Modern Recommender Systems - Part 3: Objectives

Pavel Kordik

Sep 03, 2025

Learning objectives of recommender systems and personalized search.

In the first two parts of this book, we laid the groundwork: understanding what modern recommender systems are and how data fuels them. But knowing the machinery and the raw material is not enough—you also need to know the destination. Every recommendation engine is, at its core, an optimizer: it learns patterns from interactions and then pushes its predictions toward a goal. Early systems optimized implicitly for popularity or similarity, but today’s platforms face far more complex and sometimes conflicting objectives: keeping users engaged without overwhelming them, balancing diversity with relevance, and driving business value without eroding trust. In this third part, we shine a spotlight on these objectives—the “north stars” that guide recommender systems and personalized search. By clarifying what is being optimized, we uncover the hidden logic behind why different platforms make the recommendations they do, and set the stage for translating these goals into concrete learning tasks in the next chapter.

Examples of Learning Objectives

To give you an impression how broad objectives in recommender systems and personalized search can be, we will start with examples in various domains.

Content streaming services (e.g., music, video, podcasts) prioritize objectives centered around user engagement and retention:

• Maximizing User Engagement: Keeping users actively consuming content (e.g., total view time, session duration, content completion).
• Reducing Churn Rate: Minimizing users canceling subscriptions or ceasing to use the service.
• Accelerating Content Discovery: Helping users easily find new, enjoyable content, showcasing catalog breadth.
• Balancing Mainstream vs. Niche Content Exposure: Promoting diverse content to cater to varied tastes and support a healthy content ecosystem.
• User Satisfaction and Perceived Value: Ensuring users feel recommendations are enjoyable and justify subscription costs or time spent.
• Supporting Creator Ecosystem: Ensuring fair exposure and monetization for content creators.

In general, subscription-based content streaming services focus on optimizing user satisfaction with the service. For free users, the main objective would be to convert them into subscribers (e.g., by recommending highly relevant content beyond the paywall). For ad-powered content streaming services, watch time maximization might be a good strategy to increase revenue from displaying ads. Controversies are discussed in Part 1, Introduction.

E-commerce platforms typically prioritize objectives focused on driving sales and enhancing customer value:

• Increasing Conversion Rates: Maximizing the percentage of users who make a purchase after viewing a recommendation or visiting the site.
• Increasing Average Order Value (AOV): Encouraging users to purchase more items or higher-value items per transaction.
• Reducing Cart Abandonment: Minimizing instances where users add items to their cart but leave without completing the purchase.
• Maximizing Customer Lifetime Value (CLV): Fostering long-term customer loyalty and repeat purchases through sustained relevance.
• Improving Product Discovery Across the Catalog: Helping users find relevant products beyond popular items or their immediate search.
• Optimizing Inventory Turnover: Promoting overstocked items or those nearing end-of-season, balancing business needs with user experience.
• Waking-up Inactive Customers: Offering targeted discounts or suggesting highly relevant products with limited availability.

In e-commerce, it is more about making the customer buy products rather than any other objectives. However, some e-commerce platforms and marketplaces are focusing on generating traffic for other e-shops rather than selling directly. Their focus is therefore shifted towards producing outclicks, especially when purchases associated with outclicks are not reported and rewarded by partner sites.

In other domains, objectives might be even more complex. Imagine job boards or dating sites that need to optimize for satisfaction of multiple parties under constraints.

Also, there are general level objectives that apply to most scenarios where users interact with personalized recommendations or search. User Satisfaction and Task Completion optimize for successful user sessions that result in items found in a reasonable time (Time-to-Result Optimization). One might also optimize for Relevance, Quality, Diversity, and Freshness of items. In all scenarios, we strive for Abandonment Reduction (e.g., search query, cart, watch next recommendation), which might lead to unsuccessful user sessions. Note that online platforms observe just partial user feedback signals, so all these objectives are typically optimized in a noisy environment.

One might ask how particular objectives are defined for a specific online platform and individual use cases. Typically, this is done through careful analysis of user needs, business requirements, and strategic objectives. These goals typically emerge from stakeholder discussions, user research, and business strategy sessions.

The objectives of modern recommender and search systems involve multiple stakeholders often with conflicting interests (e.g. users, content creators, editors, business). Effective optimization seeks to balance and align these. We can broadly categorize these critical objectives as follows.

Taxonomy of Learning Objectives

To make sense of the many and often competing goals in recommender systems, it helps to group them into broader categories. This taxonomy of learning objectives highlights four perspectives—user, content, business, and product—that together capture the full landscape of what modern systems are designed to optimize.

User Objectives

Focused primarily on satisfying and engaging end-users, these metrics capture how effectively the system delivers personalized experiences.

• Engagement: Measures the extent of active user interactions (clicks, views, session length, and return visits) indicating user interest and commitment to the content.
• Relevance: Ensures recommended items align closely with user preferences, past behavior, and explicit user feedback.
• Novelty and Serendipity: Goes beyond traditional relevance to introduce users to fresh, unexpected yet satisfying recommendations, keeping user experiences interesting and avoiding monotonous or predictable content.
• Trust and Transparency: Users prefer transparent, explainable recommendations that build trust and confidence in the system's decisions, especially important in sensitive or high-stakes scenarios.

Content Objectives

These objectives ensure the breadth, richness, and balanced representation of available content.

• Diversity: Guarantees variety within recommendations, preventing repetition, echo chambers, or overly similar content.
• Coverage: Refers to the proportion of the content catalog effectively recommended and utilized, ensuring both niche and popular items have a fair opportunity for exposure.
• Freshness: Prioritizes new or timely content, critical for domains where recency significantly impacts user satisfaction (e.g., news, trends, social media).
• Locality: Ensures content relevance based on geographic, cultural, or regional context, where content that's highly relevant for users in one area may be irrelevant or inappropriate for others (e.g., local news, regional events, location-specific services, cultural content).

Business Objectives

Reflecting economic and strategic goals of an online platform, these metrics typically justify the investment in a recommender system or personalized search solution.

• Profitability: Recommendations should directly or indirectly enhance revenue by increasing sales, upselling, cross-selling, or improving monetization opportunities.
• Cost Efficiency: Systems should optimize resource utilization, reducing computational costs and data processing overhead.
• User Retention and Loyalty: Strong recommendation systems support long-term customer relationships, reducing churn and boosting lifetime customer value.

Product Objectives

These objectives ensure that the recommender system contributes positively to the overall product experience, reputation, and ethical considerations.

• Speed and Responsiveness: Recommendations must be fast and timely, ensuring that latency does not degrade user experience, especially critical in real-time scenarios.
• Brand Consistency: Recommendations must align with the overall brand identity, supporting brand image and maintaining consistent messaging and quality expectations.
• Fairness and Ethics: Recommenders should proactively avoid biases, stereotypes, or unfair treatment of user groups or content providers. Fairness also encompasses equitable representation and opportunities for less prominent content providers.
• Compliance: Systems must responsibly handle user data and adhere to legal/ethical frameworks (e.g., GDPR, AI Acts), ensuring privacy and lawful processing.
• Security: Systems must be resilient to malicious activities (e.g., attacks, hacking), safeguarding integrity, data, and reliability.

Balancing Objectives in Real-World Recommender Systems

The key to operationalizing these diverse objectives is aligning them with measurable metrics that can, in turn, be optimized through specific machine learning tasks (see next chapter). For example, user engagement might be measured via session length and click-through rates, while content diversity could be quantified using intra-list similarity scores. The star plot above illustrates how different objective categories like user objectives (engagement, relevance, novelty), content objectives (diversity, coverage, freshness), business objectives (profitability, cost efficiency, retention), and product/ethical objectives (speed, brand consistency, fairness) form a multi-dimensional optimization space.

Modern recommender systems rarely optimize a single goal—they balance several at once, often through multi-stage pipelines. Netflix, for example, blends candidate generation for relevance with re-ranking for freshness and diversity, applies business filters, and enforces fairness constraints, all within milliseconds. Spotify faces similar challenges: its personalized playlists mix familiar tracks with exploration of new or lesser-known artists to keep listeners engaged without creating monotony.

The danger of over-optimizing one objective is very real. A Spotify field experiment showed that personalized podcast recommendations increased streams by 29% but reduced listening diversity by more than 11%, meaning users became more engaged but in narrower, less healthy patterns. This highlights how optimizing purely for engagement can harm long-term satisfaction. Similarly, academic and industry discussions frequently warn against filter bubbles and popularity bias, where too much focus on relevance or clicks leads to stale, repetitive experiences and reduces fairness in content exposure.

I also witnessed this problem firsthand with one customer. They boosted recommendations that generated higher out-click revenue, which initially looked like a win for the business. But over time, user engagement declined because the system was pushing too aggressively toward monetizable items at the expense of user satisfaction. In the end, they had to scale back the boost significantly to protect the overall experience.

The lesson across these cases is that objectives are interconnected: relevance, diversity, engagement, revenue, and fairness must be optimized in concert.