isotree - Isolation-Based Outlier Detection

Fast and multi-threaded implementation of isolation forest (Liu, Ting, Zhou (2008) <doi:10.1109/ICDM.2008.17>), extended isolation forest (Hariri, Kind, Brunner (2018) <arXiv:1811.02141>), SCiForest (Liu, Ting, Zhou (2010) <doi:10.1007/978-3-642-15883-4_18>), fair-cut forest (Cortes (2021) <arXiv:2110:13402>), robust random-cut forest (Guha, Mishra, Roy, Schrijvers (2016) <http://proceedings.mlr.press/v48/guha16.html>), and customizable variations of them, for isolation-based outlier detection, clustered outlier detection, distance or similarity approximation (Cortes (2019) <arXiv:1910.12362>), isolation kernel calculation (Ting, Zhu, Zhou (2018) <doi:10.1145/3219819.3219990>), and imputation of missing values (Cortes (2019) <arXiv:1911.06646>), based on random or guided decision tree splitting, and providing different metrics for scoring anomalies based on isolation depth or density (Cortes (2021) <arXiv:2111.11639>). Provides simple heuristics for fitting the model to categorical columns and handling missing data, and offers options for varying between random and guided splits, and for using different splitting criteria.

Last updated 2 months ago

anomaly-detectionimputationisolation-forestoutlier-detection

10.13 score 188 stars 4 packages 92 scripts 1.1k downloads

MatrixExtra - Extra Methods for Sparse Matrices

Extends sparse matrix and vector classes from the 'Matrix' package by providing: (a) Methods and operators that work natively on CSR formats (compressed sparse row, a.k.a. 'RsparseMatrix') such as slicing/sub-setting, assignment, rbind(), mathematical operators for CSR and COO such as addition ("+") or sqrt(), and methods such as diag(); (b) Multi-threaded matrix multiplication and cross-product for many <sparse, dense> types, including the 'float32' type from 'float'; (c) Coercion methods between pairs of classes which are not present in 'Matrix', such as 'dgCMatrix' -> 'ngRMatrix', as well as convenience conversion functions; (d) Utility functions for sparse matrices such as sorting the indices or removing zero-valued entries; (e) Fast transposes that work by outputting in the opposite storage format; (f) Faster replacements for many 'Matrix' methods for all sparse types, such as slicing and elementwise multiplication. (g) Convenience functions for sparse objects, such as 'mapSparse' or a shorter 'show' method.

Last updated 5 months ago

csrsparse-matrix

9.06 score 20 stars 25 packages 84 scripts 4.5k downloads

outliertree - Explainable Outlier Detection Through Decision Tree Conditioning

Outlier detection method that flags suspicious values within observations, constrasting them against the normal values in a user-readable format, potentially describing conditions within the data that make a given outlier more rare. Full procedure is described in Cortes (2020) <doi:10.48550/arXiv.2001.00636>. Loosely based on the 'GritBot' <https://www.rulequest.com/gritbot-info.html> software.

Last updated 3 months ago

anomaly-detectionoutlier-detection

7.63 score 56 stars 2 packages 21 scripts 509 downloads

cmfrec - Collective Matrix Factorization for Recommender Systems

Collective matrix factorization (a.k.a. multi-view or multi-way factorization, Singh, Gordon, (2008) <doi:10.1145/1401890.1401969>) tries to approximate a (potentially very sparse or having many missing values) matrix 'X' as the product of two low-dimensional matrices, optionally aided with secondary information matrices about rows and/or columns of 'X', which are also factorized using the same latent components. The intended usage is for recommender systems, dimensionality reduction, and missing value imputation. Implements extensions of the original model (Cortes, (2018) <arXiv:1809.00366>) and can produce different factorizations such as the weighted 'implicit-feedback' model (Hu, Koren, Volinsky, (2008) <doi:10.1109/ICDM.2008.22>), the 'weighted-lambda-regularization' model, (Zhou, Wilkinson, Schreiber, Pan, (2008) <doi:10.1007/978-3-540-68880-8_32>), or the enhanced model with 'implicit features' (Rendle, Zhang, Koren, (2019) <arXiv:1905.01395>), with or without side information. Can use gradient-based procedures or alternating-least squares procedures (Koren, Bell, Volinsky, (2009) <doi:10.1109/MC.2009.263>), with either a Cholesky solver, a faster conjugate gradient solver (Takacs, Pilaszy, Tikk, (2011) <doi:10.1145/2043932.2043987>), or a non-negative coordinate descent solver (Franc, Hlavac, Navara, (2005) <doi:10.1007/11556121_50>), providing efficient methods for sparse and dense data, and mixtures thereof. Supports L1 and L2 regularization in the main models, offers alternative most-popular and content-based models, and implements functionality for cold-start recommendations and imputation of 2D data.

Last updated 2 months ago

cold-startcollaborative-filteringcollective-matrix-factorization

6.91 score 117 stars 20 scripts 481 downloads

costsensitive - Cost-Sensitive Multi-Class Classification

Reduction-based techniques for cost-sensitive multi-class classification, in which each observation has a different cost for classifying it into one class, and the goal is to predict the class with the minimum expected cost for each new observation. Implements Weighted All-Pairs (Beygelzimer, Langford, & Zadrozny (2008) <doi:10.1007/978-0-387-79361-0_1>), Weighted One-Vs-Rest (Beygelzimer,Dani, Hayes, Langford, Zadrozny, (2005) <https://dl.acm.org/citation.cfm?id=1102358>) and Regression One-Vs-Rest. Works with arbitrary classifiers taking observation weights, or with regressors. Also implements cost-proportionate rejection sampling for working with classifiers that don't accept observation weights.

Last updated 5 months ago

cost-sensitive-classificationmulti-label-classification

5.30 score 47 stars 28 scripts 167 downloads

poismf - Factorization of Sparse Counts Matrices Through Poisson Likelihood

Creates a non-negative low-rank approximate factorization of a sparse counts matrix by maximizing Poisson likelihood with L1/L2 regularization (e.g. for implicit-feedback recommender systems or bag-of-words-based topic modeling) (Cortes, (2018) <arXiv:1811.01908>), which usually leads to very sparse user and item factors (over 90% zero-valued). Similar to hierarchical Poisson factorization (HPF), but follows an optimization-based approach with regularization instead of a hierarchical prior, and is fit through gradient-based methods instead of variational inference.

Last updated 5 months ago

implicit-feedbackpoisson-factorization

4.65 score 45 stars 9 scripts 344 downloads

readsparse - Read and Write Sparse Matrices in 'SVMLight' and 'LibSVM' Formats

Read and write labelled sparse matrices in text format as used by software such as 'SVMLight', 'LibSVM', 'ThunderSVM', 'LibFM', 'xLearn', 'XGBoost', 'LightGBM', and others. Supports labelled data for regression, classification (binary, multi-class, multi-label), and ranking (with 'qid' field), and can handle header metadata and comments in files.

Last updated 5 months ago

libsvmsparse-matricessvmlight

4.32 score 8 stars 13 scripts 284 downloads

nonneg.cg - Non-Negative Conjugate-Gradient Minimizer

Minimize a differentiable function subject to all the variables being non-negative (i.e. >= 0), using a Conjugate-Gradient algorithm based on a modified Polak-Ribiere-Polyak formula as described in (Li, (2013) <https://www.hindawi.com/journals/jam/2013/986317/abs/>).

Last updated 5 years ago

conjugate-gradientminimizeoptimization

3.00 score 2 stars 1 scripts 218 downloads

StratifiedRF - Builds Trees by Sampling Variables in Groups

Random Forest-like tree ensemble that works with groups of predictor variables. When building a tree, a number of variables is taken randomly from each group separately, thus ensuring that it considers variables from each group for the splits. Useful when rows contain information about different things (e.g. user information and product information) and it's not sensible to make a prediction with information from only one group of variables, or when there are far more variables from one group than the other and it's desired to have groups appear evenly on trees. Trees are grown using the C5.0 algorithm rather than the usual CART algorithm. Supports parallelization (multithreaded), missing values in predictors, and categorical variables (without doing One-Hot encoding in the processing). Can also be used to create a regular (non-stratified) Random Forest-like model, but made up of C5.0 trees and with some additional control options. As it's built with C5.0 trees, it works only for classification (not for regression).

Last updated 7 years ago

3.00 score 2 stars 6 scripts 172 downloads