Thermometer & Level

Previous Next

ThermometerEncoder and LevelEncoder

These encoders handle ordinal and categorical data respectively.

ThermometerEncoder

Overview

Property Value
Input Ordinal value (1, 2, 3, ...)
Output Hypervector
Preserves Order relationships
Best For Rankings, ratings

How It Works

Thermometer encoding accumulates similarity as values increase:

For levels 1-5 with base vectors [b₁, b₂, b₃, b₄, b₅]: - encode(1) = b₁ - encode(2) = bundle(b₁, b₂) - encode(3) = bundle(b₁, b₂, b₃) - encode(4) = bundle(b₁, b₂, b₃, b₄) - encode(5) = bundle(b₁, b₂, b₃, b₄, b₅)

Similarity Pattern

Adjacent values share most components:

sim(1, 2) = 0.71  (share 1 of 2)
sim(2, 3) = 0.82  (share 2 of 3)
sim(1, 5) = 0.45  (share 1 of 5)

Code Example

from holovec import VSA
from holovec.encoders import ThermometerEncoder

model = VSA.create('FHRR', dim=2048)

# Create thermometer encoder for ratings 1-5
encoder = ThermometerEncoder(model, levels=5)

# Encode ratings
r1 = encoder.encode(1)
r2 = encoder.encode(2)
r3 = encoder.encode(3)
r5 = encoder.encode(5)

# Adjacent ratings more similar
print(model.similarity(r1, r2))  # ~0.71
print(model.similarity(r2, r3))  # ~0.82
print(model.similarity(r1, r5))  # ~0.45

Use Cases

  • Star ratings (1-5 stars)
  • Priority levels (low, medium, high)
  • Education levels
  • Size categories (S, M, L, XL)

LevelEncoder

Overview

Property Value
Input Categorical label
Output Hypervector
Preserves Category identity
Best For Discrete categories

How It Works

Each level gets an independent random vector:

  • encode("cat") = random₁
  • encode("dog") = random₂
  • encode("bird") = random₃

All categories are approximately orthogonal.

Similarity Pattern

sim("cat", "cat") = 1.0
sim("cat", "dog") ≈ 0.0
sim("dog", "bird") ≈ 0.0

Code Example

from holovec import VSA
from holovec.encoders import LevelEncoder

model = VSA.create('FHRR', dim=2048)

# Create level encoder with categories
categories = ["red", "green", "blue", "yellow"]
encoder = LevelEncoder(model, levels=categories)

# Encode categories
red = encoder.encode("red")
green = encoder.encode("green")
red2 = encoder.encode("red")

# Same category = identical
print(model.similarity(red, red2))    # 1.0

# Different categories = orthogonal
print(model.similarity(red, green))   # ~0.0

Use Cases

  • Color categories
  • Species classification
  • Country codes
  • Product categories

Choosing Between Them

Scenario Use
Values have natural order ThermometerEncoder
Categories are unrelated LevelEncoder
Rankings (1st, 2nd, 3rd) ThermometerEncoder
Labels (A, B, C, D) LevelEncoder

Combined Example

# Product encoding with both types
model = VSA.create('FHRR', dim=2048)

# Ordinal: star rating
rating_enc = ThermometerEncoder(model, levels=5)

# Categorical: product category
category_enc = LevelEncoder(model, levels=["electronics", "clothing", "food"])

# Role vectors
RATING = model.random(seed=1)
CATEGORY = model.random(seed=2)

def encode_product(rating, category):
    return model.bundle([
        model.bind(RATING, rating_enc.encode(rating)),
        model.bind(CATEGORY, category_enc.encode(category))
    ])

# Similar ratings, same category → similar
phone_5star = encode_product(5, "electronics")
phone_4star = encode_product(4, "electronics")
shirt_5star = encode_product(5, "clothing")

print(model.similarity(phone_5star, phone_4star))  # High (same cat, close rating)
print(model.similarity(phone_5star, shirt_5star))  # Moderate (diff cat, same rating)

Comparison

Property Thermometer Level
Similarity structure Gradual Binary
Order preserved Yes No
Interpolation Natural Not applicable
Memory per level Cumulative Independent

References

  • Kanerva, P. (2009). Hyperdimensional Computing

See Also

Fractional Power Sequences