The Hidden Dimension of Relevance: How Price Shapes Recommender Outcomes

In most recommender systems, the primary objective is to surface the most relevant items to users eg : restaurants, dishes, movies, products that they are most likely to purchase. Optimizing for relevance makes practical sense as it aligns with engagement metrics like conversion rate ( CVR ) or click through rate ( CTR ). But what happens when relevance competes with an important dimension - price ?

The optimization of engagement metrics often leads to an unintended consequence : cheaper items often get ranked higher in the final lists. Lower priced items tend to have strong historical conversion rates because of the low risk attached to them. When the model learns this pattern, it begins to systematically favor these items leading to short term engagement gains but not necessarily long term value.

This creates a self reinforcing loop where cheaper items are preferred to boost engagement metrics leading to the whole marketplace/application converging towards a narrow price

band. This result is because the recommender system although technically optimized is economically imbalanced leading to an unfair marketplace. Building recommender systems that are price aware is a form of multi-objective optimization focused on finding the right balance between relevance and price awareness. This article explores how to formulate this tradeoff, combining it with the model and measuring success beyond plain engagement metrics.

This bias isn’t limited to industries like food delivery & online shopping. Many other industries face similar distortions. Past data often represents cheaper items since they receive more impressions and interactions, teaching the model that “low price means high relevance”. In reality, the model is overindexing to user friction rather than user value.

Why Price Matters in Recommendations?

Price is not just a number as it directly impacts perception, trust and users retention of the platform. For users, recommendations that truly understand their budget leads to increased satisfaction and retention. Repetitive impressions of cheap items can create a sense that the platform lacks high quality items and variety. It also can undercut discovery as users who are open to higher priced items never see them.

For the business, recommendations being price aware helps in widening participation, prevents overindexing on smaller subset of items and stabilises the marketplace dynamics. Otherwise, merchants and dishes that are expensive receive fewer impressions making the platform less attractive for them.

A balanced recommender system should consider both relevance and economic contribution as objectives. This does not just mean pushing more expensive items rather ensuring price diversity and value perception is maintained at healthy levels across all user segments. This is important because according to behavioral research - price often signals quality. A list of items dominated by low cost items can unknowingly communicate that the platform itself is lower in quality.

Price as a First Class Feature Rather Than an Afterthought

Incorporating price as a first class signal is very important, not just using price as a feature in the model and using it in post-processing adjustments. Price is not just a number as it carries context and intent. A $15 entree might be considered cheap in one neighborhood and premium in another. A $40 dinner may be appropriate for a weekend but not for a weekday lunch. Similarly, different countries have different spending patterns as well as affordability indexes that a global recommender system should be able to recognize.

There are some techniques to make your recommender systems more price aware and effective :

• Normalize prices by merchant or category to capture relative affordability.
• Model geographical and temporal price features (city averages, time of day spending patterns, etc)
• Including personalized spend embeddings representing historical average order values or price preferences.

Another important aspect is feature scaling. Price distributions often have a long tail and without proper normalizations, models can overfit to these outliers. Using techniques like z-score or log normalizations can help maintain feature stability. It is also important to incorporate seasonal effects like deals/offers, inflation etc and ensure that the model does not overindex on them, treating them as temporary fluctuations.

This enabled recommender systems to learn richer relationships i.e some users value affordability, others value quality and switching happens based on context. By embedding price context, the model can better understand when and what to recommend to different users.

Balancing Engagement and Business Value

A relevance only objective may look like the following:

������ ( ��������/������)

where you pick the item with the maximum probability of engagement. But a multi-objective system can introduce a second goal for price:

������ [ α * ��������/������ + (1 − α ) * ��(����������)]

Here, alpha represents the weight we want to place on relevance versus price. Tuning this parameter is as much a business decision as it is a technical one.

In practice, the formulation can look different than the above and can take several forms:

• Weighted loss functions combining engagement and price based objectives.
• Constraint optimization where the model tries to maximize the average order value.
• Reinforcement learning or bandit frameworks where you tradeoff price exploration and relevance.

The key principle remains the same i.e. have an objective function that is optimized for all the objectives important to the business and find the best tradeoff between them. One important step to call out here is model output score calibration. Your model output score distribution should match the empirical distribution to ensure the price objective does not interact in an unexpected fashion. This is a common scenario where subsequent model training severely changes the output score distribution requiring re-tuning of the price parameters which is time consuming.

Price-Aware Re-Ranking

Even the best and most complex trained models can overfit to relevance signals. Having a lightweight re-ranking layer provides an interpretable correction mechanism. A re-ranking module operates after the main model outputs the top N candidates. It adjusts the final ranking using price constraints or business constraints as mentioned in the equations above.

The typical methods for this are:

• Constrained re-ranking : Enforce at least X% items are in the mid range and in the high price range. This ensures a manual layer of control on the outputs which is expected to drive positive user experience.
• Penalty based re-ranking : Add a penalization on top of model score if the item is priced below a particular threshold. This ensures fairness in recommendations and also fairness in future model training by having more uniform data across price ranges.
• Marketplace based Exposure : Ensure that the item price distribution matches the rough marketplace inventory for fairness.

This step maintains stricter economic diversity that the model objective function cannot guarantee. It’s a lightweight way to prevent cheaper items from being surfaced everywhere while preserving importance to relevance.

Context Dependent Weighting

As mentioned above, the relationship between relevance and price is heavily dependent on the context. A lunch time user may prioritize convenience and affordability while a weekend dinner user is open to high quality and premium dishes. Similarly, user cohorts differ in their price elasticity where some would consistently spend more than others.

Context adaptive weighting talk about how the optimization can be adjusted dynamically:

• Increase relevance importance for user cohorts with low price elasticity.
• Increase price importance during certain time periods or premium contexts.
• Adjust for different cohorts like user cohorts, geographies, time of day etc.

This adaptive mechanism avoids the one size fits all approach that a model would take and ensures that recommendations align with contextual signals. By making the tradeoffs context aware, systems stay responsive to both context and business goals.

Evaluation of Price-Aware Recommender Systems

Evaluation becomes tricky when balancing multiple objectives. Traditional engagement metrics like CVR & CTR capture user engagement but not economic health. There are other metrics that one should monitor:

• Economic Metrics : These metrics help you understand the economics aspects of performance. Metrics in this category would include : Average Order Value, Avg Price per Item, Overall Gross Merchandise Value etc
• Price Fairness Metrics : These metrics help you understand the inputs to the economic aspects i.e what type of items are being shown to the users. Metrics in this category would include : Variance of item prices, Price range exposure across different cohorts etc
• Long Term Metrics : These are key long term metrics helping to understand whether changes made are leading to improving key long term behavior. These include Retention, Repeat Purchase rate, Lifetime value etc

Some Offline hybrid metrics like weighted NDCG ( combining price and relevance ) can be helpful in simulating tradeoffs. Along with these, evaluation frameworks that simulate price fairness across cohorts can be useful to bridge the gap between offline expectations and online performance. Online A/B tests are essential as they can provide signals on revenue change, marketplace fairness etc

Conclusion

When recommender systems optimize purely for relevance, they become victims of their own success. This leads to surfacing less expensive, most clickable items/dishes at the expense of economic value and marketplace balance.

A price-aware recommender system redefines the problem by showing users dishes that create long term value. By embedding price based features and treating it as a first class citizen, we can create recommender systems that truly understand the importance of price. Secondly, by rethinking objective functions - we can create systems serving user intent and business health.

About the Author

Madhura Raut is a Principal Data Scientist and tech leader in the San Francisco Bay Area with 9+ years of experience delivering production ML systems, with a focus on time-series forecasting, deep learning, and agentic AI. She is an IOA Fellow and Senior Member of IEEE and IET, also an active speaker, reviewer, hackathon judge, and mentor, with talks including KDD 2025, Data Science Salon, and features across Entrepreneur, The Economic Times, Mashable, and Vogue. LinkedIn : https://www.linkedin.com/in/madhuraraut/

Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE’s position nor that of the Computer Society nor its Leadership.