Distributed Processing: Training#

The distributed processing mode assumes that the data set is split in nblocks blocks across computation nodes.

Algorithm Parameters#

At the training stage, implicit ALS recommender in the distributed processing mode has the following parameters:

Training Parameters for Implicit Alternating Least Squares Computation (Distributed Processing)#

Parameter

Default Value

Description

computeStep

Not applicable

The parameter required to initialize the algorithm. Can be:

  • step1Local - the first step, performed on local nodes

  • step2Master - the second step, performed on a master node

  • step3Local - the third step, performed on local nodes

  • step4Local - the fourth step, performed on local nodes

algorithmFPType

float

The floating-point type that the algorithm uses for intermediate computations. Can be float or double.

method

fastCSR

Performance-oriented computation method for CSR numeric tables, the only method supported by the algorithm.

nFactors

\(10\)

The total number of factors.

maxIterations

\(5\)

The number of iterations.

alpha

\(40\)

The rate of confidence.

lambda

\(0.01\)

The parameter of the regularization.

preferenceThreshold

\(0\)

Threshold used to define preference values. \(0\) is the only threshold supported so far.

Computation Process#

At each iteration, the implicit ALS training algorithm alternates between re-computing user factors (\(X\)) and item factors (\(Y\)). These computations split each iteration into the following parts:

  1. Re-compute all user factors using the input data sets and item factors computed previously.

  2. Re-compute all item factors using input data sets in the transposed format and item factors computed previously.

Each part includes four steps executed either on local nodes or on the master node, as explained below and illustrated by graphics for \(\mathrm{nblocks} = 3\). The main loop of the implicit ALS training stage is executed on the master node.

Implicit Alternating Least Squares Computation: Part 1#

Implicit Alternating Least Squares Computation: Part 2#

Step 1 - on Local Nodes#

This step works with the matrix:

  • \(Y^T\) in part 1 of the iteration

  • \(X\) in part 2 of the iteration

Parts of this matrix are used as input partial models.

Training with Implicit Alternating Least Squares: Distributed Processing, Step 1 - on Local Nodes#

In this step, implicit ALS recommender training accepts the input described below. Pass the Input ID as a parameter to the methods that provide input for your algorithm. For more details, see Algorithms.

Input for Implicit Alternating Least Squares Computation (Distributed Processing, Step 1)#

Input ID

Input

partialModel

Partial model computed on the local node.

In this step, implicit ALS recommender training calculates the result described below. Pass the Result ID as a parameter to the methods that access the results of your algorithm. For more details, see Algorithms.

Output for Implicit Alternating Least Squares Computation (Distributed Processing, Step 1)#

Result ID

Result

outputOfStep1ForStep2

Pointer to the \(f \times f\) numeric table with the sum of numeric tables calculated in Step 1.

Step 2 - on Master Node#

This step uses local partial results from Step 1 as input.

Training with Implicit Alternating Least Squares: Distributed Processing, Step 2 - on Master Node#

In this step, implicit ALS recommender training accepts the input described below. Pass the Input ID as a parameter to the methods that provide input for your algorithm. For more details, see Algorithms.

Input for Implicit Alternating Least Squares Computation (Distributed Processing, Step 2)#

Input ID

Input

inputOfStep2FromStep1

A collection of numeric tables computed on local nodes in Step 1.

Note

The collection may contain objects of any class derived from NumericTable except the PackedTriangularMatrix class with the lowerPackedTriangularMatrix layout.

In this step, implicit ALS recommender training calculates the result described below. Pass the Result ID as a parameter to the methods that access the results of your algorithm. For more details, see Algorithms.

Output for Implicit Alternating Least Squares Computation (Distributed Processing, Step 2)#

Result ID

Result

outputOfStep2ForStep4

Pointer to the \(f \times f\) numeric table with merged cross-products.

Step 3 - on Local Nodes#

On each node \(i\), this step uses results of the previous steps and requires that you provide two extra matrices Offset Table i and Input of Step 3 From Init i computed at the initialization stage of the algorithm.

The only element of the Offset Table i table refers to the:

The Input Of Step 3 From Init is a key-value data collection that refers to the outputOfInitForComputeStep3 output of the initialization stage:

Training with Implicit Alternating Least Squares: Distributed Processing, Step 3 - on Local Nodes#

In this step, implicit ALS recommender training accepts the input described below. Pass the Input ID as a parameter to the methods that provide input for your algorithm. For more details, see Algorithms.

Input for Implicit Alternating Least Squares Computation (Distributed Processing, Step 3)#

Input ID

Input

partialModel

Partial model computed on the local node.

offset

A numeric table of size \(1 \times 1\) that holds the global index of the starting row of the input partial model. A part of the key-value data collection offsets computed at the initialization stage of the algorithm.

In this step, implicit ALS recommender training calculates the result described below. Pass the Result ID as a parameter to the methods that access the results of your algorithm. For more details, see Algorithms.

Output for Implicit Alternating Least Squares Computation (Distributed Processing, Step 3)#

Result ID

Result

outputOfStep3ForStep4

A key-value data collection that contains partial models to be used in Step 4. Each element of the collection contains an object of the PartialModel class.

Step 4 - on Local Nodes#

This step uses the results of the previous steps and parts of the following matrix in the transposed format:

  • \(X\) in part 1 of the iteration

  • \(Y^T\) in part 2 of the iteration

The results of the step are the re-computed parts of this matrix.

Training with Implicit Alternating Least Squares: Distributed Processing, Step 4 - on Local Nodes#

In this step, implicit ALS recommender training accepts the input described below. Pass the Input ID as a parameter to the methods that provide input for your algorithm. For more details, see Algorithms.

Input for Implicit Alternating Least Squares Computation (Distributed Processing, Step 4)#

Input ID

Input

partialModels

A key-value data collection with partial models that contain user factors/item factors computed in Step 3. Each element of the collection contains an object of the PartialModel class.

partialData

Pointer to the CSR numeric table that holds the \(i\)-th part of the input data set, assuming that the data is divided by users/items.

inputOfStep4FromStep2

Pointer to the \(f \times f\) numeric table computed in Step 2.

In this step, implicit ALS recommender training calculates the result described below. Pass the Result ID as a parameter to the methods that access the results of your algorithm. For more details, see Algorithms.

Output for Implicit Alternating Least Squares Computation (Distributed Processing, Step 4)#

Result ID

Result

outputOfStep4ForStep1

Pointer to the partial implicit ALS model that corresponds to the \(i\)-th data block. The partial model stores user factors/item factors.

outputOfStep4ForStep3

Pointer to the partial implicit ALS model that corresponds to the \(i\)-th data block. The partial model stores user factors/item factors.