Multi-objective recommendation

On this page 7

Multi-objective recommendation is the practice of training and ranking recommendations so that they serve several goals at the same time, such as click probability, purchase, watch time, long-term satisfaction, diversity and creator fairness, instead of optimizing a single metric. Recommender systems in production almost always face this problem, because what a platform values cannot be captured by one predicted number.

The core idea is that goals conflict: the items most likely to be clicked are not always the ones users enjoy or return for. Multi-objective recommendation makes these trade-offs explicit and tunable, typically by predicting each objective separately and then combining the predictions with a scoring function, constraints or a learned policy.

Why recommendation needs several objectives

A single objective tends to be gamed by the ranking itself. Optimizing only click-through rate rewards sensational thumbnails and short-lived curiosity; optimizing only watch time can favor long but unsatisfying content; optimizing only conversion can narrow the catalog to bestsellers and amplify popularity bias and the filter bubble.

Real platforms also answer to several stakeholders: users want relevance and novelty, creators and sellers want exposure, and the business wants revenue and retention. The table below lists objectives that commonly appear together and the signal each relies on.

ObjectiveTypical signalMain risk if optimized alone
EngagementClick, tap or playClickbait and shallow interest
Consumption depthWatch time or dwell timeBias toward long items
ConversionPurchase, subscription or add to cartNarrow, popularity-driven catalog
SatisfactionLikes, ratings, survey answersSparse and delayed labels
RetentionReturn visits over days or weeksVery slow, noisy feedback
Diversity and coverageCategory spread, catalog coverageLower short-term relevance
Provider fairnessExposure share per creator groupConflicts with pure relevance
Objectives commonly combined in multi-objective recommendation

How multiple objectives are modeled

Most industrial systems model objectives with multi-task learning: one network with shared layers and several output heads, each predicting a different engagement signal. Sharing representations helps sparse objectives borrow strength from dense ones, for example purchase prediction learning from abundant click data.

Simple shared-bottom networks can suffer from negative transfer, where one task degrades another. Architectures such as Multi-gate Mixture-of-Experts (MMoE) and Progressive Layered Extraction (PLE) give each task its own gating over shared expert subnetworks, which reduces interference. Cascaded objectives are often modeled jointly as well: ESMM estimates conversion through the chain impression, click, conversion to correct sample selection bias in conversion data.

ApproachHow it worksWhen it fits
Separate modelsOne model per objectiveFew objectives, independent teams
Shared-bottom multi-taskCommon layers, several headsRelated tasks with similar data
MMoE or PLETask-specific gates over shared expertsConflicting tasks, large traffic
Cascaded estimationModels chained probabilities such as click then buyFunnel objectives with selection bias
Modeling approaches for several recommendation objectives

Combining objectives into a ranking score

After prediction, the ranking stage must turn several scores into one ordering. The most common method is weighted scalarization, a weighted sum or product of predicted probabilities, sometimes with exponents per objective; the weights encode business priorities and are tuned through online A/B tests.

Scalarization is easy to operate but finds only part of the Pareto front, the set of rankings where no objective can improve without another getting worse. Alternatives include constrained optimization, which maximizes one goal subject to floors on others, and Pareto-based search for model or weight settings. The comparison below summarizes the main options.

MethodStrengthLimitation
Weighted sum or productSimple, fast, easy to explainWeights need constant retuning
Constrained rankingGuarantees minimum levels for chosen goalsHarder to solve per request
Pareto optimizationExposes the full trade-off curveStill needs a final choice of point
Re-ranking with diversityAdds variety after scoring, for example MMRWorks only on the candidate list
Reinforcement learningOptimizes long-term reward directlyHard to train and evaluate offline
Ways to combine objectives at ranking time

Where objectives enter the recommendation pipeline

In a multi-stage ranking pipeline, objectives can be applied at different points. Candidate generation usually optimizes a broad relevance signal with a two-tower model, the ranking stage applies the full multi-task model and combined score, and a re-ranking stage enforces diversity, freshness, fairness and business rules on the final slate.

Placing an objective too late limits its effect, because re-ranking cannot recover items that earlier stages discarded. Placing it too early can waste capacity on candidates that ranking would reject anyway. Teams therefore decide per objective whether it belongs in retrieval, ranking or slate construction.

Evaluation, pitfalls and governance

Evaluation must report every objective separately, not only the combined score, so that a gain in one metric does not hide a loss in another. Offline metrics such as AUC per task and beyond-accuracy metrics for diversity and coverage are useful screens, but the decisive evidence comes from online experiments that track short-term and long-term outcomes.

  • Proxy drift: optimized signals such as clicks drift away from the real goal, so satisfaction proxies need periodic validation with surveys.
  • Scale mismatch: predicted probabilities of very different magnitudes make weights hard to interpret; calibration or normalization helps.
  • Feedback loops: the combined ranking shapes future training data, which can lock in early weight choices.
  • Opaque trade-offs: weights that encode policy, such as exposure for new creators, should be documented and reviewed like any other product decision.

Understanding it simply

Multi-objective recommendation means a system tries to please several goals at once when it picks what to show you. It is not only asking whether you will click, but also whether you will enjoy it, come back tomorrow, and see some variety.

Think of a good restaurant waiter suggesting dishes. They want you to order, but also to leave happy, try something new now and then, and return next week. If they only pushed the most expensive dish, they might win tonight and lose a regular customer. The key point is that good recommendations balance short-term and long-term goals rather than chasing one number.

Applied in practice

A mid-sized video streaming service ranks its home screen by predicted click-through rate alone and notices that users click often but abandon many titles within five minutes. The team builds a multi-task model with an MMoE layer that predicts click, completion above 70 percent, and a thumbs-up, and ranks by a weighted product of the three scores.

A re-ranking step then limits any single genre to three of the first ten slots and reserves one slot for recent releases. Over a four-week A/B test, clicks fall slightly, but completion rate and weekly return visits rise, and complaints about repetitive suggestions decline. The team keeps the new ranking and schedules quarterly reviews of the objective weights.