Xerox Corp. v. X Corp., No. 23-1913 (Fed. Cir. June 17, 2025)

STOLL, Circuit Judge.

Xerox Corp. (“Xerox”) appeals the final written decision of the Patent Trial and Appeal Board holding claims 1–18 of U.S. Patent No. 7,043,475 unpatentable under 35 U.S.C. § 103. On appeal, Xerox challenges the Board’s construction of the claim term “determining proximal information,” the Board’s findings related to motivation to combine and reasonable expectation of success, and the Board’s analysis of Xerox’s alleged objective evidence of nonobviousness. For the following reasons, we affirm.

BACKGROUND

I

The ’475 patent is titled “Systems and Methods for Clustering User Sessions Using Multi-Modal Information Including Proximal Cue Information” and “relates to systems and methods for navigating through and searching for sites on a distributed network.” U.S. Patent No. 7,043,475 Title, col. 1 ll. 17–19. At a high level, the invention relates to “creating user profiles for users who visit web pages and classifying such users.” J.A. 3.

In various exemplary embodiments, (1) a plurality of user paths are selected in a collection of content portions (e.g., web pages containing links); (2) for each path, the multi-modal information for content portions associated with the user path is determined; (3) the proximal information for content portions associated with the user path is determined; (4) the multi-modal information for content portions and the proximal information for content portions associated with the user path are combined to form a user profile; and (5) the multi-modal information and proximal information of user profiles are clustered based on similarity. ’475 patent col. 1 l. 61–col. 2 l. 6. Step 3 is relevant on appeal.

3The specification discloses that weighted vectors can be used to represent multi-modal information and proximal information. ’475 patent col. 2 ll. 16–21. The “proximal cue feature vector,” the “content feature vector,” the “URL feature vector,” the “inlink feature vector,” the “outlink feature vector,”¹ and the “information need feature vector” for each document or web page on the user path are combined, and the combined “multi-modal vector” represents a user profile. ’475 patent col. 6 ll. 4–11.

The Board explained that independent claim 10 is illustrative and thus focused its analysis on that claim’s limitations. Claim 10 recites the following:

10. A computer system to cluster user sessions using multi-modal information and proximal information, comprising: [10.1] a user path determining circuit that selects a plurality of user paths in a collection of content portions; [10.2] a multi-modal information determining circuit that determines multi-modal information for content portions associated with the user path; [10.3] a proximal information determining circuit that determines proximal information for content portions associated with the user path; [10.4] a similarity determining circuit that combines the multimodal information for content portions and the proximal information for content

4portions associated with the user path to form a user profile having a unified representation; and [10.5] a cluster and similarity measure determining circuit that clusters multi-modal information and proximal information of user profiles based on similarity.

’475 patent col. 16 ll. 5–23 (emphasis added). Limitation 10.3, which involves “determin[ing] proximal information,” is relevant on appeal.

II

The Board’s decision involves five prior art references, though not all five are relevant on appeal. We thus describe only the two prior art references necessary to address Xerox’s challenges.

Chen²

Chen discloses “[a] system and method for browsing, retrieving, and recommending information from a collection [that] uses multi-modal features of the documents in the collection, as well as an analysis of users’ prior browsing and retrieval behavior.” J.A. 1584. “Multi-modal features may take on many forms, such as user information, text genre, or analysis of images.” J.A. 1587 ¶ 27. Chen’s disclosure involves (1) “methods for quantitatively representing documents in a document collection as vectors in multi-dimensional vector spaces, quantitatively determining similarity between documents, and clustering documents according to those similarities”; and (2) “methods for quantitatively representing users in a user population, quantitatively determining similarity between users, clustering users according to those similarities, and visually5 representing clusters of users by analogy to clusters of documents.” J.A. 1584.

“[E]ach document (for example, an HTML document []) chosen from a collection [] maps to a set of feature vectors [], one for each modality (for example, a text vector [] and a URL vector []).” J.A. 1589 ¶ 37. Chen discloses numerous possible document features: “text content, document link, inlinks, outlinks, text genre, image color histogram, and image complexity. The first two of the listed features are text based, inlinks and outlinks are hyperlink based, text genre is probability based, and the final two features (image color histogram and image complexity) are image-based.” J.A. 1589 ¶ 41.

Chen explains that “[a]fter the contents of a new document (which can be a text document, image, or other type of information) are isolated [], the disclosed method uses various information sources to derive the feature vectors.” J.A. 1589 ¶ 42. “Text is extracted from the document [] and used to create a corresponding text vector [] and a corresponding URL vector [].” Id. “The text feature is a term vector, where the elements of the vector represent terms used in the document itself. . . . [F]or an all-text or HTML document (or other document type actually containing text), the text vector is based on the document’s entire text content.” J.A. 1590 ¶ 51.

Chen discloses that “[a] set of distinct features (possibly from different modalities) is precomputed for each document and stored as vectors.” J.A. 1598 ¶ 101. “The text features include the words of text surrounding and associated with each image, the URL of the image, . . . [and] hyperlink text . . . The documents are clustered into groups based on each of the features.” Id.

Chen also provides that the following information is “used in a preferred embodiment of the invention to characterize pages and users: text, URLs, outlinks, inlinks, and usage logs. The availability of this information6 motivates a multi-modal approach . . . . It is desirable to be able to exploit and combine information available from all possible modalities.” J.A. 1602 ¶ 135. “Any type of information that is available about users is collected.” J.A. 1608 ¶ 165. Chen discloses that it is “useful to collect information on the pages users access, as well as the text content, inlinks, outlinks, and URLs of these pages.” Id.

Chi³

Chi “describe[s] two computational methods for understanding the relationship between user needs and user actions. . . . The algorithms use a concept called ‘information scent’, which is the subjective sense of value and cost of accessing a page based on perceptual cues.” J.A. 1649. Chi’s “underlying assumption . . . is that user behavior in the information environment is guided by information scent, which is determined from the perception of the value and cost of the information with respect to the goal of the user.” Id.

Chi explains that “[o]n the Web, users typically forage for information by navigating from page to page along Web links.” Id. “The content of pages associated with these links is usually presented to the user by some snippets of text or graphic. Foragers use these proximal cues (snippets; graphics) to assess the distal content (page at the other end of the link).” J.A. 1650.

X Corp.’s petition “relies on Chen’s proximal information as supplemented by Chi’s proximal information selected from content portions.” J.A. 67.

7III

The Board determined that claims 1–18 of the ’475 patent are unpatentable. Three aspects of the Board’s decision are relevant on appeal: (1) the Board’s construction of “determin[ing] proximal information”; (2) its determination that Chen as supplemented by Chi teaches “determin[ing] proximal information”; and (3) its analysis of Xerox’s proffered evidence of unexpected results. See Oral Arg. at 3:33–4:00 & 22:31–23:35, https://oralarguments.cafc.uscourts.gov/default.aspx?fl=23-1913_1209202 4.mp3. See generally Appellant’s Br.

First, the Board concluded that “‘determining proximal information for content portions associated with the user path,’ means ‘finding words in text near or within a link within content portions associated with the user path.’” J.A. 23. Second, the Board found that a person of ordinary skill in the art would have been motivated to modify Chen in view of Chi’s teaching of using proximal information with a reasonable expectation of success because “Chi and Chen have similar goals of recommending information based on similar techniques” and “Chi suggests using proximal information on the user path of Chen” to “provide better recommendations of information to satisfy a user’s needs.” See, e.g., J.A. 67–71. Third, the Board explained that “it is not clear what specific ‘unexpected results’ [Xerox] relies upon.” J.A. 80. The Board found that “using content words (which include words near a link) to obtain accurate clustering results was not surprising” and, “[t]herefore, little or no evidence exists to support the conclusory allegation of unexpected results.” Id.

Xerox appeals. We have jurisdiction under 28 U.S.C. § 1295(a)(4)(A) and 35 U.S.C. § 141(c).

DISCUSSION

“Claim construction is ultimately a question of law, decided de novo on review, as are the intrinsic-evidence8 aspects of a claim-construction analysis.” Intel Corp. v. Qualcomm Inc., 21 F.4th 801, 808 (Fed. Cir. 2021).

Our court reviews “the Board’s obviousness determination de novo, but its factual findings for substantial evidence.” Volvo Penta of the Ams., LLC v. Brunswick Corp., 81 F.4th 1202, 1208 (Fed. Cir. 2023). “The substantial evidence standard asks ‘whether a reasonable fact finder could have arrived at the agency’s decision,’ and ‘involves examination of the record as a whole, taking into account evidence that both justifies and detracts from an agency’s decision.’” TQ Delta, LLC v. Cisco Sys., Inc., 942 F.3d 1352, 1358 (Fed. Cir. 2019) (quoting In re Gartside, 203 F.3d 1305, 1312 (Fed. Cir. 2000)). “Whether a skilled artisan would have been motivated to combine references or would have had a reasonable expectation of success in combining references are questions of fact reviewed for substantial evidence.” Elekta Ltd. v. ZAP Surgical Sys., Inc., 81 F.4th 1368, 1374 (Fed. Cir. 2023). The Board’s finding regarding the strength of objective indicia of nonobviousness, such as unexpected results, is also a question of fact that we review for substantial evidence. Elbit Sys. of Am., LLC v. Thales Visionix, Inc., 881 F.3d 1354, 1356–57 (Fed. Cir. 2018).

I

We begin with the construction of “determining proximal information.” Xerox challenges the Board’s construction as overly broad, arguing that: (1) the Board “conflated proximal information with multi-modal information”; (2) “[p]roximal information must characterize content”; and (3) “[p]roximal information requires an association.” Appellant’s Br. 22, 26, 32 (emphases omitted). For the following reasons, we disagree with Xerox and adopt the Board’s construction of “determining proximal information” as meaning “finding words in text near or within a link within content portions associated with the user path.” J.A. 23. In our view, this construction is more consistent with the patent specification and does not conflate9 proximal information with the broader term multi-modal information.

We first turn to the claims. The customary meaning of the terms “proximal” and “information” suggests that “proximal information” is information that is near or next to something. Considering the claims as a whole, the claims recite “using multi-modal information and proximal information,” determining “multi-modal information,” and determining “proximal information.” ’475 patent cols. 15– 16. Consistent with our precedent, the claims’ use of different terms would suggest that “multi-modal information” and “proximal information” should be given different meanings. Helmsderfer v. Bobrick Washroom Equip., Inc., 527 F.3d 1379, 1382 (Fed. Cir. 2008).

The specification aids in our understanding of how the inventors understood the terms multi-modal information and proximal information. As the specification makes abundantly clear, proximal information is a subset of multi-modal information. The ’475 patent specification repeatedly discloses that “multi-modal information includ[es] proximal cue information.”⁴ See, e.g., ’475 patent Title, Abstract; id. at col. 2 ll. 56–58; id. at col. 4 ll. 48–49. Indeed, the title of the ’475 patent is “Systems and Methods for Clustering User Sessions Using Multi-Modal Information Including Proximal Cue Information.” ’475 patent Title (emphasis added).

The remainder of the ’475 patent specification includes broad examples of “proximal information.” Figure 2 of the ’475 patent (reproduced below) illustrates “a text 202 associated with a link 204” where the text 202 “may provide10 proximal cues as to the nature of the linked information.” ’475 patent col. 3 ll. 43–45.

’475 patent Fig. 2; id. col. 2 ll. 43–44.

The specification’s description of Figure 2 supports the Board’s interpretation:

Proximal terms represent information cues that convey information to the simulated user. For example, a text 202 associated with a link 204 may provide proximal cues as to the nature of the linked information. The proximal cue words may include portions of the text 202 surrounding the link 204. So, for example, if the text 202 says “Buy a gift certificate”, then the words “buy”, “a”, “gift”, and “certificate” are used as words that describe the proximal cues. If the link 204, in addition, ha[s] words 206 near it that say[] “click here to purchase gift certificates that are discounted”, then these words 206 are also used to describe the proximal cues. If the link 204 is in typical Uniform Resource Locator (URL) form, the punctuation and “/” characters associated with the URL are used to define

11word boundaries. For example, if the link 204 such as, for example, “http://www.xerox.com/products/support/index.html” is processed, the words http, www, xerox, com, products, support and index are considered as proximal cue words for the relevant link 204.

’475 patent col. 3 ll. 41–59. The specification thus indicates that words like “a” and “www” near or in a link are “proximal information.” Given the breadth of items that the specification identifies as “proximal information,” we are persuaded that the Board did not err in construing “proximal information” as “words in text near or within a link within content portions associated with the user path” without the further qualifiers urged by Xerox. J.A. 23.

Other parts of the specification further support this view, stating that “[t]ext surrounding the link structure, the title of the proximal page as well as features such as where the link is found within the document may also be analyzed for proximal cue words.” ’475 patent col. 8 ll. 34– 37. The specification broadly states that “[a]ny feature of the proximal document or web page may be used to provide proximal cue words . . . .” Id. col. 8 ll. 38–40.

Xerox’s proposed construction of “proximal information”—as information characterizing the content accessible through the link—would improperly limit the claim term to certain preferred embodiments while excluding others. As our precedent recognizes, a claim construction that omits embodiments in the specification is typically not the correct one. Oatey Co. v. IPS Corp., 514 F.3d 1271, 1276 (Fed. Cir. 2008). Xerox’s proposal would require proximal information to “describe the contents of a link,” Appellant’s Br. 27, but, as the Board aptly explained, “the word ‘a,’ the letters ‘www’ and ‘com’ are not descriptive of any content to which the link points.” J.A. 22 (emphasis added).

12We thus reject Xerox’s proposed construction and adopt the Board’s construction of determining proximal information: “finding words in text near or within a link within content portions associated with the user path.” J.A. 23.

II

Next, we turn to the Board’s obviousness analysis, and specifically the Board’s findings related to: (1) “motivation to combine Chi’s proximal information with Chen,” (2) the “use of [another reference] Chi4[⁵] . . . because it’s relevant also to motivation to combine and also reasonable expectation of success,” and (3) “secondary evidence of nonobviousness.” Oral Arg. at 22:38–23:22; see Appellant’s Br. 43–64.

At the outset, we note that Xerox failed to raise the use of Chi4 in its opening brief, so that argument is waived. See Walker v. Health Int’l Corp., 845 F.3d 1148, 1156 (Fed. Cir. 2017). See generally Appellant’s Br. For the reasons that follow, we conclude that the Board’s findings on motivation to combine and reasonable expectation of success are supported by substantial evidence. We address Xerox’s arguments regarding objective indicia of nonobviousness in Part III below.

We hold that substantial evidence supports the Board’s finding that it would have been obvious to modify Chen to include Chi’s teaching of determining proximal information because “Chi and Chen have similar goals of recommending information based on similar techniques” and “Chi suggests using proximal information on the user path of Chen” to “provide better recommendations of information to satisfy a user’s needs.” J.A. 69–70. As the Board reasonably found, Chen states that “[i]t is desirable to be able to exploit and combine information available from all possible modalities” and “[a]ny type of information that is available about users is collected.” J.A. 60 (alterations in13 original) (citations omitted). Chen provides that “[t]ext is extracted from the document [] and used to create a corresponding text vector [] and a corresponding URL vector [].” J.A. 1589 ¶ 42. “The text feature is a term vector, where the elements of the vector represent terms used in the document itself. . . . [F]or an all-text or HTML document (or other document type actually containing text), the text vector is based on the document’s entire text content.” J.A. 1590 ¶ 51. Chen also discloses that “[t]he text features include the words of text surrounding and associated with each image, the URL of the image, . . . [and] hyperlink text.” J.A. 1598 ¶ 101. Accordingly, Chen itself expressly contemplates extracting words of text surrounding and associated with each image, the URL of the image, and hyperlink text. Chi goes a step further and expressly identifies proximal information by name. Chi provides that “[t]he content of pages associated with . . . links is usually presented to the user by some snippets of text or graphic. Foragers use these proximal cues (snippets; graphics) to assess the distal content (page at the other end of the link).” J.A. 1650.

The Board also reasonably found that X Corp. “show[ed] that one of Chen’s goals is similar to Chi’s, namely to satisfy a user’s information needs.” J.A. 69. Chen discloses a system for “recommending information,” J.A. 1584, and “navigation assistance” for users. J.A. 1585 ¶ 6. And Chi discloses the desire to “infer information needs” and “[u]nderstand[] the information needs of users.” J.A. 1650. Based on this record, we conclude that substantial evidence supports the Board’s finding of motivation to combine Chi with Chen.

We also disagree with Xerox’s argument that “the Board’s expectation-of-success finding is based on an erroneous, obvious-to-try theory.” Appellant’s Br. 50. This is not a case in which a person of ordinary skill in the art would have been motivated “merely to vary all parameters or try each of numerous possible choices until one possibly14 arrived at a successful result, where the prior art gave either no indication of which parameters were critical or no direction as to which of many possible choices is likely to be successful.” Pfizer, Inc. v. Apotex, Inc., 480 F.3d 1348, 1365 (Fed. Cir. 2007) (citation omitted). Here, the prior art provides guidance on the information to use to enhance navigation, thereby supporting the Board’s reasonable expectation of success finding. Again, Chi explains that “[f]oragers use . . . proximal cues (snippets; graphics) to assess the distal content (page at the other end of the link).” J.A. 1650. And, as explained above, Chen expressly contemplates extracting words of text surrounding and associated with each image, the URL of the image, and hyperlink text. Based on this record, we conclude that substantial evidence supports the Board’s finding of reasonable expectation of success.

III

Last, we review the Board’s findings related to objective indicia of nonobviousness, specifically unexpected results. On appeal, Xerox argues that the Board erred in its analysis of nexus, as well as in its alternative finding that the evidence of obviousness outweighs the strength of Xerox’s evidence of unexpected results. Because we conclude that substantial evidence supports the Board’s weighing of the relative strength of the evidence of obviousness versus unexpected results, we affirm on the later ground and need not address Xerox’s nexus argument.

“[O]bjective evidence of nonobviousness must be objective factual evidence, and not merely argument or conclusory statements of the applicant.” In re Nilssen, 847 F.2d 841 (Table), 1988 WL 32917, at *2 (Fed. Cir. 1988) (emphasis added) (citing In re de Blauwe, 736 F.2d 699, 705 (Fed. Cir. 1984)). Such “[f]actual evidence can be in the form of patents, publications in the relevant art, data from tests or experiments and affidavits of persons skilled in the art . . . .” Id.

15As the Board recognized, Xerox relied on a “proposal submitted by the inventors of the ’475 patent”—i.e., the “Invention Proposal”—made about eleven months before the patent filing date. J.A. 79. That proposal states, among other things, that “[t]he basic idea is that terms found in and around the links actually traversed by the user are more likely to represent the actual interests of the user.” J.A. 3287. Xerox’s expert, however, conceded that this concept was known at the time of the invention. The Board thus reasonably found that using content words near or in a link to obtain accurate clustering results is not surprising. Moreover, as the Board recognized, the Invention Proposal attributed the increased accuracy to factors in addition to or other than the general use of proximal information. The invention proposal states that choosing the correct weighting schemes and data modalities allowed the inventors to achieve accuracies as high as 99%. The Board thus reasonably found that “[t]he Invention Proposal does not show that using proximal information provided anything surprising in the way of accuracy or otherwise.” J.A. 85. Accordingly, we conclude that the Board’s finding that “little or no evidence exists to support the conclusory allegation of unexpected results” is supported by substantial evidence. J.A. 80.

CONCLUSION

We have considered Xerox’s remaining arguments and find them unpersuasive. For the foregoing reasons, we reject Xerox’s challenges to the Board’s decision and affirm.

AFFIRMED