Diversity Re-Ranking

The foundational algorithm for diversity re-ranking that iteratively selects documents based on a linear combination of relevance to the query and novelty relative to already-selected documents. The formula MMR = argmax[λ * Sim(D, Q) - (1-λ) * max Sim(D, Di)] balances the trade-off, where a higher λ prioritizes relevance and a lower λ enforces diversity. Carbonell and Goldstein introduced MMR in 1998 to reduce redundancy in text summarization and information retrieval.

A diversity objective that partitions the information space into predefined facets or subtopics and ensures the re-ranked list contains at least one representative document from each. This is critical for ambiguous queries like 'jaguar' (animal, car, OS) where user intent is uncertain. The system explicitly models the aspect coverage ratio—the percentage of known subtopics represented in the top-k results—to prevent a single dominant interpretation from monopolizing the list.

An elegant probabilistic framework for diversity re-ranking that models the probability of selecting a subset of documents as proportional to the determinant of a kernel matrix. This matrix encodes both quality scores (diagonal elements) and pairwise similarity (off-diagonal elements). DPPs naturally balance relevance and diversity by assigning higher probability to subsets containing high-quality, dissimilar items. They are particularly effective for session-based recommendations where the goal is to present a visually or thematically varied set.

A modern approach that uses a query intent classifier to predict the probability distribution over user intents before re-ranking. The system then applies proportional representation: if 60% of users searching 'apple' want the fruit and 40% want the technology company, the top-10 results should reflect that 60/40 split. This is often implemented using intent-aware MMR or by training a Cross-Encoder to predict both relevance and intent coverage simultaneously.

A probabilistic framework that estimates the likelihood a document satisfies a specific aspect or subtopic given that it has not been covered by previously selected documents. The scoring function is P(D|Q) * Π(1 - P(Si|D)), where the product term penalizes documents covering already-satisfied aspects. xQuAD is highly interpretable because it explicitly models the marginal utility of each document in filling remaining information gaps in the ranked list.

The standard algorithmic approach for building a diverse ranking, since finding the globally optimal diverse subset is NP-hard. The greedy algorithm starts with the most relevant document and iteratively selects the next document that maximizes the diversity objective (e.g., MMR or DPP gain) given the current set. While suboptimal, this O(kn) approach produces high-quality approximations and is fast enough for real-time re-ranking of top-100 candidate sets in production search systems.

The foundational greedy algorithm for diversity re-ranking. MMR iteratively selects documents that maximize a linear combination of relevance to the query and novelty relative to already selected documents. The trade-off is controlled by a λ parameter: λ=1 yields pure relevance ranking, while λ=0 maximizes diversity. Similarity between documents is typically computed using cosine similarity over TF-IDF or dense embeddings. The algorithm penalizes candidates that are too similar to items already in the ranked list, effectively suppressing near-duplicate or redundant subtopics.

A diversity objective that ensures the top-k results cover as many distinct facets or interpretations of an ambiguous query as possible. For example, for the query 'jaguar', a sub-topic aware re-ranker would surface documents about the animal, the car brand, and the sports team rather than returning ten documents about the animal. This is often implemented by classifying documents into pre-defined taxonomy nodes or by clustering candidate documents in embedding space and sampling from each cluster proportionally to its relevance score.

A probabilistic model for diversity re-ranking that defines a distribution over subsets of documents, where the probability of a subset is proportional to the determinant of a kernel matrix encoding both quality (relevance) and similarity (redundancy). DPPs naturally balance relevance and diversity without a hand-tuned λ parameter. The MAP inference task—finding the subset that maximizes the determinant—is NP-hard in general, but efficient greedy approximations exist. DPPs are particularly effective for video and image recommendation where visual diversity is critical.

A diversification strategy that first predicts the distribution of user intents behind a query using a probabilistic classifier, then re-ranks documents to ensure proportional representation of each intent in the top results. For instance, if the query 'apple' has a 60% probability of referring to the technology company and 40% to the fruit, the re-ranked list should reflect this ratio. This approach requires a pre-built intent taxonomy and a model trained to estimate P(intent|query), often using click-through data as a supervision signal.

A probabilistic framework for diversity re-ranking that explicitly models query aspects (subtopics). xQuAD estimates P(d|q) — the relevance of document d to the original query q — and P(d|a_i) — the relevance of d to each aspect a_i. It then iteratively selects documents that are relevant to the query and cover aspects not yet well-represented, using a parameter to control the diversity emphasis. This provides a transparent, interpretable diversification mechanism suitable for e-commerce facets and academic search.

A critical distinction in re-ranking: diversity measures how different the top-k results are from each other (intra-list dissimilarity), while novelty measures how much new information a document provides relative to what the user has already seen (historical dissimilarity). A diverse result set avoids internal redundancy; a novel result set avoids repeating previously consumed content. Production systems often combine both objectives: penalizing similarity to items already in the ranked list (diversity) and penalizing similarity to items in the user's interaction history (novelty).