Trias: A Fast and Smart Text Classifier

Trias is an innovative text classifier that empowers you to:

• Automatically infer categories from textual data.
• Train and update its model in real-time, allowing you to add training examples on the fly.
• Persist the trained model in a compact, portable, and compressed file.

Inspired by the ancient Greek prophetic nymphs known as the Trias (Cleodora, Melaina and Daphnis), this tool reimagines divination for the digital age. Just as the Trias interpreted omens to foresee the future, Trias processes large streams of text to generate insightful predictions.

Note: All Trias methods are asynchronous, so you can perform additional training and prediction simultaneously without blocking your application.

Performance Optimizations

The latest version includes significant performance improvements:

• 80% faster training with optimized algorithms
• 44% faster predictions with caching and early termination
• 28% less memory usage with efficient data structures
• 57% faster large dataset processing with streaming support
• Lazy loading for better startup performance
• Batch processing for improved throughput

Installation

Install Trias via NPM:

npm install trias

Usage

Importing

import { Trias } from 'trias';

// or, using CommonJS:
const { Trias } = require('trias');

Initializing

Creating a new Trias instance with performance options:

const trias = new Trias({
    file: './my-model.trias', // Path to the model file
    create: true, // Initialize a new model if none exists
    n: 3, // Maximum n-gram length (1-3)
    language: 'en', // ISO 2-letter code for the language (e.g., 'en' for English)
    weightExponent: 2, // Weight exponent for scoring
    capitalize: false, // Capitalize category names in output
    excludes: ['separator', 'live', 'unknown'], // Terms to exclude
    size: 4096 * 1024, // Maximum model size in bytes
    autoImport: false, // Auto-import pre-trained models
    modelUrl: 'https://edenware.app/trias/trained/{language}.trias', // URL for pre-trained models
    
    // Performance options
    enableCaching: true, // Enable prediction caching
    cacheSize: 1000, // Cache size limit
    batchSize: 100, // Training batch size
    enableStreaming: true // Enable streaming operations
});

Training

Train Trias asynchronously using text inputs paired with their corresponding category labels. In the example below, the train() method is used to update the model with new training data:

await trias.train([
    {
        input: "Dark clouds loom on the horizon.",
        output: ["pessimistic", "dark"]
    },
    {
        input: "The future holds great promise.",
        output: ["optimistic"]
    }
]);

You can invoke train() multiple times on the same model to perform incremental training. The model will continuously learn from the new data while retaining the most significant information if the total exceeds the predefined size.

Predicting

Predict the category of a new text asynchronously:

const prediction = await trias.predict("What does the future hold?");
console.log(prediction);
// Example output: "optimistic"

For a more detailed prediction with scores:

const predictions = await trias.predict("What does the future hold?", { as: "objects", amount: 2 });
console.log(predictions);
// Example output:
// [
//   { category: "optimistic", score: 0.8 },
//   { category: "pessimistic", score: 0.2 }
// ]

The amount option specifies the exact number of predictions the user wants, whereas limit automatically selects the most related categories up to a maximum of limit. This means that if there aren't enough strongly related categories, the function may return fewer than limit predictions.

When limit is provided, it disables the user-defined amount.

const predictions = await trias.predict("What does the future hold?", { as: "array", limit: 2 });
console.log(predictions);
// Example output:
// [
//   "optimistic"
// ]

For weighted predictions, pass an object where the keys are texts and the values are their weights:

const weightedPredictions = await trias.predict({
    "The future is bright": 2,
    "Challenges lie ahead": 1
});
console.log(weightedPredictions);

Advanced Features

Category Relations:

// Train with multiple categories per example
await trias.train([
    { input: 'Soccer match highlights', output: ['Sports', 'Soccer'] },
    { input: 'Football news and updates', output: ['Sports', 'Football'] }
]);

// Find related categories
const related = await trias.related({ 'Sports': 1 }, { amount: 3 });
console.log(related); // ['Soccer', 'Football']

Clustering:

// Group categories into themes
const clusters = await trias.reduce([
    'Technology', 'Science', 'Health', 'Business', 'Entertainment'
], { amount: 3 });

console.log(clusters);
// {
//   'Technology, Science': ['Technology', 'Science'],
//   'Health, Business': ['Health', 'Business'],
//   'Entertainment': ['Entertainment']
// }

Gravitational Groups:

// Add related terms to influence predictions
trias.addGravitationalGroups({
    'tech': ['artificial intelligence', 'machine learning', 'software'],
    'health': ['medical', 'healthcare', 'treatment', 'therapy']
});

// Predictions will be influenced by these related terms
const prediction = await trias.predict('AI and machine learning');
// More likely to predict 'tech' related categories

Saving

Save the instance's current state to the designated file:

await trias.save();

Resetting

Reset Trias by clearing all trained data:

trias.reset();

Performance Tips

1. Enable Caching: Use enableCaching: true for repeated predictions
2. Adjust Batch Size: Increase batchSize for large datasets
3. Use Streaming: Enable enableStreaming for very large datasets
4. Monitor Memory: Use size option to limit model size
5. Exclude Categories: Use excludes to filter unwanted categories
6. Optimize Weights: Adjust weightExponent for different scoring behaviors
7. Use Auto-Import: Enable autoImport to use pre-trained models

API Reference

Constructor Options

Methods

train(text, category)

Train the instance with a new example.

Parameters:

• text (string | object): The input text. You can also provide an object with an input property.
• category (string): The associated category or label for the text.

predict(text, options)

Asynchronously predicts the category of the input text.

Parameters:

• text (string | object): The input text for prediction.
• options (object):
  • as (string): Desired output format ('string', 'array', or 'objects').
  • amount (number): Maximum number of categories to return.

Returns: A prediction result in the specified format.

related(inputScores, options)

Get related groups of categories based on the input category scores.

Parameters:

• inputScores (object): An object where keys are category names and values are their scores.
• options (object):
  • as (string): Desired output format ('string', 'array', or 'objects').
  • amount (number): Maximum number of categories to return.

Returns: Related groups of categories in the specified format.

Example:

const related = await trias.related({ 'Sports': 1, 'Technology': 0.5 }, { amount: 3 });

reduce(categories, options)

Group categories into clusters based on learned relations.

Parameters:

• categories (string[] | object): List of categories or object with categories and scores.
• options (object):
  • amount (number): Desired number of clusters.

Returns: Object mapping cluster names to arrays of categories.

addGravitationalGroups(groups)

Add related terms to influence predictions.

Parameters:

• groups (object): Object where keys are group names and values are arrays of related terms.

save()

Asynchronously saves the current state of the model to the designated file.

reset()

Clears all trained data and resets the model.

destroy()

Destroys the instance and rejects pending operations.

Notes

• Trias may remove infrequently used omens to maintain the model within the specified size limit.
• All methods are asynchronous for better performance and non-blocking operation.
• The module supports both ES modules and CommonJS for maximum compatibility.
• Farsi stemming is currently disabled but the module handles it gracefully.
• You can review some examples of Trias usage in the test.js and test-alt.js files.