- Category
- AI Tools
- Rank
- No. 1157Tools index
Previous survey · No. 1165 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- hkuds
- GitHub
- 538 stars
- Date
About
WSDM 2024 paper using LLMs with graph augmentation for recommendation systems, including multi-modal and side-information signals.
What it does
LLMRec enriches movie recommendation data before training. It asks a language model to choose preferred and disliked candidates, infer user profiles, and fill item attributes. A PyTorch recommender then combines those additions with interaction, image, and text features.
Why it's ranked here
This is a useful research artifact because it exposes prompts, processed data, training logic, evaluation metrics, baselines, and ablation controls. Reproduction still looks fragile: documented dataset names conflict with training branches, GPU use is effectively mandatory, and augmentation requires endpoint configuration.
What's good
The repository connects augmentation to training instead of stopping at prompt examples. It supplies processed collaborative-filtering data, original and enhanced text, visual features, and embeddings. Training supports controlled removal of interaction samples, user profiles, item attributes, and pruning for practical ablation studies.
Tradeoffs
The documented workflow covers only Netflix and MovieLens. Several tensor and model operations force CUDA despite an apparent CPU fallback. The dependency list resembles a broad environment snapshot rather than a focused package. The augmentation script contains an empty data location and a private service URL with placeholder authorization.
How to use it well
Use it as a research starting point for recommender experiments where you can inspect data assumptions, configure an LLM endpoint, and repair environment details. Generate augmented records first, then train and compare ablations. It does not provide a production recommendation service, reusable library package, or public hosted API.
Technical notes+
README.md documents a two-stage script workflow and requirements.txt pins a large Python environment. main.py loads pickled interaction and augmentation artifacts, NumPy image and text features, then sends sparse graphs, models, and a decoder directly to CUDA. Models.py projects image, text, user-profile, and item-attribute features into a shared embedding space before adding them to graph-propagated user and item embeddings. utility/parser.py advertises netflix and movieLens, but main.py initializes its attribute dictionary only for preprocessed_raw_MovieLens and netflix_valid_item, so the documented default netflix can leave augmented_total_embed_dict undefined. utility/load_data.py increments self.n_val without first initializing it. LLM_augmentation_construct_prompt/gpt_ui_aug.py builds candidate prompts and writes pickled positive and negative samples, but ships with an empty file_path and a private Baidu endpoint.
Observed
- Primary language
- Python
- Install surface
- Dependency installation through pip from requirements.txt
- Interface
- Command-line Python scripts for augmentation and recommender training
- Framework
- PyTorch training code with SciPy sparse graphs and NumPy feature files
- Documented datasets
- Netflix and MovieLens
- Distribution form
- Research source repository with code, data links, prompts, baselines, and training scripts
Read from README.md, requirements.txt, main.py, Models.py, MMSSL/MMD.py, MMSSL/main.py, MMSSL/Models.py, utility/norm.py, utility/parser.py, utility/logging.py, utility/metrics.py, utility/load_data.py, utility/batch_test.py, LLM_augmentation_construct_prompt/gpt_ui_aug.py.
What it can do
Generate personalized recommendations using large language models
User interaction data and preferences → Ranked list of recommended items
Process multi-modal content for recommendation enhancement
Text, images, and other media content → Enhanced item representations
Augment recommendations with graph-based relationships
User-item interaction graph and side information → Graph-augmented recommendation scores
Incorporate side information signals into recommendation process
Item metadata, user demographics, contextual data → Enriched recommendation predictions
Train recommendation models using LLM-based approaches
Training datasets with user-item interactions → Trained recommendation model
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.