Users, Items, Context, and Recommendation Units

The question is not which model is better but whose behavior the history belongs to. Each answer carries a price. Two of the three prices have been measured in public.

An account model bets that identity is durable, and durable identity is identifying. Netflix once released the anonymous ratings of 500,000 subscribers. In 2008 Narayanan and Shmatikov showed how little it takes to undo that. With 8 movie ratings, of which 2 may be completely wrong, and dates that may have a 14-day error, 99% of records can be uniquely identified in the dataset. For 68% of records, two ratings and dates with a 3-day error are sufficient. “We demonstrate that an adversary who knows only a little bit about an individual subscriber can easily identify this subscriber’s record in the dataset.” The file carried no names. It did not need to.

A session model bets that only the last few minutes can be trusted. For much of the web that is not a bet but the only option available. The 2016 paper that put recurrent networks on session data opens by saying why: “Many e-commerce recommender systems (particularly those of small retailers) and most of news and media sites do not typically track the user-id’s of the users that visit their sites over a long period of time.” Hidasi and his co-authors then give the session unit a contract as explicit as any account policy. Their training data is the RecSys Challenge 2015 click-stream: 7,966,257 sessions, 31,637,239 clicks, 37,483 items. Each whole session goes to either the training or the test set. The data is never split mid-session.

A household model admits from the start that several people share one record. Netflix has quantified that case in its own filings. The next section is that number.

Netflix has put a filed number on the distance between an account and a person. Its Q1 2022 shareholder letter, dated 19 April 2022, told investors: “Second, in addition to our 222m paying households, we estimate that Netflix is being shared with over 100m additional households, including over 30m in the UCAN region.” That is roughly one shared household for every two paying ones. Every one of them arrives in the logs as a single signed-in account.

On a shared television the consequence is concrete. Children, parents and guests act under one key. The embedding fitted to that key represents no actual viewer, and the slate lurches as the person in front of the screen changes. The account-equals-person assumption does not fail only in one unlucky living room. On the company's own estimate it fails at a scale of over 100m households.

Every training example links a decision subject, candidate item, context, time, exposure, and outcome. If any one of those entities is unstable, the model can learn a coherent pattern about an incoherent unit. A user can mean an account, a person, a household, a device, an organization, or an anonymous session. An item can mean a SKU, an offer, a seller listing, an article version, an episode, a bundle, or a category.

Somebody picks one of each, usually early and usually without a meeting. That pick goes on to shape sparsity, privacy, cold start, and evaluation.

The ledger runs from decision subject to outcome unit. Each layer between them fixes one entity. A model can be no more coherent than the five rows it inherits.

The last two rows are the ones most often collapsed into each other. MIND keeps them apart by construction. The impression log holds the articles displayed at a specific time; the click label holds what followed. A non-click is a 0 fastened to a particular display, not an absence in the table.

Identity merge happens at the top of the pipeline and outcome misattachment at the bottom. Neither can be fixed by a better model downstream. Check them separately, because a clean merge policy tells you nothing about whether a purchase was credited to the right exposure.

And the credit rule is something a person chooses, not something the log reports. “The industry standard for conversion attribution is to credit the advertizing platform which owns the last click in a 30 days window before the conversion,” says a 2017 paper from Criteo on attribution modeling in display advertising. Its authors released the log behind it: 30 days of Criteo production traffic, 16.5M impressions, 45K conversions, 700 campaigns. Each line is an ad impression carrying timestamp, uid, click, click position, conversion, conversion timestamp and an attribution flag. Click position is in there precisely so that a different attribution scheme can be applied to the same events. The fields are the evidence. Last-click is a choice the entity contract has to record, not a fact the data contains.

Who or what is being recommended to? The contract opens on that question and closes on attribution, because an entity nobody can trace is one the model will happily learn about anyway.

Each step has a documented price for getting it wrong. 222m paying households and over 100m shared ones are what make naming the subject unavoidable. Defining identifiers is where the Netflix Prize result lands: 8 ratings and dates with a 14-day error were enough to pick out 99% of records. Binding context and time is where the session contract lives — whole sessions to training or test, never split mid-session. Testing attribution is where the last-click default has to be written down as a rule somebody picked, with a 30-day window somebody chose.