You swore you’d be asleep by 11 p.m. However, somewhere between watching five-minute breakfast ideas and a string of true crime shorts, you lost track of time. Now it’s nearing 2 a.m., and you’re still glued to Instagram Reels, with each clip more gripping than the last. You didn’t plan to keep watching. Yet somehow, every swipe feels like it was made for you.

That’s the hallmark of a recommender system.

What pulls you in isn’t random. It’s the result of a system trained to anticipate what you’ll engage with, based on what you’ve already done.¹ Every time you pause, like, share, or linger just a moment longer than usual, the system registers that signal. From there, it recalibrates and subtly reshapes your feed to keep you engaged the next time you open the app—or even just swipe again.

To most users, this process is invisible. You’re not aware that your taps, rewatches, and hesitations are being interpreted in real time, but they are—and the system is always adapting. Behind the scenes, a recommender system is a form of artificial intelligence (AI), often powered by machine learning.² It draws on big data, which are enormous, continuously updated datasets built from millions of user interactions. The goal is to recognize patterns in behavior and serve up suggestions that feel personal, even though they’re built from collective behavior.

There are three major ways recommender systems operate. One is collaborative filtering, which looks for similarities between users. If people who watched the same dog training videos as you also spent time on a niche baking channel, that channel might now appear in your feed. The logic isn’t based on content: it’s based on clusters of shared behavior.

Another method, content-based filtering, flips the focus. It zeroes in on the characteristics of the content itself—whether it's a video’s tone, subject matter, length, or pacing—and matches those attributes to your known preferences. If you’ve recently interacted with three softly narrated cleaning tutorials, the system might queue up a fourth, even if no one else in your circle has seen it.

Increasingly, platforms use hybrid filtering, which blends the two. These systems track both who’s watching and what’s being watched, then use those layers to refine their predictions. Ideally, these systems surface content that feels familiar enough to catch your eye because it resembles something you’ve liked before, but different enough to spark curiosity so it doesn’t feel like the same thing on repeat.

Of course, the mechanics don’t stop there.

Every scroll you skip, every playlist you abandon, even the videos you replay without liking, all shape the system’s understanding of you. It doesn’t just log what you enjoyed—it infers what you avoided, what bored you, what you’re likely to binge at 1 a.m. versus what you’ll skip without a second glance. 

This process isn’t limited to a single platform. Recommender systems now shape nearly every corner of digital life, from the late-night videos that keep you swiping to the recipes, products, and playlists you didn’t plan to search for. Some rely heavily on repetition, echoing familiar formats and voices until your feed feels more confined than curated, while others aim for discovery.³ When designed with care, they can introduce unfamiliar perspectives, elevate lesser-known creators, and surface content that feels surprising yet still relevant. Sometimes they lead you somewhere unexpected and oddly helpful, even if it wasn’t what you thought you needed.

Artificial Intelligence (AI): A field of computer science dedicated to creating systems that can perform tasks traditionally requiring human intelligence, such as recognizing patterns, interpreting language, solving problems, or making decisions.

Machine Learning: A subset of AI focused on building algorithms that learn from data and improve through experience. Rather than following a static set of instructions, these models identify patterns and adjust their predictions as new information becomes available.

Big Data: Extremely large and complex datasets generated through everyday digital activity, such as online searches, location tracking, purchase behavior, sensor readings, and medical records. These data are analyzed to uncover patterns that support real-world decisions. For example, big data can be used to forecast flu outbreaks, recommend products, reroute traffic, or adjust public transit schedules in real time.

Collaborative filtering: A recommendation method that predicts what you might like based on the preferences of other users with similar behavior. It doesn’t look at the content itself, but rather at patterns in how users rate or interact with items. For example, if people who enjoyed the same podcasts as you also liked a particular show, that show may be recommended to you.

Content-based filtering: A recommendation method that analyzes the features of items you’ve liked in the past and suggests new ones with similar characteristics. For instance, if you often watch dry comedies with sarcastic dialogue and deadpan delivery, the system will prioritize shows that use a similar style of humor.

Hybrid filtering: A system that combines both collaborative and content-based filtering to improve recommendation accuracy. For example, a hybrid system might suggest a book because it matches your reading history and is also trending among readers with similar tastes.

Cold start problem: When a recommender system lacks enough data to make accurate predictions, it struggles to personalize results. This often happens with new users who haven’t interacted with the system, or with new items that lack ratings or reviews. Without enough information to identify patterns, the system cannot personalize results.

Serendipity: The ability of a recommender system to offer suggestions that are both relevant and pleasantly unexpected. Rather than reinforcing familiar choices, serendipitous recommendations introduce users to new content that aligns with their interests in surprising ways. 

Shilling Attack: Involves injecting fake user profiles into a recommender system to distort its algorithm and output. These profiles are crafted to make certain items appear more or less relevant than they truly are, influencing what gets recommended. 

Bandwagon Attack: A subtype of shilling attack that adds fake ratings to both a target item and other already-popular items. By mimicking mainstream preferences, the attack tricks the recommender system into promoting the target as widely appealing. 

Segmented Attack:  This method manipulates recommendations within a specific niche by introducing fake profiles that closely mimic real users with shared interests or behaviors. As the system looks for patterns, it begins suggesting the target item to genuine users who match the fabricated profile group.

Before algorithms could tell you what movie to watch next or which song you might love, recommendations were personal. A friend handed you a mixtape. A local bookstore owner suggested something they thought fit your taste. These choices came from memory, instinct, and conversation. No metrics, no pattern recognition. The idea of replicating that process at scale, both digitally and automatically, is what gave rise to recommender systems.

The earliest known recommender system dates back to 1979, when computer scientist Elaine Rich developed a program called Grundy.⁵ Designed as a digital librarian, Grundy asked users a series of questions and assigned them to predefined preference groups; what Rich labelled as stereotypes. These weren’t cultural labels, but functional categories based on reading habits and genre interests. After answering a few targeted questions, users were matched to a stereotype such as “science fiction fan” or “romance lover” and received recommendations aligned with that profile. The system didn’t adapt or learn in real time, but it introduced a foundational idea: that computers could use stated interests to offer tailored suggestions.

That insight gained traction in the early 1990s when more powerful systems moved beyond fixed categories into more dynamic pattern recognition. At Xerox Palo Alto Research Center (PARC), a renowned innovation hub in California, researcher David Goldberg led a team that developed Tapestry, an email and document filtering system allowing users to annotate items as “excellent” or “relevant.”⁵ Tapestry performed searches like “show documents about racing bikes that user William considered excellent.” Unlike today’s fully automated tools, Tapestry required manual input. Still, it introduced a breakthrough idea: recommendations could be based on what people had in common. If two users consistently marked the same items as helpful, the system assumed they might agree again. Instead of treating each user in isolation, Tapestry searched for shared signals, laying the groundwork for what we now call collaborative filtering.

Around the same time, the University of Minnesota’s GroupLens Research Laboratory, led by John Riedl and Paul Resnick, made major strides in collaborative filtering. The team began with a system that offered recommendations for Usenet, an early internet forum where people shared and discussed news articles. GroupLens became one of the first automated systems to generate suggestions based on large-scale user ratings. Unlike Tapestry, which relied on manual input, GroupLens used models that could learn from collective behavior without requiring feedback from every individual. Their work led to the development of MovieLens, a free platform where users rate movies, now widely used in research to evaluate recommender algorithms, tagging methods, and user interface design. This line of research helped shape the academic foundation of recommender systems and paved the way for their commercial use. By the mid-1990s, this academic momentum moved into the commercial world. Members of the GroupLens team founded Net Perceptions, a company that built early recommendation engines for major retailers, including Amazon and Best Buy.

While collaborative filtering took off, content-based filtering also gained ground. One of the earliest large-scale examples was the Music Genome Project, launched in 1999.⁵ It aimed to identify core components of songs, including tempo, instrumentation, genre, and vocal style, and connect listeners with tracks sharing those traits, even if the songs weren’t widely known.

Another major public milestone arrived in 2006 with the launch of the Netflix Prize, offering one million dollars to whoever could improve its movie recommendation algorithm significantly.⁶ The competition attracted teams worldwide and highlighted the technical and commercial importance of recommender systems. When a winning team cracked the benchmark in 2009, it marked a moment of recognition. Recommender systems were no longer tucked away in labs or online storefronts. They powered what millions watched, read, and listened to every day.

As recommender systems rose in popularity, their rapid growth across disciplines uncovered new challenges. Computer engineers quickly realized that designing these systems also meant tackling problems that don’t always have simple solutions. The cold start problem is one of the most persistent. It emerges when a system must recommend items to new users or suggest new content without sufficient data. Without a history, predictions falter. Platforms have devised strategies to address this, but it remains a hurdle. Another challenge is serendipity, which is known as the ability to surprise users with relevant but unexpected recommendations. Algorithms easily suggest more of what you already like, but balancing familiarity with novelty is critical to keeping recommendations fresh and engaging.

Today, recommender systems shape nearly every corner of digital life.⁷ From what we stream and read to what products appear in shopping carts, they guide much of what we consume online. Additionally, as the field evolves, these systems are doing more than suggesting products or media. They’re shaping education by personalizing learning paths, helping healthcare by tailoring treatment options, and influencing how we discover news and information.^8,9,10

Elaine Rich

A pioneering computer scientist, Rich developed Grundy in 1979, which is one of the earliest prototypes of a recommender system.¹¹ Her work introduced the concept of grouping users by preferences, setting the stage for personalized digital experiences.

David Goldberg

At Xerox PARC in the early 1990s, Goldberg led the creation of Tapestry, an email and document filtering system that laid the foundation for collaborative filtering.⁵ His research shifted recommendations from isolated user profiles to patterns shared across communities.

John Riedl

Co-founder and leader within the University of Minnesota’s GroupLens Research Laboratory in the mid-1990s, Riedl was instrumental in advancing automated recommender systems.⁵ His work helped develop models that learn from large-scale user ratings, powering platforms like MovieLens and transforming academic and commercial recommender systems.

Paul Resnick

A key collaborator at GroupLens during the same period as John Riedl, Resnick contributed to building scalable algorithms that harness collective user feedback to personalize recommendations.⁵ His research has influenced both theoretical frameworks and practical applications in recommender system design.

Recommender systems are no longer confined to movie queues or online shopping carts. They now shape how students are guided through digital coursework, how patients are matched with therapists, and how retailers decide which items to restock or retire.^8,9,12 As these systems move deeper into classrooms, clinics, and digital storefronts, their influence is becoming harder to ignore. 

E-commerce success

Consider the sprawling digital marketplaces where millions of products vie for attention. In this vastness, shoppers may feel overwhelmed, uncertain where to start or what to trust. Recommender systems step in as trusted navigators, analyzing behaviors such as clicks, searches, and purchases to suggest products that closely align with individual preferences. This blend of data and insight transforms sprawling selections into curated choices, smoothing the path from browsing to buying.

In their landmark 1999 paper, Schafer and colleagues identify the primary ways recommender systems boost online sales among e-commerce platforms.¹² First, they highlight products that are highly rated or popular, helping shoppers feel confident about the quality and reducing doubts. Second, recommendations cut down the time and effort spent searching by showing relevant options quickly. Third, as retailers collect more data, their algorithms improve, offering increasingly personalized suggestions that encourage customers to stay loyal and make switching to competitors less appealing.

The scale of this effect is impressive. In 2021, the major consulting firm McKinsey & Company reported that about 35% of Amazon’s sales stem from personalized recommendations.¹³ More than boosting transactions, recommender systems reshape how consumers explore digital marketplaces. By connecting products to genuine preferences, they build trust, inspire discovery, and craft shopping experiences that evolve alongside customer needs. 

More flexible learning

Recommender systems are reshaping education by making learning pathways more adaptive, personalized, and responsive.⁹ At their core are algorithms that learn from past student behavior, such as mistakes on quizzes, completed modules, or time spent on certain topics. By identifying patterns in these actions, the system can predict when a learner might struggle, adjust the level of difficulty of subsequent practice questions, and surface additional resources that fill any learning gaps. 

Everyone learns at their own rhythm. Some zip through familiar concepts, others linger where ideas feel unfamiliar. Adaptive systems accommodate this range by constructing individualized trajectories that minimize frustration and keep engagement high.⁹ When learning feels aligned with ability, both motivation and retention improve.¹⁴

One study from the Open University of the Netherlands tested how recommender systems could reshape online learning.¹⁵ In an introductory psychology course, half the students followed a traditional, linear sequence, completing each module in a set order. The other half used a hybrid recommender system that suggested which module to study next, based on a mix of their learner profile and what high-performing peers had done before them.

As the system gathered more data, its recommendations grew more refined, shifting from broad, profile-based prompts to sharper, peer-informed suggestions. This evolving precision gave students more control and clarity as they navigated the course nonlinearly, skipping familiar material or revisiting earlier topics as needed.¹⁵

While overall module completion rates stayed about the same, the experimental group consistently moved through the material faster. Students in the recommender condition reached similar learning milestones in less time, reported feeling more engaged, and expressed greater satisfaction with how the course was structured. The personalized order didn’t just feel more intuitive: it made the learning experience smoother and more motivating.¹⁵

As online education expands, recommender systems offer smarter, flexible learning paths that may empower students to navigate education with confidence and purpose.

Data-driven health decisions

In the age of information overload, identifying the next step in diagnosis, treatment, or care planning is often one of the most difficult tasks in healthcare. Recommender systems, long used to serve up music or shopping suggestions, are now being applied to help doctors, patients, and researchers make decisions that could literally save lives.⁸

One area where this shift is becoming clear is within personalized health support. As Sahoo and colleagues (2019) highlight, health recommender systems (HRS) are being used to provide tailored advice in four major areas: nutrition, exercise, diagnosis, and medication.¹⁶ For example, these systems may suggest food substitutions and meal plans to support lowering one’s cholesterol, recommend workout routines that are based on rehabilitation goals or mobility level, or help clinicians recognize symptom patterns based on similar past cases. These tools don’t just save time, but also reduce the guesswork in high-stakes environments.

Of course, the stakes get even higher in cancer care, where choosing the wrong treatment isn’t just frustrating, but can also be devastating. A study by Suphavilai et al. in 2018 introduced CaDRReS, a recommender system trained on gene expression patterns within cancer cells.¹⁷ Its job? Predict which cancer drugs a specific patient is most likely to respond to. By analyzing patterns across thousands of experiments, CaDRReS learned how different gene expression profiles relate to the success or failure of certain types of treatment. Since it uses collaborative filtering, a method that identifies similarities across many data points, the system was able to make accurate predictions for patients whose profiles weren’t part of the training data. That kind of prediction isn’t simply useful; it’s transformative. It means treatments can be more precise, come with fewer side effects, and ultimately, be better suited to the patient.¹⁷

Whether it’s helping patients choose what to eat or guiding oncologists toward effective treatments, recommender systems are becoming essential tools in modern medicine. They offer a glimpse of a future where health advice feels not only informed, but truly personalized.

Recommender systems have transformed how we shop, watch, read, and connect. Their ability to simplify decisions is part of what makes them so powerful and widely adopted. However, their efficiency comes with trade-offs. As these systems influence what we consume and view, concerns have surfaced about hidden biases, behavioral manipulation, and uneven visibility. 

Filter bubbles

Recommender systems excel at shaping choice. They comb through overwhelming volumes of data and present selections they expect will satisfy. In many cases, this feels efficient and even tailored. Yet the same system that simplifies navigation can also limit options, funneling attention toward the familiar and possibly eroding exposure to content that sits outside the algorithm’s frame.

This is the exact concern behind filter bubbles, a term coined by internet activist and author Eli Pariser in 2011.¹⁸ He argued that algorithmic personalization can create invisible walls around users, showing them only content that reinforces what they already believe or enjoy. Over time, this selective exposure may reduce diversity, not only in opinions, but in the types of information, stories, or experiences a person engages with at all.

Research on this topic has yielded more nuanced results than the filter bubble metaphor might suggest. A large-scale study using longitudinal data from the longstanding movie recommender MovieLens tracked user behavior over 21 months to see how recommendations shaped content exposure.¹⁹ The researchers asked two key questions: Do recommender systems lead to narrower content consumption over time? And do people who act on recommendations experience different outcomes than those who ignore them?

The answer to both was yes, but not in the expected direction.

As predicted, users who followed the system’s suggestions did tend to gravitate toward a narrower slice of content. They also rated that content more positively and appeared more satisfied with what they had watched overall. The surprising finding, however, was that those who consistently engaged with recommendations experienced less severe narrowing than users who ignored the suggestions entirely.¹⁹ In other words, people who picked content on their own without influence from the recommender system had actually ended up in tighter bubbles.

This nuance complicates the idea that recommender systems are inherently isolating. Algorithms do cluster content around prior choices, but it’s clear that clustering doesn’t always result in ideological rigidity or cultural isolation.

That said, the risk of deeper polarization is still present. During the 2015 refugee crisis, researchers in Denmark conducted a panel study to examine how algorithmic media exposure might influence political attitudes over time.²⁰ The study followed a group of citizens across multiple waves of data collection, tracking both their views on immigration and their patterns of social media use. What emerged was a clear trend: users who frequently encountered algorithmically filtered news feeds became more entrenched in their existing beliefs. The study indeed highlights how algorithm personalization, when left unchecked, can harden beliefs instead of broadening them.

The challenge ahead then is recognizing when recommender systems guide discovery, and when they fence us into familiar ground.

Hidden bias in recommendations

One of the greatest concerns surrounding recommender systems is not what they highlight, but what they fail to surface. These systems learn from past data—data that may mirror historical inequalities.²¹ The risk? Algorithms that appear neutral can end up repeating patterns we as a society had hoped to outgrow.

A global field study published in Marketing Science makes this clear. Researchers Anja Lambrecht and Catherine E. Tucker designed a Facebook ad campaign to promote careers in science, technology, engineering, and math (STEM).²² The ad was meant to reach men and women equally. It didn’t. They used Facebook’s algorithmic ad system, a platform built on recommender system logic, to test how a single advertisement was delivered across 191 countries. Although the campaign used gender-neutral targeting settings within Facebook’s ad platform, the ad was shown to men far more often, especially those in peak career years.

The culprit wasn’t overt bias. It was the economics of attention. On platforms like Facebook, showing ads to women costs more. Their clicks are deemed to be more valuable to advertisers because women drive more household purchases and tend to convert more often than men. As a result, the system, designed to optimize cost-efficiency, showed the ad where impressions were cheaper, which were predominantly to men. Thus, when trying to maximize return, the algorithm unintentionally limited who even got to consider a STEM career.

This isn’t a fringe case. These findings raise deeper questions about fairness, access, and how digital systems shape who gets included in tomorrow’s opportunities and who gets left out. When personalization shapes who sees what, the issue isn’t only biased content. It’s how easily opportunity can disappear from view.

Shilling attacks

Another lingering challenge with recommender systems is what happens when people try to manipulate them. Some do it to promote a product, others to bury a competitor, and still others to distort what people see. This tactic, known as a shilling attack, involves creating fake user profiles designed to game the algorithm.²³ The result? Artificial signals get treated as genuine preferences, pushing certain items into more feeds while pushing others out of view.

These attacks often fall into two broad categories: push and nuke. Push attacks aim to make a product look more popular than it really is. Nuke attacks do the opposite, lowering a product’s visibility by driving down its ratings.²³

Within the push category, a common technique is the bandwagon attack. Here, fake profiles are loaded with high ratings for already popular items—plus one very high rating for the target item.²³ Since the fake profiles resemble typical users, the algorithm starts treating the target item as part of the mainstream. 

Segmented attacks operate a bit differently. Instead of riding existing trends, they target a narrower slice of the audience by focusing on users with specific, traceable interests.²³ Imagine trying to promote a niche horror novel. Dozens of fake profiles are created and packed with strong ratings for recognizable horror titles. The algorithm, trained to detect similarities across items, begins nudging the promoted book toward real users who share those patterns. These attacks tend to be most effective in systems that organize recommendations around relationships between products rather than patterns in user behavior.

Researchers have begun to counter these attacks. One approach, developed by Alonso and colleagues, assigns a reliability score to every prediction linking a user to an item.²⁴ When those scores fluctuate in erratic patterns—particularly around items that typically attract little attention—it can suggest a shilling attack is in play. The method isn’t foolproof, but it offers a concrete step toward detection. Until recommender systems can reliably separate genuine interest from manufactured hype, what rises to prominence may owe more to manipulation than to meaningful engagement.

Nextflix: Who picks your picks?

Back in 1997, Netflix launched with a simple idea: DVD rentals by mail.²⁵ Ten years later, it pivoted to streaming, and with that shift came a new problem. How do you help people navigate a growing ocean of content without leaving them stranded at the search bar?

The solution was never about showing more. It was about showing right. To keep users from scrolling aimlessly, Netflix built one of the most influential recommender systems in the world. Their system doesn’t operate through a single algorithm. Instead, it’s a layered system of models, each optimized for a different task. Some prioritize what you’ve started but never finished. Others nudge you toward titles that spike in popularity at a particular time of day.²⁶

Every click, skip, pause, and rewatch becomes a clue. If you tend to watch stand-up comedy late at night but skip past anything longer than 40 minutes, the system will take notice. These patterns feed into a machine learning framework that treats your behavior like a constantly evolving fingerprint.

The personalization goes deeper than behavior. Netflix tags every show and film with detailed metadata, including genre, cast, mood, pace, and even the color tone of the cinematography. Then it monitors how you react. If you’re more likely to click on a thumbnail with a bright background and no faces, future suggestions will reflect that. The goal isn’t simply to match your taste—it’s to anticipate what you’ll say yes to next.

Personalized recommendations now drive over 80% of viewing on the platform.²⁷ For users, this means fewer dead ends and more of what they actually want. For Netflix, it means longer sessions, higher retention, and fewer cancellations. All of it happens in the background, but its effects are hard to miss.

How Tinder predicts your type

Dating used to start with a glance across the room. Now, it often starts with a swipe. With approximately 50 million users worldwide, Tinder has become one of the most influential matchmaking tools of the last decade.²⁸ Behind its simple interface lies a complex recommender system that’s designed not only to respond to your preferences, but to learn from them.

To make that learning possible, Tinder developed a system called TinVec. Each time a user swipes left or right, the algorithm turns that action into a data point—a “vector”—that gets mapped in a digital space.²⁹ Users who consistently swipe on profiles with similar characteristics (like beach photos, certain hobbies, or even educational backgrounds) start to cluster together. Over time, the app uses these patterns to suggest people whose swipe behavior reflects yours. Not just based on one or two interests, but on a more nuanced behavioral fingerprint.

Language plays a role as well. Tinder incorporates a model called Word2Vec to make sense of the words in profiles.²⁹ Using vector representations of words to identify their meaning and relationships, it picks up on patterns in slang, phrasing, and tone—details that signal how someone might express affection, joke around, or share values. A user who writes “looking for deep convos and cozy video game nights” might not be matched the same way as someone who says “adrenaline junkie seeking thrill partner.” It’s not about matching on keywords—it’s about learning what styles of self-expression feel familiar, attractive, or inviting to each user. 

However, some factors are excluded from the algorithm on purpose. According to Tinder’s internal team, variables like ethnicity, religion, and perceived social status aren’t used to filter recommendations.³⁰ The goal, they say, is to create a more inclusive dating pool by focusing on behavioral data rather than stereotypes.

For users, love at first swipe might feel spontaneous, but behind the screen, it’s carefully scaffolded—part chance, part pattern, and part algorithm trying to learn your personal definition of a soulmate. 

Predictive Analytics

Predictive analytics uses statistical models and historical data to forecast future outcomes. Like recommender systems, it relies on pattern recognition, but its applications extend far beyond personalized suggestions. In this reference guide, Emilie Rose Jones walks us through how predictive tools are shaping decisions in finance, healthcare, and marketing. Whether it’s anticipating risk or improving customer experience, predictive analytics helps organizations move from reacting to anticipating.

Machine Learning

Machine learning is a core driver behind today’s most adaptive technologies. This reference guide explores how algorithms learn from data to make predictions, generate content, and offer recommendations—all without being explicitly programmed.

1. Konstan, J., & Terveen, L. (2021). Human-centered recommender systems: Origins, advances, challenges, and opportunities. AI Magazine, 42(3), 31–42.
2. Jannach, D., Zanker, M., Felfernig, A., & Friedrich, G. (2010). Recommender systems: An introduction. Cambridge University Press.
3. De Marzo, G., Gravino, P., & Loreto, V. (2024). Recommender systems may enhance the discovery of novelties. Journal of Physics: Complexity, 5(4), 045008.
4. Isaacson, W. (2011). Steve Jobs. Simon & Schuster.
5. Dong, Z., Wang, Z., Xu, J., Tang, R., & Wen, J. (2022). A brief history of recommender systems. arXiv preprint arXiv:2209.01860.
6. Hallinan, B., & Striphas, T. (2016). Recommended for you: The Netflix Prize and the production of algorithmic culture. New Media & Society, 18(1), 117–137.
7. Falk, K. (2019). Practical recommender systems. Simon & Schuster.
8. Tran, T. N. T., Felfernig, A., Trattner, C., & Holzinger, A. (2021). Recommender systems in the healthcare domain: State-of-the-art and research issues. Journal of Intelligent Information Systems, 57(1), 171–201.
9. Drachsler, H., Verbert, K., Santos, O. C., & Manouselis, N. (2015). Panorama of recommender systems to support learning. In Recommender systems handbook (pp. 421–451). Springer US.
10. Raza, S., & Ding, C. (2022). News recommender system: A review of recent progress, challenges, and opportunities. Artificial Intelligence Review, 55(1), 749–800.
11. Rich, E. (1979). User modeling via stereotypes. Cognitive Science, 3(4), 329–354.
12. Schafer, J. B., Konstan, J., & Riedl, J. (1999, November). Recommender systems in e-commerce. In Proceedings of the 1st ACM Conference on Electronic Commerce (pp. 158–166).
13. Arora, N., Ensslen, D., Fiedler, L., Liu, W. W., Robinson, K., Stein, E., & Schüler, G. (2021, November 12). The value of getting personalization right—or wrong—is multiplying. McKinsey & Company. https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-value-of-getting-personalization-right-or-wrong-is-multiplying
14. Umbreit, J., Lane, K. L., & Dejud, C. (2004). Improving classroom behavior by modifying task difficulty: Effects of increasing the difficulty of too-easy tasks. Journal of Positive Behavior Interventions, 6(1), 13–20.
15. Drachsler, H., Hummel, H., Van den Berg, B., Eshuis, J., Waterink, W., Nadolski, R., Berlanga, A., Boers, N., & Koper, R. (2008, March 3). Effects of the ISIS recommender system for navigation support in self-organized learning networks. In Proceedings of the Special Track on Technology Support for Self-Organised Learners. https://research.ou.nl/ws/portalfiles/portal/12746114/drachsler.pdf
16. Sahoo, A. K., Pradhan, C., Barik, R. K., & Dubey, H. (2019). DeepReco: Deep learning based health recommender system using collaborative filtering. Computation, 7(2), 25.
17. Suphavilai, C., Bertrand, D., & Nagarajan, N. (2018). Predicting cancer drug response using a recommender system. Bioinformatics, 34(22), 3907–3914.
18. Pariser, E. (2011). The filter bubble: How the new personalized web is changing what we read and how we think. Penguin.
19. Nguyen, T. T., Hui, P. M., Harper, F. M., Terveen, L., & Konstan, J. A. (2014, April). Exploring the filter bubble: The effect of using recommender systems on content diversity. In Proceedings of the 23rd International Conference on World Wide Web (pp. 677–686).
20. Ohme, J. (2021). Algorithmic social media use and its relationship to attitude reinforcement and issue-specific political participation – The case of the 2015 European immigration movements. Journal of Information Technology & Politics, 18(1), 36–54.
21. Arora, A., Barrett, M., Lee, E., Oborn, E., & Prince, K. (2023). Risk and the future of AI: Algorithmic bias, data colonialism, and marginalization. Information and Organization, 33(3), 100478.
22. Lambrecht, A., & Tucker, C. (2019). Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads. Management Science, 65(7), 2966–2981.
23. Sundar, A. P., Li, F., Zou, X., Gao, T., & Russomanno, E. D. (2020). Understanding shilling attacks and their detection traits: A comprehensive survey. IEEE Access, 8, 171703–171715.
24. Alonso, S., Bobadilla, J., Ortega, F., & Moya, R. (2019). Robust model-based reliability approach to tackle shilling attacks in collaborative filtering recommender systems. IEEE Access, 7, 41782–41798.
25. Bansal, S., & Sharma, R. (2024, February). Revealing the evolution of Netflix recommender systems. In 2024 11th International Conference on Computing for Sustainable Global Development (INDIACom) (pp. 83–86). IEEE.
26. Hsiao, K.-J., Feng, Y., & Lamkhede, S. (2025, March 21). Foundation model for personalized recommendation. Netflix Technology Blog. https://netflixtechblog.com/foundation-model-for-personalized-recommendation-1a0bd8e02d39
27. Frey, M. (2021). Netflix recommends: Algorithms, film choice, and the history of taste. University of California Press.
28. Tinder LLC. (2025). About Tinder. https://www.tinder.com/about
29. Nan, C. (2021, November 7). Tinder recommender system. Rpubs. https://rpubs.com/nancunjie4560/832396
30. Tinder. (2022, July 11). Powering Tinder® — The method behind our matching. Tinder Newsroom.https://www.tinderpressroom.com/powering-tinder-r-the-method-behind-our-matching