v1.4.0: High-Performance Engine, Feature Importances, RFE & DocC

🚀 SwiftSci v1.4.0 Release Notes

📦 Key Additions & Features

• SystemsCSVParser (SwiftDataFrame): High-performance zero-copy memory-mapped RFC 4180 DFA byte parser, speeding up CSV ingestion by ~10× on large datasets.
• Vectorized Byte Parsers (SwiftDataFrame): Zero-allocation ASCII parsers (VectorizedByteParsers) for fast Double, Int, and string decoding from unmanaged byte buffers.
• vDSP Reductions (SwiftDataFrame): Accelerated TypedColumn<Double> mean(), variance(), and stdDev() using Apple Accelerate vDSP reductions.
• DataFrame Index Filtering & Argsort (SwiftDataFrame): Added filterRows(by:) index-based boolean mask filtering and argsort() column index sorting.
• Recursive Feature Elimination RFE (SwiftPreprocessing): Added RecursiveFeatureElimination (RFE) for iterative feature selection based on estimator feature importances.
• Feature Importances & Model Persistence (SwiftML): Added Gini featureImportances property to DecisionTreeClassifier, DecisionTreeRegressor, RandomForestClassifier, RandomForestRegressor, and added Codable serialization (save(to:) / load(from:)) across classical estimators.
• NLP Text Tokenizers & Feature Extraction (SwiftNLP): Added NGramTokenizer (word & char n-grams) and HashingVectorizer (MurmurHash3 memory-bounded token hashing).
• Time Series Windowing (SwiftForecast): Added ExpandingWindow feature transformer for cumulative time-series feature generation.
• Swift DocC Integration & Multi-Module Site: Integrated swift-docc-plugin (v1.5.0) and deployed unified 10-module static web documentation to GitHub Pages.

🛠️ Fixes & Concurrency Improvements

• GradientBoostedTrees Bounds Check (SwiftML): Enforced bounds checking in GradientBoostedTreesRegressor.
• Seeded Random Number Generators (SwiftPreprocessing): Fixed deterministic seeding in SMOTE and RandomUndersampler using LCG.
• SelectKBest ANOVA Scoring (SwiftPreprocessing): Fixed ANOVA F-statistic calculation in SelectKBest and added explicit SwiftStats dependency to SwiftPreprocessing.
• CalibratedClassifier Concurrency (SwiftML): Converted CalibratedClassifier to actor to guarantee thread safety under Swift 6 strict concurrency rules.
• Wired Memory Scaling (SwiftPreprocessing): Dynamically scaled WiredMemoryManager default limits using ProcessInfo.processInfo.activeProcessorCount.

📊 Benchmark Performance Summary (v1.4.0 vs Python)

Benchmark Test	Swift (ms)	Python (ms)	Speedup	Winner
ARIMA(1,1,1) Fit (50k pts)	2.228 ms	206.414 ms	92.66x	🟢 Swift
ARIMA(1,1,1) Forecast (horizon=24)	2.297 ms	209.709 ms	91.28x	🟢 Swift
Holt-Winters Fit (50k pts, period=12)	6.823 ms	138.947 ms	20.37x	🟢 Swift
Random Forest Fit (1k×4, 50 trees)	4.580 ms	24.217 ms	5.29x	🟢 Swift
KernelSHAP Explain (5 feats, 100 coalitions)	0.148 ms	0.452 ms	3.05x	🟢 Swift
Kalman Filter 1D (10k obs)	56.508 ms	84.692 ms	1.50x	🟢 Swift
Mean (vDSP, 1M elements)	0.083 ms	0.119 ms	1.44x	🟢 Swift

---

🌐 Live Documentation: https://nodibell.github.io/SwiftSci/

Full Changelog: v1.3.0...v1.4.0

v1.3.0 — Sklearn Parity Roadmap Release

SwiftSci v1.3.0 — Sklearn Parity Roadmap Release 🚀

SwiftSci v1.3.0 achieves complete functional parity with Core Scikit-Learn patterns, introducing estimator pipelines, subset column routing, randomized hyperparameter search, outlier detection, imbalanced dataset sampling, probability calibration, and time-series feature transformers.

🌟 What's New

🔗 Estimator Pipelines & Subset Routing (SwiftML & SwiftPreprocessing)

• ClassificationPipeline & RegressionPipeline: Chain zero or more PreprocessingTransformer steps directly into a final estimator for leak-free model training and evaluation.
• ColumnTransformer: Route specific subsets of DataFrame columns to independent preprocessing transformers and concatenate outputs seamlessly.

🌲 Outlier Detection (SwiftCluster)

• IsolationForest: Random partition isolation trees for anomaly detection and score ranking.
• LocalOutlierFactor (LOF): kNN local reachability density ratio anomaly scoring.

⚖️ Imbalanced Learning (SwiftPreprocessing)

• SMOTE: Synthetic Minority Over-sampling Technique using k-nearest neighbor interpolation.
• RandomUndersampler: Random majority class sub-sampling dataset balancer.

🎲 Hyperparameter Tuning & Probability Calibration (SwiftOptimize & SwiftML)

• RandomizedSearchCV: Parallel randomized hyperparameter search integrated with KFold cross-validation.
• CalibratedClassifier: Platt scaling logistic calibration wrapper over raw decision boundaries.

📈 Class Probabilities & Extended Metrics (SwiftML & SwiftOptimize)

• ClassifierEstimator.predictProbability: Standardized 2D class probability outputs across DecisionTreeClassifier, RandomForestClassifier, and LogisticRegression.
• Extended Metrics: balancedAccuracy, matthewsCorrelationCoefficient (MCC), cohenKappa, logLoss, brierScore, rocCurve, rocAUC, and prCurve.

⏱️ Feature Engineering, Selection & NLP (SwiftPreprocessing, SwiftForecast & SwiftNLP)

• VarianceThreshold & SelectKBest: Feature variance filtering and top K feature ranking.
• InteractionFeatures & DateFeatures: Pairwise feature product terms and date component extractions.
• LagTransformer & RollingWindow: Lagged feature matrix builder and sliding window statistics.
• CountVectorizer: Document term frequency count matrix generator.
• DatasetUtilities: Synthetic dataset generators (makeClassification, makeRegression, makeMoons).

---

📊 Benchmark Performance Summary (v1.3.0 vs Python)

Benchmark Test	Swift (ms)	Python (ms)	Speedup	Winner
ARIMA(1,1,1) Fit (50k pts)	2.280 ms	215.527 ms	94.53x	🟢 Swift
Holt-Winters Fit (50k pts, period=12)	6.811 ms	148.627 ms	21.82x	🟢 Swift
KernelSHAP Explain (5 feats, 100 coalitions)	0.075 ms	0.426 ms	5.68x	🟢 Swift
Random Forest Fit (1k×4, 50 trees)	4.516 ms	25.475 ms	5.64x	🟢 Swift
Mean (vDSP, 1M elements)	0.081 ms	0.121 ms	1.49x	🟢 Swift
Pearson Correlation (500k elements)	0.856 ms	1.256 ms	1.47x	🟢 Swift
Isolation Forest Fit (1k×10, 100 trees)	14.048 ms	16.510 ms	1.18x	🟢 Swift
StdDev (vDSP, 1M elements)	0.477 ms	0.539 ms	1.13x	🟢 Swift
Variance (vDSP, 1M elements)	0.501 ms	0.520 ms	1.04x	🟢 Swift

---

📦 Installation

Add SwiftSci to your Package.swift:

dependencies: [
    .package(url: "https://github.com/Nodibell/SwiftSci.git", from: "1.3.0")
]

Full Changelog: v1.2.0...v1.3.0

v1.2.0 — SwiftSci Package Rename, Joseph Form Kalman Filter & UTF-8 BPE Tokenizer

SwiftSci v1.2.0

📦 What's Changed

🏷️ Package Rename & Branding

• SwiftSci Rename: Main Swift Package renamed from SwiftAnalytics to SwiftSci (name: "SwiftSci") for cleaner ecosystem branding while preserving target module names (SwiftDataFrame, SwiftStats, SwiftML, SwiftForecast, etc.).

📐 Numerical Stabilization & Core Upgrades

• Joseph Form Kalman Filter (SwiftForecast): Updated covariance matrix updates in filter() and smooth() to the numerically stable Joseph form P = (I - KH) P_pred (I - KH)^T + K R K^T, preserving symmetry and positive-definiteness under long filtering runs.
• UTF-8 Byte BPE Tokenizer (SwiftNLP): Refactored BPETokenizer bpe() algorithm to operate directly on word.utf8 bytes with GPT-2 byte-to-unicode character encoding (makeByteEncoder), fixing incorrect tokenization for non-ASCII input.

📊 DataFrame Enhancements (SwiftDataFrame)

• DataFrame.addColumn(_:as:using:): Added row-closure builder for computing new columns per-row over DataFrameRow, delegating to withColumn.
• DataFrameError.partialCastFailure: Added explicit error case thrown when partial element casting fails in castColumn (failed out of total), preventing silent data loss.
• DataFrame Invariant Diagnostics: Refactored DataFrame.gathered(at:) to trap with preconditionFailure on internal length invariant violations instead of returning an empty DataFrame. Removed duplicate private rows(at:).
• DataFrame.sample Randomization: Updated sample(n:seed:ordered:) to return randomly shuffled rows by default (ordered: false) while allowing index order preservation via ordered: true.

---

📊 Benchmark Performance Summary (SwiftSci vs Python)

Benchmark Test	Swift (ms)	Python (ms)	Speedup	Winner
Mean (vDSP, 1M elements)	0.082 ms	0.121 ms	1.47x	🟢 Swift
Pearson Correlation (500k elements)	0.868 ms	1.256 ms	1.45x	🟢 Swift
ARIMA(1,1,1) Fit (50k pts)	2.323 ms	215.527 ms	92.78x	🟢 Swift
ARIMA Forecast Horizon=24	2.456 ms	211.880 ms	86.26x	🟢 Swift
Holt-Winters Fit (50k pts, period=12)	6.841 ms	148.627 ms	21.73x	🟢 Swift
Random Forest Fit (1k×4, 50 trees)	5.025 ms	25.475 ms	5.07x	🟢 Swift
KernelSHAP Explain (5 feats, 100 coalitions)	0.192 ms	0.426 ms	2.22x	🟢 Swift
Kalman Filter 1D (10k obs, Joseph Form)	62.349 ms	85.547 ms	1.37x	🟢 Swift
LLM Forward Pass (seqLen=64)	0.636 ms	0.528 ms	0.83x	🔴 PyTorch
CSV Read (100k rows, 5 cols)	177.509 ms	19.340 ms	0.11x	🔴 Pandas
CSV Stream Read (chunk=10k)	238.699 ms	22.525 ms	0.09x	🔴 Pandas

• CI Gate Status: PASSED ✅ (0 gated regressions detected).

Full Changelog: v1.1.0...v1.2.0

v1.1.0 — Streaming CSV Parser, SafeTensors & GARCH

What's Changed in v1.1.0

🚀 Highlights

• Streaming CSV Parser (SwiftDataFrame): Parallel, chunk-buffered streaming CSV reader returning AsyncThrowingStream<DataFrame, Error> via Swift TaskGroup worker threads for low-memory, high-speed parsing of large files.
• SafeTensors & GGUF Parser (SwiftLLM): Native zero-copy binary weight parsers for local AI model inference, mapping model parameters from binary headers directly into MLXArray memory.
• SARIMA & GARCH Models (SwiftForecast): Added SARIMAModel (Seasonal ARIMA) with seasonal differencing and conditional least squares, and GARCHModel (Generalized Autoregressive Conditional Heteroskedasticity) with coordinate descent MLE parameter optimization.
• Swift 6 Strict Concurrency & Package Standardization: Fully verified under -strict-concurrency=complete. Cleaned up SwiftPM build flags to allow seamless integration as a package dependency.

🧹 Refactoring

• Cleaned up legacy compiler directives across LAPACK wrappers in SwiftForecast and SwiftCluster.
• Removed legacy SwiftPrivacy primitives in preparation for apple/swift-homomorphic-encryption binding in v1.2.

Full Changelog: v1.0.0...v1.1.0

v1.0.0 Stable release

First stable release of SwiftAnalytics. Fully optimized for Apple Silicon (Accelerate vDSP/MLX) under Swift 6 strict concurrency.