December 07, 2020

by

0   Comments

It’s not been a “wonderland” for attorneys navigating the post-Alice world. 64% of patents have not survived after an Alice assertion. NPEs are feeling the sting as the federal courts continually strike down bad patents.

At AiLanthus, we wondered whether there was a better way to review patents for § 101 claims.

We started with our hypothesis: bad patents likely have similar language construction. We then figured that if we examined all cases where courts invalidated patents under § 101 we could determine whether a patent was likely to be invalidated.

We were right. We ran our algorithm across 650 cases in the Fenwick post-Alice database and found that our algorithm has 85% predictive accuracy, (we likely could have improved, but we decided not to highlight diagnostic patents out of caution. We think shortly the Supreme Court will make them patent eligible. See why here.)

How Our Algorithm Works:

We began our research with Aashish Karkhanis’ and Jenna L. Parenti’s 2016 article Toward an Automated First Impression on Patent Claim Validity. In that study, Karkhanis and Parenti found invalidated patents had frequent words that were unique to invalid patents. We originally replicated this experiment and found the same results: a 85% precision and 80% recall.

My cofounder an I thought we had cracked the code. We believed that if we expanded the data to include all of the invalidated patents post-Alice we could improve those results. We were wrong.

We realized when we expanded the dataset to include all of the invalidated patents in the Fenwick dataset, we had just created a method that would recognized every software patent as invalid.

We had to go back to the drawing board. We realized we should, instead train the model to identify the languages patterns of each of the types of the invalidated patents in the Fenwick dataset (e.g “Abstract Idea: Fundamental Economic Practice,” “Abstract Idea: Certain Methods of Organizing Human Activities,” etc.).

Without giving away our “secret-sauce,” we determined that if we could identify the language patterns and concepts that underlie each of these categories we could determine if a court would likely invalidate a patent on an Alice motion.

The results are astounding!

Take for example, U.S Patent No. 7,975,007, a method and system for facilitating marketing dialogues found here in Re: Mark Greenstein in the Fenwick dataset. The court reasoned that the patent presented nothing inventive because it merely disclosed a new abstract idea. See Synopsys, Inc. v. Mentor Graphics Corp., 839 F.3d 1138, 1151 (Fed. Cir. 2016) (“[A] claim for a new abstract idea is still an abstract idea.”).  We we ran the ‘007 patent through our platform and found the same results.  The computer brought up the same reasoning as the court did; the patent was merely an abstract idea implemented on a generic computer. We did it! We then ran this test across all of the different patents in the Fenwick dataset and found the results remarkable. We had achieved a 93% percent accuracy in determining whether the patent was invalid and a 90% accuracy for selecting the right reasoning.

This system is not bulletproof, however. Anecdotally, we found when our algorithm was incorrect, they were type-I errors with almost all of them stemming from cases decided by the Eastern District of Texas (no surprise!).

Notwithstanding, these results are truly spectacular.

Applications

We believe patent attorneys can benefit from the easy to use tool. With just a click of button, attorneys can lookup a patent to determine if it likely invalid patent. We believe that NPEs can use this tool to ensure they only assert valid patents and other patent holders can quickly determine whether they need to enter into a licensing agreement.