Learning to rank is a family of machine learning methods that produce an ordering of items tailored to a query and user context. In search, it decides which documents appear first; in ads, it balances relevance, predicted value, and policy constraints. Models are trained on signals such as clicks, conversions, and human judgments, often optimizing metrics like NDCG or MAP. This guide explains foundations and modern practice through the Top 10 Learning-to-Rank Algorithms for Search and Ads, showing how each approach fits real pipelines, from candidate generation to reranking and online experimentation, while highlighting strengths, trade offs, and deployment tips.
#1 RankSVM
RankSVM is a classic pairwise approach that learns a large margin classifier over preference pairs, such as document A should rank above document B for a query. It casts ranking as many binary comparisons and optimizes a hinge loss with regularization, yielding robust generalization on sparse, hand crafted features. In search platforms, RankSVM established strong baselines before tree ensembles and deep models rose to dominance. In ads, it can rank creatives or keywords using observed click preferences. Practical considerations include careful negative sampling, query level normalization, and kernel selection for nonlinear relations, balanced against computational cost for very large datasets.
#2 RankNet
RankNet introduced neural networks to ranking by minimizing cross entropy over pairwise preferences derived from user signals. Each item receives a score from a small multilayer perceptron; the model learns that preferred items should get higher scores. RankNet popularized differentiable surrogates for ranking metrics and enabled smooth gradients that directly target ordering quality. It is flexible with dense and sparse features, supports incremental updates, and pairs well with candidate generation from retrieval models. In ads, RankNet can integrate calibration layers to align predicted probabilities with auction goals, improving stability across traffic shifts and creative refresh cycles.
#3 LambdaRank
LambdaRank reframed training by computing pseudogradients called lambdas that approximate the change in a target metric such as NDCG when swapping two items. Instead of crafting a specific loss, it directly scales gradients by the metric impact of each pair, prioritizing top positions that matter most. This trick avoids complex listwise derivatives while aligning learning with business goals. LambdaRank can be implemented with neural scorers or with trees and supports position bias handling through gain discounting. In ad ranking, lambda weighting lets teams upweight conversions over clicks to reflect revenue, while still protecting user experience through constraints and regularization.
#4 LambdaMART
LambdaMART combines LambdaRank gradients with MART, a gradient boosted decision tree framework, to deliver state of the art accuracy on many search benchmarks. Trees capture non linear feature interactions, handle missing values, and require little feature scaling, which suits heterogeneous signals. The lambda weighting focuses learning on the highest impact swaps, driving gains in NDCG and click metrics. In ads, LambdaMART can optimize blended goals like expected value while honoring policy features through monotonic constraints. Engineers value its interpretability through split gain charts, partial dependence, and easy fallback to pointwise objectives during cold starts. Regularization, shrinkage, and shallow depth help control latency.
#5 ListNet
ListNet is a listwise method that models the entire permutation distribution of a ranked list using a probability over scores, often via a top one approximation for tractability. It minimizes divergence between predicted and target permutations, which better aligns with ranking metrics than pointwise or pairwise losses. ListNet naturally emphasizes top positions and supports graded relevance labels from human raters. In search, it provides stable improvements when feature spaces are well engineered. In ads, it can directly train on ordered sessions that reflect value, reducing mismatch between offline training and online objectives through metric aligned learning.
#6 ListMLE
ListMLE learns rankings by maximizing the likelihood of observed permutations under a Plackett Luce model, providing a probabilistic foundation for listwise learning. It treats each ranked list as an ordered factorization where higher scored items are more likely to appear earlier. This yields consistent estimators under reasonable assumptions and supports gradient based optimization with neural or linear scorers. In search, ListMLE often improves early precision by focusing on top ordering. In advertising, it can encode constraints through feature design while training on editorial judgments and de biased click logs, producing smooth updates that respect platform policies.
#7 Coordinate Ascent
Coordinate Ascent directly optimizes an offline metric such as NDCG by iteratively adjusting one weight at a time in a linear scoring function. It is simple, transparent, and surprisingly strong when you have a compact set of expert features. Because each step is one dimensional, it is easy to audit which signals help and to enforce monotonicity or sparsity. Search teams often use it as a warm start or for fast personalization. In ads, it can tune blends of relevance and value without heavy infrastructure, serving as a reliable baseline that is easy to interpret and to roll back if metrics drift.
#8 RankBoost
RankBoost adapts boosting to ranking by combining weak rankers that focus on previously misordered pairs. Each round reweights training examples, emphasizing difficult preferences that hurt metrics. The final ensemble aggregates many simple rules, which can be decision stumps or small trees, producing a strong, interpretable model. RankBoost predates several listwise methods but remains useful when feature engineering yields informative thresholds. In search, it offers competitive accuracy with modest latency budgets. In ads, it can be coupled with sampling schemes that reflect revenue impact, and it provides natural diagnostics through the distribution of example weights over time.
#9 XGBoost Ranking
XGBoost ranking extends gradient boosted trees with specialized objectives such as pairwise logistic loss and approximations of NDCG. It brings industrial grade parallelism, regularization, and handling of sparse inputs, which suits high dimensional query and user features. With ranking objectives, XGBoost focuses splits on features that correct harmful orderings at the top of results. In ads, it can optimize composite targets such as expected revenue by using instance weights and group definitions. Practitioners value strong out of the box performance, distributed training, and straightforward deployment through model export and low latency inference. Calibrators and monotonic constraints further improve reliability in production.
#10 Neural Two Tower and Deep Interaction Models
Neural two tower and deep feature interaction models power modern search and ads by learning expressive representations and flexible ranking scores. A dual encoder retrieves candidates efficiently using vector search, while a cross network or attention model reranks by modeling fine grained interactions. Architectures such as Wide and Deep, DeepFM, and transformer rankers support multi task learning that balances clicks, conversions, and policy objectives. They scale to millions of items, ingest real time features, and enable exploration through stochastic training. Careful calibration, counterfactual evaluation, and multi objective optimization make these models effective in complex auctions.