| Patent number | US12314272B2 |
| Official title | Online meta-learning for scalable item-to-item relationships |
| Assignee | Apple Inc. |
| Inventors | Dimitrios Bermperidis, Sofia M. Nikolakaki, Rabi S. Chakraborty, Rajesh Kumar, Chandrasekar Venkataraman, Natalia G. Silveira, Puja Das |
| Priority date | August 12, 2022 |
| Filed | September 23, 2022 |
| Granted | May 27, 2025 |
| Status | Active. Anticipated expiry October 15, 2042 |
| Family | US20240054138A1, WO2024035730A1 |
| Source | patents.google.com/patent/US12314272B2 |
What This Patent Covers and Why It Matters for ASO
Feature it describes: the recommendations shown on an app’s own product page, the surface developers know as “You Might Also Like”.
Claim 1 is unusually direct about the surface: a distribution platform receives a request for a seed application’s landing page, identifies candidate applications using detected item to item relationships, and presents a selection of them in the interface. So this is the patent for the competitors that appear underneath your own listing.
The most useful idea in it is that relationships come in two kinds:
- Overlapping items serve the same primary purpose. These are substitutes, and a user choosing one instead of you is a loss.
- Complementary items enhance the experience without replacing it. A user taking one of these alongside you is not a loss at all.
If the system distinguishes the two, then the composition of the shelf under your listing is a strategic fact, not decoration. A page surrounded by complements is a different competitive position from one surrounded by substitutes, and the same shelf means opposite things in those two cases.
How the relationships are learned matters just as much, because one half of it is under your control:
- Explicit signals: co-download metrics across user populations. Which apps get installed by the same people.
- Implicit signals: embedding similarity derived from content tags, descriptions and topic modeling.
That second line is the practical opening. Your description and tags feed the embedding side of the relationship model. You cannot change who co-downloads you, but you can change which apps you sit near in the text derived space, and the patent says both contribute to the composite model.
One more detail with real consequences. The model trains online from user interactions, and the patent specifies that it learns from both positive engagement and observed non-engagement. Appearing on a shelf and being ignored is not neutral. It is training data arguing against you.
Patent Summary
The system identifies relationships between items on a distribution platform by analyzing user behavior, distinguishing overlapping items that serve a similar primary purpose from complementary items that extend the experience rather than replace it. Items covered include applications, digital media such as music and video, electronic games and e-books.
Relationships are learned by a composite model that combines explicit signals, meaning co-download patterns across user populations, with implicit signals, meaning similarity between item embeddings derived from content tags, descriptions and topic modeling. The online meta-learning component continuously refines the model from user interactions with the recommendations it produces, including cases where a recommended item is shown and not engaged with, adjusting the weighting of signals over time.
At request time, when a landing page for a seed application is requested, the platform identifies candidate applications through the detected relationships and presents a selection of them within the interface.
An international application, WO2024035730A1, was filed in the same family.
A detailed breakdown of the model, the figures and the full claim language will follow in a later update to this article.
Related reading: US10394838B2 derives similarity from search query overlap instead, and WO2023235143A1 covers the suggested results surface.
Bir yanıt yazın