quanteda
Text analysis in R keeps optimising its token internals — and builds a path out to torch
Comparison · Analytics
collinear vs lstar
A side-by-side editorial comparison of collinear and lstar — release velocity, themes, recent moves, and the top alternatives to consider.
collinear vs lstar: at a glance
What is collinear?
collinear has broken its API twice to stop making the user pick thresholds.
collinear removes multicollinearity from predictor sets through pairwise correlation and VIF filtering, with a preference order deciding which variable survives each conflict. Two major versions in thirteen months each rewrote the interface: 2.0.0 extended every function to any combination of categorical and numeric responses and predictors, and 3.0.0 moved to multiple responses, restructured the output into classed objects, and made both filtering thresholds adaptive by default. Version 3.0.1 is the first release since that is purely repair.
What is lstar?
A single-cell data store commits to Zarr v3 and range-readable hosting across four language surfaces
lstar stores single-cell data behind one C++ core with Python, R and JS/WASM bindings, and ships a browser viewer that reads the store directly. Zarr v3 is now the default on-disk format across all four surfaces, with zstd compression and sharding that packs many chunks into fewer objects. Viewer stores are compressed per field and resolved at chunk granularity, so a hosted viewer fetches only what it displays. The tag stream carries both lstar and lstar-sc releases.
collinear vs lstar: editorial side-by-side
collinear
0.0ANALYTICS
collinear has broken its API twice to stop making the user pick thresholds.
◆ Current state
collinear removes multicollinearity from predictor sets through pairwise correlation and VIF filtering, with a preference order deciding which variable survives each conflict. Two major versions in thirteen months each rewrote the interface: 2.0.0 extended every function to any combination of categorical and numeric responses and predictors, and 3.0.0 moved to multiple responses, restructured the output into classed objects, and made both filtering thresholds adaptive by default. Version 3.0.1 is the first release since that is purely repair.
◆ Where it's heading
The through-line is removing decisions the user was never well placed to make. Preference-order functions were renamed twice — first onto a metric-and-model scheme in 2.0.0, then onto a response-type scheme in 3.0.0 — and f_auto() picks one when none is given; target encoding went from automatic to opt-in; max_cor and max_vif now default to NULL and trigger a data-driven threshold derived from the 75th percentile of pairwise correlations through a sigmoid and a fitted correlation-to-VIF mapping. Each change is defensible and each one broke callers, which is the cost of this approach.
◆ Prediction
3.0.1 moved the example datasets out into a separate spatialData package and fixed four crashes rather than adding anything, so the next release is most likely more consolidation on the 3.0 surface than a fourth interface.
lstar
5.0ANALYTICS
A single-cell data store commits to Zarr v3 and range-readable hosting across four language surfaces
◆ Current state
lstar stores single-cell data behind one C++ core with Python, R and JS/WASM bindings, and ships a browser viewer that reads the store directly. Zarr v3 is now the default on-disk format across all four surfaces, with zstd compression and sharding that packs many chunks into fewer objects. Viewer stores are compressed per field and resolved at chunk granularity, so a hosted viewer fetches only what it displays. The tag stream carries both lstar and lstar-sc releases.
◆ Where it's heading
The through-line is making a hosted store cheap to read. Sharding addresses the file-per-chunk explosion that makes many-chunk arrays awkward to host; per-field compression with chunk-granular resolution means colouring an embedding by one gene fetches one column rather than an array. The 0.2.x patches are the cost of maintaining four surfaces at once — a WASM heap crash that only browsers exercise, and a count-basis orientation defect where all three surfaces normalized in memory and none owned the on-disk layout.
◆ Prediction
The orientation bug's root cause — no surface owning the on-disk representation while all three normalized in memory — is the kind of gap that usually produces a validation or ownership change rather than another point fix.
Alternatives to collinear and lstar
Other Analytics products tracked by Sparkpulse, ranked by recent ship velocity. Each card links to a full editorial trajectory and lets you pivot into a head-to-head comparison with either collinear or lstar.
moderndive
The ModernDive teaching package learns to render inside the browser that runs its own textbook
glmbayes
A GPU-accelerated Bayesian GLM package buys its way into the standard R Bayesian toolchain
hydroloom
USGS puts a type system over its river network toolkit so errors surface at dispatch
chromConverter
The chromatography file-format translator keeps absorbing vendor formats one release at a time
slope
A year after gutting itself for a C++ rewrite, SLOPE is back to polishing the interface
See all collinear alternatives → · See all lstar alternatives →
Recent activity from collinear and lstar
Latest ship moves from both products, interleaved chronologically. ⚡ = editorial spark.
- 21d agolstarViewer count-basis orientation fixed; existing stores need re-prep
- 26d agolstarWASM viewer crash on resizable heaps fixed
- 1mo agolstarlstar 0.2.0 ⚡
- 1mo agolstarMeasure state inferred from content, not slot name
- 3mo agocollinearNamespace, NA and sf fixes; example data moves to spatialData
- 8mo agocollinearAdaptive thresholds, multi-response support and a new output class ⚡
- 1y agocollinearCategorical responses, f_auto() defaults and future-based parallelism ⚡
Frequently asked questions
What is the difference between collinear and lstar?
They serve adjacent needs but don't currently overlap on shipped themes. lstar is currently shipping more aggressively (velocity 5.0 vs 0.0), with 0 editorial sparks in the last 30 days against 0. See the at-a-glance table above for a side-by-side breakdown of velocity, recent sparks, and editorial themes.
Is collinear better than lstar?
Sparkpulse doesn't pick a winner — we score release velocity, not feature parity. lstar is currently shipping more aggressively (velocity 5.0 vs 0.0), with 0 editorial sparks in the last 30 days against 0. For your specific use case, the alternatives sections above list other Analytics products to evaluate alongside.
What are the best alternatives to collinear?
Top collinear alternatives in Analytics are ranked by recent ship velocity. Browse the "collinear alternatives" section above for the current picks, or visit /alternatives/collinear for the full list with editorial commentary on each.
What are the best alternatives to lstar?
Top lstar alternatives in Analytics are ranked by recent ship velocity. Browse the "lstar alternatives" section above for the current picks, or visit /alternatives/lstar for the full list with editorial commentary on each.