199
In e na ional Jou nal o Ad ance and Applied Resea ch
www.ijaa .co.in
ISSN – 2347-7075
Impac Fac o – 8.141
Pee Re iewed
Bi-Mon hly
Vol. 6 No. 38
Sep embe - Oc obe - 2025
AI Based Mind Map P og amming Tu o
Deo Sha mila Mahesh
Assis an P o esso ,
D . D. Y. Pa il Science and Compu e Science College, Aku di Pune
Co esponding Au ho – Deo Sha mila Mahesh
DOI - 10.5281/zenodo.17313221
Abs ac :
The inc easing adop ion o AI-d i en p og amming assis an s has signi ican ly ans o med he
landscape o coding educa ion. Despi e hese ad ancemen s, mos exis ing in elligen u o ing sys ems
emphasize answe gene a ion a he han diagnosing lea ne s’ unde lying misconcep ions and
p o iding adap i e emedia ion. This pape in oduces he Mind-Map P og amming Tu o (MMPT), a
no el sys em ha employs cogni i e g aphs o model lea ne s’ e ol ing knowledge s a es. In con as o
adi ional AI u o s ha equi e p e-anno a ed s uden da a o aining, MMPT le e ages ze o-sho
easoning o in e misconcep ions di ec ly om s uden code submissions. I dynamically cons uc s a
cogni i e e o g aph ha cap u es concep ual misunde s andings, knowledge gaps, and logical
inconsis encies.
Based on his cogni i e ep esen a ion, MMPT adap s he di icul y o assigned p oblems,
deli e s pe sonalized hin s, and gene a es sca olded explana ions aligned wi h he lea ne ’s cu en
knowledge g aph. The p oposed me hodology in eg a es p inciples om cogni i e psychology, g aph-
based knowledge modeling, and ze o-sho lea ning o suppo pe sonalized and scalable p og amming
ins uc ion. P elimina y simula ions indica e ha he sys em imp o es lea ning e en ion, minimizes he
ecu ence o misconcep ions, and lays a s ong ounda ion o
In oduc ion:
1.Backg ound:
P og amming educa ion emains a
c i ical skill in he digi al e a. T adi ional
u o ing me hods and online pla o ms o en
ely on s a ic p oblem se s o p e- ained
models wi h limi ed adap abili y. Recen
ad ances in na u al language p ocessing
(NLP) and g aph-based easoning p o ide
oppo uni ies o c ea e mo e adap i e, scalable
u o ing solu ions.
2.P oblem S a emen :
Cu en AI u o ing sys ems a e
p edominan ly answe -o ien ed a he han
lea ne -o ien ed. They do no main ain a
ep esen a ion o he lea ne ’s e ol ing
concep ual unde s anding, no do hey iden i y
o emedia e speci ic misconcep ions. This
lack o cogni i e awa eness hinde s hei
abili y o p o ide uly pe sonalized and
e ec i e ins uc ion. The g owing demand o
accessible and e ec i e p og amming
educa ion has led o widesp ead in e es in
in elligen u o ing echnologies. Despi e his
end, many lea ne s con inue o s uggle wi h
abs ac p og amming concep s such as
ecu sion, da a s uc u es, and debugging.
While In elligen Tu o ing Sys ems (ITS) like
CodeTu o and AI-powe ed ools such as
Gi Hub Copilo o e eal- ime code
sugges ions, hese sys ems la gely ail o
in e p e o model he lea ne ’s unde lying
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
200
cogni i e p ocesses. They p io i ize solu ion
gene a ion o e educa ional diagnosis, hus
limi ing hei e ec i eness as pedagogical
ools.
3.Objec i e:
To add ess his limi a ion, he p esen
wo k in oduces he ‘Mind-Map P og amming
Tu o (MMPT)’ , a no el in elligen u o ing
amewo k ha le e ages cogni i e g aphs o
in e lea ne s’ knowledge s a es. Unlike
adi ional sys ems ha depend on labeled
aining da a, MMPT employs ze o-sho
easoning o de ec misconcep ions and
dynamically cons uc indi idualized cogni i e
g aphs based on lea ne in e ac ions.
This pape makes h ee key
con ibu ions: (1) i p esen s a amewo k o
cons uc ing dynamic cogni i e g aphs om
s uden code submissions and in e ac ions; (2)
i applies Mind Map easoning echniques o
iden i y p og amming misconcep ions wi hou
he need o anno a ed aining da a; and (3) i
in oduces an adap i e u o ing sys em ha
modi ies he sequence and di icul y o
p og amming p oblems based on he lea ne ’s
e ol ing cogni i e g aph.
Rela ed Wo k:
In elligen Tu o ing Sys ems (ITS):
T adi ional ITS pla o ms ely hea ily
on s uc u ed knowledge bases and anno a ed
da ase s o deli e eedback and ack
p og ess. Sys ems such as Cogni i e Tu o
ha e demons a ed success in domains like
ma hema ics, bu hei eliance on p e-de ined
ules and labeled da a limi s scalabili y and
adap abili y in open-ended domains such as
p og amming.
G aph-Based Lea ning Models:
P io wo k in lea ne modeling has
explo ed echniques such as Bayesian
knowledge acing and deep knowledge
acing using Long Sho -Te m Memo y
(LSTM) ne wo ks (Piech e al., 2015). These
app oaches aim o model s uden pe o mance
o e ime bu ypically ocus on co ec ness
a he han cogni i e unde s anding. They a e
ill-sui ed o iden i ying nuanced
misconcep ions ha a ise in p og amming
asks.
Ze o-Sho Lea ning in Educa ion:
Ze o-sho lea ning has seen signi ican
applica ion in na u al language p ocessing and
compu e ision asks (B own e al., 2020),
whe e models gene alize o no el classes
wi hou explici aining examples. Howe e ,
i s use in educa ional con ex s emains limi ed.
Few s udies ha e explo ed i s po en ial o in e
lea ne s a es o misconcep ions in he absence
o labeled educa ional da a.
Resea ch Gap:
To da e, no exis ing in elligen
u o ing sys em has success ully in eg a ed
ze o-sho in e ence wi h cogni i e g aph
modeling o suppo adap i e p og amming
ins uc ion. This esea ch aims o ill ha gap
by combining hese wo me hodologies in o a
uni ied sys em capable o deli e ing scalable,
pe sonalized, and cogni i ely in o med
p og amming educa ion.
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
201
Sys em A chi ec u e:
The p oposed sys em adop s a mul i-
laye ed a chi ec u e o enable pe sonalized
u o ing by in eg a ing Ze o-Sho
Misconcep ion De ec ion and Cogni i e
G aphs. A he inpu laye , he s uden submi s
a code a emp in esponse o a p og amming
p oblem. This inpu includes he sou ce code,
execu ion esul s (such as es case ou comes),
and me ada a such as he ime aken, numbe
o a emp s, and po en ially o he beha io al
indica o s (e.g., edi his o y o submission
in e als). The aw inpu is p ocessed by he
Misconcep ion De ec o , which le e ages a
Ze o-Sho Lea ning (ZSL) model o iden i y
po en ial concep ual misunde s andings
wi hou equi ing ex ensi e ask-speci ic
anno a ed da ase s (B own e al., 2020; Wang
e al., 2021). The ZSL model gene alizes om
p e ained embeddings and easoning
capabili ies o de ec e oneous code pa e ns
and map hem o speci ic concep ual
misconcep ions. Fo example, an o -by-one
e o in loop indexing may be indica i e o a
misunde s anding o loop bounda y
condi ions. By aligning seman ic
ep esen a ions o he s uden ’s code wi h
known misconcep ion empla es, he model
can iden i y p e iously unseen e o pa e ns
and p o ide na u al language explana ions
wi h co esponding con idence es ima es
(Reime s & Gu e ych, 2019).
To pe sonalize eedback and ack
lea ne p og ess, he sys em main ains a
dynamic Cogni i e G aph o each s uden . In
his g aph, nodes ep esen co e p og amming
concep s such as ecu sion, loop cons uc ion,
o a ay indexing, while edges encode he
ela ionships and dependencies among hese
concep s. Each node main ains indi idualized
Me ic
Baseline Sys ems (S a ic /
Rule-based)
MMPT (P oposed
Sys em)
Re en ion Ra e
Mode a e – limi ed
ein o cemen o concep s
High – ein o ced ia
Cogni i e G aph and
adap i e e isi s
Misconcep ion Resolu ion
Speed
Slow – elies on epea ed
ial and e o
Fas – a ge ed g aph-based
emedia ion o weak nodes
Lea ning Pa h
Pe sonaliza ion
None – ixed o andom
sequence
S ong – adap i e
sequencing based on
knowledge s a e
F us a ion Le els
Highe – lea ne s ace
un ela ed o epea ed
p oblems
Lowe – lea ne s p og ess
h ough ele an ,
sca olded p oblems
Lea ne Con idence
Mode a e – eedback o en
gene ic
High – pe sonalized,
explana o y eedback
imp o es sel -e icacy
Scalabili y o New E o s
Limi ed – equi es
p ede ined ules/ aining
S ong – Ze o-Sho model
gene alizes o unseen
misconcep ions
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
202
in o ma ion abou he s uden ’s in e ed
mas e y le el, misconcep ion his o y, and
empo al lea ning ajec o y. When a
misconcep ion is de ec ed, he co esponding
concep node is upda ed, enabling he sys em
o moni o e ol ing knowledge gaps. This
s uc u e acili a es adap i e in e en ions by
gene a ing a ge ed, da a-d i en eedback,
including ailo ed explana ions, isual
sca olds. The sys em deli e s a scalable and
pe sonalized u o ing expe ience g ounded in
he p inciples o cogni i e science and mode n
machine lea ning.
Me hodology:
1. Da ase :
The da ase used in his s udy
comp ises p og amming p oblems, s uden
code submissions, and manually anno a ed
e o labels o e alua ion pu poses. The
p og amming p oblems span a ange o
di icul y le els—beginne , in e media e, and
ad anced—and a e designed o assess
unde s anding o undamen al p og amming
concep s such as a iables, loops, condi ionals,
ecu sion, and da a s uc u es. S uden code
submissions we e collec ed om publicly
a ailable sou ces, including open-sou ce
pla o ms such as Code o ces and Lee Code,
as well as anonymized class oom da a. To
enable quan i a i e e alua ion o he sys em's
ze o-sho misconcep ion de ec ion capabili ies,
a manually anno a ed subse o s uden
submissions was p epa ed, labeling speci ic
misconcep ions obse ed in he code.
2. En i onmen :
The expe imen s we e conduc ed on a
compu ing en i onmen equipped wi h an In el
Xeon CPU and an NVIDIA Tesla V100 GPU
o suppo model in e ence and embedding
compu a ion. The sys em was p o isioned
wi h 32 GB o RAM o manage la ge-scale
g aph s uc u es and model weigh s. The
so wa e s ack includes Py hon 3.10, along
wi h machine lea ning and g aph p ocessing
lib a ies. PyTo ch and Hugging Face
T ans o me s we e used o implemen and
ine- une he ze o-sho misconcep ion de ec o .
Fo g aph cons uc ion and analysis,
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
203
Ne wo kX was employed alongside Neo4j,
which se ed as he g aph da abase.
E alua ion me ics we e compu ed using
Sciki -lea n, and expe imen al wo k lows we e
conduc ed using Jupy e no ebooks o acili a e
analysis and isualiza ion.
3. Sys em Con igu a ion:
The co e componen o he sys em is
he Ze o-Sho Misconcep ion De ec o , which
is based on a p e ained ans o me
a chi ec u e, such as RoBERTa o a GPT-like
model. This model is ine- uned on join code
and execu ion ace embeddings o p edic
misconcep ion ca ego ies wi hou equi ing
explici p io examples. Inpu s o he model
include bo h he s uden 's submi ed code and
i s execu ion ace, while ou pu s co espond
o iden i ied misconcep ions mapped o
p ede ined concep ual ca ego ies.
The Cogni i e G aph Gene a o
models he lea ne 's unde s anding by
dynamically cons uc ing a g aph in which
nodes ep esen p og amming concep s (e.g.,
loops, ecu sion), and edges cap u e
p e equisi e o usage ela ionships be ween
hem. When misconcep ions a e de ec ed, he
co esponding nodes o edges in he g aph a e
weakened, e lec ing he lea ne ’s di icul y
wi h hose concep s.
The Adap i e P oblem Selec o
ope a es on he cogni i e g aph o de e mine
he nex mos app op ia e p oblem o he
lea ne . I inco po a es cen ali y measu es and
weakness sco es o p io i ize p oblems ha
ein o ce unde de eloped concep s while
main aining a ajec o y o inc easing
complexi y o suppo g adual skill
acquisi ion.
The Feedback Gene a o combines
empla e-based messaging wi h la ge language
model (LLM)-gene a ed explana ions. I
p o ides s uden s wi h misconcep ion-awa e
eedback, including isual o e lays on he
cogni i e g aph ha highligh weak concep s
and explain he a ionale behind selec ed
lea ning pa hs.
4. E alua ion P o ocol:
The sys em was e alua ed h ough
con olled in e ac ion wi h a g oup o s uden
pa icipan s, including bo h no ice and
in e media e p og amme s. Each pa icipan
engaged wi h he sys em o e mul iple
p oblem-sol ing sessions. Se e al e alua ion
me ics we e employed. Misconcep ion
De ec ion Accu acy was assessed using
p ecision, ecall, and F1-sco e compu ed
agains he manually anno a ed g ound u h.
Lea ning Gain was measu ed by compa ing
p e- es and pos - es sco es o e alua e
concep ual unde s anding. P oblem-Sol ing
E iciency was quan i ied by he educ ion in
he numbe o a emp s equi ed o sol e
p oblems a e ecei ing a ge ed eedback.
Use Expe ience was e alua ed h ough
su eys cap u ing pa icipan s' pe cep ions o
he sys em’s use ulness, cla i y, and
mo i a ional impac .
5. Baselines o Compa ison:
To con ex ualize he sys em's
pe o mance, compa isons we e made agains
se e al baseline app oaches. These included
ule-based eedback sys ems ha deli e
p ede ined hin s based on pa e n-ma ched
e o s, and supe ised e o classi ica ion
models ained on labeled da a. Addi ionally, a
andom p oblem selec ion s a egy was used o
e alua e he e icacy o he adap i e p oblem
sequencing employed by he p oposed sys em.
6. Expe imen al Flow:
The expe imen al p ocedu e was
di ided in o ou phases. Du ing he S uden
In e ac ion Phase, s uden s submi ed code
a emp s, and he sys em de ec ed
misconcep ions while upda ing he cogni i e
g aph in eal ime. In he Adap i e Tu o ing
Phase, he p oblem selec o used he upda ed
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
204
g aph o assign pe sonalized p oblems, and he
eedback gene a o p o ided concep -speci ic
guidance. The Da a Collec ion Phase in ol ed
logging all s uden in e ac ions, including code
submissions, de ec ed misconcep ions, and
p og ess me ics. Finally, he E alua ion Phase
compa ed he p oposed Ze o-Sho Cogni i e
G aph Tu o o baseline models using he
de ined pe o mance me ics, aiming o assess
imp o emen s in accu acy, lea ning ou comes,
e iciency, and use expe ience.
Case S udies and Applica ions:
Case S udy 1: Lea ne s s uggling wi h
ecu sion concep s we e guided h ough
p og essi ely simple p oblem
decomposi ions.
Case S udy 2: Debugging misconcep ions
we e add essed by adap i e hin gene a ion.
Applica ions:
The Mind-Map P og amming Tu o
(MMPT) p esen s nume ous applica ions
ac oss educa ional and p o essional domains.
In elligen Tu o ing Sys ems (ITS):
MMPT can be in eg a ed in o online
coding pla o ms and Massi e Open Online
Cou ses (MOOCs) such as Cou se a, edX, and
Code o ces. Unlike adi ional sys ems ha
o e gene ic eedback, MMPT deli e s
pe sonalized guidance ailo ed o he lea ne ’s
speci ic misconcep ions. This enables mo e
e ec i e concep ein o cemen and
accele a ed lea ning.
Class oom Teaching Suppo :
In adi ional class oom en i onmen s,
educa o s can le e age cogni i e g aph
dashboa ds gene a ed by MMPT o moni o
bo h indi idual and g oup-le el
misconcep ions. These isual analy ics suppo
da a-d i en in e en ions, allowing eache s o
implemen a ge ed emedial ins uc ion based
on speci ic concep ual weaknesses.
Skill Assessmen and Ce i ica ion.
MMPT enables a mo e nuanced o m
o skill assessmen by e alua ing concep ual
mas e y p og ession a he han me ely inal
p oblem-sol ing success. This app oach is
pa icula ly bene icial o ce i ica ion
pla o ms, which o en aim o assess deep
unde s anding o e o e memo iza ion.
Co po a e T aining and Upskilling:
In p o essional con ex s, MMPT can
be employed in employee aining p og ams o
acili a e he lea ning o new p og amming
languages and amewo ks. I s adap i e
lea ning pa h unc ionali y educes aining
du a ion and enhances lea ne s’ con idence in
applying skills in eal-wo ld scena ios.
C oss-Domain Adap a ion:
Al hough MMPT is designed o
p og amming educa ion, he unde lying
cogni i e g aph amewo k is ex ensible o
o he domains such as ma hema ics, logical
easoning, and language lea ning—any
domain whe e concep ual dependencies can be
mapped and dynamically upda ed.
Educa ional Resea ch and Analy ics:
MMPT also se es as a aluable ool
o educa ional esea che s. I p o ides access
o ich, eal- ime da a on lea ne
misconcep ions and concep ual ajec o ies.
This suppo s s udies in o how s uden s
in e nalize p og amming concep s and how
adap i e in e en ions in luence e en ion,
con idence, and mo i a ion.
Challenges and Limi a ions
Scalabili y: Main aining eal- ime g aph
upda es o la ge lea ne g oups emains a
signi ican challenge, as he sys em mus
p ocess and adap quickly o mul iple lea ne s
simul aneously.
In e p e abili y: The complexi y o easoning
pa hways wi hin he cogni i e g aph may
some imes con use lea ne s, making i di icul
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
205
o ace o unde s and how ce ain solu ions
a e gene a ed.
Domain Adap abili y: The cu en
implemen a ion is limi ed o p og amming
educa ion. Ex ending he amewo k o o he
subjec domains will equi e e-enginee ing o
bo h he knowledge base and easoning
componen s.
Resou ce Dependency: The app oach elies
on ex ensi e compu a ional esou ces o
na u al language p ocessing and easoning,
which may limi i s deploymen in esou ce-
cons ained en i onmen s.
Conclusion:
This esea ch in oduces he Mind-
Map P og amming Tu o (MMPT), a no el
amewo k ha in eg a es a Ze o-Sho
Misconcep ion De ec o , a Cogni i e G aph
Gene a o , and an Adap i e P oblem Selec o
o deli e pe sonalized, concep -awa e
p og amming educa ion. Unlike s a ic ule-
based u o ing sys ems, MMPT dynamically
iden i ies misconcep ions, models lea ne
unde s anding h ough e ol ing cogni i e
g aphs, and adap s ins uc ional con en
acco dingly.
The p oposed sys em demons a es he
po en ial o imp o e concep ual e en ion,
accele a e he esolu ion o misunde s andings,
and inc ease lea ne con idence by minimizing
us a ion du ing he lea ning p ocess. These
cha ac e is ics make MMPT sui able o
deploymen in bo h indi idual lea ning
scena ios and la ge-scale educa ional
pla o ms, including MOOCs and co po a e
aining en i onmen s.
Fu u e Wo k:
Se e al di ec ions a e iden i ied o u u e
esea ch and sys em enhancemen :
La ge-Scale Deploymen and Valida ion:
Fu u e e o s will ocus on conduc ing
longi udinal s udies wi h di e se lea ne
popula ions ac oss MOOCs, coding
boo camps, and adi ional class oom
en i onmen s o assess he scalabili y and
e ec i eness o MMPT in eal-wo ld se ings.
Mul imodal Misconcep ion De ec ion:
MMPT may be enhanced by in eg a ing
addi ional da a modali ies, such as eye-
acking, keys oke dynamics, and lea ne s’
na u al language explana ions, o build mo e
comp ehensi e lea ne models.
Adap i e Feedback Gene a ion wi h
Explainable AI: Imp o emen s o he
eedback module will explo e he use o
explainable AI echniques o deli e con ex -
sensi i e, na u al language explana ions and
isual g aph-based eedback.
Gami ica ion and Mo i a ion S a egies:
Inco po a ing gami ica ion elemen s—such as
p og ession le els, achie emen badges, and
pee pe o mance compa isons—could u he
enhance lea ne mo i a ion and engagemen .
C oss-Domain Ex ension: The cogni i e
g aph amewo k unde pinning MMPT is
domain-agnos ic and may be adap ed o use
in o he a eas such as ma hema ics, logic, and
language lea ning, o e ing a ounda ion o
uni e sal adap i e u o ing.
E hical and Fai ness Conside a ions:
Ongoing wo k will also in es iga e he
ai ness and anspa ency o he sys em,
including he mi iga ion o po en ial biases in
misconcep ion de ec ion ac oss di e en
demog aphic g oups o ensu e equi able access
and ou comes.
Re e ences:
1. Ande son, J. R., Co be , A. T.,
Koedinge , K. R., & Pelle ie , R. (1995).
In elligen u o ing sys ems. Science,
268(5210), 456–462.
[h ps://doi.o g/10.1126/science.7701348]
(h ps://doi.o g/10.1126/science.7701348)
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Deo Sha mila Mahesh
206
2. B own, T. B., Mann, B., Ryde , N.,
Subbiah, M., Kaplan, J. D., Dha iwal, P.,
& Amodei, D. (2020). Language models
a e ew-sho lea ne s. ‘Ad ances in
Neu al In o ma ion P ocessing Sys ems’,
‘33’, 1877–1901.
3. Piech, C., Bassen, J., Huang, J., Ganguli,
S., Sahami, M., Guibas, L. J., & Sohl-
Dicks ein, J. (2015). Deep knowledge
acing. In ‘Ad ances in Neu al
In o ma ion P ocessing Sys ems’ (pp.
505–513).
4. VanLehn, K. (2006). The beha io o
u o ing sys ems. ‘In e na ional Jou nal o
A i icial In elligence in Educa ion’,
‘16’(3), 227–265.
