repository·master·Indexed 25 days ago
https://github.com/alibaba/easyrecAn extendable and efficient framework for building industrial-scale recommendation systems. EasyRec supports candidate generation, ranking, and multi-task learning, featuring a variety of models including DeepFM, Wide & Deep, DCN, AutoInt, DSSM, and MIND. It includes tools for data preprocessing and inference via the EasyRecProcessor client, with provided examples and configurations for datasets such as Amazon Books, Criteo, and MovieLens-1M.
EasyRec is designed for industrial-scale recommendation systems with the following key features:
EasyRec provides a suite of specialized neural network layers via the easy_rec.python.layers module to facilitate the construction of recommendation models. These layers are organized by their functional purpose, such as input handling, embedding, attention mechanisms, and specific model architectures like FM or MMOE.
Available layer modules include:
easy_rec.python.layers.dnn: Deep Neural Network layers.easy_rec.python.layers.embed_input_layer: Embedding input layers for handling categorical features.easy_rec.python.layers.fm: Factorization Machine layers.easy_rec.python.layers.input_layer: Standard input layers.easy_rec.python.layers.layer_norm: Layer Normalization layers.easy_rec.python.layers.seq_input_layer: Sequence input layers for sequential recommendation tasks.easy_rec.python.layers.multihead_attention: Multi-head attention mechanisms.easy_rec.python.layers.mmoe: Multi-gate Mixture-of-Experts (MMOE) layers for multi-task learning.EasyRec provides several advanced features for recommendation workflows:
ParameterServer, Mirrored, and MultiWorker.knn algorithm on vectors in a distributed environment.EasyRec provides a variety of input modules within the easy_rec.python.input namespace to handle different data sources for recommendation tasks. Depending on your data storage and streaming requirements, you can use the following input types:
csv_input or csv_input_v2 for reading data from local or distributed CSV files.kafka_input for real-time streaming data from Kafka topics.odps_input, odps_input_v2, or odps_rtp_input for data stored in Alibaba Cloud's ODPS/MaxCompute.rtp_input or rtp_input_v2 for real-time data processing pipelines.All input modules are designed to integrate with the EasyRec framework to feed data into models for training or inference.
AITM is a multi-task learning framework designed for recommendation scenarios where users follow a sequential conversion path (e.g., Impression $\rightarrow$ Click $\rightarrow$ Conversion). It aims to improve the estimation of backend conversion rates by utilizing samples from all nodes in the conversion chain.
Key features include:
ORDER_CALIBRATE_LOSS) to correct predictions based on the sequential dependency of behaviors.FiBiNet is a recommendation model designed to improve Click-Through Rate (CTR) prediction by combining two core modules:
It is implemented as a RankModel within EasyRec.
Rocket Launching is a framework designed for online real-time response systems (like CTR prediction) where strict latency requirements prevent the use of deep, complex models during inference.
How it works:
soft target, which improves the training effectiveness of the lightweight model.RTP FG (RealTime Predict Feature Generation) is a framework designed to handle feature engineering requirements for real-time prediction. It efficiently generates complex cross-features, such as match feature and lookup feature, using C++ code to ensure consistency between offline training and online prediction.
Key capabilities:
pipeline.config files from an fg.json configuration.The Feature Generator (FG) is a data transformation module that converts raw input into the features required by a model. Its primary purpose is to ensure consistency between offline and online sample generation.
Key characteristics:
expr feature and lookup feature.EasyRec Processor.array, map), custom operators, discretization (binning), and feature dependencies.The PAIAssessor is used to compare current trial results against historical results within the same group. If the results do not meet specific criteria (e.g., falling below a certain threshold), the assessor can stop the execution of that hyperparameter group to save resources. This means the actual number of trials run may be significantly less than the configured maxTrialNumber.
Configuration Options for PAIAssessor (via classArgs in exp.yml):
| Option | Description | Values |
|---|---|---|
optimize_mode | Direction of optimization | maximize / minimize |
start_step | The step number at which early stopping decisions begin | Integer (e.g., 2) |
moving_avg | Whether to use a moving average of all historical results as the benchmark | True / False |
proportion | The ratio used to compare the current best value against historical records | Float (e.g., 0.5) |
patience | Number of consecutive decreases in the metric before stopping | Integer (e.g., 10) |
assessor:
name: PAIAssessor
classArgs:
platform: MAXCOMPUTE
optimize_mode: maximize
start_step: 1
moving_avg: true
proportion: 0.5The UNITER (UNiversal Image-TExt Representation Learning) model is a Transformer-based architecture designed to fuse multi-modal information (such as images and text) for recommendation tasks. In EasyRec, the model is extended to support item attributes and statistical features.
To use UNITER, you must organize your input features into one of four specific feature_groups. You do not need all four; providing at least one is sufficient for training. If a group is not provided, the corresponding network structure is skipped.
image: Contains image or video features (e.g., a single embedding or a set of embeddings representing multiple frames/regions).general: Contains regular features (numerical, single-value categorical, or multi-value categorical) that participate in cross-modal attention. Note: All features in this group must have the same embedding_dim.text: Contains variable-length text sequences (e.g., title, description) that participate in cross-modal attention. Note: All features in this group must have the same embedding_dim.other: Contains features that do not participate in cross-modal attention (e.g., statistical features). These are concatenated directly with the output of the cross-modal fusion module before the final MLP.Critical Requirement: For the cross-modal attention mechanism to work, the embedding_dim of all features in both the text and general groups must be consistent.
model_config: {
model_class: 'Uniter'
feature_groups: {
group_name: 'image'
feature_names: 'embedding'
}
feature_groups: {
group_name: 'general'
feature_names: 'user_id'
feature_names: 'gender'
}
feature_groups: {
group_name: 'text'
feature_names: 'title'
}
feature_groups: {
group_name: 'other'
feature_names: 'score_time'
}
}