How Office Suggests the Most Relevant Files to Users

Machine Learning Selects Most Appropriate Files

Message center notification MC255074 published on May 7 discusses a new way of highlighting recent Office documents to users on the File tab of Word, Excel, and PowerPoint. The text says that machine learning predicts which files a user is most likely to want to work on next and generates a set of cards for these files. Suggested files must be stored in OneDrive for Business or SharePoint Online. This is Microsoft 365 roadmap item 72233.

The Office MRU

Office has had a Most Recently Used (MRU) list for years. The MRU list shows the files last accessed and appears in places like when you right click on an Office app icon in the Windows toolbar. Figure 1 shows my current MRU list for Word. MRU files can be stored locally or in a cloud location.

Like many Office settings, the MRU list for an Office app is workstation-dependent and stores its data in the system registry. The exact location depends on the version of Office. For Microsoft 365 apps for enterprise (aka Office Pro Plus or Office click to run) on Windows, a set of identities used to sign into the PC is in HKLU\Software\Microsoft\Office\16.0\<app>\User MRU with the file MRU list stored in the File MRU value (at least, this is how Word and Excel works). Figure 2 shows my MRU list for Word documents in the registry. If you see ADAL in the identity name, it means that this list is for an identity signed in using the Azure Active Directory Authentication Library. LiveId means that authentication happened for a Microsoft Services Account (MSA).

Graph Based Suggestions

Storing MRU data in the system registry works acceptably well until your workstation changes. If you switch to a new PC or need to reinstall Office, the MRU list disappears. The MRU list is date based and shows files according to when they were used with the most recent file at the top.

The update to Office for Windows reflects changes previously made to Office.com and Office for Mac and uses machine learning to process Graph signals gathered for actions like edits (including updating properties in SharePoint, which show up as an edit), mentions, and comments to predict which files the user is most likely to want to open. Microsoft doesn’t say how far back the analysis of file activity looks back to suggest files. My experience is that the period covers the last week, but this might depend on how active you are in an application.

Each suggestion is in a card with a thumbnail showing some of the document (Figure 3). Two of the five files are shown because they are frequently opened; the others are due to recent edits. You can remove items from the list as a signal that you don’t want to see it again. The traditional MRU list is available as a list of recent files under the suggested documents.

Delve and Graph Privacy Settings Don’t Affect Suggestions

Suggestions are unique and personal to a user and only the owner of documents can see the set of recommended files. As such, this feature is unaffected by the privacy controls for insights applied through Delve or the Microsoft Graph. The privacy controls affect how people see insights derived from signals collected in the Office Graph for documents owned by other users (Figure 4). Although these insights also only work for files in SharePoint Online and OneDrive for Business, their focus is on making others aware of a person’s work. It wouldn’t make much sense if the privacy setting restricting visibility to others also stopped applications suggesting files to the author.

Hard to Know How Useful Suggestions Are

The value of suggestions is that they don’t depend on work done on a specific PC. The Microsoft Graph gathers signals about user activity no matter what client or device is used, so the suggestions surfaced in applications represent the totality of someone’s work rather than a snapshot from an individual device.

This change is yet another example of how Microsoft uses machine learning to process the signals gathered about user actions. Whether the suggestions prove useful will differ from person to person. If you’re the type to only open a file when you have good reason to, the suggestions probably won’t make much difference. Others who operate on a less structured basis might find them more useful.