Hylance1224/CVH-REC: v1.2

CVH-REC: A novel method for web API recommendation based on the cross-view HGNNs This repository contains the dataset and source code for the paper:"CVH-REC: A novel method for web API recommendation based on the cross-view HGNNs" 1. Environment Setup Our code has been tested under Python 3.12.3, and the following packages are required: h5py==3.11.0 scikit_learn==1.4.2 sentence_transformers==2.7.0 torch==2.7.0+cu126 We recommend setting up the environment using Conda to ensure compatibility and ease of installation. ✅ Step 1: Create and activate a new Conda environment conda create -n CVH-Rec python=3.12.3 -y conda activate CVH-Rec ✅ Step 2: Install the required Python packages Install all dependencies (except PyTorch) via requirements.txt: pip install -r requirements.txt ⚠️ Note: The requirements.txt file includes all necessary libraries except for PyTorch, which should be installed separately to match your system and CUDA version. ✅ Step 3: Install PyTorch manually Visit the official PyTorch installation page here to choose the correct version for your environment.For example, to install PyTorch with CUDA 12.6, use: pip install torch --index-url https://download.pytorch.org/whl/cu126 2. Usage 2.1 Data generation (Optional) This step can be skipped — preprocessed data is already provided.Run this only if you want to apply CVH-REC on a new dataset. To generate the dataset, start Python in command line mode and run: python generate_dataset.py This script uses the original files Original Dataset/mashups.json and Original Dataset/apis.json as input and performs the following operations: Generates training set, testing set, and validation set under the dataset folder. Creates a data folder containing: API_vectors.h5 and vectors.h5: Semantic embedding vectors for API and mashup descriptions. api_tag_vector.h5 and mashup_tag_vector.h5: Semantic embedding vectors for API and mashup tags. Each dataset/foldX/ directory includes: RS.csv # Test set TE.csv # Training set VA.csv # Validation set api_tags.csv # API–tag relationships mashup_tags.csv # Mashup–tag relationships 2.2 Reproducing Experiments in the Paper 🔍 RQ1: Performance of CVH-REC ▶️ Step 1. Train the model and generate recommendation results for a single fold (e.g., fold1) python main.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method CVH-REC --patience 5 --lr 0.002 --epoch 200 After training completes, the recommendation results for fold1 will be automatically saved in JSON format under the ../output/ directory. 🔁 Cross-Validation To perform 10-fold cross-validation, repeat Step 1 for each fold by updating the --dataset parameter: --dataset fold1 --dataset fold2 … --dataset fold10 All other parameters (e.g., learning rate, patience, alpha values) can remain unchanged. ▶️ Step 2. Compute evaluation metrics (after all 10 folds are completed) Once all recommendation result of 10 folds generated, run the following command to compute the overall performance: python metrics.py --method CVH-REC ⚠️ Important:Do not run this step until all 10 folds have finished training.If any fold’s recommendation file is missing, the script will raise an error due to incomplete data. This step calculates the 10-fold average and standard deviation for all evaluation metrics. ▶️ Calculate the p-values and Cohen’s d values (CVH-REC vs. R2API) After obtaining the 10-fold cross-validation results of CVH-REC, follow the steps below to perform the statistical analysis: Run the baseline model R2API and generate its 10-fold recommendation results. Copy the resulting files of R2API into the ../output/ folder. Execute the following command to compute the statistical differences between the two methods: python statistical.py --method1 CVH-REC --method2 R2API This script will produce the** p-values** and Cohen’s d effect sizes, which quantify the statistical significance and the magnitude of performance differences between CVH-REC and R2API. 🔄 General Note for RQ2–RQ6 The training and evaluation workflow for subsequent research questions (RQ2-RQ6) is identical to RQ1. For each variant, the process follows the same three parts: Step 1 — Train the model and generate results for each fold Cross-Validation — Repeat Step 1 for all 10 folds (fold1–fold10) Step 2 — Compute evaluation metrics after all folds are completed Below are specific commands and configurations for each RQ variant. This step calculates the 10-fold average and standard deviation for all evaluation metrics. 🔍 RQ2: Impact of sub-hypergraphs, behavioral similarity sub-hypergraphs, and semantic similarity subhypergraphs (1) CVH-subHypergraphs ▶️Step1. Train and generate results for a single fold (fold1): python main_global.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Global --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Global (2) CVH-behavioralSimilarity ▶️Step1. Train and generate results for a single fold (fold1): python main_without_call_view.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_call_view --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_call_view (3) CVH-semanticSimilarity ▶️Step1. Train and generate results for a single fold (fold1): python main_without_tag_view.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_tag_view --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_tag_view 🔍 RQ3: Impact of contrastive learning, global contrastive learning, and local contrastive learning (1) CVH-contrastive ▶️Step1. Train and generate results for a single fold (fold1): python main_without_CL.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_CL --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_CL (2) CVH-globalContrastive ▶️Step1. Train and generate results for a single fold (fold1): python main_without_global_CL.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_global_CL --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_global_CL (3) CVH-localContrastive ▶️Step1. Train and generate results for a single fold (fold1): python main_without_local_CL.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_local_CL --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_local_CL 🔍 RQ4: Impact of multi-task learning (1) CVH-multi-task ▶️Step1. Train and generate results for a single fold (fold1): python main_without_multi.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Without_multi --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Without_multi 🔍 RQ5: Impact of requirement vector transfer (1) CVH(semantic) ▶️Step1. Train and generate results for a single fold (fold1): python main.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_semantic.py --dataset fold1 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method CVH-REC_semantic (2) CVH(R2API) ▶️Step1. Train and generate results for a single fold (fold1): python main.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_R2API.py --dataset fold1 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method CVH-REC_R2API 🔍 RQ6: Impact of assigning the semantic vectors (1) CHV-semantic ▶️Step1. Train and generate results for a single fold (fold1): python main_random_init.py --dataset fold1 --patience 5 --epoch 200 --lr 0.002 --alpha1 0.6 --alpha2 0.6 --alpha3 0.6 --alpha4 0.6 python main_text_encoder.py --dataset fold1 --method Random_init --patience 5 --lr 0.002 --epoch 200 🔁 Cross-Validation. Repeat for folds 2–10 by changing the --dataset parameter. ▶️Step 2. Compute metrics: python metrics.py --method Random_init 3. Hyperparameter setting and implementation details The hyperparameters of this project are defined in utility/parser.py. Below is the hyperparameter setting and implementation details. Variable name Description Setting alpha1 Mashup behavioral similarity threshold 0.6 a