How Well Does the Algorithm Know Me?

Amazon Personalize

User

Interaction patterns¹

User identification²

¹“User interaction data on the website/ application is captured in the form of events and is sent to Amazon Personalize often via an integration that involves a single line of code. This includes key events such as click, buy, watch, add-to-shopping cart, like, etc.” [4]

²“User profile data including user demographic data such as gender and age. This data is optional.” [4]

Live learning of user needs and preferences¹ through observational behavior²

User-adaptivity adjusts over time based on continuous learning¹

¹“Amazon Personalize takes care of machine learning for you. […] As your data set grows over time from new metadata and the consumption of real-time user event data, your models can be retrained to continuously provide relevant and personalized recommendations. […] Make your recommendations relevant by responding to the changing intent of your users in real time.” [4]

²See input data

Context Device¹

Spatio-temporal¹

¹“Improve relevance of recommendations by generating them within a context, for instance device type, time of day, and more.” [4]

System

Meta data^{1 2}

¹“This can be any type of catalog including books, videos, news articles or products. This involves item ids and metadata associated with each item. This data is optional.” [4]

²“Consider what’s relevant to your users and what is important for your business when generating recommendations. You can define an objective, in addition to relevance, to influence recommendations.” [4]

Google Maps Matches

User

User identification¹

Interaction patterns²

Personal preference settings³

¹“To see your matches, you need to sign in to your account” [21]

²“Whether you’ve saved, rated, or visited a place or somewhere similar, your interactions with places on Google maps or Google Search” [21]

³“We use your saved preferences […] to suggest locations and content you might like” [21]

Live learning of user needs and preferences¹ through observational behavior²

User-adaptivity adjusts over time based on continuous learning³

²“As you use Google Maps more often, you’ll see matches for various places. These matches show how well a place matches your preferences.” [21]

²See used data

³“They’re unique to you and improve over time.” [21]

Context

Spatio-temporal¹

¹“Your matches are based on your Google Location History” [21]

System

Meta data¹

¹“The types of cuisine and restaurants you typically visit or avoid” [21]

Spotify Music recommendations

User

User Identification¹

Interaction patterms²

¹[45]

²“Brands like Spotify accumulate a mountain of implicit customer data comprised of song preferences, keyword preferences, playlist data, geographic location of listeners, most used devices and more.” [1]

Live learning of user needs and preferences based on directly measurable data¹

User-adaptivity adjusts over time based on continuous learning²

¹“On every Monday, each and every users are presented with a customized list of thirty songs. And the recommended playlist comprises tracks – which user might have not heard before” [1]

²“Machine learning enables the recommendations to improve over a period of time” [1]

Context

Device¹

Spatio-Temporal¹

¹“Brands like Spotify accumulate a mountain of implicit customer data comprised of song preferences, keyword preferences, playlist data, geographic location of listeners, most used devices and more.” [1]

System

Meta Data¹

¹“Spotify’s AI scans a track’s metadata as well as blog posts and discussion about specific musicians and news articles about songs or artists on the internet.” [1]

Fitbit daily readiness

User

User identification¹

Physiological parameters²

¹“Daily Readiness requires a Fitbit Premium membership.” [19]

²“The score compares your recent activity, sleep and heart rate variability (HRV) levels against your personal baseline.” [19]

Live learning of user needs and preferences¹ through observational behavior²

User-adaptivity adjusts over time based on continuous learning³

¹“Before you receive your first daily readiness score, you wear your Fitbit device continuously for 4 days to get your baseline. Your readiness score is then based on how your stats in each category compare to your personal baseline. This data helps you understand how your activity levels, sleep patterns, and stress from the previous days contribute to your energy levels. Wear your device consistently for at least 2 weeks to develop a more accurate personal baseline.” [19]

²See input data

³“Your daily goals and recommendations adapt with you over time” [19]

Context

—

System

—

Instagram Intelligent Explore Page

User

User identification¹

Interaction Patterns²

¹Instagram requires login to personal profile

²“[…] we leverage accounts that people have interacted with before (e.g., liked or saved media from an account) on Instagram to identify which other accounts people might be interested in.” [33]

Live learning of user needs and preferences¹ through observational behavior²

¹“[…] we leverage accounts that people have interacted with before (e.g., liked or saved media from an account) on Instagram to identify which other accounts people might be interested in.” [33]

²See data

Context

—

System

Meta data¹

¹“The most important signals we look at, in rough order of importance, are: Information about the post. […] These are signals like how many and how quickly other people are liking, commenting, sharing, and saving a post. […] we have rules for what we recommend in places like Explore. […] These include things like avoiding potentially upsetting or sensitive posts, for example, we aim to not show content that promotes tobacco or vaping use in Explore.” [36]

Gmail Smart Reply

User

User Identification¹

Interaction Patterns²

¹Login to account required

²“If I’m more of a You’re welcome’ kind of person, instead of a don’t mention it person, over time, it’ll suggest the response that better suits me” [25]

Live learning of user needs and preferences¹ through observational behavior²

User-adaptivity adjusts over time based on continuous learning³

¹“Smart Reply utilizes machine learning to give you better responses the more you use it.” [10]

²See data

³“Over time, it will suggest the response that better suits me.” [25]

Context

—

System

Meta data1

¹“The responses suggested by Smart Replay are tailored to the content of the email […]” [25]

Google Search

User

User Identification^{1 2}

Interaction patterns³

Personal preference settings⁴

¹“Info in your Google Account, like your age range and gender” [23]

²“If you’re not signed in, your ad settings are saved to your device or browser.” [20]

³“Your current search query, previous search activity, your activity while you were signed in to Google, your previous interactions with ads, types of websites you visit, types of mobile app activity on your device” [23]

⁴“Under “How your ads are personalized,” select your personal info or interests.” [20]

Live learning of user needs and preferences¹ through observational behavior²

User-adaptivity adjusts over time based on continuous learning³

¹“Google can personalize ads so they’re more useful to you.” [23]

²See data

³“Your previous interactions with ads” [23]

Context

Device¹

Spatio-temporal²

¹“Your activity on another device” [23]

²“The time of day.” [23]

System

Meta data¹

¹“Personalized ads aren’t shown or hidden from you based on sensitive categories, like race, religion, sexual orientation, or health.” [23]

Netflix

Movie recommendations

User

User identification¹

Interaction patterns²

User preference settings¹

¹“When you create your Netflix account, or add a new profile in your account, we ask you to choose a few titles that you like. We use these titles to ‘jump start’ your recommendations.” [37]

^2“We estimate the likelihood that you will watch a particular title in our catalog based on a number of factors including: Your interactions with our service (such as your viewing history and how you rated other titles), other members with similar tastes and preferences on our service […]” [37]

Live learning of user needs and preferences¹ based on directly measurable data²

User-adaptivity adjusts over time based on continuous learning³

¹“Once you start watching titles on the service, this will ‘supercede’ any initial preferences you provided us, and as you continue to watch over time, the titles you watched more recently will outweigh titles you watched in the past in terms of driving our recommendations system” [37]

²See data

³“We take feedback from every visit to the Netflix service and continually re-train our algorithms with those signals to improve the accuracy of their prediction of what you’re most likely to watch.” [37]

Context

Device¹

Spatio-temporal¹

¹“In addition to knowing what you have watched on Netflix, to best personalize the recommendations we also look at things like: the time of day you watch, the devices you are watching Netflix on, and how long you watch” [37]

System

Meta data¹

¹“We estimate the likelihood that you will watch a particular title in our catalog based on a number of factors including: […] information about the titles, such as their genre, categories, actors, release year, etc.” [37]

Google Pixel 6

Material You

User

User identification¹

¹Android Phones require login to Google accounts

Saving and connecting inputs to a specific user¹

Pre-defined adaptations that are similar to every user and do not evolve over time¹

¹“When a user chooses a wallpaper, the entire UI and theme update to match.” [14]

Context

Device¹

¹“The UI reacts to screen changes” [22]

System

Direct manipulation¹

Meta data¹

¹“We built upon this insight to generate unique Material palettes for everyone, derived from a personal signal wallpaper that can be applied to their entire experience” [22]

Nio ES8

Nomi

User

User identification¹

Interaction patterns¹

Physiological parameters²

¹“NOMI learns user preferences over time to understand the specific context of the car in relation to its owner. For instance, NOMI can set the personal seating and steering wheel positions whenever it senses a driver approaching the vehicle.” [16]

²“NOMI incorporates artificial intelligence with a charming face-like interface that swivels and blinks its oval “eyes” to address each vehicle occupant directly, depending on their location in the digital cockpit.” [16]

Live learning of user needs and preferences¹ based on directly measurable data²

User-adaptivity adjusts over time based on continuous learning¹

¹“NOMI learns user preferences over time to understand the specific context of the car in relation to its owner.” [16]

²See data

Context

Spatio-temporal¹

¹“NOMI further engages with users by understanding their daily routines and travel patterns, giving suggestions about the best real-time transit routes, providing notices about family or business appointments” [16]

System

—

Philips Hue with IFTTT

User

User identification¹

Physiological parameters²

¹IFTTT requires a Log in to use their applets

²“Connect your wearables to Philips Hue through IFTTT and it will automatically turn on your lights in the morning when you have reached your sleep goal.” [43]

Saving and connecting inputs to a specific user¹

Pre-defined adaptations that are similar to every user and do not evolve over time²

¹“You can create your own Applets or enable published Applets by clicking on them.” [27]

²“IIFTTT, also known as “if this then that,” gives you creative control over the products and apps you love. IFTTT Applets are simple connections between products and apps. When something happens in one application, an IFTTT Applet can update it in another.” [43]

Context

Spatio-temporal¹

Environmental conditions²

¹“Automatically turn on your Philips Hue lights when you arrive home. You can also create an Applet to turn off your lights when you leave.” [43]

²“Get a rainy weather update and the IFTTT Applet will turn your lights blue so you remember to take an umbrella with you to work.” [43]

System

Direct manipulation¹

System state²

¹“This makes a widget you can add to the home screen on your phone. Then you can press the button to make your Hue lights dance as they do a color loop of rainbow colors.” [43]

²“When something happens, you’ll be the first one to know by connecting IFTTT with Philips Hue. No matter if it’s a phone call, e-mail or weather update.” [43]