This repo provides notebooks to (i) identify green/CCMT patent families via CPC (Y02/Y04), (ii) match to non-green patents using text embeddings and ANN (HNSW), (iii) cluster/name green tech groups (HDBSCAN/UMAP), (iv) aggregate to firms/cities/countries and (v) analyze data in the paper "Green building blocks reveal the complex anatomy of climate change mitigation technologies".

• OS: Linux / macOS / Windows (WSL2 recommended on Windows)
• Python: 3.11 (see requirements.txt for pinned packages)
• R: 4.4.3 (for 5.3_entry_reg_R.ipynb, 7.3_complement_reg_R.ipynb)
• Optional: Stata (for 5.4_marginplot_stata.ipynb).

Install with conda:

conda create -n gbb python=3.11 -y
conda activate gbb
conda install r=4.4.3 r-irkernel
pip install -r requirements.txt

Typical install time on a normal desktop: 10–25 minutes (most time is Python wheel downloads; nmslib may compile on some platforms).

Open notebooks to see the processing codes and results on whole datset.

A tiny demo dataset is in sampledata/, which are not the whole data used but a few sampled rows to show the structure. Execution of notebooks on these data will not get the same results.

Paths are now configured via an auto-inserted setup cell. You can also set an environment variable to relocate the project:

export GBB_PROJECT_ROOT=/path/to/this/repo

A download of Patentsview dataset and PATSTAT (license needed) is necessary, and the raw data is recommended to store in Parquet files for faster loading.

Run in order: 0_get_familiy_id.ipynb → 1_ccmt_matching.ipynb → 2_cpc.ipynb → 3.1_cluster_GBB.ipynb → 3.2_cluster_sourcefields.ipynb → 3.3_name_clusters.ipynb → 4_assocition_source_gbb.ipynb → 5.1_firm_agg.ipynb → 5.2_data4reg.ipynb → 5.3_entry_reg_R.ipynb (R) → 5.4_marginplot_stata.ipynb (Stata) → 6_viz_firm_cities.ipynb → 7.1_firm_primary_ctry.ipynb → 7.2_complement_firm_ctry.ipynb → 7.3_complement_reg_R.ipynb (R) → 7.4_fig_complement.ipynb

• Parquet IO requires pyarrow (already pinned).
• R notebooks require IRkernel if run in Jupyter.
• Stata notebook can be run in Stata directly or using a Jupyter kernel configured for Stata.
• The matching and clustering step requires large memory, typically need memory of 64GB or larger for the whole patent dataset.

See LICENSE.