SMM

Teammates

王竣樺 梁致銓 曾繁斌

Patent Crawler

Build Env

1. python -m venv venv
2. venv/Script/Activate
3. pip install -r requirements.txt

Proxy Pool Usage

We currently utilizing 10 proxies provided by the free account of Webshare.

Warning

REMEMBER del the save page part of get_page() or your computer will be filled with htmls

Run Crawler

cd Patent_Search_Crawler
# crawling
python main.py
# merging files
python merger.py

Data

You can find the complete raw data in this link: link

• We use the data merge_data.sqlite for training.

Tip

The merge_data.sqlite should be put at SMM folder.

Explanation of the preprocessed data (in /SMM/EDGPAT/data):

Tip

The datasets below are for training, no need to preprocess.

• 2-1-level.csv: IPC Level mapping from 1 to 2.
• 3-2-level.csv: IPC Level mapping from 2 to 3.
• 4-3-level.csv: IPC Level mapping from 3 to 4.
• 5-4-level.csv: IPC Level mapping from 4 to 5.
• real-data.json: Illustrate the company's patents within the current year.

Note

we encode year 2018 as 0.

IPC Patent level example:

Final Project: Exploring Patent Trends in Taiwan with Event-based Graph Techniques

Our Goal

Our project focuses on developing a patent prediction model specifically for forecasting Taiwan's future patent trends. Utilizing Event-based Graph techniques, this model analyzes historical patent data to identify emerging trends and patterns.

• Data-Driven: Uses real-world patent data (Taiwan) to identify trends.
• Dynamic: Adapts to changes in technology and innovation.
• Predictive: Forecasts areas likely to see growth in patent filings.

Model FrameWork

Framework of the proposed model. We just show the calculations of the patent classification codes and one of the related companies for simplicity.

Code

We utilized the code from EDGPAT

Warning

The Python env should be Python 3.6!

Required packages:

• PyTorch 1.7.1
• tqdm
• numpy
• pandas
• matplotlib

Preprocessing

Just run the build_input.ipynb

We split the data into three parts: training, validation and testing by year.

Training

Run the code:

sh EDGPAT/train.sh

Note

This code will ouptut the training result in EDGPAT/out.txt

Our Results

Origin Paper Results	Our Results
Recall NDCG PHR Top-10 0.1175 0.1725 0.5491 Top-20 0.1646 0.1742 0.6304 Top-30 0.1868 0.1769 0.6612 Top-40 0.2006 0.1769 0.6800	Recall NDCG PHR Top-10 0.1577 0.3395 0.8381 Top-20 0.2349 0.3364 0.8995 Top-30 0.2837 0.3401 0.9217 Top-40 0.3150 0.3427 0.9321

Droput 0.5	Dropout 0
Recall NDCG PHR Top-10 0.1668 0.3552 0.8407 Top-20 0.2363 0.3496 0.8851 Top-30 0.2762 0.3487 0.9073 Top-40 0.3051 0.3490 0.9164	Recall NDCG PHR Top-10 0.1577 0.3395 0.8381 Top-20 0.2349 0.3364 0.8995 Top-30 0.2837 0.3401 0.9217 Top-40 0.3150 0.3427 0.9321