KBest: Top-k Selection Methods & Applications
- KBest is a family of methods that involve selecting the top k candidates, with each domain tailoring the strategy to ranking, enumeration, or adaptive sampling.
- In Bayesian network structure learning, KBest uses dynamic programming to compute multiple high-scoring network structures for improved model averaging over MAP selection.
- Across feature selection, bandits, and vector search, KBest provides fast, domain-specific strategies, balancing efficiency with challenges like redundancy and hardware constraints.
Searching arXiv for papers on “KBest” across domains to ground the article in published sources. {"query":"KBest arXiv KBest Bayesian network structures feature selection bandits vector search", "max_results": 10} {"query":"KBest Bayesian Model Averaging Using the k-best Bayesian Network Structures (Tian et al., 2012)", "max_results": 5} KBest is not a single canonical method but a recurrent designation for top- selection, enumeration, or search procedures whose semantics depend on the domain. In Bayesian network structure learning, it denotes the -best highest-scoring network structures used for Bayesian model averaging; in classical filter feature selection, it denotes the univariate Max-Rel procedure that ranks features by relevance and selects the top ; in bandit theory, closely related Best- and Best-of- formulations concern identification of the top arms or subsets; and in systems work, KBest is also the name of a vector search library tailored for Kunpeng 920 CPUs (Tian et al., 2012, Ebiele et al., 30 Mar 2026, Jiang et al., 2017, Simchowitz et al., 2016, Ma et al., 5 Aug 2025). The common motif is retention of a restricted set of high-scoring candidates rather than commitment to a single optimum, but the search space, statistical objective, and computational trade-offs differ substantially across these usages.
1. Terminological scope and core abstraction
Across the cited literature, KBest denotes one of three closely related ideas: a ranking rule, an enumeration strategy, or a system name. In the feature-selection literature, KBest is explicitly identified with a univariate filter feature selection algorithm, also referred to as SelectKBest or Max-Rel, that uses sorting in descending order after computing a relevance score for each feature (Ebiele et al., 30 Mar 2026). In Bayesian network learning, the phrase “-best” denotes the top- Bayesian network structures under a scoring criterion and is used to approximate Bayesian model averaging over an otherwise infeasible DAG space (Tian et al., 2012). In systems work, KBest is a proper noun for a high-performance approximate nearest neighbor search library optimized for Huawei Kunpeng 920 ARM CPUs (Ma et al., 5 Aug 2025).
| Domain | Meaning of KBest | Primary operation |
|---|---|---|
| Bayesian networks | -best structures | Dynamic-programming enumeration |
| Feature selection | Max-Rel / SelectKBest | Relevance ranking and top- selection |
| Bandits | Best-0 / Best-of-1 related formulations | Adaptive identification under uncertainty |
| Vector search | KBest library | Hardware-aware ANN search |
| Quantization | K-best sampling | Candidate generation and residual minimization |
This multiplicity matters because claims about optimality, efficiency, or robustness for one KBest formulation do not transfer automatically to another. A common misconception is that KBest always means simple top-2 sorting; the literature instead uses the term for deterministic dynamic programming, confidence-based adaptive sampling, and hardware-specialized search pipelines, depending on context.
2. 3-best Bayesian network structures and Bayesian model averaging
In Bayesian network structure learning, KBest refers to the problem of finding the 4-best Bayesian network structures for a dataset and then using those structures for Bayesian model averaging rather than relying on a single MAP network (Tian et al., 2012). The method extends earlier dynamic programming algorithms for optimal structure search to maintain the top 5 solutions at each subproblem.
The algorithm proceeds in two nested stages. First, local scores are precomputed for all variable–parent set pairs, written as 6. Second, for each variable 7 and candidate parent set 8, the algorithm computes the 9-best parent sets for 0 from 1 through a recursive best-first construction over subsets 2. Network construction then uses the standard sink decomposition: for a variable set 3, and each sink 4, the 5-best networks over 6 are assembled from the 7-best parent sets for 8 from 9 and the 0-best networks over 1, with combination score
2
The resulting time and memory costs remain exponential in 3, but the extension from top-1 to top-4 is described as efficient. The reported complexities are 5 for finding 6-best parent sets and 7 for finding 8-best networks (Tian et al., 2012).
The same paper proposes to approximate Bayesian model averaging by restricting the sum to the 9-best DAGs: 0 For structure features 1, the approximation becomes
2
and for predictive likelihood on new data 3,
4
Model selection is recovered as the special case 5 (Tian et al., 2012).
A distinctive contribution is the quality analysis of this approximation. Let
6
Then the posterior approximation error is bounded by
7
The paper also introduces
8
the ratio of posteriors between the best and worst networks in the 9-best list, and notes that a small 0 indicates that the 1-best list sufficiently dominates the plausible posterior mass (Tian et al., 2012).
Empirically, averaging over the 2-best networks is reported to outperform model selection and state-of-the-art MCMC methods, including the DP+MCMC method of Eaton and Murphy, especially for structural discovery tasks such as recovering the true network structure and edge prediction. The paper states that Top 100 networks matched or exceeded the best DP+MCMC performance, with much lower variability than MCMC, and that with synthetic 3 networks and 4 samples the 5-best method produced AUC curves nearly indistinguishable from exact full BMA (Tian et al., 2012).
3. KBest as univariate Max-Rel feature selection and as an interactive trainer
In feature selection, KBest is a classic univariate filter method. The procedure computes a relevance score 6 for each feature 7, sorts features in descending order, and selects the top 8 (Ebiele et al., 30 Mar 2026). The relevance score can be instantiated by mutual information, the F-value, or Gini importance. The Max-Rel formulation is given as
9
Its primary advantages are simplicity, repeatability, and computational efficiency; its principal limitation is that it does not consider redundancy, so it may select highly correlated or overlapping features (Ebiele et al., 30 Mar 2026).
The comparison with mRMR and KGroups clarifies the algorithmic role of KBest. KBest is univariate and uses sorting; mRMR is multivariate and uses incremental search with explicit redundancy control; KGroups is proposed as a new univariate “mRMR” that employs clustering for selection and is reported to outperform KBest while remaining very fast (Ebiele et al., 30 Mar 2026). The KGroups paper states that KBest is parameterisable only via the number of selected features 0, whereas KGroups leaves room for further predictive performance improvement through hyperparameter finetuning (Ebiele et al., 30 Mar 2026).
A second feature-selection usage appears in the IRFS framework for automated feature selection, where a “KBest based trainer” acts as an external advisor to reinforcement learning agents (Fan et al., 2020). Each feature 1 is assigned a corresponding agent 2, and agents first issue initial select/deselect actions. Among the currently participated features, agents are divided into assertive and hesitant subsets. The KBest-based trainer intervenes only with hesitant agents. If 3 is the number of assertive features and 4 the number of hesitant features, the framework computes
5
runs KBest on the participated features 6 to obtain 7, and advises a hesitant agent to keep its feature selected if that feature belongs to the top-8 set (Fan et al., 2020). The final action rule is
9
Within IRFS, the KBest-based trainer is used alongside a Decision Tree based trainer in a hybrid teaching strategy designed to diversify training. The paper reports that IRFS with a KBest-based trainer consistently outperformed pure MARLFS and classic KBest in both accuracy and search efficiency, and that hybrid teaching further improved performance on datasets such as FC and Musk (Fan et al., 2020). This usage preserves KBest’s identity as a fast filter heuristic but embeds it within a sequential decision process rather than using it as a standalone selector.
4. Best-0 and Best-of-1 formulations in bandit theory
In bandit theory, the terminology shifts from KBest as a deterministic ranking rule to Best-2 identification and Best-of-3 subset selection. The Best-4-Arm problem considers 5 stochastic bandit arms with unknown reward distributions and asks for identification of the 6 arms with the largest means with high confidence by adaptive sampling (Jiang et al., 2017). The paper “Practical Algorithms for Best-K Identification in Multi-Armed Bandits” introduces lil'RandLUCB and lil'CLUCB, both based on law-of-iterated-logarithm confidence bounds. Their stated sample complexity is
7
matching the lower bound up to logarithmic factors (Jiang et al., 2017). The paper reports that lil'RandLUCB requires significantly fewer samples than prior provably correct algorithms and outperforms prior algorithms even for 8 in practice (Jiang et al., 2017).
A related but distinct problem is the Best-of-9 Bandit game. Here, at each time the player chooses a subset 0 among all 1 options and observes reward 2 where 3 is drawn from a joint distribution (Simchowitz et al., 2016). The objective is to identify the subset
4
with high probability using as few queries as possible. The paper studies bandit, marked-bandit, and semi-bandit feedback models and develops distribution-dependent lower bounds showing that, for certain hard dependent distributions, a learner is forced to consider all 5 subsets, matching naive subset-as-arm upper bounds up to lower-order and logarithmic factors (Simchowitz et al., 2016).
The independent-arm case exhibits a different structure. If the arms are independent Bernoulli variables with means 6, then the optimal set consists of the 7 largest means and the expected reward of a set 8 is
9
The paper gives an algorithm and analysis for this setting and identifies “information occlusion” as the central difficulty in max-only feedback: the contribution of an arm may be masked by the presence of stronger arms in the queried subset (Simchowitz et al., 2016). This distinction between Best-0-Arm and Best-of-1 is conceptually important: the former is a top-2 arm identification problem, whereas the latter is a combinatorial pure-exploration problem over subsets.
5. KBest as a vector search library on Kunpeng CPUs
KBest is also the name of a vector search library tailored for the latest Kunpeng 920 CPUs (Ma et al., 5 Aug 2025). In this usage, the term does not denote a generic top-3 abstraction but a specific software system for approximate nearest neighbor search on ARM hardware. The library targets in-memory graph-based indexes such as HNSW, NSG, Vamana, and SSG, and its architecture includes a CSR-formatted proximity graph, linear vector storage, and a dynamic thread pool (Ma et al., 5 Aug 2025).
The defining characteristic of this KBest is hardware-aware optimization. The paper lists single-instruction-multiple-data accelerated distance computation, data prefetch, index refinement, early termination, and vector quantization as core optimizations (Ma et al., 5 Aug 2025). It emphasizes ARM NEON/SVE utilization rather than x86 AVX2/AVX-512, batch prefetch during graph traversal, huge pages and cache-line alignment for memory management, graph reordering based on a minimum spanning tree, and plug-in support for Product Quantization and Scalar Quantization (Ma et al., 5 Aug 2025).
The reported evaluation compares KBest on Kunpeng 920 with strong x86 baselines on Glove, Deep, Text-to-Image, and BigANN. At Recall4, the reported QPS values are 170k on Glove versus a best baseline of 145k, 628k on Deep versus 459k, 113k on Text2Image versus 94k, and 357k on BigANN versus 305k, corresponding to relative improvements of 1.17x, 1.37x, 1.20x, and 1.17x respectively (Ma et al., 5 Aug 2025). The paper further states that the optimizations can improve query throughput by over 2x, that KBest attains a 1.04–1.45x QPS advantage over the best x86 baseline at high recall, and that comparisons against less optimized graph libraries can yield improvements up to 12x (Ma et al., 5 Aug 2025).
This KBest is also described as deployed in production. The paper states that it serves internal business and external enterprise clients with tens of millions of queries on a daily basis and lists major social media, e-commerce, and food delivery platforms as application settings. It also reports direct integration into vector databases such as Milvus and openGauss and describes industrial uses including real-time personalized recommendations, product search, food menu matching, satellite imagery, and telecom analysis (Ma et al., 5 Aug 2025).
6. Extensions, limitations, and recurrent misunderstandings
A recurrent misunderstanding is to treat KBest as a universally interchangeable method. The literature instead presents domain-specific instantiations with distinct objectives, complexity profiles, and failure modes. In Bayesian network learning, the 5-best method is a deterministic approximation to full Bayesian model averaging, but its computational cost is still exponential in 6 and is described as practical only for moderately-sized networks, up to approximately 7; the paper also notes redundancy because many of the 8-best networks are Markov equivalent or nearly so, and equivalence-class enumeration remains open (Tian et al., 2012). In univariate feature selection, KBest is fast and model-agnostic, but it ignores redundancy and is not parameterisable beyond 9 (Ebiele et al., 30 Mar 2026). In Best-of-00 bandits, hard dependent instances can force exhaustive exploration of all 01 subsets, so the phrase “best-02” does not imply an easy top-03 screening problem (Simchowitz et al., 2016). In the Kunpeng vector-search system, the reported gains are inseparable from ARM-specific engineering choices and therefore should not be read as architecture-independent statements about approximate nearest neighbor search (Ma et al., 5 Aug 2025).
A further extension of the top-04 idea appears in post-training quantization. OJBKQ, “Objective-Joint Babai-Klein Quantization with K-Best Sampling,” formulates weight quantization as a multiple-right-hand-side box-constrained integer least squares problem and, for each column, runs 05 randomized Klein paths plus one Babai path before selecting the minimum-residual Babai-Klein point (Wang et al., 9 Feb 2026). The paper states that OJBKQ achieves lower perplexity at 3–4 bits compared to existing PTQ approaches while maintaining comparable computational cost, and that increasing 06 rapidly improves perplexity with diminishing returns after 07 (Wang et al., 9 Feb 2026). This suggests that “KBest” functions not only as a selection rule or system name but also as a reusable search primitive for generating and evaluating multiple high-quality candidates under a constrained optimization objective.
Taken together, the literature presents KBest as a family resemblance rather than a single algorithmic identity. The common structure is explicit retention of several strong candidates, but the substantive meaning ranges from top-08 relevance ranking, to dynamic-programming enumeration over DAGs, to adaptive pure exploration, to hardware-specialized ANN search, and to randomized candidate generation in quantization.