-
During regression and classification training prefer smaller models if performance is similar (See {ml-pull}1516[#1516].)
-
Add a response mechanism for commands sent to the native controller. (See {ml-pull}1520[#1520], {es-pull}63542[#63542], issue: {es-issue}62823[#62823].)
-
Speed up anomaly detection for seasonal data. This is particularly effective for jobs using longer bucket lengths. (See {ml-pull}1549[#1549].)
-
Fix an edge case which could cause typical and model plot bounds to blow up to around max double. (See {ml-pull}1551[#1551].)
-
Estimate upper bound of potential gains before splitting a decision tree node to avoid unnecessary computation. (See {ml-pull}1537[#1537].)
-
Improvements to time series modeling particularly in relation to adaption to change. (See {ml-pull})1614[#1614].)
-
Warn and error log throttling. (See {ml-pull}1615[#1615].)
-
Soften the effect of fluctuations in anomaly detection job memory usage on node assignment and add
assignment_memory_basistomodel_size_stats. (See {ml-pull}1623[#1623], {es-pull}65561[#65561], issue: {es-issue}63163[#63163].)
-
Fix potential cause for log errors from CXMeansOnline1d. (See {ml-pull}1586[#1586].)
-
Fix scaling of some hyperparameter for Bayesian optimization. (See {ml-pull}1612[#1612].)
-
Fix missing state in persist and restore for anomaly detection. This caused suboptimal modelling after a job was closed and reopened or failed over to a different node. (See {ml-pull}1668[#1668].)
-
Calculate total feature importance to store with model metadata. (See {ml-pull}1387[#1387].)
-
Change outlier detection feature_influence format to array with nested objects. (See {ml-pull}1475[#1475], {es-pull}62068[#62068].)
-
Add timeouts to named pipe connections. (See {ml-pull}1514[#1514], {es-pull}62993[#62993], issue: {ml-issue}1504[#1504].)
-
Fix progress on resume after final training has completed for classification and regression. We previously showed progress stuck at zero for final training. (See {ml-pull}1443[#1443].)
-
Avoid potential "Failed to compute quantile" and "No values added to quantile sketch" log errors training regression and classification models if there are features with mostly missing values. (See {ml-pull}1500[#1500].)
-
Correct the anomaly detection job model state
min_version. (See {ml-pull}1546[#1546].)
-
Report significant changes to anomaly detection models in annotations of the results. (See {ml-pull}1247[#1247], {pull}56342[#56342], {pull}56417[#56417], {pull}57144[#57144], {pull}57278[#57278], {pull}57539[#57539].)
-
Add support for larger forecasts in memory via max_model_memory setting. (See {ml-pull}1238[#1238] and {pull}57254[#57254].)
-
Don’t lose precision when saving model state. (See {ml-pull}1274[#1274].)
-
Parallelize the feature importance calculation for classification and regression over trees. (See {ml-pull}1277[#1277].)
-
Add an option to do categorization independently for each partition. (See {ml-pull}1293[#1293], {ml-pull}1318[#1318], {ml-pull}1356[#1356] and {pull}57683[#57683].)
-
Memory usage is reported during job initialization. (See {ml-pull}1294[#1294].)
-
More realistic memory estimation for classification and regression means that these analyses will require lower memory limits than before (See {ml-pull}1298[#1298].)
-
Checkpoint state to allow efficient failover during coarse parameter search for classification and regression. (See {ml-pull}1300[#1300].)
-
Improve data access patterns to speed up classification and regression. (See {ml-pull}1312[#1312].)
-
Performance improvements for classification and regression, particularly running multithreaded. (See {ml-pull}1317[#1317].)
-
Improve runtime and memory usage training deep trees for classification and regression. (See {ml-pull}1340[#1340].)
-
Improvement in handling large inference model definitions. (See {ml-pull}1349[#1349].)
-
Add a peak_model_bytes field to model_size_stats. (See {ml-pull}1389[#1389].)
-
Fix numerical issues leading to blow up of the model plot bounds. (See {ml-pull}1268[#1268].)
-
Fix causes for inverted forecast confidence interval bounds. (See {ml-pull}1369[#1369], issue: {ml-issue}1357[#1357].)
-
Restrict growth of max matching string length for categories. (See {ml-pull}1406[#1406].)
-
Better interrupt handling during named pipe connection. (See {ml-pull}1311[#1311].)
-
Trap potential cause of SIGFPE. (See {ml-pull}1351[#1351], issue: {ml-issue}1348[#1348].)
-
Correct inference model definition for MSLE regression models. (See {ml-pull}1375[#1375].)
-
Fix cause of SIGSEGV of classification and regression. (See {ml-pull}1379[#1379].)
-
Fix restoration of change detectors after seasonality change. (See {ml-pull}1391[#1391].)
-
Fix potential SIGSEGV when forecasting. (See {ml-pull}1402[#1402], issue: {ml-issue}1401[#1401].)
-
Speed up anomaly detection for the lat_long function. (See {ml-pull}1102[#1102].)
-
Reduce CPU scheduling priority of native analysis processes to favor the ES JVM when CPU is constrained. This change is only implemented for Linux and macOS, not for Windows. (See {ml-pull}1109[#1109].)
-
Take
training_percentinto account when estimating memory usage for classification and regression. (See {ml-pull}1111[#1111].) -
Support maximize minimum recall when assigning class labels for multiclass classification. (See {ml-pull}1113[#1113].)
-
Improve robustness of anomaly detection to bad input data. (See {ml-pull}1114[#1114].)
-
Adds new
num_matchesandpreferred_to_categoriesfields to category output. (See {ml-pull}1062[#1062].) -
Adds mean squared logarithmic error (MSLE) for regression. (See {ml-pull}1101[#1101].)
-
Adds pseudo-Huber loss for regression. (See {ml-pull}1168[#1168].)
-
Reduce peak memory usage and memory estimates for classification and regression. (See {ml-pull}1125[#1125].)
-
Reduce variability of classification and regression results across our target operating systems. (See {ml-pull}1127[#1127].)
-
Switched data frame analytics model memory estimates from kilobytes to megabytes. (See {ml-pull}1126[#1126], issue: {issue}54506[#54506].)
-
Added a {ml} native code build for Linux on AArch64. (See {ml-pull}1132[#1132] and {ml-pull}1135[#1135].)
-
Improve data frame analysis runtime by optimising memory alignment for intrinsic operations. (See {ml-pull}1142[#1142].)
-
Fix spurious anomalies for count and sum functions after no data are received for long periods of time. (See {ml-pull}1158[#1158].)
-
Improve false positive rates from periodicity test for time series anomaly detection. (See {ml-pull}1177[#1177].)
-
Break progress reporting of data frame analyses into multiple phases. (See {ml-pull}1179[#1179].)
-
Really centre the data before training for classification and regression begins. This means we can choose more optimal smoothing bias and should reduce the number of trees. (See {ml-pull}1192[#1192].)
-
Trap and fail if insufficient features are supplied to data frame analyses. This caused classification and regression getting stuck at zero progress analyzing. (See {ml-pull}1160[#1160], issue: {issue}55593[#55593].)
-
Make categorization respect the
model_memory_limit. (See {ml-pull}1167[#1167], issue: {ml-issue}1130[#1130].) -
Respect user overrides for
max_treesfor classification and regression. (See {ml-pull}1185[#1185].) -
Reset memory status from
soft_limittookwhen pruning is no longer required. (See {ml-pull}1193[#1193], issue: {ml-issue}1131[#1131].) -
Fix restore from training state for classification and regression. (See {ml-pull}1197[#1197].)
-
Improve the initialization of seasonal components for anomaly detection. (See {ml-pull}1201[#1201], issue: {ml-issue}#1178[#1178].)
-
Fixed background persistence of categorizer state (See {ml-pull}1137[#1137], issue: {ml-issue}1136[#1136].)
-
Fix classification job failures when number of classes in configuration differs from the number of classes present in the training data. (See {ml-pull}1144[#1144].)
-
Fix underlying cause for "Failed to calculate splitting significance" log errors. (See {ml-pull}1157[#1157].)
-
Fix possible root cause for "Bad variance scale nan" log errors. (See {ml-pull}1225[#1225].)
-
Change data frame analytics instrumentation timestamp resolution to milliseconds. (See {ml-pull}1237[#1237].)
-
Fix "autodetect process stopped unexpectedly: Fatal error: 'terminate called after throwing an instance of 'std::bad_function_call'". (See {ml-pull}1246[#1246], issue: {ml-issue}1245[#1245].)
-
Add instrumentation to report statistics related to data frame analytics jobs, i.e. progress, memory usage, etc. (See {ml-pull}906[#906].)
-
Multiclass classification. (See {ml-pull}1037[#1037].)
-
Improve computational performance of the feature importance computation. (See {ml-pull}1005[1005].)
-
Improve initialization of learn rate for better and more stable results in regression and classification. (See {ml-pull}948[#948].)
-
Add number of processed training samples to the definition of decision tree nodes. (See {ml-pull}991[#991].)
-
Add new model_size_stats fields to instrument categorization. (See {ml-pull}948[#948] and {pull}51879[#51879], issue: {issue}50794[#50749].)
-
Improve upfront memory estimation for all data frame analyses, which were higher than necessary. This will improve the allocation of data frame analyses to cluster nodes. (See {ml-pull}1003[#1003].)
-
Upgrade the compiler used on Linux from gcc 7.3 to gcc 7.5, and the binutils used in the build from version 2.20 to 2.34. (See {ml-pull}1013[#1013].)
-
Add instrumentation of the peak memory consumption for data frame analytics jobs. (See {ml-pull}1022[#1022].)
-
Remove all memory overheads for computing tree SHAP values. (See {ml-pull}1023[#1023].)
-
Distinguish between empty and missing categorical fields in classification and regression model training. (See {ml-pull}1034[#1034].)
-
Add instrumentation information for supervised learning data frame analytics jobs. (See {ml-pull}1031[#1031].)
-
Add instrumentation information for outlier detection data frame analytics jobs.
-
Write out feature importance for multi-class models. (See {ml-pull}1071[#1071])
-
Enable system call filtering to the native process used with data frame analytics. (See {ml-pull}1098[#1098])
-
Use largest ordered subset of categorization tokens for category reverse search regex. (See {ml-pull}970[#970], issue: {ml-issue}949[#949].)
-
Account for the data frame’s memory when estimating the peak memory used by classification and regression model training. (See {ml-pull}996[#996].)
-
Rename classification and regression parameter maximum_number_trees to max_trees. (See {ml-pull}1047[#1047].)
-
Add feature importance values to classification and regression results (using tree SHapley Additive exPlanation, or SHAP). (See {ml-pull}857[#857].)
-
Improve performance of boosted tree training for both classification and regression. (See {ml-pull}775[#775].)
-
Reduce the peak memory used by boosted tree training and fix an overcounting bug estimating maximum memory usage. (See {ml-pull}781[#781].)
-
Stratified fractional cross validation for regression. (See {ml-pull}784[#784].)
-
Added
geo_pointsupported output forlat_longfunction records. (See {ml-pull}809[#809] and {pull}47050[#47050].) -
Use a random bag of the data to compute the loss function derivatives for each new tree which is trained for both regression and classification. (See {ml-pull}811[#811].)
-
Emit
prediction_probabilityfield alongside prediction field in ml results. (See {ml-pull}818[#818].) -
Reduce memory usage of {ml} native processes on Windows. (See {ml-pull}844[#844].)
-
Reduce runtime of classification and regression. (See {ml-pull}863[#863].)
-
Stop early training a classification and regression forest when the validation error is no longer decreasing. (See {ml-pull}875[#875].)
-
Emit
prediction_field_namein ml results using the type provided asprediction_field_typeparameter. (See {ml-pull}877[#877].) -
Improve performance updating quantile estimates. (See {ml-pull}881[#881].)
-
Migrate to use Bayesian Optimisation for initial hyperparameter value line searches and stop early if the expected improvement is too small. (See {ml-pull}903[#903].)
-
Stop cross-validation early if the predicted test loss has a small chance of being smaller than for the best parameter values found so far. (See {ml-pull}915[#915].)
-
Optimize decision threshold for classification to maximize minimum class recall. (See {ml-pull}926[#926].)
-
Include categorization memory usage in the
model_bytesfield inmodel_size_stats, so that it is taken into account in node assignment decisions. (See {ml-pull}927[#927], issue: {ml-issue}724[#724].)
-
Fixes potential memory corruption when determining seasonality. (See {ml-pull}852[#852].)
-
Prevent prediction_field_name clashing with other fields in ml results. (See {ml-pull}861[#861].)
-
Include out-of-order as well as in-order terms in categorization reverse searches. (See {ml-pull}950[#950], issue: {ml-issue}949[#949].)
-
Improve performance and concurrency training boosted tree regression models. For large data sets this change was observed to give a 10% to 20% decrease in train time. (See {ml-pull}622[#622].)
-
Upgrade Boost libraries to version 1.71. (See {ml-pull}638[#638].)
-
Improve initialisation of boosted tree training. This generally enables us to find lower loss models faster. (See {ml-pull}686[#686].)
-
Include a smooth tree depth based penalty to regularized objective function for boosted tree training. Hard depth based regularization is often the strategy of choice to prevent over fitting for XGBoost. By smoothing we can make better tradeoffs. Also, the parameters of the penalty function are mode suited to optimising with our Bayesian optimisation based hyperparameter search. (See {ml-pull}698[#698].)
-
Binomial logistic regression targeting cross entropy. (See {ml-pull}713[#713].)
-
Improvements to count and sum anomaly detection for sparse data. This primarily aims to improve handling of data which are predictably present: detecting when they are unexpectedly missing. (See {ml-pull}721[#721].)
-
Trap numeric errors causing bad hyperparameter search initialisation and repeated errors to be logged during boosted tree training. (See {ml-pull}732[#732].)
-
The {ml} native processes are now arranged in a .app directory structure on macOS, to allow for notarization on macOS Catalina. (See {ml-pull}593[#593].)
-
Only trap the case that more rows are supplied to outlier detection than expected. Previously, if rows were excluded from the data frame after supplying the row count in the configuration then we detected the inconsistency and failed outlier detection. However, this legitimately happens in case where the field values are non-numeric or array valued. (See {ml-pull}569[#569].)
-
Upgrade to a newer version of the Apache Portable Runtime library. (See {ml-pull}495[#495].)
-
Improve stability of modelling around change points. (See {ml-pull}496[#496].)
-
Remove hard limit for maximum forecast interval and limit based on the time interval of data added to the model. (See {ml-pull}214[#214].)
-
Use hardened compiler options to build 3rd party libraries. (See {ml-pull}453[#453].)
-
Only select more complex trend models for forecasting if there is evidence that they are needed. (See {ml-pull}463[#463].)
-
Improve residual model selection. (See {ml-pull}468[#468].)
-
Stop linking to libcrypt on Linux. (See {ml-pull}480[#480].)
-
Improvements to hard_limit audit message. (See {ml-pull}486[#486].)