likely

Recommendation engine for Node.js applications.

likely.js

A javascript library for collaborative filtering and recommendation engines designed for node.js

Likely.js is a library used for collaborative filtering and making recommendation. It takes an input training matrix where the rows represent users and the columns represent items. Each entry in the matrix is the rating the user has given to that item. After training, you can retrieve a list of recommended items and their estimated ratings for any given user.

For example, you may provide as input a training matrix where each row is a customer and the columns are the ratings they give to various movies. Then you might use this model to get recommendations for movies a given customer might like that they have not yet seen.

npm install likely

To use Likely, require the likely module and follow the 3 steps below.

var Recommender = require('likely.js');

Create an input matrix where rows are users and the columns are items A cell represents the rating of that item by the entity. inputMatrix[0][1] is the rating of item 1 by user 0

var inputMatrix = [ [ 1, 2, 3, 0 ],
                    [ 4, 0, 5, 6 ],
                    [ 7, 8, 0, 9 ]
                  ];

Labels to provide more context to the input. Row 0 of the input matrix corresponds to label rowLabels[0]

var rowLabels = ['John', 'Sue', 'Joe'];
var colLabels = ['Red', 'Blue', 'Green', 'Purple'];

Using these values, John rates Red 1 while Joe rates Red 7. Sue has no rating for Blue.

Using the inputMatrix you build a model, which estimates the ratings for all entities for all users.

var Model = Recommender.buildModel(inputMatrix, rowLabels, colLabels);

or, if you don't have or care about labels

var Model = Recommender.buildModel(inputMatrix);

The Model object now contains a matrix of the same size as inputMatrix but with estimates ratings for all items for all entities.

There are a few ways to retrieve recommendations from the model you have built.

Example 1: Retrieve a list of all items not already rated by a user, sorted by estimated ratings using labels.

var recommendations = model.recommendations('John');

// recommendations = [['Purple', 1.34]];

Example 2: Retrieve a list of all items not already rated by a user, sorted by estimated ratings without labels.

var recommendations = model.recommendations(0);

// recommendations = [[3, 1.34]];

Example 3: Retrieve a list of all items, sorted by the ratings for a given user (both estimated and actual), using labels.

var allItems = model.rankAllItems('John');

// allItems = [['Green', 3.00], ['Blue', 2.00], ['Purple', 1.34], ['Red', 1.00]];

Example 4: Retrieve a list of all items, sorted by the ratings for a given user (both estimated and actual) without labels.

var allItems = model.rankAllItems(0);

// allItems = [[2, 3.00], [1, 2.00], [3, 1.34], [0, 1.00]];

As with any machine learning, it's important to test to make sure it works with your data. The best way to do this is to take your input data (known ratings) and divide it into two groups: training and cross-validation (CV). To see if the model is working you should train the model using your training set and check the error using the CV set. If it looks good then you can apply the model built using the training set to your entire data and have confidence in the estimates.

In practice, this would look as follows:

var Recommender = require('likely.js');

// Build the model using the training set
var model = Recommender.buildModel(trainingSet);

// Calculate the error from the produced model against the CV set of known values
var totalError = Recommender.calculateTotalError(model.estimate, crossValidationSet);

This is common for any machine learning application. In this case it will be necessary to tune the parameters of the Likely.js learning algorithm to better fit your data. The available options that you can adjust are as follows:

var Recommender = require('likely.js');

// Number of iterations it will use to try and learn the model, the larger the better. 
//  However, the larger the steps the longer it will take to train the model.
Recommender.DESCENT_STEPS;  // DEFAULT = 5000 

// The rate of learning. If your error is very large you can try making this larger.
//  However, if the learning rate is too large the error will increase.
Recommender.ALPHA;          // DEFAULT = 0.0005

// The regularization factor, this prevents over-fitting. If your error is not affected by changing
//  the steps or learning rate try adjusting this. It should never be a large value.
Recommender.BETA;           // DEFAULT = 0.0007

// The number of features to learn. In theory the more the better but a larger value will slow down
//  the training. 
Recommender.k;          // DEFAULT = 5

Likely has both unit and regression tests. To make sure your version is working you should run both.

Running the unit tests:

./scripts/run_unit_tests.sh

Running the regression test:

node regression_test.js

Note that the unit test requires you have installed the expresso testing framework which can be done easily:

npm install expresso

Or install the npm with the -dev flag

npm install -dev likely

Likely.js uses matrix factorization to estimate the rating values not provided. MF attempts to find two matrices, P and Q such that the product is equal to the input matrix.

P x Q = inputMatrix

P and Q are initialized as random matrices and Gradient Descent is used to minimize the error over a number of iterations. Regularization is used to make sure that there is no overfitting, which means the actual result of P x Q is not exactly the inputMatrix but something that closely approximates it. The effect of this approximation is estimates for the values not provided in the inputMatrix.

These estimated values are what are used to provide the recommendations.

Likely.js uses Stochastic Gradient Descent to estimate P and Q which, while simple, can be slow for very large input matrices. This is the most basic but effective method of recommendations that was highlighted in the Netflix prize: http://www2.research.att.com/~volinsky/papers/ieeecomputer.pdf