[00:00:00] Speaker 04: Appeal 15, 1848 is concerned. [00:00:03] Speaker 04: That case is submitted. [00:00:05] Speaker 04: We'll turn to the next one, which is 1649, Software Rights Archive against Facebook and others. [00:00:15] Speaker 04: Mr. Hardy, and I think to the extent that the issues are the same and overlap, you might just proceed to whatever new issues you wish to raise. [00:00:27] Speaker 01: Sure. [00:00:28] Speaker 01: OK. [00:00:28] Speaker 01: So I want to continue on. [00:00:30] Speaker 01: In fact, for claim 18 in the 494 patent, it also has a structure that requires you to create a first numerical representation and direct relation. [00:00:38] Speaker 04: So you're saying that the same issue exists in this next case? [00:00:41] Speaker 04: That's correct. [00:00:42] Speaker 04: I don't think we need to repeat ourselves. [00:00:49] Speaker 01: So along with that same issue, I want to point something out and address something that she said. [00:00:54] Speaker 01: She seemed to say that a few points being negative [00:00:58] Speaker 01: That doesn't mean anything. [00:01:00] Speaker 01: That means they're all still useful. [00:01:01] Speaker 01: Just because they're negative, it's not useful. [00:01:04] Speaker 01: And I think that I really want to address that. [00:01:06] Speaker 01: And she's citing the Mutei argument, citing the case Mutei. [00:01:11] Speaker 01: I'd like to direct the court to page 31 of SRA's reply brief. [00:01:20] Speaker 01: And what we have depicted on page 31 is table 8.11. [00:01:30] Speaker 04: Okay, this is the reply brief in this case. [00:01:32] Speaker 01: That's correct. [00:01:33] Speaker 01: Okay. [00:01:34] Speaker 01: Oh, wait a minute. [00:01:36] Speaker 01: Let me... I apologize, Your Honor. [00:01:38] Speaker 01: That's the 352 citation, but the same point is in both groups. [00:01:48] Speaker 01: It's found on page 29 of the Fortnight Portrait. [00:01:55] Speaker 03: Page 29 of your grade brief in the 1648. [00:01:59] Speaker 01: So it's page 29 of the 494 numeric representation group. [00:02:06] Speaker 01: What is the number of that? [00:02:09] Speaker 01: 1649. [00:02:14] Speaker 01: If we turn to that table, I just want to show exactly what we're talking about when we say these things, harm versus direct links. [00:02:25] Speaker 04: The table on page 29? [00:02:27] Speaker 01: Page 29 of the reply brief. [00:02:29] Speaker 04: Pellet and response, brief, the case. [00:02:37] Speaker 04: It looks like this. [00:02:47] Speaker 01: There's a big table there. [00:02:53] Speaker 01: OK. [00:02:54] Speaker 03: It's the yellow brief. [00:02:59] Speaker 04: Okay. [00:03:00] Speaker 04: Okay. [00:03:01] Speaker 04: Proceed. [00:03:02] Speaker 01: Okay. [00:03:02] Speaker 01: If we look at this table, I want to direct you to where the first arrow is and show you how to read these tables because they have a very clear picture here. [00:03:11] Speaker 01: What you see is terms and LN, right? [00:03:16] Speaker 01: And that produced an 8.8 improvement versus term or precision of .3431. [00:03:23] Speaker 01: And that's using just terms and direct links to search. [00:03:27] Speaker 01: Now we look for the second result, and what you'll see is terms BC and LN are being tested. [00:03:35] Speaker 01: And the only difference between the other result is that BC, an indirect relationship, is added to the search vector, and then the test is run. [00:03:44] Speaker 01: And what it shows is the impact of adding BC by itself is negative 7.7%. [00:03:52] Speaker 01: It's not that it's less effective. [00:03:54] Speaker 01: It's actually harmful to the search. [00:03:57] Speaker 01: It's contributing negatively to the search. [00:04:00] Speaker 01: Because terms and direct links are in both subvectors, and the only difference is BC. [00:04:06] Speaker 01: You are actually isolating BC's impact, and it's harmful. [00:04:10] Speaker 01: Not a viable alternative, not a less effective alternative. [00:04:13] Speaker 01: Now, in this particular case, we see it's negative 7.7%. [00:04:17] Speaker 01: I want to point out that if you look at Salton's work, and Salton is the thesis advisor of Fox, [00:04:23] Speaker 01: What his work demonstrated is... I've come further down. [00:04:26] Speaker 03: There's another one where there's the same lineup of an X next to the BC and an X next to the LN. [00:04:34] Speaker 03: And it says plus 10.0. [00:04:37] Speaker 03: Okay. [00:04:38] Speaker 01: That one is using direct links and LN using regression coefficients. [00:04:42] Speaker 01: And how it differs than the previous test is the regression coefficients, they look at the correct answers and then they try to rejigger weights to optimize the outcome. [00:04:53] Speaker 01: In this case, by using LN, they got a 10% improvement, which is even a stronger case of just using direct links by themselves. [00:05:04] Speaker 03: I don't understand that. [00:05:07] Speaker 03: I mean, we've got a plus 10. [00:05:11] Speaker 03: That's better than plus 8.8. [00:05:14] Speaker 01: Right. [00:05:14] Speaker 01: And that's from terms and direct links. [00:05:17] Speaker 01: I understand the one you're pointing at. [00:05:19] Speaker 01: I'm pointing to the one from the fifth from the bottom. [00:05:23] Speaker 01: Fifth from the bottom. [00:05:24] Speaker 03: One, two, three, four, five. [00:05:27] Speaker 03: See how there's a BC and there's an LN with an X next to each? [00:05:32] Speaker 01: BC, LN. [00:05:34] Speaker 01: OK, so they added BC there. [00:05:35] Speaker 01: If you compare BC, that's the exact point. [00:05:38] Speaker 01: What you see here in this particular instance with the regression coefficients turned on, when you compare direct links and [00:05:46] Speaker 01: and BC and LN to using just terms and LN which is the result three of you see the plus ten that's the third result in this section or one two three four above again you showed that the impact of BC in this particular test was actually zero meaning that well actually it's a slightly negative [00:06:10] Speaker 01: 0.01 percent, negative 0.01 when you know the comparisons. [00:06:15] Speaker 03: This is for the CACM database, right? [00:06:18] Speaker 03: That's correct. [00:06:19] Speaker 03: Isn't this database where Fox himself opined that maybe BC isn't as good for this particular database just because of the particular data that's in this database? [00:06:30] Speaker 01: No, he actually says something much stronger than that. [00:06:32] Speaker 01: What he says to this, based upon the experiments on this database, [00:06:37] Speaker 03: And there were limitations to the database, in his view, which is why he didn't feel like the BC was as effective as it could have been otherwise. [00:06:46] Speaker 03: First off, you need to understand that half of all... Am I wrong in my understanding of what he said? [00:06:50] Speaker 01: Yes, I believe it's incorrect. [00:06:52] Speaker 01: What he's actually saying is this. [00:06:54] Speaker 01: He's saying, based upon these tests, and this is the quote, and it's actually, if you flip the page, you will see a series of these quotes on page 30. [00:07:04] Speaker 01: Just flip the next page, it'll direct you to where to find these. [00:07:07] Speaker 01: The first thing I want to point from this is what he says is that if the other subvectors are present, cc is not really needed, and bc is probably not either. [00:07:18] Speaker 01: One of the subvectors we're talking about is the ln subvector. [00:07:22] Speaker 01: So he's directly stating is if you have the ln subvector, you don't need bc, and you don't need cc. [00:07:28] Speaker 01: And then he goes on. [00:07:29] Speaker 01: The very next thing to say he says about this is that he says his ultimate recipe of what subvectors you should use [00:07:36] Speaker 01: are direct links, terms, and no BC and no CC. [00:07:41] Speaker 01: And if I could get you to look at page 12 of the opening brief, we can see these quotes in full. [00:07:47] Speaker 01: And so this isn't a mixed bag. [00:07:50] Speaker 03: Because the CAC... I was looking at JA5594. [00:07:54] Speaker 03: I'm sorry, this is going back to the 352 patents. [00:08:03] Speaker 03: There's a sentence in there. [00:08:05] Speaker 03: I'm sure you're familiar with it. [00:08:07] Speaker 03: Since the CACM collection of bibliographic data based only on internal references, that is between pairs of articles in the same journal, it is not surprising that the BC subvector seems rather sparse and not very useful. [00:08:20] Speaker 03: So it's a commentary more on the nature of the data being mined rather than on the nature of the algorithm being used. [00:08:28] Speaker 01: Well, I would suggest to you that it's directly suggesting what I'm saying. [00:08:32] Speaker 01: What he's saying is, and I don't have the quote in front of me, but I think what I just heard you say is that the CACM is a closed collection where the analyzed links are direct relationships between the objects in the database. [00:08:50] Speaker 01: That's different than the ISI test that these positive quotes come from, where it's only using source-sided pairs outside the database. [00:08:58] Speaker 01: That's a direct reference. [00:08:59] Speaker 01: When direct citations in the database are there, BC is not very needed or not very important. [00:09:05] Speaker 01: And that's exactly what we're saying here. [00:09:07] Speaker 01: And indeed, this is far more explicitly said. [00:09:11] Speaker 01: If I could get the court to turn to page 12 of the opening brief, which is, if I can pull that out. [00:09:27] Speaker 01: Could we turn to page? [00:09:28] Speaker 01: If we could turn to page 15 of the opening brief on the 494, here are the quotes directly on point. [00:09:48] Speaker 01: And this is what I'm talking about. [00:09:49] Speaker 01: When it comes to the specific issue of whether indirect relationships perform, I'll perform direct relationships. [00:09:56] Speaker 01: There's only one story being told in these papers. [00:09:59] Speaker 01: On page 15, what we can see here are two quotes that I think the court should particularly focus on. [00:10:04] Speaker 01: If we're there, I will read them too. [00:10:07] Speaker 01: It can be inferred, however, that if the other subvectors present, CC is not really needed and BC is probably not either. [00:10:14] Speaker 01: In other words, if you created a direct link subvector, just like the claim 26 requires you to do, then you don't need BC and CC. [00:10:22] Speaker 01: And why? [00:10:23] Speaker 01: Because they harm the search results. [00:10:25] Speaker 01: Then look at the next quote we see. [00:10:27] Speaker 01: The recipe proposed is to at least employ terms, some manually assigned category scheme, that's a CR, and direct links between the documents. [00:10:36] Speaker 01: In other words, use only the direct relationships to search. [00:10:40] Speaker 01: That's what these empirical results that I'm saying are teaching. [00:10:44] Speaker 01: And he's explicitly stating it right here that that is what you should do. [00:10:49] Speaker 04: He says at least. [00:10:50] Speaker 04: You're saying that excludes any alternative. [00:10:54] Speaker 01: I'm sorry, I didn't understand the question. [00:10:56] Speaker 04: What you just read to us is at least employ terms. [00:11:00] Speaker 04: Right. [00:11:00] Speaker 04: Saying that that's the limitation. [00:11:03] Speaker 01: That's correct. [00:11:04] Speaker 01: And then he goes on to say that the LN subvectors are longer than the other two, are easier to obtain, so use of them is encouraged by practicality considerations as well as effectiveness tests. [00:11:15] Speaker 01: The effectiveness test, he's also saying it's difficult to do BC and CC because it took them two years to do this, but he's also referring to the effectiveness test. [00:11:24] Speaker 01: These are the words that encapsulate what his experimental results actually say about BCNCC when you take a harder look at it. [00:11:31] Speaker 01: And again, I just can't stress enough. [00:11:33] Speaker 03: It's about... It's confusing though, because he has these sum up statements at the end of various chapters that says, [00:11:41] Speaker 03: You know, BC and CC, they can be used too. [00:11:45] Speaker 03: And there's still some optimism about using them. [00:11:48] Speaker 03: It's not, you know, you can only use terms and direct links and categorization values. [00:11:58] Speaker 03: And don't ever try this PCCC business. [00:12:03] Speaker 01: Well, your honor, I don't believe it's confusing because [00:12:07] Speaker 01: The reason why is yes. [00:12:09] Speaker 01: And BC and CC can be useful when you don't have direct links present. [00:12:13] Speaker 01: That's why he says that, because the ISI where there's no direct links, it is useful. [00:12:17] Speaker 01: But when it comes to a situation where direct links are present, as required by our claim, it's not confusing. [00:12:23] Speaker 01: It's never useful. [00:12:25] Speaker 01: The only teachings that are addressed to that is don't use BC and CC, and the only experimental results are harmful. [00:12:30] Speaker 01: So yeah, there's some statements that generally BC and CC in some contexts might be useful, but not the context that [00:12:37] Speaker 01: to claim specifically choir, where first numerical representation has been created. [00:12:42] Speaker 01: And Your Honor, if I may, I want to save some time for rebuttaling. [00:12:46] Speaker 04: Yes, you have a little rebuttal time. [00:12:48] Speaker 04: Let's hear from Ms. [00:12:49] Speaker 04: Keefe. [00:12:50] Speaker 01: OK. [00:12:51] Speaker 01: Actually, I would like to make one last point, very quickly. [00:12:55] Speaker 01: And that is, I want them to have a chance to respond to this. [00:13:01] Speaker 01: The petitioners seem to say, if there's any substantial evidence the indirect relationships are useful, [00:13:07] Speaker 01: their entitled affirmance. [00:13:09] Speaker 01: But that is absolutely incorrect. [00:13:11] Speaker 01: Because this is not a jury trial where one presumes all fact findings in favor of the petitioner necessary to support the prevailing party or the judgment. [00:13:21] Speaker 01: This is an IPR with a rule of con, RUFET, which are case decided on page 27 of my opening brief. [00:13:30] Speaker 01: as well as being affirmed recently by this court on cuts first versus motive power, a January 22, 2016 opinion that has come out, where the board must affirmatively explain the motivation to combine in its opinion, or did you presume to be arrived at a hindsight bias? [00:13:48] Speaker 01: And if I can, I can show you that there is no analysis directed to the issues that we're talking about here in the opinion. [00:13:56] Speaker 01: And they have failed to do so. [00:13:57] Speaker 01: Therefore, SRA is entitled to a reversal, or at least a remand on this issue, regardless if there's substantial evidence in this record. [00:14:09] Speaker 01: Because we have no idea what this board actually thought about that substantial evidence. [00:14:13] Speaker 01: And I want to turn to page six, where the board does address indirect relationships. [00:14:20] Speaker 01: If we can turn to page six of the reply brief. [00:14:25] Speaker 01: I'm going to try to get this as fast as I can. [00:14:28] Speaker 01: And what is said here is the board only confines its analysis, if we turn to page six, to whether the pre-elements limitation of electronic database is taught away by the negative empirical results. [00:14:45] Speaker 01: What it fails to do is determine where the empirical evidence would teach away from the last three elements [00:14:56] Speaker 01: of the claim embodied in searching with indirect relationships. [00:15:01] Speaker 01: There's no attempt to explain why somebody would search using indirect relationships, even though empirically it hurts within the claimed arrangement. [00:15:10] Speaker 01: Rather, what the board solely focuses on is we say that the Fox papers, in the sum total of that language, if I get the court to look at it, it's found on page six of our reply brief. [00:15:22] Speaker 01: It says, it's evidence of degraded results does not teach away from the combination of the Fox papers, but rather from the modification of the teachings of the Fox papers to incorporate electronic database. [00:15:32] Speaker 01: Never does it explain why someone would be motivated to search and view the indirect relationships. [00:15:37] Speaker 01: They never refer to any of these quotes they now say is sequential evidence. [00:15:42] Speaker 01: We don't know what the board thinks about those quotes or whether they interpret them the way they're now alleging. [00:15:46] Speaker 01: It's just not part of the record. [00:15:48] Speaker 01: And under the rule of Kahn, what Kahn says in that situation, [00:15:52] Speaker 01: is that you must, when the board does not explain the motivation or suggestion or teaching that would have led a skilled artist at the time of invention to declaim the combination as a whole, we infer that the board used hindsight to conclude the invention was obvious. [00:16:08] Speaker 01: And you set aside the verb. [00:16:10] Speaker 01: Here, they did not provide any explanation of why someone would search and view the massive amount of evidence we put on the indirect relationships and other information in the record. [00:16:19] Speaker 01: And that failure to explain affirmatively in their opinion requires that this court vacate the opinion. [00:16:26] Speaker 01: Of course, we would say that there's no point to even go back and remand, because there's no substantial evidence as to this claimed arrangement. [00:16:32] Speaker 01: We could resolve this with a decision in favor of SRA right here and right now. [00:16:37] Speaker 01: But again, there's a complete failure to explain in the opinion. [00:16:39] Speaker 01: It only confines its analysis to direct links. [00:16:43] Speaker 01: And it does not provide a sufficient explanation of why someone would search [00:16:48] Speaker 01: in view of the negative empirical results. [00:16:51] Speaker 04: To the extent that you've divided your time, it's up to you as to when you want to yield to your colleague. [00:17:05] Speaker 00: I appreciate it, your honor. [00:17:06] Speaker 00: Thank you. [00:17:06] Speaker 00: To briefly address, the board absolutely gave [00:17:12] Speaker 00: its rationale for why you would use the Fox papers. [00:17:15] Speaker 00: As I explained when I first stood up, the board said that they found the Fox papers to be useful. [00:17:21] Speaker 00: They cited directly to the pages in the Fox thesis that said that you would want to use BC and CC to enhance results, as I explained earlier. [00:17:30] Speaker 00: There's another quotation which has been ignored, which is on JA 05588 from the first record and 5718 for the 494 case. [00:17:40] Speaker 00: which specifically says, Fox says, co-citations, indirect, do much better, though still not as well as terms, when only direct co-citations are considered, they are better than authors, and here's the critical point, but still not as good as the combination of direct and indirect co-citations. [00:18:00] Speaker 00: So Fox is saying, it's still best when you use indirect as well as direct. [00:18:05] Speaker 00: And so, in fact, you would be motivated to use the [00:18:11] Speaker 00: claim to create indirect relationships because they will, in fact, enhance. [00:18:16] Speaker 00: The record is almost identical. [00:18:18] Speaker 00: The only thing that changes the 494 versus the 352 is that in the 494, it's actually an even broader claim that talks specifically about the fact that all you have to do is create these relationships and use them for searches. [00:18:31] Speaker 00: It doesn't link back up to the computer. [00:18:33] Speaker 00: And so it's even broader and therefore even more abstract. [00:18:37] Speaker 00: But all of the same motivations, [00:18:39] Speaker 00: Although the quotations that say using BC and CC improves can be found at 5693 in the original appendix, at 5583, which encourages further investigation, 5588 saying the combination of direct and indirect is incredibly positive, which we just saw, 5728, which says the indirect alone with terms yields a positive 5.6%. [00:19:03] Speaker 03: Which JAs are these sites coming from? [00:19:06] Speaker 00: I do apologize. [00:19:07] Speaker 00: Those were from. [00:19:08] Speaker 00: The original, they are co-cited for 5728 is from the 494. [00:19:14] Speaker 00: So 5728 is the chart that you had been looking at earlier, which if you look through that entire chart, what it shows, this is the chart 8.11 at 5728 for the 494, co-cited at 5598 for the 352. [00:19:31] Speaker 00: Every time that BC and CC are added to terms, they improve. [00:19:37] Speaker 00: the results. [00:19:39] Speaker 00: When BC is added, it's improved. [00:19:41] Speaker 00: When BC and CC are added, it's improved. [00:19:43] Speaker 00: But even more critically, the best result of all is when you add in all of the terms, including the indirect relationships. [00:19:50] Speaker 00: So absolutely, you would want to create indirect links, indirect couplings, in order to add them so that you have the best possible result. [00:20:01] Speaker 00: Therefore, the board does, in fact, have substantial evidence. [00:20:04] Speaker 00: The board indicated that it found these to be useful. [00:20:07] Speaker 00: and therefore found that Fox said what it said and had all of the elements. [00:20:12] Speaker 00: And very last, I'd like to point out that the reason that the board focused so specifically on whether or not you would modify the references in order to include the electronic database and focused its attention on that in terms of the teaching away was because of a question posed directly to patent owner during the oral argument. [00:20:31] Speaker 00: The board asked patent owner, so you're saying, and this is directly in the board's decision, [00:20:36] Speaker 00: You're saying that the teaching away only goes to whether or not to modify the reference to include an electronic database. [00:20:44] Speaker 00: The answer was yes. [00:20:45] Speaker 00: The board therefore focused its attention on explaining that there was an adequate teaching to combine the references and use them with an electronic database, and that is replete through the board's analysis. [00:20:57] Speaker 00: But that doesn't negate the fact that earlier in [00:21:00] Speaker 00: The board's analysis, the board specifically said that it found what Fox was teaching to be useful, citing to a page that said adding BCNCC is useful, and we want to keep doing that, as affirmed by the results in Chapter 8. [00:21:13] Speaker 00: The end of Chapter 8 says, using all of these subvectors, including indirect relationships, has a marked improvement, and we want to keep looking at it. [00:21:21] Speaker 00: And with that, Your Honor, I'll pass my time over to my colleague. [00:21:27] Speaker 04: I don't think we're ready for you as yet. [00:21:30] Speaker 04: We haven't completed the response time. [00:21:35] Speaker 04: They've divided up the time. [00:21:38] Speaker 00: Thank you, Your Honor. [00:21:39] Speaker 02: Good morning, Your Honors. [00:21:40] Speaker 02: And I may have lost track of the time. [00:21:43] Speaker 02: I was going to address, with the court's permission, Petitioner's Cross Appeal. [00:21:49] Speaker 02: I believe, if I followed the clock correctly, Mr. Hardy had used all of his 15 minutes for this proceeding. [00:21:56] Speaker 02: But maybe I was confused about that. [00:21:58] Speaker 04: Well, in any case, too. [00:22:00] Speaker 02: If the court is ready to hear argument on the cross-appeal, that's what I'd like to speak to. [00:22:04] Speaker 04: Thank you, Your Honor. [00:22:08] Speaker 02: So petitioner's cross-appeal concerns a single claim limitation. [00:22:13] Speaker 02: And that claim limitation, which is deriving actual cluster links from the set of candidate links, recites what the specification itself essentially describes as a routine step [00:22:27] Speaker 02: in the process of clustering, which is to say grouping similar documents together, which is in turn and long has been a routine technique in information retrieval. [00:22:39] Speaker 02: And I think very significantly for this appeal, the parties are in agreement in their briefing that in finding that the Fox papers failed to disclose that single step deriving actual cluster links from candidates, [00:22:55] Speaker 02: the board disregarded certain teachings within the Fox papers. [00:23:01] Speaker 02: And we contend, and I would argue that SRA only half-heartedly disputes, that those particular teachings within the Fox papers, that the parties agree the board... Do they have cluster splitting? [00:23:15] Speaker 02: Yes, specifically cluster splitting. [00:23:16] Speaker 03: I thought I saw the board actually talking about the notion [00:23:20] Speaker 03: clusters splitting in its tree formation analysis. [00:23:24] Speaker 03: It didn't talk about it? [00:23:26] Speaker 02: The board mentioned splitting, and it mentioned specifically the uncore, what are called the uncore and concentration tests. [00:23:36] Speaker 02: And it said that these are a part of forming the tree, which is itself going a little bit astray. [00:23:44] Speaker 02: What the board did not do was consider at all [00:23:48] Speaker 02: The position that was presented by petitioners in our opening petition, opening declaration, and then consistently thereafter, that during the cluster splitting process, what occurs is, and I say cluster splitting and that may be not the best abbreviation because a little bit more accurately descriptive is cluster splitting and reformation of new clusters. [00:24:12] Speaker 02: What happens is a cluster has a size limit. [00:24:16] Speaker 02: This is what Fox teaches. [00:24:17] Speaker 02: He gives the example of 20 documents. [00:24:19] Speaker 02: Otherwise, it becomes too unwieldy to be useful. [00:24:22] Speaker 02: When the cluster exceeds that size limit, the cluster is split, which is to say it's abolished. [00:24:28] Speaker 02: And what the system then does is it now has 21 documents that it wants to cluster. [00:24:35] Speaker 02: It wants to form new, smaller clusters out of those 21 documents. [00:24:41] Speaker 02: What we presented below and what the board never engaged with is the following, that what happens in Fox Smart at that point and Fox Thesis is that the system creates what Fox describes as a complete similarity matrix, which consists of, again using Fox's term, [00:25:03] Speaker 02: pairwise similarity values. [00:25:05] Speaker 02: So all that means is the system has 21 documents that it wants. [00:25:09] Speaker 03: The claim calls for selecting a node for analysis. [00:25:12] Speaker 03: Correct. [00:25:12] Speaker 03: And then all of the analysis is done with reference to that node. [00:25:18] Speaker 03: So the fact that a particular cluster on the tree during the formation process starts growing to a point where it needs to be cleaved into two, it doesn't really [00:25:31] Speaker 03: It's not really relevant to the question of the selected node to what extent are the candidate cluster links being derived down to actual cluster links for that particular selected node. [00:25:44] Speaker 02: Well, they are. [00:25:45] Speaker 02: And I think the first thing to say about that is that was not the basis for the board's rejection. [00:25:50] Speaker 02: But independently, the way that the process works is it picks a document of the 21. [00:25:56] Speaker 02: Say they're numbered and you have document 1. [00:25:59] Speaker 02: And it calculates what is this document's similarity to every single other document in the grouping. [00:26:05] Speaker 02: So a node is a document. [00:26:08] Speaker 02: It's selecting a document. [00:26:10] Speaker 02: And it's performing a calculation to create a similarity value between what's the similarity between document one and document two. [00:26:18] Speaker 02: What's the similarity between document one and document three? [00:26:20] Speaker 03: I thought the node is being selected for creating a tree. [00:26:23] Speaker 03: The overall tree for the node. [00:26:27] Speaker 03: And then you start populating the tree. [00:26:30] Speaker 03: And you're creating clusters. [00:26:32] Speaker 03: And some clusters are, I don't know, on the visual tree closer to the node than other clusters. [00:26:40] Speaker 02: I think this is a little bit where the confusion arises, but the clustering [00:26:45] Speaker 02: A node is selected for analysis in the claim to be clustered with other nodes based on partner and direct relationships. [00:26:53] Speaker 02: That is what's taught in the Fox system. [00:26:56] Speaker 02: When documents are added to the tree in the first place, they are clustered. [00:27:01] Speaker 02: But what Fox teaches is that clusters need to have size limits to be useful. [00:27:05] Speaker 02: If you have 300 documents in a cluster, it's too big to be helpful. [00:27:09] Speaker 02: So a size limit is set. [00:27:11] Speaker 02: Say it's 20 documents. [00:27:13] Speaker 02: When a cluster exceeds that, it's abolished and new clusters are formed. [00:27:18] Speaker 02: And each document is selected as a node. [00:27:22] Speaker 02: And analysis is performed with respect to each document as to how similar is this document to every other document in the group. [00:27:30] Speaker 02: And then the algorithm looks at what are the closest relationships, what Fox calls highly correlated pairs, the strongest similarity values. [00:27:38] Speaker 02: And it puts those into clusters. [00:27:41] Speaker 02: I see my time is up. [00:27:43] Speaker 04: I believe... Yes, you've saved some rebuttal time now that we're on the cross appeal. [00:27:48] Speaker 04: So let's hear from Mr. Hardy and let's limit the discussion to the new issues that have been raised on the cross appeal. [00:27:56] Speaker 01: Sure. [00:27:56] Speaker 01: I would like to... I think you, Judge Timm, hit it right on the head, is they never explain how this process generates candidates for a selected node or derives actuals for a selected node. [00:28:08] Speaker 01: You see the clusters being split has nodes that are not linked to each other or have a common link. [00:28:13] Speaker 01: They have orphans in them. [00:28:14] Speaker 01: You can all through the fox discuss it. [00:28:17] Speaker 01: Indeed, in the initial assignment of the tree, you just placed nodes in without respect to any particular relationship between the nodes. [00:28:26] Speaker 01: So what you have in what the complete similarity matrix or the 20 nodes that he's referring to analyze it to every other node in the database, you'll have orphans among that. [00:28:36] Speaker 01: which are not linked to the selected node. [00:28:39] Speaker 01: Generating candidate cluster links for the selected node is referring to finding a set of candidates linked [00:28:46] Speaker 01: to one particular node. [00:28:48] Speaker 03: But I think what the opposing council is trying to say is if there's, say, a hard limit on the size of any given cluster, maybe it's 20 nodes. [00:28:57] Speaker 03: And then now you locate a new node in the database you want to insert in that cluster because that new node best fits with this particular family of nodes in one cluster. [00:29:09] Speaker 03: Now you've got to divide up that cluster. [00:29:11] Speaker 03: and now each of the now 21 nodes could be regarded as a selected node and now you've got to understand and evaluate that individual selected node's relationship with all the other quote unquote selected nodes and then by doing that you [00:29:29] Speaker 03: pair down into two smaller clusters and so everybody now is in a tighter, more closely linked family than the original 20 node cluster. [00:29:41] Speaker 01: But it just doesn't pan out that way that you actually, as I walk through this, derive actual cluster links for the selected node. [00:29:49] Speaker 01: The derived actual cluster links has to come from the candidate set for a selected node. [00:29:55] Speaker 01: So the actual cluster links also have to link to the selected node. [00:29:59] Speaker 01: Now, just because you pick one of those nodes and says that node is a conducted analysis, the problem is what they identified as the candidates is the 20 matrix, all 20 other nodes that are going to be analyzed in this matrix. [00:30:17] Speaker 01: And the problem with that is that can't be a candidate set for the selected node because it'll have five or six or some amount of those nodes are orphans. [00:30:25] Speaker 01: and therefore have no link relationship to the selected node, therefore are not a candidate set for the selected node. [00:30:32] Speaker 01: And so, in other words, it includes an analysis of nodes without any respective position to the selected node. [00:30:39] Speaker 01: But this problem becomes far more acute when you move to the deriving step. [00:30:43] Speaker 01: Because what you're going to do in a deriving step, what you have to do is derive a set of actual cluster links for a selected node because they come from the candidates. [00:30:51] Speaker 01: Well, when you start picking out highly correlated pairs, you can pick out A and B as being highly correlated, and B and D as highly correlated. [00:30:59] Speaker 01: But B and D aren't related to each other, nor are they related to the same selected node. [00:31:05] Speaker 01: So at the end of this process, you also do not get actual clustering links for the selected node. [00:31:11] Speaker 01: And they've never attempted to explain this. [00:31:14] Speaker 01: And I submit that they still have not attempted to explain this anywhere in their papers. [00:31:19] Speaker 01: The more important point that I want to bring up, however, is that none of these processes are used to display nodes. [00:31:26] Speaker 01: In the end, these aren't actual cluster links being used to display nodes. [00:31:30] Speaker 01: The display procedures use document to similarity, query similarity to display, and they're not based on cluster splitting. [00:31:36] Speaker 01: And I guess I will take that up. [00:31:37] Speaker 01: Since I have 30 seconds, it's a heady topic. [00:31:40] Speaker 01: And it's also in the 571. [00:31:41] Speaker 01: I'll take that up at that time. [00:31:43] Speaker 04: OK. [00:31:44] Speaker 04: Thank you. [00:31:44] Speaker 04: Now let's see. [00:31:45] Speaker 04: Mr. Silver, do you have some rebuttal time? [00:31:51] Speaker 02: Thank you very much, Your Honor. [00:31:52] Speaker 02: I do not see what the possibility of orphan nodes, which is to say documents in the original cluster that's being abolished that end up not being placed into one of the smaller clusters, therefore they're orphans, has to do with anything. [00:32:12] Speaker 02: So to repeat the process again, which is described in the Fox papers, and it's essentially exactly as Your Honor described it, Judge Chen, [00:32:21] Speaker 02: A document number 21 is added to a cluster. [00:32:27] Speaker 02: System says, now the cluster's too big. [00:32:29] Speaker 02: It's exceeded the size limit. [00:32:31] Speaker 02: Therefore, the system abolishes the cluster. [00:32:35] Speaker 02: And it knows it wants to create smaller clusters. [00:32:38] Speaker 02: There are now 21 documents that the system is analyzing that it wants to figure out how to put into clusters. [00:32:47] Speaker 02: It proceeds in an iterative process, document by document, which means selecting a node. [00:32:53] Speaker 02: It selects each document individually as a node. [00:32:56] Speaker 02: And it executes an algorithm to calculate the similarity value between the document it's examining and every one of those 20 other documents. [00:33:08] Speaker 02: Creates what Dr. Fox calls a complete similarity matrix. [00:33:13] Speaker 02: And so it's got a value for the similarity between every other document. [00:33:17] Speaker 03: What council is essentially saying is... Let me ask a different question. [00:33:21] Speaker 03: Let's assume for a second that this theory, we're not persuaded by it. [00:33:26] Speaker 03: And so it really comes down to this tree formation process that Fox teaches. [00:33:32] Speaker 03: And we also conclude that [00:33:35] Speaker 03: you create the tree, you populate the tree for the selected node, and there really isn't any final step where you say, pair back that tree, or you identify some portion of the tree as the actual cluster links. [00:33:55] Speaker 03: And so therefore, what we have here is not the 102, and so therefore the board was [00:34:02] Speaker 03: right in denying the proposed 102 rejection. [00:34:09] Speaker 03: Why wasn't there a 103 case here? [00:34:11] Speaker 03: Because maybe it is obvious and routine to, once you have the tree, to say, well, instead of having a tree representing the entire database, [00:34:25] Speaker 03: of basically ranking all the documents in the database. [00:34:28] Speaker 03: Now I'm going to try to pair back in some way for relevance purposes just to some more tidier, less unwieldy sub-portion of that entire database. [00:34:43] Speaker 03: Why isn't there a 103 in here? [00:34:46] Speaker 02: I believe there unquestionably would be and I think it's [00:34:51] Speaker 02: It's important in considering this issue to take a step back. [00:34:54] Speaker 03: But I guess what I'm asking, you're the petitioner. [00:34:56] Speaker 03: So why didn't you bring a 103? [00:35:00] Speaker 02: So in retrospect, and I'd have to look at this particular issue we may have, but the board instituted on particular grounds. [00:35:09] Speaker 02: And in this particular ground, it instituted on 102, clearly only preliminarily, but preliminarily finding that [00:35:18] Speaker 02: every claim was there explicitly taught. [00:35:21] Speaker 02: Every claim limitation was. [00:35:23] Speaker 02: We do believe that that was correct. [00:35:26] Speaker 02: Certainly, I think it goes without saying that if this limitation, which in all candor, I don't think anyone thought would be much of a subject of dispute vis-a-vis the other things we've been talking about this morning, isn't there explicitly, then it is there. [00:35:42] Speaker 02: It's clearly an obvious thing to do. [00:35:46] Speaker 02: Just to take a step back, the idea of if you accept that using indirect relationships to calculate similarity, et cetera, was known, which it was and taught, what you get to in the cross appeal is the quote inventive idea is I'm going to have a set of candidates. [00:36:06] Speaker 02: And to pick the actual ones, I'm going to pick the strongest ones from among the candidates based on some criteria. [00:36:15] Speaker 02: notion in and of itself is such a basic way to go about selecting anything, that it's hard to say that if everything else was known, to say that I'm going to pick the actual cluster. [00:36:30] Speaker 03: But that's 103. [00:36:31] Speaker 03: That's 103. [00:36:33] Speaker 03: That's not 102. [00:36:37] Speaker 02: If your honor is asking why at the outset did we not petition alternatively [00:36:43] Speaker 02: on 103 grounds. [00:36:44] Speaker 02: That's my only question. [00:36:49] Speaker 02: First of all, I'm not positive that we didn't for this particular issue. [00:36:52] Speaker 02: We did petition on multiple grounds, and the board selected the ones that it was going to institute on. [00:36:59] Speaker 02: But certainly, I don't think that should decide the outcome of this case. [00:37:10] Speaker 04: Thank you. [00:37:14] Speaker 04: Let's see. [00:37:16] Speaker 04: That concludes the argument. [00:37:19] Speaker 04: I think that concludes the argument on this issue. [00:37:23] Speaker 04: And let's the appeal on