summary | demo web app | usage | walk through notebooks | license

Machine learning-based patent signals for technology decisions

Techniche is a recommendation engine-based decision support tool to help business users surface technology ideas from patent documents for machine learning inventions.

Technology decision-makers - in engineering, people and product - require data to make choices in markets shaped by machine-learning technologies. Techniche recommends technology ideas based on the pipeline of underlying machine learning technologies in patents.

Dependencies are specified in requirements.txt

pip install -r requirements.txt

To run the notebooks locally, you will need to have python installed, preferably through anaconda .

You can then clone the techniche repository. From a command line, run

git clone https://github.com/glmack/techniche.git

Move into the techniche directory:

cd techniche

Set up software environment with the provided conda environment:

conda env create -f environment.yml
conda activate techniche_env

Launch Jupyter in your web browser

jupyter notebook

Walk-through notebooks are available in the model selection directory.

Techniche learns from public patent documents of the United States Patent Organization (USPTO) that are made available through the PatentsView API, dump files of the PatentsView backend database, and supplementary files containing full patent documents not available throuh the API. Users can explore the data used in Techniche via the PatentsView graphical user interface.

Natural language pre-processing techniques such as word tokenization and punctuation cleaning are applied to raw text data from patent titles and summary descriptions prior to introduction to models.

Explore notebooks detailing data preparation and modeling in topic_model.ipynb and rec_system.ipynb of the model selection directory.

Techniche predicts technology recommendations using a hybrid recommender system. At the current stage of development, a collaborative filtering recommender component uses matrix factorization based on the Spark implementation of alternating least squares (ALS). A content-based recommender component, currently under development, addresses the cold start problem associated with making predictions for new users and items. The recommender will use text-based document (item) similarity metrics and also elicit user preferences through the web app. Latent Dirichlet Allocation (LDA), an unsupervised set of topic models is explored to generate the probable range of topics expressed in patent documents for machine learning-based inventions.

Recommendations are evaluated in terms of relevance to technology decision-makers. Intermediate intrinsic evaluation metrics, such as coherence and perplexity metrics for LDA provide additional diagnostics.

Techniche will be made available for user experimentation as a Flask web app that offers a search interface where - as an intermediate demo step - users can input text strings describing technical areas and return predicted topics and their associated word co-occurences.