属性约简（Attribute Reduction）与维度约简（Dimensional Reduction）的差异及技术划分

Great question! Let's unpack the key differences between attribute reduction and dimensionality reduction, then clarify which methods fall squarely into the attribute reduction category (and aren't just dimensionality reduction in disguise).

1. Definition & Core Approach #

• Attribute Reduction: This is a subset selection technique focused on discrete, symbolic, or categorical attributes (think "color", "customer segment", "product type"). The goal is to eliminate redundant or irrelevant attributes while preserving the dataset's ability to support tasks like classification or decision-making. Crucially, it doesn't modify existing attributes—it just picks a smaller, more meaningful subset of the original ones.
• Dimensionality Reduction: This is a space transformation technique that maps high-dimensional data into a lower-dimensional space. It often works with continuous numerical features (like pixel values, sensor readings) and creates new, synthesized features (not just picking from the original set) that capture the majority of the data's variance or information.

2. Input Data Focus #

• Attribute reduction is designed primarily for discrete, nominal, or ordinal data. It’s commonly used in rule-based systems, knowledge discovery, and scenarios where attribute interpretability is critical.
• Dimensionality reduction shines with continuous numerical data (though it can handle discrete data after encoding). It’s a go-to for high-dimensional data like images, text embeddings, or sensor arrays where raw features are too dense to work with directly.

3. Outcome Interpretability #

• Attribute reduction outputs a subset of the original attributes, each with clear, unchanged business or semantic meaning. For example, if you reduce a customer dataset to "age", "income", and "purchase frequency", you can still directly explain what each attribute represents.
• Dimensionality reduction outputs new, composite features that lack direct interpretability. For instance, PCA’s principal components are linear combinations of raw features—you can’t easily say "Principal Component 1 = 0.3age + 0.7income" has a concrete real-world meaning.

4. Primary Objective #

• Attribute reduction’s top priority is removing redundancy and simplifying model/rule structure without losing critical decision-making power. It’s about making the dataset leaner while keeping its predictive or descriptive utility intact.
• Dimensionality reduction’s main goal is combating the curse of dimensionality—reducing computational cost, reducing noise, and making data easier to visualize or model—by preserving as much of the data’s core information as possible in a smaller space.

These methods exclusively focus on selecting subsets of original attributes, with no feature transformation or synthesis:

• Rough Set-Based Attribute Reduction: The gold standard for attribute reduction in symbolic data. Methods like dependency-based reduction or entropy-based reduction calculate attribute importance to filter out redundant attributes while maintaining the dataset’s decision consistency.
• Decision Tree Attribute Selection & Pruning: Algorithms like ID3, C4.5, or CART inherently perform attribute reduction by selecting features with the highest information gain/gain ratio during tree construction. Post-pruning further removes non-critical attributes that don’t improve predictive performance.
• Filter Methods for Discrete Data: Techniques like chi-squared test (for categorical attribute-label relationships) or mutual information selection (measuring the dependence between attributes and the target) rank and select original attributes based on statistical relevance—no new features are created.
• Wrapper Methods for Attribute Subset Search: Approaches like genetic algorithms, particle swarm optimization, or forward/backward selection evaluate subsets of original attributes using a model’s performance (e.g., accuracy on a classifier) to find the optimal, smallest subset.
• Consistency-Based Reduction: This method removes attributes that don’t affect the dataset’s classification consistency—if removing an attribute doesn’t change how any sample is labeled, it’s deemed redundant and discarded.

内容的提问来源于stack exchange，提问作者Rubiks