Course outline
Module 1
Module 1: Introduction to Recommendation Systems
Introduction to Recommendation Systems
Knowledge check
Q1. What is the primary goal of a recommendation system on a social media platform?
- A. To randomly display posts to all users equally
- B. To predict and surface content a user is most likely to engage with ✓
- C. To delete content that has fewer than 100 views
- D. To store every post a user has ever seen in chronological order
Why: Recommendation systems aim to predict which items a user will find relevant or engaging, such as posts, videos, or accounts, and rank them accordingly to keep the user interested.
Q2. Which type of signal is considered an implicit feedback signal in a social media recommendation system?
- A. A user explicitly rating a post 5 stars
- B. A user typing a comment on a post
- C. A user spending 30 seconds watching a video without interacting ✓
- D. A user filling out a survey about their interests
Why: Implicit feedback is inferred from user behavior, such as watching time, scrolling pauses, or clicks, rather than being directly stated by the user. Explicit feedback includes ratings, comments, and surveys.
Q3. What does collaborative filtering rely on to make recommendations?
- A. The text content of the item only, such as keywords in a caption
- B. Patterns of behavior across many users with similar tastes ✓
- C. The user's physical location and device model only
- D. Manually written rules set by platform engineers for each post
Why: Collaborative filtering assumes that users who agreed in the past will agree in the future, so it uses similarities among users' interaction patterns to suggest items one user hasn't seen yet.
Q4. What is the 'cold start' problem in recommendation systems?
- A. A server overheating due to high traffic during peak hours
- B. Difficulty recommending content for new users or new items that have little interaction data ✓
- C. Users getting bored after viewing the same content too many times
- D. A delay in loading images when the internet connection is slow
Why: The cold start problem occurs when the system lacks enough historical data about a new user or a newly uploaded item to make accurate predictions, which is why platforms often ask new users about their interests or show popular content at first.
Q5. Why do social media algorithms often rank content using a combination of multiple predicted signals rather than a single score?
- A. Because the law requires platforms to use at least three ranking factors
- B. Because user engagement is multi-dimensional, so predicting likes, shares, watch time, and comments separately can give a more balanced ranking ✓
- C. Because a single score would take up too much storage space
- D. Because multiple signals make posts load faster on mobile devices
Why: Engagement comes in many forms, and a post might be likely to earn a like but unlikely to be shared. Combining several predicted outcomes lets the system balance these aspects into a single ranking that reflects overall value to the user.
Module 2
Module 2: How Feeds Are Ranked
How Feeds Are Ranked
Knowledge check
Q1. What is the primary purpose of a social media feed ranking algorithm?
- A. To display posts in strict chronological order
- B. To predict and order content by how relevant or engaging it is likely to be for each user ✓
- C. To randomly shuffle posts so every user sees different content
- D. To show only posts from accounts the user has never interacted with
Why: Feed ranking algorithms score each candidate post for a specific user and order them by predicted relevance or engagement, rather than by time alone or at random.
Q2. Which of the following is an example of an engagement signal a ranking algorithm might use?
- A. The user's physical location in the real world
- B. The time the user spends viewing a post before scrolling past ✓
- C. The user's password strength
- D. The device's battery level
Why: Dwell time, likes, comments, and shares are common engagement signals. Location, password strength, and battery level are not typically used to rank content relevance.
Q3. What does it mean when a feed is described as 'personalized'?
- A. Every user sees exactly the same posts in the same order
- B. Ranking is tailored to an individual user's past behavior and inferred preferences ✓
- C. Only the account owner's own posts are displayed
- D. Posts are hand-selected by human editors for each user
Why: Personalization means the ranking system uses an individual's history and inferred interests to decide what they see, so different users get different feeds.
Q4. Why do ranking systems often predict multiple outcomes for each post, such as the probability of a like, comment, or share?
- A. Because each action is weighted and combined into an overall score that determines placement ✓
- B. Because the law requires platforms to track each action separately
- C. Because the post can only display one type of interaction
- D. Because predictions reduce the total number of posts stored on servers
Why: Models estimate several interaction probabilities and combine them with weights into a single relevance score, which is then used to rank the post.
Q5. What is a common reason a post from a friend might appear lower in your feed than expected?
- A. The algorithm predicts you are unlikely to engage with it based on your past behavior ✓
- B. The friend's account was created recently
- C. Friends' posts are always placed at the bottom by default
- D. The post contains fewer than ten words
Why: Because ranking is predictive, a post from a close connection can still rank low if the model expects little engagement from you, regardless of the relationship.
Module 3
Module 3: Engagement Signals and Metrics
Engagement Signals and Metrics
Knowledge check
Q1. Which of the following is generally considered a strong engagement signal for a social media algorithm?
- A. The number of times a user scrolls past a post without pausing
- B. The amount of time a user spends watching or reading a post ✓
- C. The total number of followers the poster has
- D. The time of day the post was created
Why: Dwell time, or how long a user spends viewing or consuming content, signals genuine interest. Algorithms use it to judge whether a post is worth showing to more people.
Q2. What is the main difference between a 'like' and a 'share' as engagement signals?
- A. A like is weighted more heavily than a share because it is easier to give
- B. A share typically signals stronger endorsement because the user is putting their own reputation behind the content ✓
- C. Likes are only counted on video posts, while shares apply to all content
- D. Shares are ignored by algorithms because they take users off the platform
Why: Sharing requires the user to republish content to their own audience, which signals stronger approval than a single tap of a like. Many algorithms therefore weight shares more heavily.
Q3. Why might a post with many comments be ranked higher by an algorithm, even if it has fewer likes?
- A. Comments are always free of spam and therefore more valuable
- B. Comments indicate active back-and-forth interaction, which platforms treat as a strong sign of meaningful engagement ✓
- C. Algorithms only count comments that include emojis
- D. Comments reduce the number of ads shown to users
Why: Comments take more effort than likes and often create conversation threads. Platforms interpret this as higher-quality engagement and may boost posts that spark discussion.
Q4. Which metric is most directly described as a 'negative' engagement signal?
- A. Saving a post for later
- B. Sending a post to a friend
- C. Hiding a post or marking it as 'not interested' ✓
- D. Clicking a profile picture
Why: Hiding or marking content as not interested tells the algorithm the user did not want to see it, so similar content is likely to be shown less often.
Q5. What is a common reason engagement metrics can be misleading when judging content quality?
- A. Engagement is only measured in private messages, which are not public
- B. High engagement can come from outrage or controversy, not necessarily from content users find valuable ✓
- C. Algorithms never measure engagement after a post is published
- D. Engagement metrics are the same for every type of content
Why: Content that provokes anger or strong reactions can generate lots of comments and shares without being helpful or high quality, so raw engagement does not always reflect value.
Module 4
Module 4: User Profiling and Personalization
User Profiling and Personalization
Knowledge check
Q1. What is the primary purpose of user profiling in social media algorithms?
- A. To verify the real identity of each user
- B. To predict what content a user is most likely to engage with
- C. To delete accounts that have been inactive for a long time
- D. To store users' passwords securely
Q2. Which type of data is most commonly used to personalize a user's social media feed?
- A. The user's physical location only
- B. The user's browsing and engagement history, such as likes, shares, and watch time
- C. The user's credit card information
- D. Random content selected with no reference to user behavior
Q3. What is meant by 'implicit' data in user profiling?
- A. Data the user explicitly provides, such as profile bio or stated interests
- B. Data inferred from behavior, such as how long a user pauses on a post
- C. Data that has been encrypted and cannot be read
- D. Data that is deleted after 24 hours
Q4. How does personalization often lead to what is called a 'filter bubble'?
- A. By showing users only content from advertisers they have never interacted with
- B. By repeatedly showing content similar to what a user already engages with, limiting exposure to differing views
- C. By automatically blocking all content from friends and family
- D. By requiring users to log in every time they open the app
Q5. Why might a platform use a user's past engagement to recommend new content?
- A. Because past engagement is the only data the platform is legally allowed to store
- B. Because past behavior is treated as a signal of future interests, helping keep the user active on the platform
- C. Because recommendations must be identical for all users under platform rules
- D. Because the algorithm is designed to ignore user preferences entirely
Module 5
Module 5: Content Classification and Filtering
Content Classification and Filtering
Knowledge check
Q1. What is the primary purpose of content classification in social media algorithms?
- A. To delete all user posts automatically
- B. To categorize content by topic, type, and characteristics so it can be ranked and filtered appropriately ✓
- C. To encrypt messages between users
- D. To count the number of followers each user has
Why: Content classification labels posts by topic, format, and other features so that algorithms can decide what to show, hide, or rank for each user.
Q2. Which of the following is a common signal used to classify a post's topic?
- A. The user's local weather forecast
- B. The text, hashtags, and detected objects or speech within the content ✓
- C. The battery level of the user's device
- D. The time the user last changed their password
Why: Platforms analyze text, hashtags, images (via computer vision), and audio transcripts to infer what a piece of content is about.
Q3. What is a 'false positive' in content filtering?
- A. A post that is correctly flagged as harmful
- B. A post that is wrongly flagged as violating a policy when it actually does not ✓
- C. A user who reports a post that is later removed
- D. A filter that runs faster than expected
Why: A false positive occurs when the filter incorrectly identifies benign content as violating, such as flagging a medical article as graphic content.
Q4. Why do platforms often combine automated classifiers with human reviewers?
- A. Automated classifiers cannot process any data at all
- B. Humans can catch context and nuance that automated systems may miss, improving accuracy on borderline cases ✓
- C. Human reviewers are required by law to read every post
- D. Classifiers only work on text, never images
Why: Hybrid systems use automation for scale and humans for nuanced or borderline decisions, since automated tools often struggle with sarcasm, cultural context, and evolving slang.
Q5. What is the main reason classification models can show bias against certain groups?
- A. They are programmed to dislike specific users
- B. Training data may underrepresent or mislabel certain dialects, languages, or communities ✓
- C. They only run on weekends
- D. They ignore all user feedback entirely
Why: If training data contains more examples from some groups or labels certain dialects as negative, the model learns and reproduces those skews, leading to uneven enforcement.
Module 6
Module 6: Feedback Loops and Filter Bubbles
Feedback Loops and Filter Bubbles
Knowledge check
Q1. What is a feedback loop in a social media recommendation algorithm?
- A. A technical error that causes the app to crash repeatedly
- B. A cycle where a user's engagement shapes the content shown, which in turn shapes future engagement ✓
- C. A method for users to send direct messages to each other
- D. A process that deletes old posts after a set period of time
Why: A feedback loop occurs when the content a user engages with informs what the algorithm shows them next, and that new content shapes their subsequent behavior, creating a self-reinforcing cycle.
Q2. What is a filter bubble?
- A. A security feature that blocks spam accounts
- B. A setting that lets users hide their profile from search engines
- C. A state in which a user is mostly exposed to information that matches their existing views and interests ✓
- D. A type of advertisement that follows users across websites
Why: A filter bubble describes the isolation that can result when algorithms prioritize content similar to what a user already likes or believes, limiting exposure to differing perspectives.
Q3. Which signal is typically weighted most heavily by engagement-based recommendation algorithms?
- A. The user's account age
- B. The user's location on a map
- C. The user's spelling accuracy in comments
- D. Behavioral signals such as likes, watch time, shares, and comments ✓
Why: Recommendation systems rely heavily on behavioral data like how long someone watches, what they like, and what they share, because these signals indicate interest and predict future engagement.
Q4. Why can feedback loops cause a user's feed to gradually narrow over time?
- A. Because the algorithm deletes all content after a user views it once
- B. Because the platform limits each user to a fixed number of posts per day
- C. Because each click on similar content signals the algorithm to show even more of it, reducing variety ✓
- D. Because users are required by law to follow only one topic
Why: Each interaction with a certain type of content reinforces the algorithm's belief that the user wants more of it, so the feed becomes progressively more concentrated on those topics and less diverse.
Q5. Which of the following is a common way users can reduce the effects of a filter bubble?
- A. Turning off their phone's notifications entirely
- B. Deliberately engaging with diverse sources and perspectives, and adjusting feed or interest settings where available ✓
- C. Changing their username every week
- D. Only using the platform at night
Why: Intentionally seeking out and interacting with a range of viewpoints sends the algorithm different signals, and many platforms offer settings like 'not interested' or chronological feeds that can broaden what users see.
Module 7
Module 7: Auditing and Transparency Tools
Auditing and Transparency Tools
Knowledge check
Q1. Which of the following best describes the primary purpose of an algorithm audit on a social media platform?
- A. To increase the number of ads shown to users
- B. To systematically examine how a platform's ranking and recommendation systems behave and what outcomes they produce ✓
- C. To delete accounts that violate community guidelines
- D. To encrypt user messages so they cannot be read by third parties
Why: Algorithm audits aim to systematically test and examine how ranking and recommendation systems behave, including what content they amplify or suppress and whether outcomes are fair or harmful.
Q2. What is a 'sock puppet' account in the context of auditing social media algorithms?
- A. An account that a platform automatically generates to test its own servers
- B. A fake or controlled account created by researchers to observe how the algorithm responds to specific behaviors ✓
- C. An account that has been permanently banned for spam
- D. A verified business account used for advertising
Why: Researchers create sock puppet accounts with controlled histories and behaviors so they can observe and compare how the recommendation system responds to different inputs.
Q3. Which transparency tool allows the public to see the ads a page is running, including who paid for them?
- A. A platform's public ad library or ad transparency center ✓
- B. The platform's terms of service document
- C. A browser's privacy dashboard
- D. The platform's daily trending topics list
Why: Many platforms maintain public ad libraries that display active and past ads, the advertiser's identity, and sometimes targeting and spend information, enabling external scrutiny.
Q4. Why is access to platform data important for independent auditing of algorithms?
- A. It allows auditors to change the algorithm's code directly on the live site
- B. Without it, external researchers must infer algorithmic behavior indirectly, which limits the reliability and scope of findings ✓
- C. It guarantees that the platform will remove all harmful content
- D. It eliminates the need for user consent in research
Why: Independent auditors often rely on limited or indirect methods when platforms restrict data access, which constrains how thoroughly and accurately they can evaluate algorithmic effects.
Q5. What does a 'chronological feed' option offer users in terms of transparency?
- A. It shows posts ranked purely by algorithmic engagement prediction
- B. It displays posts in the order they were published, offering a simpler and more predictable view of content ✓
- C. It hides all posts from accounts the user does not follow
- D. It automatically translates all posts into the user's language
Why: A chronological feed presents posts in publication order rather than ranked by predicted engagement, making the feed's ordering easier for users to understand and compare against algorithmic defaults.