MassiveClicks: A Massively-parallel Framework for Efficient Click Models Training

MassiveClicks is a multi-node multi-GPU framework for training click models using expectation maximization (EM). The framework supports heterogeneous GPU architectures, variable numbers of GPUs per node, and allows for multi-node multi-core CPU-based training when no GPUs are available. The following click models are currently supported:

1. Position-based Model (PBM).
2. User Browsing Model (UBM).
3. Click Chain Model (CCM).
4. Dynamic Bayesian Network Model (DBN).

MassiveClicks builds upon the generic EM-based algorithm for CPU-based single-node click model training, ParClick.

Requirements

• CUDA version: 12.1
• MPI version: 3.1
• C++ version: C++11 or higher

Installation

1. Clone the repository: git clone https://github.com/skip-th/MassiveClicks.git
2. Navigate to the project directory: cd MassiveClicks
3. Run the installation commands: cmake . && make

Usage

Here is a basic example of how to run MassiveClicks:

./mclicks --raw-path 'dataset.txt' --max-sessions 40000 --itr 50 --model-type 0 --partition-type 0 --test-share 0.2

The above command will train a PBM click model on 40000 sessions from the dataset dataset.txt for 50 iterations. Sessions are assigned to all available GPUs in a round-robin fashion. 20% of the dataset is used as the test set.

The dataset queries are assumed to be in the following format:

<session_id> <time_passed> <event_type> <query_id> <region_id> <document_id_0> ... <document_id_9>

And the clicks that follow a query are assumed to be in the following format:

<session_id> <time_passed> <event_type> <document_id>

Only clicks containing a document ID occurring in the immediately preceding query are considered.