Loading…

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ1TZ2

Session: May 2023
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2023 TZ1 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

Cloud deployment architectures and distinguishing PaaS from SaaS

Networks
Resource management

(a)

2 marks

State two standard cloud deployment models utilized for hosting enterprise applications:
1) Public cloud: Computing resources, servers, and storage infrastructure are owned and operated by a third-party cloud vendor (e.g., AWS, GCP, Azure) and delivered over the public internet to multiple client organizations on a shared multi-tenant basis.
2) Private cloud: Cloud infrastructure is provisioned for exclusive use by a single organization, operated on-premises or hosted by a service provider behind private firewalls for maximum data isolation.

Mentor tip

Name two distinct cloud deployment paradigms such as public, private, hybrid, or community cloud.

(b)

2 marks

Identify two differences between Platform as a Service (PaaS) and Software as a Service (SaaS):
1) Target audience and purpose: PaaS provides application runtime engines, database tools, and software development frameworks for developers to build and deploy custom code. SaaS delivers fully finished end-user applications hosted and maintained entirely by the vendor for direct business consumption.
2) Control over data and application configuration: In PaaS, client developers retain full ownership, architecture design, and management over their application logic and stored database records. In SaaS, the cloud vendor manages both the software platform and the underlying application data storage.

Mentor tip

Contrast developer-facing runtime environments (PaaS) with fully managed consumer applications (SaaS).

Question 2

8 marks

User privacy governance with behavioural data and strategies to mitigate overfitting

System fundamentals
Modelling and simulation

(a)

4 marks

Explain how NextStar can ensure that users' fundamental right to privacy is protected when collecting implicit and explicit behavioural data:
1) Data pseudonymization and anonymization: NextStar should strip direct identifiers (such as user names, government IDs, and email addresses) from interaction logs, replacing them with salted cryptographic pseudonyms or synthetic IDs linked via an encrypted, access-restricted lookup table.
2) Transparent disclosure and granular consent controls: The platform must clearly inform users about tracking mechanisms (such as watch duration, search history, and clickstreams) upon registration, providing an intuitive settings dashboard allowing users to opt out of targeted advertising without losing core functionality.
3) Data minimization and strict purpose limitation: NextStar should gather only the behavioural telemetry strictly necessary to generate recommendations and deliver adverts, avoiding invasive monitoring tools (such as keyloggers or microphone access).
4) Automated retention schedules and secure cryptographic storage: Stored interaction databases must be encrypted at rest and in transit, with automated data retention policies that permanently scrub obsolete behavioural logs after a defined operational period.

Mentor tip

Structure the explanation around pseudonymization/encryption, transparent consent/opt-out mechanisms, data minimization, and automated retention limits.

(b)

4 marks

Outline two strategies for reducing the risk of overfitting in supervised machine learning models:
1) Regularization penalty integration: Introducing an analytical penalty term (such as L1 LASSO or L2 Ridge regularization) into the loss function penalizes overly large weight coefficients, constraining model complexity and preventing the algorithm from fitting noise in the training set.
2) Cross-validation and early stopping: Partitioning the dataset into $k$ distinct training and validation folds (k-fold cross-validation) enables continuous validation tracking during training; early stopping halts optimization as soon as validation loss begins to diverge or rise, preventing the model from over-adapting to the training slice.

L_{regularized} = L_{base} + λ j \sum w_{j}^{2}

Mentor tip

Outline two established machine learning strategies such as L1/L2 regularization, k-fold cross-validation with early stopping, or feature dimensionality reduction.

Question 3

6 marks

Evaluating recommender system accuracy and harmonic precision-recall using the F-measure

Modelling and simulation

6 marks

Explain how the F-measure metric is formulated and applied to evaluate NextStar's video audition recommender system:
1) Categorizing audition predictions into the confusion matrix:
- Total audition video clips evaluated: $N = 20$ .
- Recommended by system (predicted positive): 5 clips.
- Not recommended by system (predicted negative): 15 clips.
- Actual roles assigned by casting recruiters (actual positive): 4 actors.
- Actual actors rejected (actual negative): $20 - 4 = 16$ actors.
Map outcomes across the four contingency categories:
- True Positives ( $T P$ ): Recommended by system AND received a role: $T P = 3$ .
- False Positives ( $F P$ ): Recommended by system BUT did not receive a role: $F P = 5 - 3 = 2$ .
- False Negatives ( $F N$ ): Not recommended by system BUT received a role: $F N = 4 - 3 = 1$ .
- True Negatives ( $T N$ ): Not recommended by system AND did not receive a role: $T N = 15 - 1 = 14$ .
2) Calculating precision:
Precision measures the proportion of recommended videos that were genuinely relevant:
$Precision = \frac{T P}{T P + F P} = \frac{3}{3 + 2} = \frac{3}{5} = 0.60 (60%)$
3) Calculating recall:
Recall measures the proportion of successful actors that the system successfully identified:
$Recall = \frac{T P}{T P + F N} = \frac{3}{3 + 1} = \frac{3}{4} = 0.75 (75%)$
4) Calculating the balanced $F_{1}$ -measure:
The $F_{1}$ -score is the harmonic mean of precision and recall, balancing false discovery against missed talent:
$F_{1} = 2 \times \frac{Precision \times Recall}{Precision + Recall} = 2 \times \frac{0.60 \times 0.75}{0.60 + 0.75} = \frac{0.90}{1.35} = \frac{2}{3} \approx 0.667 (66.7%)$
5) Strategic application in system tuning: NextStar uses this $F_{1}$ -score benchmark to evaluate algorithmic updates. When engineers adjust hyperparameter thresholds or similarity algorithms, they compare resulting $F_{1}$ values; an increased $F_{1}$ score confirms improved talent discovery without overwhelming casting agents with irrelevant clips.

Precision = \frac{T P}{T P + F P} = \frac{3}{5} = 0.60 Recall = \frac{T P}{T P + F N} = \frac{3}{4} = 0.75 F_{1} = \frac{2 \cdot Precision \cdot Recall}{Precision + Recall} = \frac{2 \cdot 0.60 \cdot 0.75}{0.60 + 0.75} = \frac{2}{3} \approx 0.667

Mentor tip

Explicitly identify $T P = 3$ , $F P = 2$ , $F N = 1$ , compute precision (60%) and recall (75%), and calculate the harmonic mean ( $F_{1} = 66.7%$ ).

Question 4

12 marks

Architectural evaluation of content-based, collaborative, and hybrid recommender systems

Modelling and simulation

12 marks

Discuss the relative advantages and disadvantages of content-based filtering, collaborative filtering, and hybrid architectures for the NextStar media discovery platform:
Content-based filtering analysis:
- Advantages: Does not require extensive historical interaction records from other platform users, making it effective for recommending newly uploaded niche videos provided metadata tags are available. User privacy is preserved because individual recommendation models depend solely on that specific user's interaction profile rather than aggregating community data.
- Disadvantages: Highly vulnerable to the 'filter bubble' or over-specialization; the algorithm only recommends items closely matching previously viewed features (e.g., suggesting only rap music if the user watched hip-hop), failing to introduce novel creative genres. Furthermore, it depends entirely on comprehensive, accurate item feature extraction (e.g., manual tags, video genre annotations, audio features).
Collaborative filtering analysis:
- Advantages: Disregards low-level item metadata entirely, operating purely on user-item interaction matrices. This enables serendipitous discovery: if User A and User B share similar viewing preferences, the platform can suggest an unexpected genre to User A simply because User B enjoyed it.
- Disadvantages: Suffers severely from the 'cold start' dilemma: new users with zero watch history cannot receive meaningful recommendations, and newly uploaded artist videos with zero ratings cannot be recommended to anyone. It also faces matrix sparsity (users rate only a tiny fraction of total content), non-linear computational scaling when calculating pairwise user/item similarities, and popularity bias.
Hybrid recommender system architectures:
- Feature-weighted combining: Combines predictions from both content-based algorithms (evaluating video genre and audio descriptors) and collaborative filtering (matrix factorization) using a weighted ensemble score.
- Switching and cascading pipelines: Deploys content-based filtering for new users and fresh uploads to bypass cold-start barriers, transitioning to collaborative matrix factorization as interaction logs accumulate.
- Context-aware deep embeddings: Utilizes deep neural networks to project textual artist tags, video feature embeddings, and collaborative interaction vectors into a unified latent space.
Strategic conclusion and recommendation for NextStar:
Neither standalone content-based filtering nor standalone collaborative filtering can fully satisfy NextStar's dual mission of discovering new talent and monetizing user engagement. NextStar should build a hybrid recommender architecture: deploying a lightweight content-based engine using automated video feature tagging to support immediate upload discovery and new user onboarding, combined with collaborative matrix factorization trained periodically on popular user interactions to drive long-term engagement and serendipitous media exploration.

Mentor tip

Address both approaches systematically: content-based (metadata-driven, immune to user cold-start, but suffers over-specialization) vs collaborative (interaction-driven, serendipitous, but suffers cold-start and sparsity), concluding with a hybrid solution.

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Modelling and simulation carries the most marks on this paper.

Modelling and simulation22 marks · 73%
Q2, Q3, Q4
System fundamentals4 marks · 13%
Q2
Networks2 marks · 7%
Q1
Resource management2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing cloud deployment models with cloud service models · Q1
Public, private, and hybrid are cloud deployment models; IaaS, PaaS, and SaaS are cloud service delivery tiers.
02
Calculating arithmetic mean instead of harmonic mean for F-measure · Q3
The F-measure is the harmonic mean of precision and recall ( $2 P R / (P + R)$ ), which penalizes extreme imbalances between precision and recall far more heavily than a simple arithmetic average.
03
Assuming collaborative filtering analyzes video metadata or text descriptions · Q4
Pure collaborative filtering inspects only the user-item interaction matrix and user ratings; it is entirely agnostic to video content, genres, or artist descriptions.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ1TZ2

Session: May 2023
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2023 TZ1 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

Cloud deployment architectures and distinguishing PaaS from SaaS

Networks
Resource management

(a)

2 marks

State two standard cloud deployment models utilized for hosting enterprise applications:
1) Public cloud: Computing resources, servers, and storage infrastructure are owned and operated by a third-party cloud vendor (e.g., AWS, GCP, Azure) and delivered over the public internet to multiple client organizations on a shared multi-tenant basis.
2) Private cloud: Cloud infrastructure is provisioned for exclusive use by a single organization, operated on-premises or hosted by a service provider behind private firewalls for maximum data isolation.

Mentor tip

Name two distinct cloud deployment paradigms such as public, private, hybrid, or community cloud.

(b)

2 marks

Identify two differences between Platform as a Service (PaaS) and Software as a Service (SaaS):
1) Target audience and purpose: PaaS provides application runtime engines, database tools, and software development frameworks for developers to build and deploy custom code. SaaS delivers fully finished end-user applications hosted and maintained entirely by the vendor for direct business consumption.
2) Control over data and application configuration: In PaaS, client developers retain full ownership, architecture design, and management over their application logic and stored database records. In SaaS, the cloud vendor manages both the software platform and the underlying application data storage.

Mentor tip

Contrast developer-facing runtime environments (PaaS) with fully managed consumer applications (SaaS).

Question 2

8 marks

User privacy governance with behavioural data and strategies to mitigate overfitting

System fundamentals
Modelling and simulation

(a)

4 marks

Explain how NextStar can ensure that users' fundamental right to privacy is protected when collecting implicit and explicit behavioural data:
1) Data pseudonymization and anonymization: NextStar should strip direct identifiers (such as user names, government IDs, and email addresses) from interaction logs, replacing them with salted cryptographic pseudonyms or synthetic IDs linked via an encrypted, access-restricted lookup table.
2) Transparent disclosure and granular consent controls: The platform must clearly inform users about tracking mechanisms (such as watch duration, search history, and clickstreams) upon registration, providing an intuitive settings dashboard allowing users to opt out of targeted advertising without losing core functionality.
3) Data minimization and strict purpose limitation: NextStar should gather only the behavioural telemetry strictly necessary to generate recommendations and deliver adverts, avoiding invasive monitoring tools (such as keyloggers or microphone access).
4) Automated retention schedules and secure cryptographic storage: Stored interaction databases must be encrypted at rest and in transit, with automated data retention policies that permanently scrub obsolete behavioural logs after a defined operational period.

Mentor tip

Structure the explanation around pseudonymization/encryption, transparent consent/opt-out mechanisms, data minimization, and automated retention limits.

(b)

4 marks

Outline two strategies for reducing the risk of overfitting in supervised machine learning models:
1) Regularization penalty integration: Introducing an analytical penalty term (such as L1 LASSO or L2 Ridge regularization) into the loss function penalizes overly large weight coefficients, constraining model complexity and preventing the algorithm from fitting noise in the training set.
2) Cross-validation and early stopping: Partitioning the dataset into $k$ distinct training and validation folds (k-fold cross-validation) enables continuous validation tracking during training; early stopping halts optimization as soon as validation loss begins to diverge or rise, preventing the model from over-adapting to the training slice.

L_{regularized} = L_{base} + λ j \sum w_{j}^{2}

Mentor tip

Outline two established machine learning strategies such as L1/L2 regularization, k-fold cross-validation with early stopping, or feature dimensionality reduction.

Question 3

6 marks

Evaluating recommender system accuracy and harmonic precision-recall using the F-measure

Modelling and simulation

6 marks

Explain how the F-measure metric is formulated and applied to evaluate NextStar's video audition recommender system:
1) Categorizing audition predictions into the confusion matrix:
- Total audition video clips evaluated: $N = 20$ .
- Recommended by system (predicted positive): 5 clips.
- Not recommended by system (predicted negative): 15 clips.
- Actual roles assigned by casting recruiters (actual positive): 4 actors.
- Actual actors rejected (actual negative): $20 - 4 = 16$ actors.
Map outcomes across the four contingency categories:
- True Positives ( $T P$ ): Recommended by system AND received a role: $T P = 3$ .
- False Positives ( $F P$ ): Recommended by system BUT did not receive a role: $F P = 5 - 3 = 2$ .
- False Negatives ( $F N$ ): Not recommended by system BUT received a role: $F N = 4 - 3 = 1$ .
- True Negatives ( $T N$ ): Not recommended by system AND did not receive a role: $T N = 15 - 1 = 14$ .
2) Calculating precision:
Precision measures the proportion of recommended videos that were genuinely relevant:
$Precision = \frac{T P}{T P + F P} = \frac{3}{3 + 2} = \frac{3}{5} = 0.60 (60%)$
3) Calculating recall:
Recall measures the proportion of successful actors that the system successfully identified:
$Recall = \frac{T P}{T P + F N} = \frac{3}{3 + 1} = \frac{3}{4} = 0.75 (75%)$
4) Calculating the balanced $F_{1}$ -measure:
The $F_{1}$ -score is the harmonic mean of precision and recall, balancing false discovery against missed talent:
$F_{1} = 2 \times \frac{Precision \times Recall}{Precision + Recall} = 2 \times \frac{0.60 \times 0.75}{0.60 + 0.75} = \frac{0.90}{1.35} = \frac{2}{3} \approx 0.667 (66.7%)$
5) Strategic application in system tuning: NextStar uses this $F_{1}$ -score benchmark to evaluate algorithmic updates. When engineers adjust hyperparameter thresholds or similarity algorithms, they compare resulting $F_{1}$ values; an increased $F_{1}$ score confirms improved talent discovery without overwhelming casting agents with irrelevant clips.

Precision = \frac{T P}{T P + F P} = \frac{3}{5} = 0.60 Recall = \frac{T P}{T P + F N} = \frac{3}{4} = 0.75 F_{1} = \frac{2 \cdot Precision \cdot Recall}{Precision + Recall} = \frac{2 \cdot 0.60 \cdot 0.75}{0.60 + 0.75} = \frac{2}{3} \approx 0.667

Mentor tip

Explicitly identify $T P = 3$ , $F P = 2$ , $F N = 1$ , compute precision (60%) and recall (75%), and calculate the harmonic mean ( $F_{1} = 66.7%$ ).

Question 4

12 marks

Architectural evaluation of content-based, collaborative, and hybrid recommender systems

Modelling and simulation

12 marks

Discuss the relative advantages and disadvantages of content-based filtering, collaborative filtering, and hybrid architectures for the NextStar media discovery platform:
Content-based filtering analysis:
- Advantages: Does not require extensive historical interaction records from other platform users, making it effective for recommending newly uploaded niche videos provided metadata tags are available. User privacy is preserved because individual recommendation models depend solely on that specific user's interaction profile rather than aggregating community data.
- Disadvantages: Highly vulnerable to the 'filter bubble' or over-specialization; the algorithm only recommends items closely matching previously viewed features (e.g., suggesting only rap music if the user watched hip-hop), failing to introduce novel creative genres. Furthermore, it depends entirely on comprehensive, accurate item feature extraction (e.g., manual tags, video genre annotations, audio features).
Collaborative filtering analysis:
- Advantages: Disregards low-level item metadata entirely, operating purely on user-item interaction matrices. This enables serendipitous discovery: if User A and User B share similar viewing preferences, the platform can suggest an unexpected genre to User A simply because User B enjoyed it.
- Disadvantages: Suffers severely from the 'cold start' dilemma: new users with zero watch history cannot receive meaningful recommendations, and newly uploaded artist videos with zero ratings cannot be recommended to anyone. It also faces matrix sparsity (users rate only a tiny fraction of total content), non-linear computational scaling when calculating pairwise user/item similarities, and popularity bias.
Hybrid recommender system architectures:
- Feature-weighted combining: Combines predictions from both content-based algorithms (evaluating video genre and audio descriptors) and collaborative filtering (matrix factorization) using a weighted ensemble score.
- Switching and cascading pipelines: Deploys content-based filtering for new users and fresh uploads to bypass cold-start barriers, transitioning to collaborative matrix factorization as interaction logs accumulate.
- Context-aware deep embeddings: Utilizes deep neural networks to project textual artist tags, video feature embeddings, and collaborative interaction vectors into a unified latent space.
Strategic conclusion and recommendation for NextStar:
Neither standalone content-based filtering nor standalone collaborative filtering can fully satisfy NextStar's dual mission of discovering new talent and monetizing user engagement. NextStar should build a hybrid recommender architecture: deploying a lightweight content-based engine using automated video feature tagging to support immediate upload discovery and new user onboarding, combined with collaborative matrix factorization trained periodically on popular user interactions to drive long-term engagement and serendipitous media exploration.

Mentor tip

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Modelling and simulation carries the most marks on this paper.

Modelling and simulation22 marks · 73%
Q2, Q3, Q4
System fundamentals4 marks · 13%
Q2
Networks2 marks · 7%
Q1
Resource management2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing cloud deployment models with cloud service models · Q1
Public, private, and hybrid are cloud deployment models; IaaS, PaaS, and SaaS are cloud service delivery tiers.
02
Calculating arithmetic mean instead of harmonic mean for F-measure · Q3
The F-measure is the harmonic mean of precision and recall ( $2 P R / (P + R)$ ), which penalizes extreme imbalances between precision and recall far more heavily than a simple arithmetic average.
03
Assuming collaborative filtering analyzes video metadata or text descriptions · Q4
Pure collaborative filtering inspects only the user-item interaction matrix and user ratings; it is entirely agnostic to video content, genres, or artist descriptions.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers