Fact-checked by Grok 2 weeks ago

Lazy learning

Lazy learning, also known as , is a paradigm in in which algorithms defer the generalization or processing of training data until a specific query or prediction is required, typically by storing the raw training examples in memory and performing computations based on local similarities to the query instance at runtime. This approach contrasts with eager learning methods, such as decision trees or neural networks, which construct an explicit global model during the training phase and use it for subsequent predictions. Key examples of lazy learning algorithms include the k-nearest neighbors (k-NN) algorithm, which classifies or a query point by aggregating the labels or values of its k most similar training instances, and locally weighted regression, which fits a local model weighted by proximity to the query for . These methods are particularly suited for tasks involving non-linear decision boundaries where global modeling may be inefficient. Lazy learning gained prominence in the through research on instance-based techniques, with foundational work exploring their applications in , , and . One of the primary advantages of lazy learning is its minimal training overhead, as it requires only storage of the data without complex , making it computationally efficient during the learning phase and resistant to by avoiding premature . Additionally, these algorithms can produce flexible, interpretable predictions that adapt easily to new training data by simply updating the stored instances, and they handle noise well through local averaging in methods like k-NN. However, lazy learning suffers from high memory requirements to retain all training examples, potentially leading to issues with large datasets, and prediction times can be slow due to the need for similarity computations over the entire dataset for each query. It is also vulnerable to the curse of dimensionality, where performance degrades in high-dimensional spaces due to sparse data distributions, and irrelevant features can distort similarity measures.

Introduction

Definition and Core Concepts

Lazy learning, also known as , is a paradigm that defers the induction or generalization process until a classification or prediction query is received. Unlike parametric methods that construct an explicit model from the training data, lazy learning algorithms store the raw training examples directly in memory without deriving abstractions or compact representations during a dedicated training phase. This approach enables predictions to be generated by interpolating or approximating based solely on specific instances relevant to the query. At its core, lazy learning relies on the non-parametric nature of the stored data, avoiding the learning of fixed global parameters and instead performing computations . Key characteristics include deferred learning, where the bulk of the occurs at time, and memory-based computation that leverages the full or a subset of instances. This facilitates to instance-specific neighborhoods, allowing for localized approximations that can capture complex, non-linear patterns without assuming a predefined functional form. The process in lazy learning typically involves two main steps: first, computing similarity measures between the new instance and the stored training examples, often using distance metrics such as ; second, aggregating these similarities to form a , such as through majority voting for tasks or weighted averaging for . This method contrasts with eager learning techniques, which perform upfront to enable faster queries but may overlook local data nuances.

Historical Development

The origins of lazy learning can be traced to early work in during the mid-20th century, where non- methods began to emerge as alternatives to rigid models. In 1951, Evelyn Fix and Joseph Hodges introduced the nearest-neighbor approach for nonparametric classification, proposing a method that classifies new instances based on the closest training examples without assuming an underlying distribution. This foundational idea addressed limitations in early statistical classifiers by avoiding the need for explicit model fitting, laying the groundwork for instance-based techniques in . A key milestone came in 1967 with the development of the k-nearest neighbors (k-NN) algorithm by Thomas Cover and Peter Hart, who extended the single-nearest neighbor rule to use multiple neighbors for improved robustness and provided theoretical bounds on its error rate relative to the Bayes optimal classifier. This work formalized the nearest-neighbor rule within , demonstrating its asymptotic consistency and practical utility for pattern classification tasks. By the , as computational resources grew, these methods gained traction in research, enabling the storage and retrieval of large instance sets that parametric models struggled to handle due to assumptions about . The formalization of as a broader occurred in 1991, when David , Dennis Kibler, and Marc Albert introduced the IB1 , an incremental variant of 1-NN that incorporated tolerance and drift handling. This work formalized (IBL) as deferring until prediction time, emphasizing similarity-based reasoning from stored examples. The term "lazy learning" was later applied to such methods, notably in a 1997 collection edited by . Influential in early , these lazy methods highlighted the drawbacks of parametric approaches—such as to incorrect assumptions—in domains like and , where data variability was high. During the , advancing hardware and memory capabilities further propelled lazy learning's adoption, allowing memory-intensive algorithms to scale beyond small datasets. By the , integration into accessible libraries, such as scikit-learn's implementation of k-NN starting with its public release in 2010, democratized these techniques for practitioners. This evolution underscored lazy learning's role in complementing models, particularly in non-stationary environments where flexibility outweighed the computational cost of deferred processing.

Comparison to Eager Learning

Fundamental Differences

Lazy learning, also known as , fundamentally differs from eager learning in its approach to model construction. During the training phase, lazy learning algorithms do not build an explicit global model; instead, they simply store the training instances, resulting in minimal computational effort at this stage, often described as constant for processing beyond storage. In contrast, eager learning methods, such as those used in decision trees or neural networks, perform extensive computations during training to derive a generalized model, such as fitting parameters or constructing hierarchical structures from the entire dataset. This deferral of abstraction in lazy learning allows it to avoid premature generalization, but it requires retaining the raw data for later use. The phase highlights another key contrast. In lazy learning, predictions are computed on-the-fly for each new query by examining relevant instances, such as the k nearest neighbors, which can involve scanning a portion or all of the stored data, leading to potentially linear per prediction. Eager learning, however, leverages the pre-trained model for rapid inference, where computations are typically independent of the set size after the initial , enabling faster query responses at the expense of upfront investment. This local, instance-specific computation in lazy learning contrasts with the global application of the abstracted model in eager approaches. Lazy learning methods are inherently non-parametric, meaning they do not assume a fixed number of parameters or a predefined functional form; instead, their complexity grows with the training data size, adapting flexibly without imposing distributional assumptions. For example, unlike the parametric form of , expressed as y = \beta x + \epsilon, where \beta represents fixed learned coefficients, lazy learners rely directly on instance similarities without estimating such parameters. Eager learning often employs parametric models that assume a specific structure, facilitating compact representations but potentially limiting adaptability to data variations. In terms of data handling, lazy learning retains all or a of the original instances, enabling perfect of the and the ability to revisit specifics for each prediction, which supports to the . This contrasts with eager learning's strategy of abstracting the into a condensed model during , which promotes but can introduce risks of underfitting or by discarding instance-level details. As a result, lazy methods excel in scenarios where data locality is crucial, while eager methods prioritize scalable abstraction.

Trade-offs in Model Building and Prediction

Lazy learning algorithms exhibit minimal computational overhead during the training phase, as they primarily involve storing the input data instances without performing any generalization or model construction, resulting in near-constant time complexity often denoted as O(1) per instance added. This makes lazy learning particularly advantageous for environments with dynamic datasets that update frequently, where retraining eager models would otherwise demand substantial recomputation. In contrast, eager learning methods, such as decision trees or neural networks optimized via gradient descent, incur high training costs due to iterative processes like multiple passes over the data for parameter adjustment, leading to time complexities that can scale polynomially or worse with dataset size. At prediction time, the reverses sharply: lazy learners must compute responses on-the-fly by searching or scanning the stored for relevant instances, often resulting in O(n) time per query in basic implementations, which becomes prohibitive for large n. This per-query latency can be mitigated with indexing structures but remains inherently higher than eager approaches. Eager learning, having distilled the data into a fixed model during , enables rapid , typically in O(1) or logarithmic time, facilitating applications post-initial setup. Space requirements further highlight the paradigm's differences, with lazy methods demanding O(n \cdot d) storage to retain the full of n instances each with d features, precluding and potentially leading to bottlenecks in resource-constrained settings. Eager learners, by contrast, encode learned patterns into compact representations—such as weight matrices in linear models or layered architectures in neural networks—that occupy far less space, often independent of n once trained. These resource profiles influence profiles distinctly: lazy learning scales well to small-to-medium sets (n \lesssim 10^5) or streaming scenarios where evolves without full retraining, avoiding the upfront optimization burden of eager methods. However, for massive-scale problems, eager learning's precomputation enables efficient deployment on distributed systems, as the model's size remains manageable even as n grows large. Lazy approaches face inherent challenges in high-dimensional or contexts due to the quadratic growth in search costs without advanced approximations. In terms of predictive accuracy, lazy learners achieve low through their non-parametric flexibility, adapting locally to patterns without assumptions, but they exhibit high variance in sparse or noisy regions where instance retrieval may yield inconsistent neighbors. Eager methods counter this via regularization techniques—such as weight decay or —that explicitly balance and variance, often yielding more stable error rates across diverse distributions.

Properties

Advantages

One key advantage of lazy learning is its adaptability to non-stationary environments. By storing raw instances and deferring until time, these algorithms can incorporate new without retraining a fixed model, allowing them to respond dynamically to changes in distribution. This contrasts with eager learning approaches that may become outdated as they rely on a static global model built during . Lazy learning avoids training-phase biases inherent in model selection processes of eager methods. Without committing to a specific form or early on, it constructs local approximations tailored to individual query instances using similarity measures, reducing the risk of to global assumptions that may not hold locally. This unbiased approach enables more accurate representations of complex, irregular decision boundaries. The interpretability of lazy learning stems from its reliance on stored training examples for predictions. Each decision can be traced back to a subset of specific instances, such as the k nearest neighbors, facilitating , validation, and understanding of the reasoning behind classifications or regressions. This transparency is particularly valuable in domains requiring explainable , where users can inspect the contributing examples directly. Lazy learning algorithms exhibit simplicity in implementation and tuning. They typically require minimal preprocessing and few hyperparameters—often just the number of neighbors k or a metric—making them straightforward to deploy compared to eager methods that demand extensive model optimization. This ease extends to handling diverse data types by selecting appropriate similarity functions.

Disadvantages

Lazy learning algorithms, while adaptable to new data without extensive retraining, exhibit several key limitations that can hinder their practical deployment, particularly in resource-constrained environments. A primary disadvantage is computational inefficiency during , as these methods defer all formation to query time, requiring repeated calculations across the entire set for each new instance. This results in high , especially with large datasets, where the scales linearly with the number of stored examples. For instance, in k-nearest neighbors, the need to compute similarities for every can make applications infeasible without additional indexing structures. Storage demands represent another significant drawback, as lazy learners must retain the full training in memory to enable local generalizations at prediction time, leading to substantial space requirements that grow directly with dataset size. This becomes problematic for scenarios, where maintaining millions of instances can exceed available resources and complicate deployment. Lazy learning demonstrates robustness to through local averaging in methods like k-NN, which smooths out isolated erroneous examples, but it remains vulnerable to s that can distort local neighborhoods and influence predictions. Techniques such as robust distance weighting or outlier detection can help mitigate this issue. The curse of dimensionality further exacerbates these issues, as distance metrics used for similarity—such as —degrade in high-dimensional spaces, causing most points to appear and diminishing the effectiveness of local approximations. In dimensions exceeding 10 or so, the volume of the space concentrates near boundaries, making nearest neighbor searches unreliable and often leading to poorer performance compared to lower-dimensional cases. Finally, the absence of a compact model in lazy learning precludes insights into global data patterns, as predictions are derived solely from instance-specific interpolations rather than a summarized representation of the target function. This local focus limits interpretability and the ability to extract overarching trends or rules from the data, contrasting with eager learners that produce explicit, analyzable structures.

Algorithms

k-Nearest Neighbors (k-NN)

The k-nearest neighbors (k-NN) algorithm represents the canonical example of a lazy learning method, deferring all computation until prediction time by storing the entire training dataset without building an explicit model. For a given query instance \mathbf{x}, the algorithm identifies the k training instances closest to \mathbf{x} based on a chosen distance metric, such as the defined as d(\mathbf{x}, \mathbf{y}) = \sqrt{\sum_{i=1}^{d} (x_i - y_i)^2}, where d is the dimensionality of the feature space. In tasks, the query is assigned the class label held by the majority of these k neighbors; for , the output is the average of the target values from the k neighbors. This instance-based approach relies on the assumption that similar instances cluster together in the feature space, enabling local without global assumptions about the data distribution. The key hyperparameter k governs the algorithm's bias-variance trade-off, with smaller values (e.g., k=1) producing complex decision boundaries that capture fine-grained patterns but are sensitive to noise, while larger k yields smoother boundaries by averaging over more instances, reducing variance at the cost of potentially underfitting. Optimal k is typically selected via cross-validation to minimize prediction error on held-out data, balancing these effects based on dataset characteristics. To mitigate ties in binary classification, k is often chosen as an odd integer, ensuring a decisive majority vote. A common variant, the weighted k-NN, addresses limitations in uniform neighbor aggregation by assigning weights inversely proportional to distances, such as w_i = 1 / d_i for the i-th neighbor, thereby emphasizing closer instances in the or averaging . This distance-weighted scheme improves performance in scenarios with varying local densities, as demonstrated in early empirical evaluations on pattern tasks. The algorithm's highlights its simplicity, with "training" merely involving storage and prediction entailing computations and aggregation: Training Phase:
Store the training [dataset](/page/Data_set) D = {(x_1, y_1), ..., (x_n, y_n)}
No further [processing](/page/Processing)
Prediction Phase (for ):
For query instance x:
  Compute distances d(x, x_i) for all x_i in D
  Select the k smallest distances and corresponding neighbors N = {x_{i1}, ..., x_{ik}}
  Assign [class](/page/Class) label: argmax_c (sum over i in N of I(y_{i} = c)), where I is the [indicator function](/page/Indicator_function)
  (For [regression](/page/Regression): average y_i over N)
This structure underscores the lazy nature, as all effort occurs at query time. In terms of , k-NN incurs negligible cost during (O(1), just ), but each requires O(n d) time to compute distances across the n instances in d dimensions, followed by O(k log k) for or selection—dominating for large n. This linear scaling with dataset size motivates approximate indexing techniques in practice, though the basic form remains exact and straightforward.

Other Instance-Based Methods

Beyond k-nearest neighbors, several other instance-based methods embody lazy learning principles by deferring until query time, often incorporating specialized mechanisms for , , or reasoning while storing instances as core references. These approaches adapt the core idea of using stored examples for but introduce variations such as prototype refinement or weighted local modeling to handle specific tasks like or continuous function approximation. Case-Based Reasoning (CBR) applies to complex problem-solving domains, such as medical diagnostics or legal advice, by maintaining a case base of past experiences (instances) and cycling through retrieval, reuse, revision, and retention to address new queries. In the retrieval phase, similar cases are fetched using similarity metrics; reuse adapts solutions from those cases to the current problem; revision validates and repairs the adapted solution if needed; and retention incorporates the new solved case back into the base for future use. This four-stage cycle, formalized by Aamodt and Plaza, ensures lazy generalization by relying on instance similarity rather than abstract rules, making CBR particularly effective for domains where experiences are heterogeneous and adaptation is key. Locally Weighted Regression (LWR) extends instance-based principles to tasks by constructing a local weighted by proximity to the query instance, avoiding global fitting and instead performing lazy optimization at prediction time. For a query point, LWR selects relevant training instances and assigns weights that decay with distance, commonly via a Gaussian w_i = \exp\left( -\frac{d_i^2}{2\sigma^2} \right), where d_i is the distance to instance i and \sigma controls the locality. The prediction is then the weighted average of local fits, as surveyed by Atkeson et al., enabling flexible approximation of nonlinear functions without preprocessing the entire dataset. Radial Basis Function (RBF) networks configured in lazy mode treat training instances directly as the centers of radial basis functions, deferring network evaluation and weight computation until a query arrives, thus aligning with instance-based learning paradigms. In this setup, the output is a distance-weighted sum of basis functions centered at stored instances, often using Gaussian kernels similar to LWR, but structured as a single hidden layer for approximation; no eager training of centers or weights occurs, allowing on-demand generalization. This lazy RBF variant, explored by Bermejo and Gamez, bridges memory-based methods and neural architectures by avoiding fixed prototypes and instead querying the instance set dynamically. In contrast to k-NN's static retrieval and voting, these methods like CBR, LWR, and RBF incorporate lazy adaptation—such as case revisions or local parameter estimation—to enhance representation without full eager precomputation, improving efficiency in structured data spaces.

Applications and Extensions

Real-World Applications

Lazy learning algorithms, particularly instance-based methods like k-nearest neighbors (k-NN), have found practical utility in recommendation systems through , where user preferences are matched to similar profiles to suggest items such as movies on platforms like . In these systems, k-NN computes similarity metrics, such as cosine distance, between user-item interaction vectors to generate personalized recommendations without building a global model, enabling efficient handling of sparse data in large-scale environments. In , case-based reasoning (CBR) leverages lazy learning by retrieving and adapting solutions from historical patient cases stored in electronic health records to support clinicians in identifying similar conditions. For instance, CBR systems compare new patient symptoms and test results against a database of prior cases using similarity measures to propose diagnoses, aiding in domains like or infectious disease management where patterns vary widely. Lazy methods contribute to image recognition via in , allowing models to classify new image categories with minimal examples by comparing them to stored prototypes using k-NN in feature space. This approach excels in scenarios requiring adaptation to novel classes, such as in , where nearest-neighbor matching on embedded representations achieves high accuracy without retraining on vast datasets. In , instance-based techniques like k-NN support for prevention by scoring transactions as outliers based on their distance to known legitimate patterns in historical data. Applied to transactions, k-NN identifies unusual spending behaviors, such as atypical merchant locations or amounts, achieving high accuracy in hybrid setups when combined with sampling methods on imbalanced datasets. Robotics employs nearest-neighbor methods for path planning by mapping environmental to previously explored , enabling real-time navigation in dynamic spaces like warehouses or uneven terrains. In sampling-based planners, k-NN accelerates collision-free generation by querying similar poses from a , reducing computation time in high-dimensional spaces. Notable case studies include k-NN in spam filtering, where email content is classified as by proximity to labeled examples in , attaining over 95% accuracy on benchmark corpora like Enron-Spam. Similarly, (LVQ), an instance-based variant, enhances by quantizing acoustic features into prototypes for digit or word classification, improving error rates in isolated word tasks through competitive learning on datasets.

Modern Extensions and Variants

Modern extensions of lazy learning have focused on enhancing scalability for large-scale datasets by incorporating approximate nearest neighbor (ANN) techniques. Traditional k-nearest neighbors (k-NN) suffers from linear query time complexity, but methods like k-d trees partition the data space hierarchically to enable logarithmic-time searches in low dimensions, achieving sublinear performance for exact or approximate queries. Similarly, locality-sensitive hashing (LSH) hashes similar points into the same buckets with high probability, allowing for efficient ANN retrieval in high-dimensional spaces by reducing the search space from O(n) to O(1 + n^{1/c}) for some constant c > 1, as demonstrated in query-aware variants that adapt to query patterns. These approaches address the computational bottlenecks of lazy learners, enabling their use in applications with millions of instances while maintaining high recall rates, often exceeding 90% in benchmarks on datasets like SIFT1M. Integration with has extended lazy learning through metric learning frameworks that learn spaces where distances reflect , deferring to time much like k-NN. In prototypical networks, support set embeddings are averaged into class prototypes, and test points are classified via to these prototypes in the learned space, achieving state-of-the-art few-shot learning performance on Omniglot (98.0% 5-way 1-shot) by leveraging lazy computation of prototypes. Siamese networks complement this by training twin architectures to minimize distance between similar pairs and maximize for dissimilar ones, often using contrastive loss; when augmented with k-NN, they boost nearest neighbor classifiers' accuracy on by refining the metric. These hybrids combine the non-parametric flexibility of lazy methods with deep feature extraction, mitigating the curse of dimensionality in high-dimensional data. Ensemble methods have been adapted to lazy learning to improve robustness and accuracy without eager . Lazy random forests, for instance, defer tree construction until query time, using neighborhood projections to select relevant features dynamically, which enhances performance on high-dimensional data by reducing and achieving up to 15% accuracy gains over standard random forests. Boosting variants for lazy classifiers, such as those based on structures, iteratively weight instances and predictions via bagging-like resampling, yielding superior results on UCI datasets compared to single k-NN (e.g., 5-10% error reduction on and Wine). These ensembles preserve the instance-based nature while leveraging diversity for better generalization. To handle big data streams, streaming lazy learners employ online k-NN with to maintain a fixed-size of representative instances without storing the entire stream. randomly selects and replaces elements with probability inversely proportional to stream position, ensuring unbiased approximation of the data distribution; integrated into incremental k-NN (RW-kNN), it supports concept drift adaptation on electricity datasets, achieving comparable accuracy to batch k-NN (around 85%) with O(1) update time per instance. This enables lazy learning in resource-constrained environments like sensor networks. Post-2015 developments include graph-based lazy methods for semi-supervised learning, where label occurs lazily via graph convolutions that update node representations only during , reducing computational overhead; for example, lazy random walks on graphs preserve structure while enabling semi-supervised on Cora by minimizing steps. Quantum-inspired variants further accelerate computations, such as enhanced quantum k-NN using polar metrics on qubit-encoded features, which outperform classical k-NN on (95% accuracy) with speedup in query time for small datasets, as benchmarked in 2024 studies. These innovations, current as of 2025, hybridize lazy learning with emerging paradigms to tackle evolving challenges in and efficiency.

References

  1. [1]
    Lazy Learning | SpringerLink
    This edited collection describes recent progress on lazy learning, a branch of machine learning concerning algorithms that defer the processing of their inputs.
  2. [2]
    [PDF] Locally Weighted Learning - CMU Robotics Institute
    Oct 12, 1996 · Lazy learning methods defer processing of training data until a query needs to be answered. This usually involves storing the training data in ...
  3. [3]
    [PDF] CS 536: Machine Learning Instance-based learning Lazy and Eager ...
    Lazy Learning. IBL Advantages: • Learning is trivial. • Works. • Noise Resistant. • Rich Representation, Arbitrary Decision Surfaces. • Easy to understand.
  4. [4]
    [PDF] Instance Based Learning
    Mar 4, 2023 · During training phase, lazy algorithms have fewer computational costs than eager algorithms. During testing phase, lazy algorithms have greater ...
  5. [5]
    (PDF) Instance-Based Learning Algorithms - ResearchGate
    Instance-based learning algorithms do not maintain a set of abstractions ... Kibler, Aha,. and. Albert (1989). This lemma estab-. lished. the. sample ...
  6. [6]
    What is k-Nearest Neighbor (kNN)? - Elastic
    Brief history of the kNN algorithm. kNN was first developed by Evelyn Fix and Joseph Hodges in 1951 in the context of research performed for the US military1.
  7. [7]
    Nearest neighbor pattern classification | IEEE Journals & Magazine
    Abstract: The nearest neighbor decision rule assigns to an unclassified sample point the classification of the nearest of a set of previously classified points.
  8. [8]
    [PDF] Nearest neighbor pattern classification - Semantic Scholar
    The nearest neighbor decision rule assigns to an unclassified sample point the classification of the nearest of a set of previously classified points, ...
  9. [9]
    Instance-based learning algorithms | Machine Learning
    Aha, D.W., Kibler, D. & Albert, M.K. Instance-based learning algorithms. Mach Learn 6, 37–66 (1991). https://doi.org/10.1007/BF00153759. Download citation.
  10. [10]
    [PDF] Instance-Based Learning Algorithms - Semantic Scholar
    Instance-Based Learning Algorithms · D. Aha, D. Kibler, M. K. Albert · Published in Machine-mediated learning 3 January 1991 · Computer Science.
  11. [11]
    Instance-based learning algorithms
    We introduced the coverage lemma in Kibler, Aha, and Albert (1989). This lemma estab- lished the sample sizes that IBL algorithms require to perform well. We ...
  12. [12]
    [PDF] Scikit-learn: Machine Learning in Python
    The k-nearest neighbors classifier implementation constructs a ball tree (Omohundro, 1989) ... scikit-learn's k-means algorithm is implemented in pure Python.<|control11|><|separator|>
  13. [13]
    1.6. Nearest Neighbors — scikit-learn 1.7.2 documentation
    scikit-learn implements two different nearest neighbors classifiers: KNeighborsClassifier implements learning based on the k nearest neighbors of each query ...KNeighborsClassifier · Sklearn.neighbors · Nearest Neighbors Classification
  14. [14]
    [PDF] chapter instance-based learning
    Instance-based methods are sometimes referred to as. "lazy" learning methods because they delay processing until a new instance must be classified. A key ...
  15. [15]
    [PDF] Chapter 6 - Instance-Based Learning: A Survey - Charu Aggarwal
    Instance-based learning is also sometimes referred to as lazy learning, since most of the computational work is not done upfront, and one waits to obtain the.
  16. [16]
    [PDF] Scalable Nearest Neighbor Search based on kNN Graph - arXiv
    Feb 3, 2017 · In this paper, a scalable solution based on hill-climbing strategy with the support of k-nearest neighbor graph (kNN) is presented. Two major ...
  17. [17]
    [PDF] Instance Based Learning - The University of Texas at Dallas
    • Disadvantages. – distance function must be carefully chosen. – Irrelevant or ... • Instance-based learning. – non-parametric. – trade decreased learning ...
  18. [18]
    Instance-Based Learning Algorithms
    Although they summarize data, they do not attempt to maximize classification accuracy on novel instances. This ignores real- world problems such as noise.Missing: lack global
  19. [19]
    [PDF] Reduction Techniques for Instance-Based Learning Algorithms
    Instance-based learning (IBL) algorithms (Aha, Kibler & Albert, 1991; Aha, 1992) are a subset of exemplar-based learning algorithms that use original instances ...
  20. [20]
    Lecture 2: k-nearest neighbors / Curse of Dimensionality
    The curse of dimensionality has different effects on distances between two points and distances between points and hyperplanes. An animation illustrating the ...
  21. [21]
    The Distance-Weighted k-Nearest-Neighbor Rule - IEEE Xplore
    The Distance-Weighted k-Nearest-Neighbor Rule weights evidence of nearby sample observations more heavily than those further away, using a neighbor weighting ...
  22. [22]
    Improved versions of learning vector quantization | Semantic Scholar
    Improved versions of learning vector quantization · T. Kohonen · Published in IJCNN International Joint… 17 June 1990 · Computer Science · 1990 IJCNN International ...Missing: seminal | Show results with:seminal
  23. [23]
    A review of learning vector quantization classifiers - ResearchGate
    Aug 6, 2025 · In this work, we present a review of the state of the art of learning vector quantization (LVQ) classifiers. A taxonomy is proposed which integrates the most ...
  24. [24]
    Case-Based Reasoning: Foundational Issues, Methodological ...
    Aug 10, 2025 · This paper gives an overview of the foundational issues related to case-based reasoning, describes some of the leading methodological approaches within the ...
  25. [25]
    Learning radial basis neural networks in a lazy way: A comparative ...
    In this paper, a lazy strategy is applied to train radial basis neural networks. That strategy incorporates a dynamic selection of patterns, and that selection ...
  26. [26]
    [PDF] Use of KNN for the Netflix Prize )K - CS229
    This paper analyzes the performance of various KNNs techniques as applied to the netflix collaborative filtering problem. 1. Introduction. In the Netflix ...
  27. [27]
    [PDF] Collaborative Filtering Recommender Systems - Michael Ekstrand
    Abstract. Recommender systems are an important part of the information and e-commerce ecosystem. They represent a powerful method for enabling.<|separator|>
  28. [28]
    (PDF) Case-Based Reasoning in the Health Sciences - ResearchGate
    Aug 9, 2025 · Case-based reasoning (CBR) is an Artificial Intelligence (AI) approach with broad applicability to building intelligent systems in health sciences domains.
  29. [29]
    [PDF] Using Case-Based Reasoning for Diagnosis in Medical Field
    ABSTRACT. The Case-Based Reasoning (CBR)is a method of problem solving that uses prior experiences (cases) to understand and solve new problems.
  30. [30]
    [PDF] Few-Shot Image Recognition With Knowledge Transfer
    The goal of few-shot learning is to learn an image classification model by using the training set and the support set that can accurately classify images in the.
  31. [31]
    [PDF] K-NN Embeded Space Conditioning for Enhanced Few-Shot Object ...
    Few-shot learning has attracted significant scientific in- terest in the past decade due to its applicability to visual tasks with a natural long-tailed ...<|separator|>
  32. [32]
    Financial Fraud: A Review of Anomaly Detection Techniques and ...
    May 1, 2022 · This survey aims to investigate and present a thorough review of the most popular and effective anomaly detection techniques applied to detect financial fraud.
  33. [33]
    A Combination of Deep Neural Networks and K-Nearest Neighbors ...
    May 27, 2022 · This paper uses under-sampling, KNN, and Deep Neural Networks (DNN) for credit card fraud detection, achieving 98.12% accuracy with the DNN ...
  34. [34]
    [PDF] Efficient Nearest Neighbor Searching for Motion Planning
    We present and implement an efficient algorithm for per- forming nearest-neighbor queries in topological spaces that usually arise in the context of motion ...
  35. [35]
    Path planning techniques for mobile robots: Review and prospect
    Oct 1, 2023 · In this paper, a systematic review of mobile robot path planning techniques is presented. Firstly, path planning is classified into global path ...
  36. [36]
    Spam Filtering with KNN: Investigation of the Effect of k Value on ...
    Abstract: In this study, it is aimed to filter spam e-mails by using machine learning and text mining techniques. K-Nearest Neighbor (KNN) algorithm which ...
  37. [37]
    DP matching-based spoken digit recognition using LVQ | IEEE ...
    This paper deals with a method of speech recognition using dynamic programming (DP) matching and LVQ techniques. It is well known that learning vector ...
  38. [38]
    [PDF] Scalable Nearest Neighbor Algorithms for High Dimensional Data
    In this paper we describe a modified k-means tree algorithm that we have found to give the best results for some data sets, while randomized k-d trees are best ...
  39. [39]
    [PDF] Query-Aware Locality-Sensitive Hashing for Approximate Nearest ...
    Locality-Sensitive Hashing (LSH) and its variants are the well-known indexing schemes for the c-Approximate Nearest. Neighbor (c-ANN) search problem in high- ...
  40. [40]
  41. [41]
    [PDF] Augmenting Nearest Neighbor-Based Algorithms With Siamese ...
    By learning a dis- tance metric for a specific dataset using Siamese layers in convolutional neural networks (CNNs), we can boost the performance of nearest ...Missing: lazy | Show results with:lazy<|separator|>
  42. [42]
    Improving random forests by neighborhood projection for effective ...
    In this article, we propose a lazy version of the traditional random forest (RF) classifier (called LazyNN_RF), specially designed for highly dimensional ...
  43. [43]
    [PDF] Ensemble Techniques for Lazy Classification Based on Pattern ...
    In [6] [5] a model of lazy (query-based) classification using pattern structures was proposed. The model shows quite good performance, which, how- ever, is ...
  44. [44]
    Incremental k-Nearest Neighbors Using Reservoir Sampling for ...
    In this paper, we propose, RW-kNN, a k-Nearest Neighbors (kNN) algorithm that employs a practical way to store information about past instances.
  45. [45]
    Benchmarking quantum versions of the kNN algorithm with a metric ...
    Jul 19, 2024 · This research focuses on comparing the performance of the two quantum kNN algorithms using the original Hamming distance with qubit-encoded features and our ...