[ROADMAP] version 1.0.0

[Mec-iS]

Plan to version 1.0.0

1. close issues that are irrelevant or too far-fetched
2. aggregate issues into macro-issues with sub-issues
3. prioritise and order for potential implementation

https://github.com/smartcorelib/smartcore/issues?page=1

Here is a comprehensive plan to organise the 52 open issues in smartcorelib/smartcore.

---

Issues to Close

These issues are either too vague, exploratory dead-ends, near-duplicates, or out of scope for a focused ML library:

#	Title	Reason to Close
^302	CUDA support request	Out of scope; WASM-first posture conflicts with GPU backend; author stated they'll fork regardless
^300	Geometric algebra / geonum	Highly speculative, O(1) claims are contested; no concrete impl path
^244	RL model-free algorithms	Massive scope, unrelated to supervised/unsupervised ML focus; no body beyond a screenshot
^195	BED file format support	Domain-specific (genomics), too niche for core library
^188	String support via anyinput	Superseded by the type system; no concrete use case defined
^84	Multi-target training (question)	This is a usage question, not a feature issue; no actionable request
^233	"Improve trees"	Completely empty body; too vague to action
^198	Generic cache system	Subsumed by #317 (memory pooling), can be closed as duplicate
^242	std::mem size/alignment analysis	Exploratory with no clear outcome; low ROI

---

Macro-Issues with Sub-Issues

Group actionable issues into 7 macro-themes:

🧱 1. Core API & Type System Cleanup

Foundation work that unblocks everything else.

• (Review usages for numbers traits #219) — Consolidate FloatNumber / RealNumber traits
• (Trait bounds of RandomForest::fit conflict with those of predict and other metrics #322) — Fix trait bound conflicts in RandomForest::fit vs metrics
• (Use Default::default instead of new in traits #230) — Replace new() with Default::default() in traits
• (Defaults only exist for Euclidean impls #174) — Defaults only exist for Euclidean distance impls
• (Serde JSON defaults don't match Default::default() #175) — Serde JSON defaults don't match Default::default()
• (Add getters to the structs in order to check internal state #73) — Add getters to structs to expose internal state

⚡ 2. Performance & Memory

Optimisation work for real-world scale.

• (tech-debt: replace unsafe raw-pointer iterator_mut in DenseMatrix (basic/matrix.rs) with safe split_at_mut approach #368) — Remove unsafe from iterator_mut (safe split_at_mut)
• (Consider implementing memory pooling system #317) — Memory pooling for tree construction
• (SIMD and Optimath #292) — SIMD / Optimath investigation
• (Look into slipstream and ILP #284) — Investigate slipstream and ILP
• (Compare performance of argsort and algorithm::sort::quick_sort #213) — Compare argsort vs quick_sort performance
• (Performance #261) — General performance gap vs scikit-learn
• (Consider the possibility to use [T] instead of Vec<T> in trait Array #199) — Use [T] slices instead of Vec in Array trait

🌲 3. Tree Models & Ensembles

High-demand improvements to the most-used model family.

• (Feature request: sample weights support in RandomForest (and other tree-based models) #356) — Sample weights in RandomForest
• (Variable Importance in Random Forest Analysis #124) — Variable importance in RandomForest
• (feature request: extra-trees classifier and regressor #183) — Extra-Trees classifier/regressor
• (Implement LEURN #254) — LEURN (neural net for tabular data, tree-based)

📐 4. Linear Models & Solvers

Completing the linear algebra and regression stack.

• (Add new solver that supports L1 penalty to Logistic Regression #70) — L1 solver (SAGA/LIBLINEAR) for Logistic Regression
• (Add L1 and Elastic Net regularization terms to the Logistic Regression #63) — L1 / ElasticNet regularisation in Logistic Regression
• (Request to add further summary stats, like Python OLS to regression #106) — F-stat, adjusted R², t-stat for regression
• (Add Linear Discriminant Analysis (LDA) support #136) — Linear Discriminant Analysis (LDA)
• (Support for rust_decimal  #49) — rust_decimal support (non-float arithmetic)

🔧 5. Data I/O & Interoperability

Making smartcore work in real pipelines.

• (Implement a convenience method DenseMatrix::from_ndarray2 (in case the ndarray feature is on) #326) — DenseMatrix::from_ndarray2 convenience method
• (Array conventions: support CSR #156) — CSR (Compressed Sparse Row) array support
• (Allow read/write of parquet files #249) — Parquet / Arrow file format support
• (Support streaming/incremental fitting #279) — Streaming / incremental fitting
• (Implement SGDClassifier #62) — SGDClassifier (incremental learning)

🖥️ 6. DX, Display & Documentation

Developer experience and discoverability.

• (Define default formatter for linalg::basic::arrays::{Array, Array2} #197) — Display impl for Array/Array2
• (Improve Display for naive_bayes #225) — Improve Display for NaiveBayes
• (No implementation of Display for Dataset #77) — Display for Dataset
• (Document features in Cargo.toml #68) — Document Cargo features
• (Have a guide for Wasm compilation #259) — WASM compilation guide
• (Consider applying derive_builder #189) — derive_builder for model parameters
• (index out of bounds error using logistic_regression #256) — Index out of bounds bug in logistic regression (large input)

🧩 7. New Algorithms

Expanding coverage for ML practitioners.

• (Make SVC multiclass #224) — Multiclass SVC (OvR/OvO)
• (Embed HDBSCAN #291) — HDBSCAN clustering
• (OPTICS algorithm #52) — OPTICS clustering
• (Implement Recursive Feature Elimination (RFE) algorithm #61) — Recursive Feature Elimination (RFE)
• (Implement OrdinalEncoder #59) — OrdinalEncoder / preprocessing
• (Add feature extractions tools #170) — Feature extraction (CountVectorizer, HashVectorizer)
• (Survival analysis #248) — Survival analysis
• (Feature request: time series functionality #91) — Time series functionality
• (Investigate possible usages of const contexts #226) — const context investigation

---

Priority Order for Implementation

Priority	Macro-Issue	Rationale
P0	🧱 Core API & Type System	Blocking: trait conflicts (#322) cause user-facing breakage today; type unification unblocks everything
P1	⚡ Performance & Memory	The unsafe removal (#368) is actively tracked; perf gap vs scikit-learn (#261) erodes adoption
P1	🖥️ DX, Display & Docs	Low-effort, high-visibility; good first issues; #256 (logistic regression panic) is a bug
P2	🌲 Trees & Ensembles	Most-used models; sample weights and variable importance are highly requested
P2	📐 Linear Models & Solvers	Completes the regression/classification surface; solvers (#70, #63) are well-specified
P3	🔧 Data I/O & Interop	ndarray bridge (#326) and parquet (#249) open up real pipeline usage; streaming is complex
P4	🧩 New Algorithms	Valuable but requires stable foundation; HDBSCAN, multiclass SVC, and RFE are highest demand

The single highest-impact starting point is #322 + #219 (trait bound and number trait cleanup), since they directly break user workflows and their resolution would unblock correct cross-validation for tree models. Immediately after, the #256 logistic regression panic should be treated as a bug fix given its user impact.