopting Interactive AI is crucial for staying competitive in a rapidly evolving digital landscape.

Looking forward, the integration of Interactive AI in loyalty programs is not just an improvement—it’s a game-changer that promises a more personalized, efficient, and engaging user experience. This shift heralds a new era in the loyalty industry, where technology and user experience converge to drive innovation and growth.

The following research papers are interesting to read if you want to push further this topic on a scientific level:

Neural collaborative filtering

He, X., Liao, L., Zhang, H., Nie, L., Hu, X., & Chua, T.-S. (2017). Neural collaborative filtering. arXiv:1708.05031v2 [cs.IR]. https://doi.org/10.48550/arXiv.1708.05031

A contextual bandit bake-off

Bietti, A., Agarwal, A., & Langford, J. (2018). A contextual bandit bake-off. arXiv:1802.04064v5 [stat.ML]. https://doi.org/10.48550/arXiv.1802.04064

Mehdi B.

Mehdi B.

Mehdi is the founder of reco-genius.com, an AI agency specializing in performance solutions for reward platforms. He brings over a decade of private equity experience and a flair for innovative tech solutions. Mehdi is a software engineer, a graduate of École Polytechnique (aka "The French MIT"). He also holds a Professional Certificate in AI from Stanford and the AWS Machine Learning Certification.