ModelING Litigation in USPTO patent Claims using Doc2Vec

“If you don't look/think too hard this too will make sense”

-Someone

Zachary D White Feb - 2019

Capstone

How to do a capstone

Your Problem My Project

Initial INQUIRY

Can machine learning inform and return better search results for an FTO analysis thereby decreasing costs?

Where do we begin?

Revised Definition

The scope of a patent is all the items (invented or not) that are stated directly in the claims section of the patent.

Practical Definition

The universe of inventions that infringe on the patent.

WHAT IS PATENT INFRINGEMENT?

Going Deeper

Oldest Method - 1990 -2001
- How Broad Should the Scope of Patent Protection Be? - Klemperer
- Characteristics of Patent Litigation - Lanjouw. & Schankerman
Stanford Method 2015
- THE UNINFORMED TOPOGRAPHY OF PATENT SCOPE - Freilich
USPTO Method - 2016
- Patent Claims and Patent Scope - Marco & Sarnoff
- Patent Litigation and USPTO Trials - Marco & Miller
Kuhn's Methods - 2017
- How to Measure and Draw Causal Inferences with Patent Scope - Kuhn & Thompson
- Patent-to-Patent Similarity: A Vector Space Model - Kuhn

Patent Claim

Terms

ICC - Independent Claim Count is the number of independent claims made by a patent

ICL - Independent Claim Length is the number of words in said claim**.

PC - Primary Claim is the first independent Claim listed in a patent.

Project Goal

Use patent claim's probability of litigation to rank search results.

Obtain

DF.INFO()

Number of patent claims

The claim number for a particular patent

Int64Index: 8216031 entries, 0 to 8216149
Data columns (total 3 columns):
pat_no       object
claim_no     int64
claim_txt    object
dtypes: int64(1), object(2)
memory usage: 250.7+ MB

The text contained in the claim

The Patent ID number of a particular claim

DF.head()

df = df = pd.read_csv('claim_2000_2014_v001.csv')
df.head()


 	pat_no 	    claim_no 	    claim_txt
0 	8697278 	17 	17. Battery comprising an interior of the batt...
1 	7385756 	81 	81. A catadioptric projection objective for im...
2 	7387146 	1 	1. A heavy duty tire comprising a tread portio...
3 	7387253 	43 	43. A system comprising: (a) a optical reader ...
4 	7387278 	17 	17. A parachute ripcord pin for holding a para...

Scrub DATA &Create FEATURES

     pat_no 	claim_no      claim_txt 	                                ICL    ICC   litigation
     8697278    17        battery comprising interior battery active ele...     106    2     0
     8697278    1         battery cell casing comprising first casing el...     97     2     0
     7385756    81        catadioptric projection objective imaging patt...     108    33    0
     7385756    94        catadioptric projection objective imaging patt...     116    33    0
     7385756    79        catadioptric projection objective imaging patt...     103    33    0

Text Cleaning
- Removed legal stop words
- Lower cased
- Standard text cleaning
Removed Dependent Claims

Computed
- Independent Claim Length (ICL)
- Independent Claim Count (ICC)
Remove patents prior to 2000
Merged Patent-wise Litigation Data from Stanford

explore

df['litigation'].value_counts()

0    8185679
1      30352
Name: litigation, dtype: int64

0 = number of non-litigated claims
1 = number of litigated claims

df['litigation'].value_counts(normalize=True)

Non-Litigated Claims  99.63%
Litigated Claims       0.37%
Name: type, dtype: float64

All Claims

DF.describe()

All Claims 
df.loc[:,['ICL','litigation']].groupby(['litigation']).agg(['count','min','max','mean','std'])

 	ICL
	count 	min 	max 	mean 	        std
litigation 					
0 	8185679 1 	18433 	104.827044 	73.717400
1 	30352 	1 	1509 	94.007380 	62.130354

 	ICC
	count 	 min 	max 	mean 	        std
litigation 					
0 	2907222  1 	276 	2.814694 	2.326064
1 	7526 	 1 	358 	4.030162 	7.705218

#Two Tailed Independent T-Test
 	statistic 	pvalue
ICL 	30.260286 	3.155879e-198
ICC 	-13.683259 	4.067142e-42

Unsupervised Learning

Supervised Learning

Modeling Techniques

Gensim Doc2Vec Language Model

Doc2vec

## Sample of a tagged document
next(corpus_for_doc2vec.__iter__())

>TaggedDocument(words=['battery', 'comprising', 'interior', 'battery', 'active', 'elements', 'battery',
 'cell', 'casing', 'said', 'cell', 'casing', 'comprising', 'first', 'casing', 'element', 'first', 'contact', 
'surface', 'second', 'casing', 'element', 'second', 'contact', 'surface', 'wherein', 'assembled', 'position',
 'first', 'second', 'contact', 'surfaces', 'contact', 'first', 'second', 'casing', 'elements', 'encase',
 'active', 'materials', 'battery', 'cell', 'interior', 'space', 'wherein', 'least', 'one', 'gas', 'tight',
 'seal', 'layer', 'arranged', 'first', 'second', 'contact', 'surfaces', 'seal', 'interior', 'space', 
'characterized', 'one', 'first', 'second', 'contact', 'surfaces', 'comprises', 'electrically', 'insulating',
 'void', 'volume', 'layer', 'first', 'second', 'contact', 'surfaces', 'comprises', 'formable', 'material', 
'layer', 'fills', 'voids', 'surface', 'void', 'volume', 'layer', 'hermetically', 'assembled', 'position',
 'form', 'seal', 'layer'], tags=['8697278-17'])

Build a Language Model

import gensim
from gensim.models.doc2vec import Doc2Vec, TaggedDocument
assert gensim.models.doc2vec.FAST_VERSION > -1 #parallelize doc2vec
from gensim.utils import simple_preprocess


class MyDataframeCorpus(object):
    def __init__(self, source_df, text_col, tag_col):
        self.source_df = source_df
        self.text_col = text_col
        self.tag_col = tag_col
      
    def __iter__(self):
        for i, row in self.source_df.iterrows():
            yield TaggedDocument(words=simple_preprocess(row[self.text_col]), 
                                 tags=[row[self.tag_col]])

#Model 1
corpus_for_doc2vec = MyDataframeCorpus(df, 'claim_txt', 'claim_no')

#Model 2
corpus_for_doc2vec = MyDataframeCorpus(df, 'claim_txt', 'litigation')

%%time

#Params
model_smple = Doc2Vec(vector_size=100, # 100 should be fine based on the standards
                window=8, #number of context words
                alpha=.025, #initial learning rate
                min_alpha=0.00025, #learning rate drops linearly to this
                min_count=2, #ignores all words with total frequency lower than this.
                dm =1, #algorith 1=distributed memory , 0=distributed bag of words (PV-DBOW)
                epochs=30,
                workers=workers)#cores to use


model_smple.build_vocab(corpus_for_doc2vec,progress_per=500000)

## Build vocab from streaming tagged document Show progress every 500k rows
## Should take ~ 30 minutes

# Train Language Model
model_smple.train(corpus_for_doc2vec, total_examples=model_smple.corpus_count, epochs=model_smple.epochs)


>>Wall time: 22h 02min 24s

Model In Action

def model_infer_test(df,model):
    #grab a random claim from random_claim function
    tagged_claim = next(random_claim(df))
    
    #Infer a vector from that claim
    inferred_vector = model.infer_vector(tagged_claim.words)
    sims = model.docvecs.most_similar([inferred_vector], topn=len(model.docvecs)) 
    
    #Print top 5 similar claims
    print(sims[:5])

    # Compare and print the most/median/least similar documents from the train corpus
    print('Test Document ({}): «{}»\n'.format(tagged_claim.tags, ' '.join(tagged_claim.words)))
    print(u'SIMILAR/DISSIMILAR DOCS PER MODEL %s:\n' % model)
    for label, index in [('MOST', 0),('Second', 1) ,('MEDIAN', len(sims)//2), ('LEAST', len(sims) - 1)]:
        
        print(label,index)
        #print(u'%s %s: «%s»\n' % (label, sims[index], ' '.join(df[sims[index][0]].words)))
        print(u'%s %s: «%s»\n' % (label, sims[index], df.loc[df['claim_no']==sims[index][0],['claim_txt']].values))

# This will let you get random tagged documents
def random_claim(df):
    doc_index =  np.random.randint(0,len(df))
    #row = [df.iloc[doc_index,:]['claim_txt'],df.iloc[doc_index,:]['claim_no']]
    yield TaggedDocument(words=simple_preprocess(df.iloc[doc_index,:]['claim_txt']), 
                             tags=df.iloc[doc_index,:]['claim_no'])

model_infer_test(df,model_smple)

[('6208475-7', 0.9243185520172119), ('6208475-3', 0.883671224117279), ('6680491-1', 0.6106572151184082),
 ('8913211-14', 0.5911911725997925), ('8913211-1', 0.5858227610588074)]

Test Document (6208475-7): «optical member inspection apparatus obtaining image data used inspections applying
 illumination light optical member one side thereof photographing optical member side comprising holder 
comprising frame shaped base portion plurality spaced optical member holding portions provided base portion
 enclose space supporting optical member formed base portion said optical member holding portions receiving 
face outer margin optical member mounted restriction wall restrict movement optical member mounted receiving
 face contacting outer margin optical member said receiving face contacting outer margin face optical member
 light passes operation optical member said restricting wall located said moving face»


SIMILAR/DISSIMILAR DOCS PER MODEL Doc2Vec(dm/m,d100,n5,w5,mc2,s0.001,t16):

MOST 0

MOST ('6208475-7', 0.9243185520172119): «[['optical-member inspection apparatus obtaining image data used 
inspections applying illumination light optical member one side thereof photographing optical member side,
 comprising:a holder comprising frame shaped base portion plurality spaced optical-member holding portions
 provided base portion enclose space supporting optical member formed base portion, said optical-member 
holding portions receiving face outer margin optical member mounted restriction wall restrict movement 
optical member mounted receiving face contacting outer margin optical member, said receiving face contacting
 outer margin face optical member light passes operation optical member, said restricting wall located said 
moving face.']]»

#### OUTPUT CONTD

Second 1

Second ('6208475-3', 0.883671224117279): «[['holder holding outer margin optical member, comprising:a frame
 shaped base portion; anda plurality spaced optical-member holding portions provided base portion enclose 
space supporting optical member formed base portion, said optical-member holding portions receiving face 
outer margin optical member mounted restriction wall restrict movement optical member mounted receiving face
 contacting outer margin optical member, said receiving face contacting outer margin face optical member 
light passes operation optical member, said restricting wall located said receiving face.']]»

MEDIAN 4108015

MEDIAN ('6592902-1', 0.08781707286834717): «[['pharmaceutical composition comprising (a) particulate 
eplerenone d90 particle size 25 400 microns, amount 10 mg 1000 mg, (b) one pharmaceutically acceptable 
carrier materials; said composition controlled release oral dosage form wherein 50% said eplerenone 
dissolved vitro least 1.5 hours 1% sodium dodecyl sulfate solution 37� c.']]»

LEAST 8216030

LEAST ('6823942-1', -0.3702787756919861): «[['tree saver apparatus preventing excess pressure christmas tree
 well, wherein christmas tree comprises tubing, master valve, top valve, second master valve, wing valve, 
wherein apparatus comprises:a. hydraulic system comprising: i. piston connected piston rod; ii. cylinder 
connected piston; iii. upper plunger connected piston; iv. hand wheel connected upper plunger; b. connection
 least 3 outlets connected cylinder; c. frac wing valve connected connection least 3 outlets; d. frac valve
 first port second port, wherein first port connected connection least 3 outlets; e. mandrel first end 
second end, wherein first end connected piston rod; f. landing bowl connected top valve.']]»

Data transformation

claim_no       ICL    ICC    litigation    0 	              1 	      2 	      3 	      4      5 	6 	... 	90 	91 	92 	93 	94 	95 	96 	97 	98 	99 																					
8697278-17 	106 	2 	0 	-0.021915 	0.124170 	-0.467477 	1.725972 	1.353391 	-0.197958 	1.234415 	... 	-1.201108 	-0.272672 	0.649778 	-1.023098 	0.146432 	0.122571 	0.215042 	-1.076661 	-0.508896 	-1.690101
8697278-1 	97 	2 	0 	-0.271349 	0.260269 	-0.315466 	1.746054 	0.874254 	-0.225060 	0.971804 	... 	-1.190130 	-0.417508 	0.587534 	-1.157797 	0.314950 	0.189521 	0.474607 	-1.119020 	-0.376173 	-1.699343
7385756-81 	108 	33 	0 	1.429105 	0.936555 	-0.296551 	-1.073354 	-0.421059 	-0.114407 	-0.881726 	... 	0.885844 	0.934247 	0.678008 	-0.752642 	-0.121980 	-0.623260 	-2.384536 	0.471751 	0.874406 	1.251178
7385756-94 	116 	33 	0 	1.027478 	0.484034 	-0.535233 	-1.124252 	-0.382812 	-0.101879 	-0.348340 	... 	0.772912 	0.951951 	0.901475 	-0.721144 	-0.304449 	-0.278908 	-2.655865 	0.323664 	0.943390 	0.687369
7385756-79 	103 	33 	0 	1.396978 	0.419591 	-0.072397 	-0.642861 	-0.482944 	0.496666 	-1.178749 	... 	0.074441 	1.187986 	0.909484 	0.278953 	-0.483475 	-0.028059 	-2.268470 	0.986993 	0.039141 	0.374451

Modeling Techniques

Gensim Doc2Vec Language Model

XGBOOST Logistic Regression

Supervised Learning

XGBOOST

Multi Experiment TTS

### Test Train Split
y = df['litigation']
X = df.drop(['litigation','claim_no'],axis=1)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30,random_state=42)

#free up memory
del df,y,X

#Class balance test and train
y_test.value_counts(normalize=True)

0    0.996355
1    0.003645
Name: litigation, dtype: float64

y_train.value_counts(normalize=True)
0    0.996285
1    0.003715
Name: litigation, dtype: float64

# Convert input data from numpy to XGBoost format
dtrain = xgb.DMatrix(X_train, label=y_train)
dtest = xgb.DMatrix(X_test, label=y_test)

#again to conserve memory
del X_train, X_test, y_train, y_test

from sklearn.model_selection import train_test_split 
import xgboost as xgb

##ICL  - Independent Claim Length
##ICC  - Independent Claim Count
##100F - One Hundred Length Feature Vector from Doc2Vec language model

# Model ICL_100F
df= pd.read_csv('claims_vectors_litigation.csv',
                usecols=lambda x : x not in ['claim_txt','ICC'])
# Model ICC_100F
df= pd.read_csv('claims_vectors_litigation.csv',
                usecols=lambda x : x not in ['claim_txt','ICL'])
# Model ICL_ICC
df= pd.read_csv('claims_vectors_litigation.csv',
                usecols=['claim_no','ICL','ICC','litigation'])
#Model ICL_ICF_100F
df= pd.read_csv('claims_vectors_litigation.csv',
                usecols=lambda x : x not in ['claim_txt'])
# Model 100F
df= pd.read_csv('claims_vectors_litigation.csv',
                usecols=lambda x : x not in ['ICL','ICC','claim_txt'])

Training the Model

# Specify sufficient boosting iterations to reach a minimum
num_round = 5000

param = {'objective'   : 'binary:logistic', # Specify binary classification
         'eta'         : .5,
         'max_depth'   : 8,
         'predictor'   : 'gpu_predictor',
         #'verbosity'   : 3,
         'tree_method' : 'gpu_hist', # Use GPU accelerated algorithm
         'random_state': 42
        }

gpu_res = {} # Store accuracy result
tmp = time.time()
# Train model
tst_model = xgb.train(param, dtrain, num_round, evals=[(dtest, 'test')], evals_result=gpu_res)
print("GPU Training Time: %s seconds" % (str(time.time() - tmp)))

INITIAL Results

Turning water into wine

 model 	    % non-lit claims 	% lit claims 	RMSE 	    % predicted  #predicted
ICC_100F 	99.6355 	0.3645 	        0.003138 	0.0507    15
ICL_ICC_100F 	99.6355 	0.3645 	        0.003142 	0.0503    15
ICL_100F 	99.6355 	0.3645 	        0.003267 	0.0378    11
100F 	        99.6355 	0.3645 	        0.003283 	0.0362    11
ICL_ICC 	99.6355 	0.3645 	        0.003490 	0.0155    5

*predicted values are based on probability >.5

ICC - Independent Claim Count

ICL - Independent Claim Length

100_F - Embedded Language Vector

What do our results mean?

Text

Politely laugh at the image below

Test model on Less data!

.. A LOT LESS DATA

list_of_litigated_claims = []

For each element in list of litigated claims:
    Infer a vector
    Calculate Similarity for all other claims
    Return all Primary Claims of the 100 most similar claims

df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 205501 entries, 0 to 205500
Columns: 105 entries, Unnamed: 0 to 99
dtypes: float64(100), int64(4), object(1)
memory usage: 164.6+ MB


df['litigation'].value_counts()
0    197975 - 96.34%
1      7526 - 03.66%
Name: litigation, dtype: int64

Train a new model

Model prediction w/o language features

Try Try Again...one final time

Art of the Law Suit

Rich Patent Poor Patent


#Model 2
corpus_for_doc2vec = MyDataframeCorpus(df, 'claim_txt', 'litigation')

:(

 	model 	            % lit-predicted
0 	ICL_ICC_100F_LS_NLS 	0.007204
1 	ICL_100F_LS_NLS 	0.007104
2 	ICC_100F_LS_NLS 	0.006804
3 	100F_LS_NLS 	        0.007204
4 	ICL_ICC_LS_NLS 	       -0.039496

What do we do now?

Subtitle

Lessons learned

Medium Data is still hard to work with
Don't be afraid to go back to the basics
Use Joblib
Fancy Graphs don't save money, but they make us all warm inside.
~~Read research papers~~ Don't believe everything you read

Questions??

Oldest Method - 1990 -2001
- How Broad Should the Scope of Patent Protection Be? - Klemperer
- Characteristics of Patent Litigation - Lanjouw. & Schankerman
Stanford Method 2015
- THE UNINFORMED TOPOGRAPHY OF PATENT SCOPE - Freilich
USPTO Method - 2016
- Patent Claims and Patent Scope - Marco & Sarnoff
- Patent Litigation and USPTO Trials - Marco & Miller
Kuhn's Methods - 2017
- How to Measure and Draw Causal Inferences with Patent Scope - Kuhn & Thompson
- Patent-to-Patent Similarity: A Vector Space Model - Kuhn

Sources listed below