Imagine asking for the best map.
Best for what?
Driving?
Comparing country area?
Understanding subway connectivity?
Showing terrain?
The question is incomplete until the job is named.
A projection objective is the rule for deciding which relationships deserve to survive when the lower-dimensional representation cannot preserve everything.
This is the fourth pillar beneath How Lossy Works | Projection. The master owns the general loss. This article owns the preservation contract: variance, neighbourhoods, global distance, topology, interpretability, prediction, fairness or another task-specific structure.
Quick Read
There is no universally best low-dimensional projection because different methods optimise different notions of faithfulness. PCA finds linear directions that preserve large amounts of variance. t-SNE represents pairwise neighbourhood similarities and is primarily a visualisation tool for local structure. UMAP builds a neighbourhood graph and offers parameters that trade local detail against broader organisation and compactness. Other methods may prioritise global distances, class separation, reconstruction error, sparsity or interpretability. A projection is fit for purpose when the structure it preserves matches the decision the receiver needs to make and the lost structure is not decision-critical.
receiver job → preservation objective → projection method + parameters → retained relationships → discarded relationships → validation → decision boundary
The Best Shadow Is a Task Statement
A shadow can preserve silhouette.
It cannot preserve full three-dimensional shape.
If the job is recognising a distinctive outline, that may be enough.
If the job is machining the object, it is not.
Projection quality is conditional on downstream need.
PCA Preserves Variance Under a Linear Model
Current scikit-learn documentation describes PCA as linear dimensionality reduction that projects centred data onto principal axes ordered by explained variance.
Its implicit preservation priority is:
keep directions along which the data vary most.
This can produce compact representations with low reconstruction error under a linear subspace model.
It does not ask whether those directions are the most meaningful for a particular human decision.
Variance Preservation Is Excellent for Some Jobs
Data compression.
Noise reduction where low-variance dimensions are plausibly noise.
Visualising dominant linear structure.
Preprocessing for models that benefit from decorrelated compact features.
These are legitimate objectives.
Variance Can Be the Wrong Objective for Rare Events
A rare fault state may explain little total variance.
A small subgroup may occupy a direction that barely affects the population covariance.
If the task is anomaly detection, safety or minority subgroup analysis, low variance can still be high importance.
t-SNE Preserves a Different Kind of Relationship
Scikit-learn describes t-SNE as converting similarities between points into probability distributions and minimising divergence between high- and low-dimensional neighbourhood relationships.
Its purpose is not to make every global distance numerically faithful.
It is especially useful for visualising local neighbourhood structure.
That Changes How You Read the Plot
Who sits near whom can be meaningful.
The distance between far-away islands may not be.
Cluster size may not be.
Absolute orientation certainly is not.
The objective tells you which visual claims are licensed.
UMAP Makes Scale Choice Part of the Objective
UMAP constructs local neighbourhood relationships and then optimises a low-dimensional embedding intended to represent that structure. Its n_neighbors parameter controls how local or broader the neighbourhood model becomes.
Official UMAP documentation makes the trade-off concrete: lower neighbour counts focus more on local detail, while higher values allow more global structure to influence the representation.
There is no parameter-free definition of “the” correct scale.
min_dist Adds a Packing Objective
UMAP’s min_dist influences how tightly points may pack in the low-dimensional space.
Low values make tight local clumps easier to form.
Higher values spread points more broadly.
The visible density of a cluster is therefore partly a projection design choice.
Global Distance Is Another Possible Objective
Suppose the job is preserving pairwise distances among all points as faithfully as possible.
Methods in the multidimensional-scaling family can make global distance distortion the central criterion.
This can be better for some geometry questions than an embedding designed mainly around local neighbourhoods.
Topology Can Matter More Than Metric Distance
For some tasks, the crucial question is:
which points are connected through neighbourhood structure?
Exact kilometre-like distance may be less important than preserving local continuity, branches or neighbourhood graph structure.
A topological objective and a metric objective are related but not identical.
Reconstruction Error Is a Different Objective Again
Encode high-dimensional data into a small latent representation.
Decode it back.
Now optimise how closely the reconstruction matches the input under a chosen loss.
This can produce a representation useful for compression or denoising even if a two-dimensional scatterplot of the latent coordinates is not especially interpretable.
Visual Interpretability Can Compete With Reconstruction
A dense latent representation can reconstruct beautifully while remaining semantically opaque.
A sparse or constrained representation may reconstruct slightly worse but give humans clearer axes.
The better model depends on whether the receiver needs faithful decoding or intelligible concepts.
Interpretability Is Not One Number
An interpretable axis might mean:
- few original variables contribute;
- the direction has a clear domain meaning;
- changes along it are monotonic;
- the receiver can explain why a point moved.
A mathematically efficient projection can be operationally poor if nobody can explain the representation at the decision boundary.
Class Separation Can Be an Explicit Objective
If labelled categories already exist, a supervised method can favour dimensions that distinguish those labels.
This may be appropriate for classification.
It is dangerous if the resulting separation is later presented as evidence that the categories emerged naturally without label information.
A projection trained to separate classes cannot later be used as independent proof that the classes separate.
Prediction Can Be the Objective Instead of Visual Geometry
A low-dimensional representation may be judged by downstream predictive performance rather than by whether its scatterplot looks elegant.
If the job is forecasting demand, a latent representation that preserves predictive information can be useful even when human geometry is unintuitive.
Fairness Can Be a Preservation Constraint
A projection that preserves majority structure while collapsing a small subgroup may be efficient on average and harmful for equitable decision-making.
If subgroup distinctions matter for the action boundary, fairness becomes part of the projection objective or validation constraint.
Privacy Can Be the Opposite Objective
Sometimes we deliberately want identity-relevant dimensions to disappear.
A privacy-preserving representation may sacrifice reconstruction or individual distinctiveness while retaining aggregate utility.
Anonymisation owns the specific privacy trade-off.
Compression and Decision Objectives Can Conflict
The smallest representation may omit a low-frequency feature that drives an expensive failure.
Optimising average reconstruction error can sacrifice tails.
Rate efficiency and decision safety should be separated rather than assumed to align automatically.
Local and Global Structure Are a Budget Allocation
Preserve every local neighbourhood perfectly and large-scale distances may distort.
Preserve broad global arrangement and tiny local structures may blur.
Recent dimensionality-reduction research continues to investigate this balance because lower-dimensional spaces cannot generally satisfy all structural objectives at once.
Scale Is Embedded in the Objective
What counts as “local”?
Five neighbours?
Fifty?
A neighbourhood graph is already a statement about the scale at which similarity matters.
Projection objectives inherit scale assumptions even when the final picture has no scale bar.
Distance Metric Defines What “Near” Means
Euclidean distance.
Cosine distance.
Correlation distance.
Domain-specific metrics.
Change the metric and the neighbourhood objective can change before the embedding algorithm begins.
Feature Scaling Defines Metric Weight
A variable measured in thousands can dominate Euclidean distance over one measured between zero and one.
Normalisation can make variables numerically comparable.
But whether they should carry equal weight is a domain question.
The Objective Should Be Written Before the Plot Is Seen
If the analyst first sees the picture and only later decides what it was meant to preserve, the objective can become retrofitted to the aesthetically pleasing result.
A stronger workflow declares:
- the receiver;
- the decision;
- the structure that matters;
- the acceptable loss;
- the validation test.
Projection Quality Needs Multiple Metrics
One metric can reward local neighbourhood preservation.
Another can measure global distance distortion.
Another can evaluate reconstruction.
Another can measure downstream classification or prediction.
The evaluation suite should mirror the objective rather than rely on one generic “embedding score.”
A Receiver Can Need More Than One Projection
One view for global structure.
One view for local neighbourhoods.
One interpretable table for decision explanation.
Trying to force every job into one two-dimensional picture may create more loss than necessary.
Projection Direction Is Downstream of Objective
The first pillar, Projection Direction, explains how the chosen view changes visible structure.
This pillar explains why that direction should be chosen: because a receiver values some source relationships over others.
Projection Collisions Must Be Judged Against Objective
The second pillar, Projection Collisions, shows that distinct states can become indistinguishable.
If colliding states require the same action, the collision may be acceptable.
If they require different actions, the objective has failed.
Stability Is Objective-Relative Too
The third pillar, Projection Stability, asks what survives perturbation.
But stability should be measured on the relationship the objective cares about—not on raw coordinates simply because they are easy to compare.
Education: One Projection Cannot Represent a Whole Learner
A total examination score may be useful for one administrative decision.
A teacher repairing misconceptions needs a richer projection:
- topic accuracy;
- error type;
- retrieval stability;
- working speed;
- transfer performance.
The “best” learner representation changes when the teaching job changes.
Dashboards Are Projection Objectives in Business Clothing
A dashboard chooses a few metrics from a larger organisation.
Revenue.
Retention.
Conversion.
The dashboard is good if those axes support the decision.
It is dangerous if people begin optimising the visible axes while hidden quality collapses.
Policy Indicators Have the Same Problem
One index compresses many social dimensions.
Weights decide what counts.
Normalisation decides comparability.
Aggregation decides which trade-offs are invisible.
A single ranking can be useful while remaining a projection of a much larger policy world.
The Objective Should Include a Failure Test
What distinction must never disappear?
What subgroup must remain visible?
What distance can be distorted safely?
What false cluster would mislead the receiver?
A preservation goal without a failure boundary is incomplete.
A Better Projection-Objective Model
receiver → task → decision-critical relationships → acceptable loss → method objective → parameters/metric → projection → objective-specific validation → stability + collision tests → use or escalate to richer representation
A 30-Lens Projection Objective Audit
- Receiver: who will use the representation?
- Task: visual exploration, compression, prediction, classification or explanation?
- Decision: what action follows?
- Source: what richness exists upstream?
- Target dimension: how small must representation become?
- Variance: should dominant variability be preserved?
- Local neighbourhoods: must nearest neighbours survive?
- Global distance: must far-apart relationships remain meaningful?
- Topology: must continuity and branching survive?
- Reconstruction: must source features be recoverable approximately?
- Interpretability: must axes have human meaning?
- Sparsity: should few features explain each axis?
- Class separation: are labels allowed to influence projection?
- Prediction: must target-relevant information survive?
- Fairness: which subgroup distinctions are decision-critical?
- Privacy: which distinctions should deliberately disappear?
- Metric: what definition of similarity applies?
- Scale: what counts as local?
- Preprocessing: how are feature units weighted?
- Collision tolerance: which source states may safely merge?
- Stability: which invariant should survive parameter changes?
- Visual aesthetics: are they being mistaken for objective quality?
- Validation metric: how will preserved structure be measured?
- Source-space check: can claims be confirmed upstream?
- External validation: does projection predict or correspond to reality?
- Failure test: what lost distinction would make the view unsafe?
- Multiple views: would two projections serve different jobs better?
- Currentness: is the objective still aligned with the current task?
- Disclosure: are unpreserved structures stated clearly?
- Escalation: when must the receiver return to richer data?
Laboratory 1: One Dataset, Three Jobs
Take one multivariable dataset. Define three goals: preserve variance, preserve nearest neighbours and separate a known label. Explain why the optimal projection can differ for each.
Laboratory 2: Write the Failure Test First
Before making a projection, write one distinction that must not disappear. After projecting, verify that the representation still supports it.
Laboratory 3: Two Views Instead of One
Create one local-structure view and one global-structure view. Ask whether the pair gives a safer mental model than forcing both objectives into a single plot.
For Primary Readers
Ask for the “best drawing” of a playground. A child finding the slide needs a different drawing from a builder measuring the fence. The job decides what the picture should keep.
For Secondary Readers
For every reduced representation, state what it was designed to preserve and one relationship it may distort.
For Advanced Readers
Model dimensionality reduction as constrained optimisation over competing fidelity functionals. Method choice is justified only relative to downstream utility, distortion tolerance and the invariants required at the receiver/action boundary.
Common Misconceptions
- “There is one best projection of a dataset.” Different objectives produce different legitimate representations.
- “PCA preserves the most important information.” It prioritises variance under a linear model, not universal importance.
- “t-SNE distances are ordinary map distances.” Its objective focuses on neighbourhood similarities; global geometry requires caution.
- “A projection with separated known classes proves natural categories.” If labels shaped the projection, separation is partly built into the objective.
- “If the plot looks intuitive, the objective is correct.” Visual intuition is a receiver response, not a fidelity proof.
Research Corridor
- scikit-learn — PCA — principal axes and explained variance.
- scikit-learn — t-SNE — neighbourhood similarities, KL-divergence objective and parameter considerations.
- UMAP — Basic Parameters — local/global neighbourhood scale and min_dist.
- Comprehensive Review of Dimensionality Reduction Algorithms — comparative trade-offs, stability and generalisation.
Frequently Asked Questions
What is a projection objective?
It is the criterion that defines which source-space relationships a lower-dimensional representation should preserve as faithfully as possible.
Which objective does PCA use?
PCA selects linear principal directions ordered by explained variance and provides the best low-rank linear reconstruction under common squared-error formulations.
How do I choose between PCA, t-SNE and UMAP?
Start with the task and the relationship that must survive. PCA is useful for dominant linear variance and compact representation; t-SNE is often useful for local-neighbourhood visualisation; UMAP provides configurable local/global neighbourhood trade-offs. Validate the chosen method against the actual claim rather than selecting by appearance alone.
Final Thought: Loss Becomes Intelligent Only After the Job Is Named
A smaller representation cannot keep everything.
So it must know what it is for.
The best projection is not the shadow that looks most complete. It is the shadow that preserves the relationships the receiver needs while making its losses impossible to forget.
PROJECTION · FOUR PILLAR LEGS
Return to Projection, or continue through Projection Direction, Projection Collisions and Projection Stability. Return to the Information & Representation Hub.