Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
DOI:10.5121/mlaij.2020.7402 15
A DEEP LEARNING MODEL TO PREDICT
CONGRESSIONAL ROLL CALL VOTES FROM
LEGISLATIVE TEXTS
Jona han Wayne Ko n and Ma k A. Newman
Depa men o Da a Science, Ha isbu g Uni e si y, Ha isbu g, Pennsyl ania
ABSTRACT
De elopmen s in na u al language p ocessing (NLP) echniques, con olu ional neu al ne wo ks (CNNs),
and long-sho - e m memo y ne wo ks (LSTMs) allow o a s a e-o - he-a au oma ed sys em capable o
p edic ing he s a us (pass/ ail) o cong essional oll call o es. The pape in oduces a cus om hyb id
model labeled "P edic Tex Classi ica ion Ne wo k" (PTCN), which inpu s legisla ion and ou pu s a
p edic ion o he documen 's classi ica ion (pass/ ail). The con olu ional laye s and he LSTM laye s
au oma ically ecognize ea u es om he inpu da a's la en space. The PTCN's cus om a chi ec u e
p o ides elemen s enabling adap a ion o he inpu 's a iance om adjus men o he ke nel weigh s o e
ime. On he documen le el, he model epo ed an a e age e alua ion o 67.32% using 10- old c oss-
alida ion. The esul s sugges ha he model can ecognize cong essional o ing beha io s om he
associa ed legisla ion's language. O e all, he PTCN p o ides a solu ion wi h compe i i e pe o mance o
ela ed sys ems a ge ing cong essional oll call o es.
KEYWORDS
Deep Lea ning (DL), Con olu ional Neu al Ne wo ks (CNNs), Long-Sho -Te m Memo y Ne wo ks
(LSTMs), Na u al Language P ocessing (NLP), Cong essional Roll Call Vo es
1. INTRODUCTION
P edic ing he s a us (pass/ ail) o cong essional oll call o es has been poli ical scien is s' goal
o decades. The e a e pa e ns o cong essional o ing beha io cap u ed in he legisla i e ex ,
which has shown signi icance when p edic ing cong essional o es' s a us. Unde s anding he
u u e s a us o legisla ion p o ides i al insigh s in o go e nmen and indus y ma e s.
Analyzing oll-call da a allows insigh in o in o ma ion de ailing he legisla ion's o e s a us and
can p edic u u e o es [1].
O he app oaches using quan i a i e oll call da a and legisla i e ex ha e shown success in he
pas , howe e only unde ce ain condi ions. Thei success exp esses limi a ions due o he
dimensionali y o he da a and un o eseen condi ions in Cong ess's complex social en i onmen .
Fo ins ance, using wo sepa a e da ase s diminishes a model's lexibili y o p edic and adap o
he e en space's changing condi ions.
Poli ical en i onmen s a e complex social ne wo ks ha o en c ea e noisy da a. The scale o
opics ha he go e nmen conside s in legisla ion is a di e se subjec ma e , and he language is
spa se. Mos pas app oaches ely on an a emp o use ex a dimensionali y om bo h ex and
quan i a i e da a. Howe e , using ex a dimensionali y p oduces limi a ions, o p edic i e
bo leneck, in app oaches using ex and quan i a i e cong essional da a. Many si ua ions occu
when cong essional oll call da a is no a ailable o ep esen s uncon olled condi ions c ea ing
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
16
issues p edic ing he e en . Fo ins ance, a pu e exp ession o mixed bu non-pa isan suppo in
he o ing da a could limi exis ing models' o e all accu acy.
A mo e obus model is equi ed o add ess he cu en limi a ions exp essed in p io models
add essing he p oblem discussed abo e. S a is ical modeling seeks o lea n he join p obabili y
unc ion om wo ds
con ained in he ex s [2]. Howe e , i is di icul o ob ain his goal because o he "... cu se o
dimensionali y" [2]. Howe e , he ise o deep lea ning makes i possible o compu e sys ems o
ecognize pa e ns in complex ex ep esen a ions. Ad ancemen s in na u al language p ocessing
echniques allow ex o con e in o okenized wo d ec o spaces. The con e sion p o ides he
p ope dimensions o embed he ex s in o di e en deep neu al ne wo ks. A cus om hyb id
a chi ec u e p o ides he abili y o inpu ex s and p o ide accu a e ou pu s om ecognized
pa e ns. Con olu ional neu al ne wo ks (CNNs) and long- sho e m memo y neu al ne wo ks
(LSTMs) can ecognize hese in ica e pa e ns wi hin he da a's dimensions. Each laye in he
ne wo k p o ides bene i s in ecognizing empo al and spa ial ea u es om abs ac da a
ep esen a ions. Mainly, CNNs e lec success ul esul s in iden i ying abs ac da a pa e ns
mainly because o de elopmen in max-pooling laye s.
Howe e , due o long lag pe iods om he da a's complexi y, he CNN a chi ec u e canno alone
cap u e he ex 's pa e ns. A p ima y eason o implemen ing LSTM laye s in he model's
a chi ec u e is o o e come he long lag ime p oblem ha occu s when p ocessing high
dimensional da a. O e coming long lag pe iods allows o he neu al ne wo k's dep h o con inue,
enabling minu e ea u es o be ecognized.
Combining CNN and LSTM algo i hms p o ide an a chi ec u e capable o ecognizing ea u es
in spa se wo d ec o spaces o e long pe iods, known as a Con olu ional Long-Sho Te m
Memo y Neu al Ne wo k (C-LSTM). Adap able laye s du ing ke nel ini ializa ion help il e ou
he non-signi ican ea u es om he inpu s dimensional space. The adap abili y in each laye o
he ne wo k p o ides s abili y in he p edic ions and obus ness agains he dynamic social
en i onmen c ea ing he da a. The cus om a chi ec u e ensu es he ne wo k dep h is sui able o
accu a e p edic ion om highly spa se ex samples. The pape p esen s a cus om deep lea ning
solu ion o accomplish he bina y classi ica ion o legisla i e ex s.
1.1. O ganiza ional S uc u e
The emainde o he pape is o ganized in o he ollowing sec ions, including Sec ion 2.
Li e a u e Re iew, Sec ion 3. Me hods and Ma e ials, Sec ion 4. Resul s, Sec ion 5. Conclusion,
and Sec ion 6. Fu u e Wo k. Sec ion 3. Me hods and Ma e ials includes wo sec ions including
3.1. The Da a, and 3.2. The Deep Lea ning App oach. Sec ion 3.1. includes Sec ion 3.1.1. P e-
p ocessing ex and 3.1.2. Equal Dis ibu ion o Samples. Sec ion 3.2. includes 3.2.1. The Cus om
PTCN A chi ec u e and 3.2.2. The PTCN Modelling P ocess.
2. LITERATURE REVIEW
A ocus o quan i a i e poli ical scien is s is using oll-call da a o unde s and legisla i e o ing
beha io s be e . Few models ha e a emp ed o use supplemen a y da a such as he ex o
legisla ion o unde s and cong essional o ing ends be e .
In 1991, esea ch exp essed ha spa ial posi ions cap u ed in oll call da a is s able and con ains
eliable ea u es o ecognize o ing pa e ns [3]. The pa y discipline is p esen in he spa ial
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
17
dimensions o oll-call o es o he legisla o 's ideal poin s cap u ed in he da a [4]. Spa ial
dimensions in he con ex o he domain e e o he p obabilis ic wo d occu ence con ained
ac oss he ex samples. Resea che s u ilized a me hod using Bayesian simula ion models o
cap u e he ideal poin s o he legisla o s. The app oach allows esea che s o ensu e he belie s
inco po a ed in o he inpu s' dimensional space o oll call analysis [5]. The me hod enables
esea che s o handle he inc easing complexi y in highe -dimensional con ex s [5]. Policy ideas
a e s anda d ea u es in legisla ion ha p o ide insigh in o legisla o s' beha io simila o
quan i a i e oll call da a [6].
In 2011, Ge ish and Blei in oduced a p obabilis ic model capable o in e ing a pe son's
poli ical posi ion o speci ic opics [7]. The model ocuses on cap u ing he indi idual
ep esen a i e's iew on speci ic poli ical in e es s om he ex alone. The au ho s ha e used 12
yea s o cong essional legisla i e da a in hei expe imen o cap u e signi ican pa e ns. The
pa e ns exp ess he lawmake s' o e beha io and on which ype o documen [7]. Ge ish and
Blei in eg a ed he analysis o ex in o quan i a i e models such as he ideal
poin model wi h success in e ing "... ideal poin s and bills' loca ions..." om oll call da a
leading o he p edic ion o legisla i e o e s a us (pass/ ail) [8]. The in eg a ion o he bill ex s
in o he ideal poin model helped mi iga e a limi a ion o only p edic ing on o e da a alone,
which may, a imes, be inconsis en in i s a ailabili y. The au ho s de eloped a supe ised ideal
poin opic model capable o p edic ing pending bills using o es. I also is a me hod o explo ing
he connec ion be ween language and poli ical suppo [8]. O he models de eloped by Ge ish
and Blei o p edic he "...in e ed ideal poin s using di e en o ms o eg ession on ph ase
coun s." [8]. Howe e , lead exis ing models canno p edic when mixed bu non-pa isan suppo
is p esen in he da a. Many exis ing models canno expand beyond one-dimensional limi a ions,
such as he ideal poin model [8]. The au ho s also based hei model's pe o mance on he
baseline ha 85% o all o es a e 'yea', limi ing he model's pe o mance esul s. The au ho s
epo ed hei ideal opic model p edic ed 89% o he o es wi h 64 opics, and hei L2 model
p edic ed 90% [8]. In sequen ial p edic ions, bo h hei models p edic ed 87% and 88.1%
accu a e a p edic ing u u e o es, espec i ely [8]. Thei s udy only a emp s o unde s and a
ew opics wi h a one-dimensional poli ical space, which c ea es a p edic i e bo leneck [8].
In e es ingly, he ideal poin model e lec s ep esen a i es' p e e ences, cons i uency p e e ences,
o any o he ea u e indica ing a p e e ence o pa icula legisla ion [1]. In 2013, Spa io- empo al
modeling exp essed success in using ex o p edic cong essional oll call o es wi h ela i e
success o he ideal poin s model [9]. In 2015, ideal poin s we e ede ined as wo
cha ac e iza ions, including wo d and o e choice [10]. These cha ac e ize he ideal poin s and
he dimensions o he policy. The esea che s u ilize Spa se Fac o Analysis o combine bo h
o es and ex ual da a o es ima e he ideal poin s [10].
In 2016, Yang and o he s in oduced hie a chical a en ion ne wo ks o documen classi ica ion
[11], he esul s ou pe o med p e ious models by a la ge ma gin, which can be indica ed in he
au ho 's esul s using he Yelp 2013, Yelp 2014, Yelp 2015, IMBD e iew, yahoo Answe , and
Amazon e iew da ase s o es hei algo i hm and compa e i o o he me hods [11]. On a e age,
he HN-ATT pe o med wi h abou a 70% accu acy a e ac oss he asks. They we e gene alizing
he ne wo k o add ess mul iple ypes o asks ha limi hei abili y o iden i y he ex 's c i ical
ea u es, such as he poli ical ideology in he cong essional legisla i e ex s ac oss spa ial and
empo al dimensions. A simila app oach ha uses mul i-dimensional bill ex s o p edic oll
calls' s a us is K a , Jain, and Rush's app oach es ablished in 2016. Using an embedding model
wi h p epa ed bill ex s, hey compe ed wi h Ge ish and Blei's app oach. Mainly he app oach
u ilized ideal ec o s a he han ideal poin s [12]. The app oach elies on quan i a i e da a ha
leads o he same p edic i e bo lenecks as p io models due o he complex e en space.
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
18
In 2016, he Go 2Vec me hod exp essed success in cap u ing opinions om legal documen s.
The ex 's ans o ma ion allows wo ds o be embedded in a model o lea n he ep esen a ions o
indi iduals' opinions [13]. In 2017 i became possible o use quan i a i e da a and ex o p edic
i a bill will become law. The esea che s a es ha hey always pe o m be e han using ex o
con en indi idually [14]. In e es ingly, he au ho conduc ed h ee expe imen s, including
"... ex -only, ex , and con ex , o con ex -only..." o es he p edic i e powe o each ype o
model [14]. Nay's app oach uses a language model ha p o ides a p edic ion on he ex 's
sen ence le el, p o iding a p obabili y o he sen ence con ibu ing o he bill ex 's (pass/ ail)
s a us [14]. The da a included a ew pe o mance measu es and wo da a condi ions spanning 14
yea s.[14]. Nay concluded ha a ex alone app oach is undamen al o be e esul s. Fo newe
da a, he use o bill ex ou pe o ms bill con ex -only. Con ex -only only ou pe o ms bill ex
alone o olde da a [14]. The mos impo an inding is ha ex adds p edic i e powe o he
model. Howe e , he app oach elies on wo da a sou ces, which c ea es a p edic i e bo leneck.
As he mos success ul app oach o da e using ex alone, he model can p edic a 65% accu acy
[14].
A hyb id C-LSTM Neu al Ne wo ks a chi ec u e can help mi iga e he p edic i e bo leneck in
exis ing models by cap u ing mo e ea u es om he bill ex s alone. In s a is ical language
modeling, he p ima y goal is lea ning. The main objec i e o lea n is he "... join p obabili y
unc ion ..." o each sequence o wo ds con ained in he language [2]. Howe e , he e is di icul y
in his ype o ask because o he cu se o dimensionali y. By lea ning dis ibu ed ep esen a ions
o language, he cu se mi iga es. The main issue is add essing he a iance om he aining da a
o he es ing da a. Pas esea ch disco e ed "... he simila i ies
be ween wo ds o ob ain gene aliza ion om aining sequences o new sequences..." [15] [16]
[17] [18]. Much o his ype o wo k is hanks o con ibu ions by Schu ze, 1993, whe e ec o -
space ep esen a ions o wo ds can be lea ned based on he p obabili y o he wo d co-occu ing
in documen s [19]. Mos o he expe imen al wo kings can be summed up om lea ning
dis ibu ed ea u e ec o s o ep esen hei simila i ies be ween wo ds, which is discussed in
[15] [17] [20]. Pas esea ch has shown ha he hyb idiza ion o CNNs and LSTMs asked o
sol e ex classi ica ion p oblems is success ul [21]. The combina ion o he wo di e en ypes
o neu al ne wo ks builds a C-LSTM [21]. In 2015, Zhou e al. in oduced a success ul C- LSTM
in sen ence and documen modeling [21]. The CNN ex ac s a sequence o high-le el ph ase
ep esen a ions, which a e in- e u n ed h ough LSTM laye s se o ob ain he sen ence-
ep esen a ions [21]. O e all, he hyb id model pe o med be e han exis ing sys ems in
classi ica ion asks. The esul s indica ed ha he local ea u es o ph ases and he sen ence's
global/ empo al seman ics a e ecognizable by hei model [21]. In 2019, a eam o esea che s
ook ad an age o con olu ional ecu en neu al ne wo ks o ackle ex classi ica ion asks [22].
Thei expe imen s showed ha he me hod o using a C-LSTM achie es be e success han o he
ne wo ks.
3. METHODS AND MATERIALS
The ollowing sec ions discuss key componen s o de eloping he PTCN and he associa ed
esul s e e ed o in Sec ion IV. The me hodology is summa ized below in Figu e 1.
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
19
Fig. 1: Visualiza ion o he me hodology.
3.1. The Da a
The o iginal da a is ex ac ed and o ganized om (h ps://www.Go ack.us) consis ing o
legisla i e ex s and associa ed quan i a i e oll call da a. The o iginal da a con ained samples
om he yea 2000 o 2019, including 3668 samples om he house and sena e. Re e o Figu e 2
o an example o o iginal legisla i e ex s. No all he samples om he o iginal popula ion will
be included in he expe imen due o limi ing ac o s as discussed below.
Fig. 2: Legisla i e ex samples.
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
20
3.1.1. P e-P ocessing he Tex
A se ies o NLP echniques a e equi ed o supe ise he cus om C-LSTM aining o classi y he
legisla i e ex based on o ing s a us. The ex samples con ain noise, which is in he o m o
s op wo ds, uppe casing, special cha ac e s, NA alues, punc ua ion, numbe s, and whi espace.
The emo al o noise om he ex samples helps mi iga e he C-LSTM om ecognizing
pa e ns wi hin he ex s ha c ea e bias p edic ions.
The C-LSTM is a deep lea ning algo i hm ha au oma ically ex ac s ea u es om an inpu
ec o . Each ex unde goes augmen a ion using he ollowing condi ions, including con e ing
da a ypes om cha ac e o a s ing, lowe case con e sion, s op-wo ds emo al using he
“SMART” unc ion, punc ua ion emo al, numbe emo al, whi e-space emo al, and documen
s emming. The abo e augmen a ion esul ed in Figu e 3.
Fig. 3: Sample o p e-p ocessed Legisla i e ex s.
A e p e-p ocessing, he ex goes h ough a okeniza ion and ec o iza ion s ep esul ing in each
wo d, symbol, o any o he cha ac e ep esen ed as a unique numbe . Fo ins ance, he wo d
‘bill’ is ep esen ed by he alue o 3109 ac oss all he documen s. No e he ex is limi ed o
10000 max ea u es du ing he okeniza ion p ocess. Re e o Figu e 4 o an example o he
ec o ized ocabula y om he legisla i e documen s.
Fig. 4: Sample o okenized and ec o ized Vocabula y.
The p e-p ocessing o he ex esul ed in ec o ized legisla i e documen s, as seen in Figu e 5. A
me hod o padding is implemen ed o ans o m all he ex s o he same leng h. The dimensions
o he ex a e con e ed in o a okenized, ec o ized, and padded o ma wi h a max leng h o
10,000.
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
21
Fig. 5: Sample o okenized, ec o ized, and padded ex .
3.1.2. Equal Dis ibu ion o Samples
The labeled legisla i e ex samples a e balanced in o an equal dis ibu ion o each o ing s a us
o educe bias. I is essen ial o mi iga e bias in he da a. Deep neu al ne wo ks ans o m inpu s
o e ex ensi e c edi assignmen pa hways (CAPs). The mo e complica ed he dimensions o an
inpu , hen he mo e complex he CAPs.
Each o e’s s a us was cap u ed by de e mining he numbe o “yea” o “Aye” esponses eached
o each legisla ion. Each ansla es in o a alue o 1, ep esen ing a o e o pass he documen .
All o he esponses a e conside ed a o e agains he documen , labeled as a alue o 0.
In e es ingly, he numbe o documen s labeled one is g ea e han 0 labeled documen s. No e ha
some o es equi e a special condi ion o pass, which is mo e han 2/3 o he o es. The s udy
igno es he special o ing condi ion because i only ocuses on an equal dis ibu ion o (pass/ ail)
ex ep esen a ions.
An equal dis ibu ion o he o ing s a uses in he sample is necessa y o mi iga e bias in he
ne wo k o ei he classi ica ion. The ex samples a e andomly selec ed om each class o
ep esen an equal dis ibu ion, which educes he o iginal numbe o samples. Each class o o e
is ep esen ed equally by 98 andomly selec ed legisla i e ex s. 98 is he maximum numbe o 0
labeled samples a ailable in he da a due o he beha io o cong ess. Each ex is a max leng h o
10,000 ea u es. I should be no ed ha mos o he legisla i e ex s a e close o highe o he
maximum numbe o ea u es. Cus om ea u es in he PTCN’s a chi ec u e a e implemen ed o
deal wi h he small numbe o samples du ing aining as discussed below.
3.2. The Deep Lea ning App oach
The ollowing sec ion discusses he cus om deep lea ning a chi ec u e and modelling p ocess
implemen ed in he s udy.
3.2.1. The Cus om PTCN A chi ec u e
The sequen ial deep lea ning model is a s ack o di e en laye s se wi h se e al pa ame e s,
including d opou a e, hidden con olu ional nodes, LSTM hidden nodes, L1 egula iza ion a e,
L2 egula iza ion a e, ba ch size, inpu max leng h, max ea u es, embedding dimensions, leaky
Relu a e, ke nel size, epochs, max-pooling size, lea ning a e, and alida ion spli . In Figu e 6,
an example o he PTCN models a chi ec u e:
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
22
Fig. 6: PTCN a chi ec u e
The i s laye in he model is he embedding laye , which embeds he okenized wo ds. A he
nex laye , he inpu s pipe in o a 1-dimensional con olu ional laye wi h only a 4 h o he se
con olu ional hidden nodes. The inpu s a e ep esen a ions o leng hy and highly spa se wo d
ec o spaces, making he model weigh 's c i ically impo an . The model equi es a me hod o
adjus o he high a iance be ween samples due o he low numbe o samples. The con ol o he
ke nels will help he model s eadily lea n signi ican ea u es con ained in he ex .
Th oughou he con olu ions, he model can explo e deepe o ecognize mo e signi ican
ea u es. A a iance scaling ke nel ini ialize p o ides an “ini ialize capable o adap ing i s scale
o he shape o weigh s. ” [23]. Va iance scaling is a ke nel ini ializa ion s a egy ha encodes an
objec wi hou knowing he shape o a a iable [23]. The Con olu ional laye s’ ini ialize makes
he deep ne wo k adap o he inpu weigh s. I is impo an o egula ize he ke nels when
u ilizing an adap i e ini ialize . Se ing a duel egula iza ion echnique ha deploys L1 and L2
egula iza ion helps mi iga e high luc ua ions while
ba ching samples h ough he laye s. L1 is a Lasso Reg ession (LR). L2 is Ridge Reg ession. The
main di e ence be ween he wo me hods is he penal y e m. The laye includes s ides se a 1L
h ough he con olu ions a e an “...an in ege o lis o a single in ege , speci ying he s ide
leng h o he con olu ion” [24]. The con olu ional laye is ac i a ed using a Leaky Relu unc ion.
The leaky Relu unc ion allows o a “... small g adien when he uni is no ac i e ...”, p o iding
“... sligh ly highe lexibili y o he model han adi ional Relu. ” [25].
Machine Lea ning and Applica ions: An In e na ional Jou nal (MLAIJ) Vol.7, No.3/4, Decembe 2020
23
The i s con olu ional laye ex ac s lowe le el ea u es om he inpu s due o he dec ease in
he hidden nodes. The educ ion o he numbe o ans o ma ions p o ides con ol o he
adap able ea u es ini ializing he weigh s. The second laye pipes he inpu s h ough ano he 1-
dimensional con olu ional laye wi h he same pa ame e s se , excep he numbe o hidden
nodes se o 32. Inc easing he numbe o hidden nodes p o ides mo e ans o ma ions ex ac ing
highe -le el ea u es om he inpu s. The second con olu ional laye is ac i a ed using he leaky
Relu unc ion. The nex laye in he s ack is ano he con olu ional laye se a hal he se numbe
o hidden nodes. Reducing he numbe o nodes and ollowing wi h a max-pooling laye helps
mi iga e o e i ing du ing aining. In he s udy, he max-pooling laye is se o 4. The ollowing
a e wo mo e laye s o 1-dimensional con olu ional laye s se o hal he se hidden nodes and a
4 h o he se hidden nodes.
All pa ame e s a e se he same as he p io con olu ional laye s. A second max-pooling laye ,
ba ch no maliza ion, and d opou laye help mi iga e o e i ing u he . The nex laye is an
LSTM laye se a 32 hidden nodes. Va iance Scaling ke nel ini ialize s and L1 and L2 ke nel
egula ize s con ol he model's explo a ion o he ea u e space. The LSTM laye is ac i a ed
using a Leaky Relu unc ion. A d opou laye o 0.5 is in he s ack be o e he ou pu laye . The
ou pu laye is ac i a ed using a sigmoid unc ion. The model compiles using a loss unc ion o
bina y c oss-en opy. The PTCN uses a s ochas ic g adien descen (SGD) op imize . The hype -
uning sessions de e mine he lea ning a e.
3.2.2. The PTCN Modelling P ocess
To ensu e he model is p oducing he bes esul s, he pa ame e s o he PTCN a e hype - uned.
The da a is spli 80/20% o aining and alida ion o he model du ing he hype - uning
sessions. The hype - uning model’s pe o mance e alua es a andom selec ion o 100 samples.
The bes pa ame e s a e cap u ed and se o he inal model aining. Once he inal pa ame e s
a e iden i ied, he PTCN is ained o a inal aining session. The inal aining session o he
PTCN uses he bes pa ame e s iden i ied du ing he hype - uning sessions, and 10- old c oss-
alida ion is implemen ed o e alua e he model’s pe o mance.
4. RESULTS
A e implemen ing he model using 10- old c oss- alida ion, he PTCN Model a e aged 67.32%
e alua ion accu acy wi h a s anda d de ia ion o 9.11. The bes model pe o med a 76.32%
e alua ion on old 6, as depic ed below in Figu e 7.
Fig. 7: Sample o PTCN aining and alida ion pe o mance including ea ly s opping.
