dissenting in part.
I join the court’s opinion with the exception of parts V, VI, and VII, which all relate to the cross-appeals filed by Face-book, Inc., Linkedln Corporation, and Twitter, Inc. (collectively “Cross-Appellants”). In those parts of the opinion, the majority reverses the Patent Trial and Appeal Board’s decisions, which found that claims 1 and 5 of U.S. Patent No. 5,832,494 were not anticipated by Fox SMART; claims 15 and 16 of the ’494 patent were *644not anticipated by Fox Thesis; and claim 21 of U.S. Patent No. 6,233,571 would not have been obvious in view of Fox SMART, Fox Thesis, and Fox Envision. I would hold that the Board’s findings that the prior art references failed to disclose “deriving actual cluster links from the candidate cluster links”1 were supported by substantial evidence. I would therefore affirm the Board’s Final Written Decisions in full.
In support of its unpatentability holding, the majority identifies the cluster-splitting process disclosed in Fox SMART and Fox Thesis as disclosing the “deriving” limitation. The majority describes the steps in this process as involving (1) the creation of a “similarity matrix” representing the calculated similarity between individual pairs of documents in a cluster; (2) the identification of “highly correlated pairs” of documents within the cluster based on the calculated similarity values found in the similarity matrix; and (3) the formation of new, smaller clusters containing those “highly correlated pairs” of documents. Op. at 639 - 40 (citations omitted). The majority equates these new, smaller clusters with the “actual cluster links” described in the claims and contends that they are derived from “candidate cluster links” in the form of the calculated similarity values. Id. at 640.
•The Board found to the contrary. It first construed “actual cluster links” to mean “a subset of the candidate cluster links, which meet certain criteria.” Board Decision III, 2015 WL 456539, at *5. It then reasoned that under the construction—which Cross-Appellants did not challenge on appeal— derivation of actual cluster links required deletion of at least some candidate cluster links. Id. at *9-10. The Board reviewed the Fox references and found that the only cluster links deleted in the cluster-splitting process are “those that simply overlap, or duplicate, other clusters,” which means that, in fact, no candidate cluster links are deleted at all. Id. at *10. As a result, Fox SMART and Fox Thesis failed to disclose the “deriving” step.
I believe the Board’s conclusion is based on a reasonable reading of those references. The ’494 patent’s description of deriving candidate cluster links is clear:
Once the candidate cluster link set has been generated, deriving the actual cluster links is a simple matter of selecting or choosing the T top rated candidate links, and eliminating the rest.
’494 patent, 24:1-4 (emphasis added) (diagram numbering omitted); see also Op. at 639. That description is consistent with the Board’s construction of “actual cluster links.” The cluster-splitting process disclosed in Fox SMART and Fox Thesis, in contras];, uses a qualitatively different approach. As the Board found, that process does not involve eliminating any candidate cluster links. Instead, cluster-splitting simply redistributes the candidate cluster links among two or more clusters. Such redistribution cannot result- in a “subset of the candidate cluster links” as required by the claims.
Further, the majority, in my view, conflates two different aspects of information *645retrieval to reach its conclusion: creating a database and searching a database.
The Fox references use cluster-splitting in database creation. Fox describes adding documents to a database as akin to placing leaves on a tree. J.A. I 10272 (“The clustering algorithm produces a hierarchy where all N documents in a collection end up as leaves of a multilevel tree.”). One by one each document is placed on the particular branch of the tree (i.e., the “cluster”) containing other documents with which it is most similar. Id. When a branch is assigned more documents than it can bear, that branch splits in two. Id. It is at this step that the cluster-splitting process occurs. Documents continue to be added to the tree—and branches/clusters are split as necessary—until no documents remain. J.A. I 10275 (“Adding documents ... simply involves finding the proper place to insert, attaching the incoming entry appropriately, and recursively splitting overly large nodes from that point up the tree as needed.”).
In contrast to creating the database, the claims at issue are directed to analyzing (i.e., searching) an already populated database. Claims 1 and 5 of the ’494 patent are exemplary. They expressly claim a “method of analyzing a database.” ’494 patent, 51:38. The claimed method begins with the selection of a document (i.e., a “node”) for analysis and results in the display of other, similar documents as determined by the “actual cluster links.” Id. at 51:38-49. During oral argument, counsel for Cross-Appellants Linkedln and Twitter compared the claimed methods to a Lexis search for “more cases that are like this case.”2 Oral Argument Tr., No. 2015-1652, 19:13-20:30.
Fox SMART confirms that, while the process used to search a database is similar to the process used to create the database, the processes are different. J.A. I 10281 (“Searching a clustered tree is very much like finding the correct place to add a new document.... However, searching has a different objective; instead of finding the single twig where insertion should follow[,] one would like to retrieve and rank documents so that all relevant documents, regardless of what cluster they appear in, are retrieved as soon as possible.”). Neither Fox SMART nor Fox Thesis suggests that the cluster-splitting process used in database creation is part of search. Cross-Appellants’ reliance on Fox’s cluster-splitting in support of their unpatentability arguments is therefore misplaced.
Moreover, the search process described in Fox SMART and Fox Thesis necessarily follows database creation. See J.A. I 10270 (“Clustering is simply a process for creating groupings, and clustered search allows retrieval of groups in response to query submission.”). By the time a user searches the database, the document tree is already full. It is thus the links within this fully formed tree that must serve as the claimed “candidate cluster links,” a proposition that counsel for Cross-Appellants Linked-ln and Twitter admitted at oral argument results in a “dead end.” Oral Argument Tr., No. 2015-1652, 22:34-22:42. This is because Fox SMART makes clear that a search returns all documents in the database. As the Board explained, “Fox *646SMART states that ‘one would like to retrieve and rank documents so that all relevant documents, regardless of what cluster they appear in, are retrieved as soon as possible.’ ” Board Decision III, 2015 WL 456539, at *10 (emphasis added) (quoting J.A. 1 10281). No documents are deleted or otherwise removed before presentation to the user. As such, the Board was correct to find that Fox SMART’S search process cannot satisfy the unchallenged construction of “actual cluster links.” Id, at *10 (“We are not persuaded that Fox SMART’S description of ranking documents discloses deriving a subset because a set of ranked documents provides an indication of an order of presentation, but is not a subset.... Because Petitioner does not point to disclosure of deriving a subset, Petitioner has not shown by a preponderance of the evidence display using links of that subset.”). I therefore would find that substantial evidence supports the Board’s decision that the Fox references do not disclose a key limitation of claims 1, 5, 15, and 16 of the ’494 patent and claim 21 of the ’571 patent and, as a result, do not render those claims unpatentable.
. Though claims 15 and 16 of the '494 patent are directed to “method[s] of representing the relationships between nodes,” '494 patent, 52:51-52, Cross-Appellants make no argument that this difference impacts the unpa-tentability analysis. Cross-Appellants' Br. 75 (“For the same reasons explained above with respect to claim 1, the Board's decision [with respect to claims 15 and 16] was erroneous and should be reversed.”). And, like claims 1 and 5 of the '494 patent, claim 21 of the '571 patent involves a method that begins with the selection of a node for analysis and results in the display of data as determined by "actual, cluster links.” '571 patent (Reexamination Certificate), 2:38-55.