US 11085792

Search scope and a necessary caveat

I ran targeted searches for the literal identifier 11085792 (title/assignee/claims), plus a Federal Circuit 2026 docket sweep. Two things to flag up front:

• I could not retrieve the verbatim claim set for US11085792B2. The authoritative full text supplied to me (Google Patents scrape of https://patents.google.com/patent/[US11085792](/patent/US11085792)/en) is truncated at the beginning of the description and does not include the "Claims" section. My claim overview below is therefore reconstructed from the Abstract, the "Summary" section (which recites the same three statutory aspects), and the specification — it is a faithful paraphrase of claim scope but not a verbatim claim quote.
• No CAFC 2026 docket and no district-court litigation surfaced in my searches tying the string "11085792" (or US11085792) to any Federal Circuit appeal. Treat this as a negative search result, not a certification — my tooling returned only unrelated 2026 Federal Circuit opinions (e.g., Ollnova v. ecobee, Ironburg v. Valve, AGI v. Farmers Edge, TJTM v. Google), none of which involve this patent.

Bibliographic data (from the authoritative patent text)

Abstract (as published)

A method for determining an estimated time of arrival (ETA) includes receiving a start location and a destination from a user device via a network. The method also includes obtaining a machine learning model for determining an ETA, which is generated according to a process including: obtaining historical data related to an on-demand service order; determining a high-dimensional sparse feature based on the historical data; and determining a machine learning model based on the high-dimensional sparse feature. The method further includes determining an ETA for a target route based on the machine learning model, the start location, and the destination, and transmitting the determined ETA to the user device via the network.

Plain-language overview of the independent claims

The Summary of the Invention recites three aspects, which correspond to the three independent claims (one system claim, one method claim, one non-transitory machine-readable storage medium claim). Substantively all three cover the same four-step pipeline:

1. Receive a start location and a destination from a user device via a network (the system claim frames this as a processor receiving from "the user device"; the method claim frames it as a server; the medium claim frames it as instructions causing a processor to do so).
2. Obtain a machine learning model for determining ETA, where the model is generated by a training process that itself has three sub-steps: (a) obtaining historical data related to an on-demand service order; (b) determining a high-dimensional sparse feature based on that historical data; and (c) determining (training) the machine learning model based on the high-dimensional sparse feature.
3. Determine an ETA for a target route based on the machine learning model, the start location, and the destination.
4. Transmit the determined ETA back to the user device via the network.

The system claim is additionally characterized by a storage device storing a set of instructions and at least one processor of an online on-demand service platform configured to communicate with the storage device, the processor being directed by the instructions to perform the operations above.

What the specification says makes the claims concrete

• "High-dimensional sparse feature": a mathematical expression (vector/matrix) describing characteristics of a route as a whole, potentially aggregating hundreds/thousands/millions of features, including both features of individual road sections and features reflecting interactions between road sections (and optionally between different routes).
• How it is built (FIG. 5 flow): features are extracted from historical data and classified as categorical (e.g., start location, destination, road level, vehicle type, driving mode, road condition, weather) versus continuous (e.g., driving distance, travel duration, number of traffic lights, vehicle speed, traffic status, number of road sections). Categorical features are encoded with one-hot encoding (dimension set by the feature's "attribution," i.e., the number of mutually exclusive possible states); continuous features are encoded using a Gradient Boosting Decision Tree (GBDT) producing leaf-node encodings; a mixed case combines both, optionally through a factorization machine (FM). The resulting encodings form the high-dimensional sparse feature / training set.
• Model: the machine learning model is described as a factorization machine (FM) or a field-aware factorization machine (FAM/FFM) — i.e., a model that factorizes high-order (e.g., second-order) parameters over the sparse feature. The specification notes FM may be viewed as a particular case of FFM where the sparse feature's items belong to a single field.
• Updating: the model may be updated using an online learning method (e.g., FTRL, OGD, SGD, FOBOS, RDA) on recent historical data from new service orders — described in connection with FIG. 7.
• Inference (FIG. 8): receive start location/destination, build a target high-dimensional sparse feature (which may have fewer dimensions than the training feature if features are missing, or retain N dimensions with defaults/nulls for missing features), feed it to the model, obtain the ETA, and transmit/display the ETA on the requester and/or provider terminal.

Uncertainty notes

• The exact number of claims, wording, and dependency structure are not verifiable from the text I hold; Google Patents does not show a "20 Claims" indicator for this document in the retrieved material, so I will not guess a claim count.
• I have no authoritative evidence of any CAFC 2026 appeal, PTAB/IPR proceeding, or district-court case involving US11085792. If a docket exists, my searches did not surface it, so this should be verified directly in USPTO PatentCenter / PACER / the CAFC docket rather than relied on as a definitive absence.
• Related-but-distinct DiDi patents appeared in the search results (e.g., US10948301, US11087253, US20200116504, US20200160225). These are different documents and their subject matter should not be conflated with US11085792; I did not return results keyed to those numbers in place of the requested patent.

I'll search for litigation involving US patent 11085792 across patent litigation databases.

The initial results don't reference this patent. Let me search more specific litigation databases.

No litigation hits yet. Let me check dedicated litigation databases directly.

Based on my searches across patent litigation resources (including Unified Patents' litigation portal, CourtListener/Justia federal dockets, and general web results), I was unable to identify any litigation involving US Patent 11085792. Here is my honest, grounded assessment.

No litigation found for US 11085792

Result: No district court litigation, ITC action, or PTAB proceeding referencing US 11085792 was found in the sources searched.

What I searched

• General queries for the patent number paired with "litigation," "lawsuit," "complaint," and "case docket"
• Unified Patents' litigation caselist / portal (portal.unifiedpatents.com/litigation/caselist)
• CourtListener / Justia federal district court RECAP documents
• Queries directed at the assignee, Beijing Didi Infinity Technology and Development Co., Ltd.

What I found instead (not litigation)

The searches surfaced only patent-family and prosecution materials, which are not litigation:

• US 11085792 B2 itself — "Systems and methods for determining estimated time of arrival," assignee Beijing Didi Infinity Technology and Development Co., Ltd., priority date 2017-05-22, continuation of PCT/CN2017/085375.
• Related Didi family members for ETA technology (e.g., US 10,816,352 B2; US 10,948,301 B2; US 11,580,451 B2; EP 3461310; GB 2564181). These are sibling/related patents by the same assignee — not evidence of litigation over US 11085792.
• Unified Patents' general reporting on the litigation landscape — no case naming this patent.

Important caveats on this conclusion

I want to be explicit about the limits of this answer rather than imply a guaranteed negative:

1. Absence of evidence is not conclusive proof of no litigation. The patent issued 2021-08-10 and its family has a 2017 priority date (anticipated expiration 2037-05-22). A patent this recent may simply not yet have been asserted, or any assertion may not be surfaced by the general-web searches I could run.
2. My search tool reached its step limit on the two most targeted follow-up queries (specific docket-number format and Didi-defendant searches), so those specific searches did not complete.
3. Authoritative databases I could not fully query directly include PACER (behind a paywall) and Docket Navigator / RPX Insight. A definitive negative would require a direct query of those systems, plus the CAFC docket and ITC EDIS.

Recommendation

For a definitive check, I suggest directly querying:

• PACER (pacer.uscourts.gov) — search "11085792" across district courts
• Unified Patents litigation portal — portal.unifiedpatents.com/litigation/caselist (search by patent number)
• RPX Insight — insight.rpxcorp.com
• USPTO PTAB — portal.unifiedpatents.com/ptab or the USPTO Patent Trial and Appeal Board End-to-End system

If you can confirm whether you have access to PACER or Docket Navigator, I can refine the query strategy (for example, searching by the assignee's litigating entities, since Didi's patent portfolio has been involved in disputes in other contexts) — but I will not invent case names, numbers, or party names that I cannot verify. At present, I have no verified litigation to report for US 11085792.

I'll start by verifying the structured data against live sources, since the ODP block reports no proceedings.

Let me run a few more targeted checks before concluding.

PTAB / AIA Trial Proceedings — US 11,085,792 B2

Proceedings overview

Total AIA trial proceedings on file: 0. The structured PTAB dataset supplied in this prompt (USPTO Open Data Portal ingest) returns no AIA trial proceedings for US 11,085,792 — no active proceedings, no claims invalidated, no claims sustained, no settlements, and no institution denials. Bottom line for a defendant: there is no claim-cancellation leverage available off the shelf. Nothing has been canceled, nothing has been sustained through trial, and every claim of the patent (claims 1–20 as issued) remains untested and fully live. There is no FWD to cite, no IPR estoppel to exploit, and no prior-art record to free-ride on — any invalidity case you want is a case you will have to build yourself.

---

No individual proceeding sections

The output template calls for one ### {PROCEEDING_NUMBER} block per proceeding. There are none to report, and I will not manufacture proceeding numbers, panels, or dispositions to fill the template. For the avoidance of doubt, the fields that would normally appear resolve as follows:

• Proceeding number: none exists.
• Type: n/a.
• Filed / Status / Judge panel / Grounds / Institution / FWD / Settlement / Appeal: not applicable — no petition has been filed, so no § 314 institution decision, no § 318 FWD, and no Federal Circuit appeal (CAFC docket) exists to report.

Search corroboration and its limits

• The canonical source for this task (the structured ODP block in the prompt) reports zero proceedings. That is the controlling answer.
• Web search surfaced no IPR, PGR, or CBM petition naming US 11,085,792 or its application number 16/684,693, and no Final Written Decision, institution decision, or CAFC opinion referencing it.
• The only near-hit was a different, sibling Didi patent — US 11,580,451 B2, "Systems and methods for determining estimated time of arrival" (Didi's ETA family with a 2017-07-27 priority date), which appeared only on Unified Patents' patent-analytics portal. That page is a patent profile, not evidence of a proceeding, and it relates to a different patent from US 11,085,792. Flag it as a monitoring target, not as PTAB activity on this patent.
• Verification caveat, stated plainly: two of my follow-up searches (IPR-number-prefix queries and a district-court/assertion query) were cut off before returning results, and I could not directly query PTAB E2E / the PTAB API in this session. I therefore cannot claim an exhaustive negative with absolute certainty. The honest characterization is: ODP says zero, and my independent web checks found nothing that contradicts it. The default stands — no PTAB activity on file. To close the residual gap, confirm at PTAB E2E / PTAB Center by searching application 16/684,693 and patent 11,085,792, and check CourtListener for any CAFC appeal captioned with this patent number.

---

Strategic summary

Claim status: all issued claims UNTESTED. Because no IPR or PGR was ever instituted, there is no claim-level outcome to report. Unlike a hardened patent that survived trial or a troll patent with a dead independent claim, US 11,085,792 has simply never been through an AIA trial. Every claim of the patent — including the independent claims directed to receiving a start location and destination, obtaining a machine-learning model generated from historical on-demand service order data, determining a high-dimensional sparse feature from that data, determining an ETA for a target route, and transmitting the ETA — remains in force and presumptively valid under § 282. Do not let opposing counsel tell you the patent is "narrowed" or "partially invalidated"; it is neither. Nothing has been canceled, and nothing has been confirmed as patentable over art either.

Estoppel landscape: no § 315(e) estoppel exists, and that cuts both ways. Because no petitioner ever reached a Final Written Decision, no § 315(e)(2) estoppel bars anyone from raising any ground. That means every prior-art ground is still available to you — § 102 and § 103 combinations on the machine-learning, feature-encoding, and ETA-prediction aspects are all on the table, subject only to the ordinary § 315(b) one-year bar running from service of a complaint on you (which will matter if you later want to file your own petition). The downside is equally real: there is no third-party petitioner's expert declarations, no institution decision framing the disputed claim terms, and no FWD to reverse-engineer. You will be doing the prior-art search and claim-construction work from zero. Note also that PGR was available only for nine months from grant (grant date 2021-08-10) and that window has long closed; IPR under §§ 102/103 remains the only AIA vehicle.

Pattern signals: no pattern to report. No petitioner has filed even one petition against this patent, so there is no serial petitioner to infer a strategy from and no defensive aggregator (Unified Patents, RPX, etc.) in the chain on this patent. The patent owner, Beijing Didi Infinity Technology and Development Co., Ltd., has not been forced to defend a PTAB appeal here, so there is no evidence of aggressive or non-aggressive POP/appeal behavior to read. One practical observation: the claims have a 2017-05-22 priority date and an anticipated expiration of 2037-05-22, so the patent has roughly a decade of life left — plenty of runway for someone (you, a co-defendant, or a defensive aggregator) to file the first petition if assertion activity materializes.

---

Recommended next steps

• Do not represent that any claim has been invalidated. There is no FWD, no institution decision, and no certificate of cancellation. Any argument to a court or a counterparty that a claim of US 11,085,792 is "dead" would be unsupported and sanctionable. If you need the authoritative negative, pull it from PTAB E2E for application 16/684,693 and preserve the screenshot as your evidentiary record.
• Treat the absence of PTAB activity as a signal about assertion posture, not about patent strength. Well-asserted patents tend to attract IPRs from defendants and defensive aggregators. A clean PTAB docket suggests this patent has not yet been the subject of sustained litigation — which is consistent with the ODP block's silence. If you have received a demand letter, you may be among the first targets, which means no one has done the invalidity work for you and no one has paid for claim construction.
• Build the invalidity case from scratch and check the pre-patent-owner work. The specification itself frames the invention over known techniques — gradient boosting decision trees (GBDT), one-hot encoding, and factorization machines (FM) / field-aware factorization machines (FFM) — as the core of the high-dimensional sparse feature pipeline. Those are textbook, pre-2017 machine-learning techniques, which makes a § 103 obviousness theory grounded in prior-art GBDT-plus-FM recommendation/CTR-prediction literature the natural starting point, particularly against the independent claims' "generated according to a process including" recitations.
• If you intend to file your own IPR, watch the § 315(b) clock. The one-year bar runs from service of a complaint alleging infringement of US 11,085,792. Because no other petitioner has triggered § 315(e) estoppel, a well-timed petition preserves all § 102/§ 103 grounds while avoiding the § 315(b) bar.
• Diarize the PTAB statutory timeline if a petition is filed. Once instituted, PTAB has a one-year deadline from institution to FWD, with a § 314(b) institution decision due within six months of the petition's filing. Set those milestones now so you can prepare a stay motion and coordinate any parallel district-court invalidity contentions.
• Monitor the sibling ETA family. US 11,580,451 B2 (Didi, different family, 2017-07-27 priority) is the patent that shows up in third-party analytics tooling. If assertion activity begins, expect the family to be asserted together and expect IPRs against the sibling to yield useful prior-art and claim-construction work product that is not estopped against US 11,085,792.

I'll research this patent's ownership chain. Let me start with targeted searches.

Let me try more specific sources for the assignment record and any litigation.

Let me try to surface the actual assignment record data and any litigation history.

US 11,085,792 B2 — Ownership & Assignment Analysis

Patent: Systems and methods for determining estimated time of arrival
App. No.: 16/684,693 · Filed: 2019-11-15 (continuation of PCT/CN2017/085375, filed 2017-05-22) · Granted: 2021-08-10
Priority: 2017-05-22 · Anticipated expiration: 2037-05-22

Sourcing caveat up front. The authoritative full text plus Google Patents' legal-events table was available to me (fetched 2026-09-24). I was not able to open the USPTO Assignment Center record itself, so I do not have the reel/frame number, execution date, or correspondent of record for the single recorded assignment. I will not invent them. Every field below that I could not verify is explicitly marked not retrievable. Verify the reel/frame at the Assignment Center search page: https://assignmentcenter.uspto.gov/ (patent number search) — the same records are indexed at https://assignment.uspto.gov/patent/index.html.

---

Inventors

Pattern notes:

• All three inventors are co-assignors to the same corporate assignee in a single instrument — the conventional employee-invention take-back.
• I see no evidence of inventor departure, no individual-to-third-party conveyances, and no inventor-retained rights. The "all inventors exit within 12 months" fire-sale tell is not observable here — but note I could not retrieve execution dates, so I cannot affirmatively date-check departures. Marked unclear.
• Zheng Wang is a prolific Didi inventor (multiple Didi ETA/ML patents, including a later continuation in this same family that adds an LSTM/RNN cell-state architecture). This is a deep, continuously-staffed corporate R&D bench, not a one-shot patent.

---

Original assignee

Beijing Didi Infinity Technology and Development Co., Ltd.
Building 34, No. 8 Dongbeiwang West Road, Haidian District, Beijing 100193, China

• Primary line of business: the R&D and IP-holding arm of DiDi (DiDi Global Inc.), the dominant Chinese ride-hailing / on-demand mobility operator. This entity appears as applicant/owner across DiDi's global patent portfolio (CN, WO, US, EP, AU, GB, MO, HK records all name it).
• Does it ship a product embodying the claims? Yes. US 11,085,792 claims a server that receives a start location and destination from a user device, applies a machine-learning model trained on high-dimensional sparse features extracted from historical on-demand-service orders, computes an ETA for a target route, and transmits it back to the device. That is literally the DiDi rider/driver app's ETA function, running on DiDi's dispatch platform. This is core operating-company subject matter, not a paper patent.
• Current status: operating. The corporate parent went public on the NYSE in 2021 and was later delisted from the NYSE; those are corporate-structure events at the listed parent level, not assignment events on this patent. The Chinese operating/IP entity remains active — its 2025 annual report (Tianyancha) shows active status, ~8,700+ insured employees, and 小桔科技香港有限公司 (Xiaoju HK) as sole shareholder. (Parent-level listing/delisting details are from my background knowledge and should be re-verified against SEC EDGAR before being relied on.)

---

Assignment timeline

Chronological list of every recorded assignment appearing in the legal-events record:

• Execution date: not retrievable / recorded 2021-05-05 — Reel NNNNNN/NNNN (not retrievable from available sources)
  • Conveyance: Assignment — "ASSIGNMENT OF ASSIGNORS' INTEREST (SEE DOCUMENT FOR DETAILS)"
  • Assignor: Zhiyuan Zhong; Qing Luo; Zheng Wang (all three inventors)
  • Assignee: Beijing Didi Infinity Technology and Development Co., Ltd.
  • Correspondent: not retrievable. I could not open the recording instrument, so I cannot name the attorney/firm of record. I will not guess and will not name a firm on a single-appearance basis, since the signal requires recurrence.
  • Context: Internal — inventor-to-employer assignment confirming title in the original operating company. Not an acquisition, not a fire-sale, not a securitization, not a transfer-to-asserter.

That is the entire recorded chain. There is no post-issuance assignment to any LLC, no security agreement, no merger, no change of name, and no release on record as surfaced in my sources.

If the Assignment Center shows additional entries I could not reach, the most likely candidates would be (a) a Change of Name reflecting a DiDi entity renaming, or (b) a reconfirmation of the same inventor→Didi grant. Neither would change the substance of this analysis.

---

Timeline diagram

timeline
    title Ownership of US 11085792
    2017 : PCT application filed by Didi
    2019 : US continuation filed by Didi
    2021 : Patent issued to Didi
         : Inventors assign rights to Didi

---

NPE / troll-pattern signals

1. Shell-entity transfer — NOT PRESENT.
No assignment to any entity bearing "IP / Patents / Licensing / Holdings / Ventures." The only recorded transfer runs toward the operating company (2001-style inventor take-back, recorded 2021-05-05), not away from it. Current assignee is an operating company with 8,000+ employees, not a single-member Delaware or Texas LLC. No registered-agent-service address appears.

2. Known asserter in the chain — NOT PRESENT.
The chain contains exactly two nodes: three individual inventors, and Beijing Didi Infinity Technology and Development Co., Ltd. None of the listed NPE families (Acacia, Marathon, IV, IPNav, Wi-LAN, Mosaid/Conversant, Vringo, Pendrell, Innovatio, MPHJ, Lumen View, Round Rock, DGC, Spangenberg entities) appears as assignee or assignor. I found no Unified Patents or RPX high-frequency-plaintiff record tying this patent or this assignee to assertion activity.

3. Repeat correspondent across the chain — NOT PRESENT / INDETERMINATE.
There is only one recorded instrument in the chain. A single appearance cannot establish recurrence by definition, and I could not retrieve the correspondent of record for that one instrument. No finding. (Contrast: the prosecution agents visible across DiDi's foreign filings — Shelston IP in Australia, Spruson & Ferguson in Australia, China Patent Agent (H.K.) Ltd. in Hong Kong — are ordinary operating-company foreign-associate firms, not NPE recording counsel. Note that these are prosecution agents of record, not assignment correspondents, so they should not be conflated with the field the task asks for.)

4. Cascading transfers — NOT PRESENT.
Zero consecutive assignments through chained LLCs. Zero transfers in any 24-month window. Nothing to cascade.

5. Pre-litigation transfer — NOT PRESENT.
No infringement suit naming US 11,085,792 surfaced in my searches, and no assignment dated within 6 months of any such suit exists. The single recorded assignment (2021-05-05, ~3 months pre-grant) predates the grant itself and is an internal title confirmation, not a standing-cleaning transfer to an asserter.

6. Bankruptcy fire-sale — NOT PRESENT.
No Chapter 7/11 of the assignee, no court-supervised patent sale, no 363 sale order in the record. DiDi Infinity Technology is an active operating entity; the parent-level NYSE delisting is a securities-market event, not an insolvency of the patent-holding entity.

7. Privateering — NOT PRESENT.
For privateering to exist, the operating company must transfer the patent to an NPE that asserts it against the operating company's competitors. No transfer out of DiDi exists at all, so there is no NPE vehicle to assert on DiDi's behalf. No Patent Progress / EFF / SEC coverage linking this patent to a proxy asserter was found.

8. Defensive aggregator — NOT PRESENT.
The chain does not terminate at RPX, AST, LOT Network, Unified Patents, or OIN. No neutralizing transfer occurred. The patent is simply still held by its original owner.

---

Verdict

Insufficient data — meaning, precisely, that the record contains only the original assignment (three inventors → Beijing Didi Infinity Technology and Development Co., Ltd., recorded 2021-05-05, reel/frame not retrievable from my sources), with no post-issuance transfers of any kind.

This is a clean negative finding, not a gap: an absence of assignment activity here is affirmative evidence that the original operating assignee never sold, securitized, or spun the patent out. All eight NPE signals score not-present, and there is no litigation, no shell entity, and no repeat NPE correspondent to point to. I would flag only that I could not independently pull the reel/frame, execution date, and correspondent from the Assignment Center — if you need those three fields nailed down for a docket, run the patent-number search directly at https://assignmentcenter.uspto.gov/ and cross-check the legal-events entry at https://patents.google.com/patent/US11085792/en.

Bottom line for portfolio tracking: this patent is not an NPE asset and should not be flagged in any NPE-screen. It is organic IP of an operating ride-hailing company, held by the same entity that filed the underlying PCT in 2017.

I'll research the prior art references cited for US 11085792. Let me search multiple sources.

Let me look for the specific reference citations on the patent's front page and its PCT search report.

Prior-Art Analysis — US 11085792 B2

⚠️ Threshold limitation (please read first)

I was able to confirm the requested patent and its claims, and I identified a body of highly relevant prior art from the same technical field and patent family. However, my tool session hit its step limit before I could retrieve the verbatim "(56) References Cited" front-page list printed on US 11085792 B2 itself. The reference lists I did retrieve come from:

• the International Search Reports of sibling Does—Didi applications in the same ETA patent family (PCT/CN2017/088089 and PCT/CN2017/086530), and
• forward citations ("Cited By") to the '792 patent.

I am flagging this explicitly rather than presenting family references as if I had transcribed them from the '792 face. Where a reference is a sibling patent's citation, I say so. I will not invent front-page citations I could not verify.

---

1. The patent at issue (confirmed)

Independent claims to map against (from US 2020/0096361 A1, same disclosure):

• Claim 1 (system): receive start location + destination; obtain a machine-learning model generated by obtaining historical data related to an on-demand service order → determining a high-dimensional sparse feature → determining a machine-learning model on that feature; determine ETA for a target route; transmit the ETA.
• Claim 11 (method): mirror of claim 1.
• Dependent claims: 2/12 (GBDT or one-hot encoding); 3–5/13 (leaf-node encoding / categorical-feature one-hot handling); 6 (FM or FFM); 7–9 (online-learning update: FTRL, OGD, SGD, FOBOS, RDA); 10 (basic route/vehicle/provider info).

The novelty hook of this patent is the high-dimensional sparse feature generated from historical order data (via GBDT leaf-node encoding and/or one-hot encoding) fed into a factorization machine / field-aware factorization machine, with online-learning updates.

---

2. Prior art references identified

Group A — References cited in the International Search Report of the closest sibling Didi ETA application (PCT/CN2017/086530, pub. EP 3479306 / GB 2564181)

These were flagged X/A/I (novelty-relevant) by the EPO/ISA and are the most technically on-point art in the family:

Group B — References cited in the ISR of a different Didi application (PCT/CN2017/088089 — "System and method for determining estimated arrival time," pub. JP 2019527871A)

All were cited as Category "A" (general state of the art, not anticipatory):

Because these were all "A" references in that case, they are background art and are not strong § 102 anticipation candidates for '792.

Group C — Other ETA art surfaced in search

---

3. Forward citations — NOT prior art (for completeness only)

Search results show "Cited By (29)" entries for '792, including US 2018/0137594 A1 (GT Gettaxi Limited), US 2022/0101273 A1, and US 2020/0160225 A1 (Didi). These are later documents citing '792 and therefore are not § 102 prior art against it. I mention them only to correct any impression that "cited" lists on Google Patents are prior art.

---

4. § 102 anticipation assessment — and its limits

Mapping the identified art to '792's claims:

• Claims 1 / 11 (independent): The preambles ("obtain historical data… determine an ETA… transmit to user device") are broadly taught by the travel-time-prediction art (Group A: US 6,317,686; TomTom; Horvitz). No reference I located, however, teaches the specific ordered combination of (i) determining a high-dimensional sparse feature from order data and (ii) building the ML model on that sparse feature — which is the core of claims 1/11.
• Claims 2 / 12 (GBDT or one-hot): No anticipation found; these are implementation-specific.
• Claims 3–5 / 13 (tree info → leaf-node encoding; categorical one-hot; combination): No anticipation found in the located art.
• Claim 6 (FM / FFM): No anticipation found — factorization machines are a distinct technical element this patent leans on.
• Claims 7–9 (online learning: FTRL/OGD/SGD/FOBOS/RDA): No anticipation found.
• Claim 10 (route/vehicle/provider info): Broadly evident in the field; § 102 alone unlikely, but relevant to § 103.

Conclusion: The prior art I could verify is relevant background/obviousness art, but I found no verified § 102 anticipation of the independent claims of US 11085792 B2. A definitive § 102/§ 103 opinion requires the actual front-page references (which I could not transcribe) plus claim charts.

---

5. What I could not do, and how to close the gap

1. Retrieve the verbatim (56) References Cited list from US 11085792 B2. This is the single most important missing piece for the task as framed and should be pulled directly from:
   • USPTO Patent Center / Patent Full-Text (patft) for patent number 11085792,
   • the FreePatentsOnline and Google Patents "References Cited" tables for US11085792B2, and
   • the PCT/CN2017/085375 International Search Report (the parent's cited art is the most likely source of the '792 face references).
2. Confirm dates/jurisdictions for each reference against INPADOC/Espacenet.

I hit my tool-step ceiling on the three follow-up queries intended to pull exactly that front-page list, so those did not complete. If you re-run this with a higher step budget (or supply the '792 PDF), I can produce a reference-by-reference § 102 chart keyed to the printed front-page citations rather than to family-level art.

Bottom line: I can confirm the patent, its claims, and the technically closest family-level prior art (US 6,317,686/Bin Ran; TomTom WO 2012/062760 and WO 2014/170434; Horvitz US 2008/004794; plus same-assignee siblings), but I cannot yet assert specific § 102 anticipations because I could not verify the exact references printed on the '792 front page. I will not fabricate that list.

I'll gather the prior-art landscape before building the § 103 analysis. Let me run the key searches.

Obviousness Analysis — US 11,085,792 B2 under 35 U.S.C. § 103

0. Two threshold disclosures before the analysis

(a) There is no "Prior Art section" in the material supplied to me. The sections handed off from earlier in this workflow are: Patent summary, Litigation summary, PTAB challenges, Strategic summary, and Recommended next steps. None of them contains a prior-art (references cited / IDS / search-report) section, and the authoritative Google Patents scrape is truncated before the "Claims" and "Referenced Cited" sections. I therefore could not "use the results from the Prior Art section of this page" — that input does not exist. Rather than stop, I have built the § 103 case from (i) the prior art the patent itself admits (its Background/Description framing of GBDT, one-hot encoding, FM/FFM and online learning as known techniques), (ii) the single cited reference visible in my retrieval (US 6,317,686 — the first entry of the "Referenced Cited → U.S. Patent Documents" list at https://patents.justia.com/patent/[11085792](/patent/11085792)), and (iii) independent searches for pre-2017-05-22 art. Everything below is flagged by confidence level.

(b) Contradiction with the earlier "Patent summary" section — resolved in favor of the search result. The Patent summary states: "I could not retrieve the verbatim claim set for US11085792B2… does not include the 'Claims' section." My search did surface substantial verbatim claim text from https://patents.justia.com/patent/11085792. Per the operating rule to prefer search results over prior working assumptions, I treat that claim text as authoritative and supersede the earlier "unverifiable claims" caveat. This materially changes the analysis: the independent claims contain limitations (notably the "distance information of a route" requirement in claim 11) that the earlier reconstruction from the Summary missed entirely.

---

1. Claims under analysis (as retrieved — partially verbatim)

From the Justia rendering of US 11,085,792:

• Claim 1 (system) — storage medium with instructions + at least one processor of an online on-demand service platform configured to: receive a start location and a destination from a user device via a network; obtain a machine learning model for determining an ETA, the model generated by a process including obtaining historical data related to an on-demand service order, determining a high-dimensional sparse feature based on the historical data, and determining a machine learning model based on the high-dimensional sparse feature; determine an ETA for a target route based on the model, the start location, and the destination; and transmit the determined ETA to the user device via the network.
• Claims 2–6 — not recovered. I will not guess at them. (Given the system/method parallelism, claims 2–6 are most likely directed to GBDT/one-hot encoding, leaf-node encoding, categorical-feature/attribution handling, and FM/FFM — i.e., mirroring method claims 12–16 — but that is inference, not evidence.)
• Claim 7 — update the machine learning model with an on-line learning method.
• Claim 8 — on-line learning method involves at least one of FTRL, OGD, SGD, FOBOS, or RDA.
• Claim 9 — updating includes: obtain recent historical data related to another on-demand service order; determine a recent high-dimensional sparse feature; update the model based on it.
• Claim 10 — historical data includes at least one of basic route information, vehicle information, or provider information.
• Claim 11 (method) — receiving (server, via network) a start location and a destination from a user device; obtaining a machine learning model generated by a process including obtaining historical data related to an on-demand service order, determining a high-dimensional sparse feature based on the historical data, wherein the high-dimensional sparse feature includes at least one feature corresponding to distance information of a route, and determining a machine learning model based on the high-dimensional sparse feature; determining an ETA for a target route based on the model, the start location, and the destination; transmitting the ETA to the user device.
• Claim 12 — determining the high-dimensional sparse feature involves at least one of a GBDT algorithm or a one-hot encoding algorithm.
• Claim 13 — obtain a feature from the historical data → determine tree information → determine a leaf node encoding → determine the high-dimensional sparse feature based on the leaf node encoding.
• Claim 14 — determine whether the feature is a categorical feature; if so, determine an attribution; determine a one-hot encoding based on the attribution; determine the high-dimensional sparse feature based on the one-hot encoding.
• Claim 15 — if the feature is not categorical: determine tree information → leaf node encoding → determine the high-dimensional sparse feature based on the leaf node encoding and the one-hot encoding.
• Claim 16 — determining the model involves at least one of FM or FFM.
• Claim 17–19 — parallel to claims 7–9 (server-side online updating).
• Claim 20 (CRM) — parallel to claim 1 in non-transitory medium form.

Effective filing date / applicable law. US 11,085,792 is a continuation of PCT/CN2017/085375 filed 2017-05-22. That is the § 102/§ 103 critical date. The AIA applies. Prior art must therefore be (i) publicly available before 2017-05-22, or (ii) a US patent/application publication "effectively filed" before 2017-05-22 under § 102(a)(2)/§ 102(d).

Claim construction notes that matter. "High-dimensional sparse feature" is a coined term; the specification defines it functionally as "a mathematical expression (e.g., a vector or a matrix) to describe characteristics of a route as a whole," potentially with "hundreds… thousands… millions of features." It is not a term with an accepted art meaning, so it must be construed in light of the spec — i.e., as a sparse vector/matrix encoding of route features. That construction is what makes the He/FM art so dangerous to the patent.

---

2. Level of ordinary skill in the art (PHOSITA)

A person with at least a bachelor's degree in computer science, electrical engineering, or a related quantitative field, plus 2–3 years of experience (or a master's degree) building and deploying supervised machine-learning models for prediction tasks, including experience with (i) decision-tree ensembles such as GBDT/XGBoost, (ii) feature encoding of categorical and continuous data (one-hot / 1-of-K), and (iii) large-scale sparse linear or latent-factor models. By May 2017 this profile was common in the ride-hailing, mapping, ad-tech, and recommendation industries. Significantly, the patent's own specification confirms this skill level — it treats GBDT, one-hot encoding, FM, FFM, FTRL, OGD, SGD, FOBOS, and RDA as off-the-shelf techniques requiring no explanation of their internals, only of how they are strung together.

---

3. Prior art identified (with § 102 status)

Not prior art (flag to avoid conflation): US 11,580,451 B2 (Didi sibling, priority 2017-07-27 — after 2017-05-22); US 10,948,301 B2; US 10,816,352 B2; EP 3461310; GB 2564181. These post-date the critical date and cannot be used.

---

4. Grounds of rejection

Ground 1 — Tencent + He 2014 → claims 1, 11, 12, 13, 20

Claim 12 (GBDT and/or one-hot encoding) — He 2014 discloses GBDT and 1-of-K (one-hot) coding, literally, in the same sentence.
Claim 13 (feature → tree information → leaf node encoding → sparse feature) — He 2014's leaf-index encoding is a verbatim structural match.
Claim 1 / Claim 20 — Tencent's server 400 (CPU 401, ROM/RAM, mass storage) and Tencent's own boilerplate disclosure of "a non-transitory computer readable storage medium including an instruction … a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, or an optical data storage device" supply the storage-medium/processor wrapper limitations.

Motivation to combine. Tencent supplies the application (ride-hailing/route travel-time prediction) and the shortcut (model rather than segment-by-segment accumulation). He 2014 supplies the feature-engineering technique that solves the problem Tencent leaves open — how to convert a heterogeneous route feature vector (mix of continuous distance/speed features and categorical road-type/vehicle-type features) into a representation a predictive model can consume, and how to capture non-linear and cross-feature effects automatically instead of by hand. He 2014 itself frames the technique as general-purpose ("supervised feature encoding"), which makes it reasonably pertinent to a travel-time modeler's problem. The result — a more accurate ETA — is exactly the predictable improvement both references seek (Tencent: "more accurate travel time prediction will improve order-assignment accuracy and increase the order conversion rate").

---

Ground 2 — Ground 1 + Rendle 2010 and/or Juan 2016 → claims 16 (and the FM/FFM reading of the model step)

Rendle: FM factorizes second-order parameters so that "even with very sparse data" meaningful predictions are possible via shared latent vectors. Juan: FFM, one-hot inputs, field grouping, "FM can be viewed as a special case of FFM when all features belong to one field."

Why the combination. The sparse binary vector produced by He's GBDT 1-of-K encoding is precisely the data regime FM and FFM were invented for. Juan 2016 expressly diagnoses the problem ("one-hot encoding … causes serious data sparsity") and offers FM/FFM as the fix; He/GBDT encodings are high-dimensional binary vectors of the same character. Using a factorizing model in place of He's linear classifier is a substitution of one known sparse-data model for another, with a predictable result (better handling of under-sampled feature combinations — exactly what route-section interactions are). Where the only distinction between FM and FFM is the field grouping, KSR's "finite number of identified, predictable solutions" rationale applies directly: FFM is FM plus the field concept, and Juan tells you which to pick based on whether the sparse feature's items span more than one field — which the patent's own specification copies verbatim ("if feature items of the high-dimensional sparse feature belong to more than one field, the processing engine 112 may determine FFM as the machine learning model").

---

Ground 3 — Tencent + Avazu/Youdao 2015 (+ He 2014) → claims 11, 12, 16

This is the strongest ground because it collapses the combination into near-single-reference teaching. The Youdao/Avazu submission:

• takes numerical features into a gradient boosting tree model;
• uses the tree indices as output features (19 trees × depth 5 → 19 new features);
• feeds those GBDT features into both FTRL and FFM.

That is the patent's FIG. 5/FIG. 6 pipeline (GBDT leaf encoding → sparse feature → FM/FFM) plus the claim-8/18 FTRL online-learning limitation, in one 2015 document. Pair it with Tencent for the travel-time/route domain and the "receive a start location and destination → transmit ETA" hardware/network framing. The nexus to ETA is supplied by Tencent; the technique library is supplied in toto by the CTR literature the patent itself treats as the wellspring of its methods.

---

Ground 4 — Ground 1 + Juan 2016 / Rendle 2010 / He 2014 → claims 14 and 15

Claim 14 (categorical → attribution → one-hot → sparse feature). Juan 2016 states one-hot encoding is the standard input encoding for categorical variables and that "the N features [generated by one-hot encoding a categorical attribute] are put in the same field — features describing the same attribute." The patent's "attribution" step (number of mutually exclusive possible states → vector length) is nothing more than sizing the one-hot vector by the cardinality of the categorical variable — an inherent, definitional property of one-hot encoding, not an inventive contribution (cf. In re Kubin / Pfaff — reciting the known technique plus the result obtained is not inventive).

Claim 15 (non-categorical feature → tree info → leaf node encoding and one-hot encoding). He 2014 discloses exactly the hybrid: the boosted-tree leaf features are concatenated with the other (categorical, one-hot) features into the single input vector for the linear classifier. The claim is a verbatim description of He's Figure-1 hybrid input.

---

Ground 5 — Ground 1 + McMahan 2013 / Duchi 2010 / Xiao 2010 / Zinkevich 2003 (+ Tencent's retraining) → claims 7, 8, 9, 17, 18, 19

• "Update the machine learning model with an on-line learning method" (7, 17): He 2014 describes continuous/online training of the hybrid model; Tencent describes adjusting the training sample set and retraining the travel-time model when the quality parameter fails the precision threshold.
• The enumerated algorithms (8, 18 — FTRL, OGD, SGD, FOBOS, RDA): each is a named, published online-learning method from 2003–2013, and the patent's own specification lists them without elaboration — an admission of their conventionality. FTRL in particular is already tied to GBDT-derived features by the Avazu reference.
• "Obtain recent historical data related to another on-demand service order; determine a recent high-dimensional sparse feature; update the model based on it" (9, 19): this is the definition of incremental/streaming online learning applied to the same feature pipeline; nothing more than applying the known online-update technique to the known feature-generation step. Tencent's "adjust the training sample set" and its express design goal of prediction accuracy under real-time traffic at the current moment supply the motivation (concept drift in traffic conditions).

---

Ground 6 — claim 10

"Basic route information, vehicle information, or provider information" as the historical data content: Tencent discloses route information (lengths, road classes, traffic-light density, speeds, departure/arrival times); He 2014 discloses user/context features; the patent's own Background and Description classify these as conventional. Design-choice breadth, not a patentable distinction.

---

Ground 7 — the "on-demand service order" wrapper

If an examiner or litigant prefers an explicit on-demand-order-teaching reference in addition to Tencent, US 2020/0042885 A1 names the on-demand service/platform context directly — but see the § 103(c) caveat above. Alternatively, the phrase "on-demand service order" is a label for a completed service-request record; Tencent's historical travel data, keyed to trips and applied to order assignment, is functionally the same data.

---

5. Motivation-to-combine synthesis (KSR / Graham framework)

Articulated rationales, each tied to record evidence:

1. Known technique applied to a known device ready for improvement, yielding predictable results. (KSR prong; Pacific Bell v. AT&T.) The known ETA-modeling device (Tencent) was plainly "ready for improvement": Tencent itself concedes accuracy is the goal and that its model uses only route-level features. Applying GBDT-based feature encoding (He) and a factorizing sparse model (Rendle/Juan) is the predictable next step.
2. Combination of known elements according to known methods. Both references are in the same field of endeavor (large-scale supervised prediction over transactional/behavioral feature logs) and, at a minimum, the CTR/feature-engineering references are reasonably pertinent to the problem the inventor faced — how to encode hundreds-to-millions of route features into a sparse representation for a predictive model. That is the In re Bigio / In re Klein test, and it is satisfied.
3. Simple substitution of one known sparse-data model for another. FM/FFM for the linear classifier; FFM for FM where fields exist. Predictable improvement in handling under-observed feature combinations.
4. Design incentives / market forces. The 2016–2017 ride-hailing industry competed on ETA accuracy; Tencent's own text states improved travel-time prediction "will improve order-assignment accuracy and increase the order conversion rate." That is a concrete, documented business driver.
5. "Obvious to try" with a finite set of identified solutions. For claims 8 and 18 the claim literally enumerates five known online learners; for claim 16 it enumerates two related factorizing models. Both are closed sets with predictable selection criteria (number of fields; data volume; sparsity).
6. Technology-is-the-same rationale. The Youdao/Avazu submission shows the industry already combining GBDT features with both FFM and FTRL — i.e., the patent's own two "novel" building blocks were already used together, in a document published two years before the critical date.

---

6. Anticipated counterarguments (and why they should fail)

---

7. Honest limits on this analysis

1. Claims 2–6 are not in my retrieved text. Any ground aimed at those claims (most plausibly the GBDT/one-hot/leaf-node/FM-FFM family, mirroring 12–16) is provisional until the full claim set is pulled from USPTO PatentCenter for application 16/684,693 or the granted-PDF claim column.
2. The complete "Referenced Cited" list is not in my source — I recovered only US 6,317,686. The applicant's own IDS may contain art that is materially better or worse than the references I have identified; a proper analysis must start from that list plus the PCT/CN2017/085375 international search report and written opinion, which I have not seen.
3. The Tencent US filing particulars are reconstructed from the justia/Google snippets ("filed on … 12, 2017, which claims priority to Chinese Patent Application No. 201610893915.9 … filed on Oct. 13, 2016"). The 2016-10-13 effective date under § 102(d)(2) is what makes Tencent § 102(a)(2) art; verify the exact PCT/national-stage filing date before relying on it. Note the sources differ slightly on the CN filing date (2016-10-12 vs. 2016-10-13) — the discrepancy does not affect the outcome.
4. US 2020/0042885 A1 is probably unusable in a § 103 combination because of common ownership with the patentee (§ 103(c)(1)). Anyone building this case should not lean on it.
5. US 2016/0202074 A1 came from a secondary aggregator, not a primary source I could verify in this session. Treat as corroborative, not load-bearing.
6. I have not run a prior-art search against the acronym-dense claim language in the Chinese-language literature (CNIPA, Baidu/Didi technical papers), where the closest § 102(a)(1) art against a Beijing-filed 2017 application is most likely to live.

---

8. Bottom line

The independent claims (1, 11, 20) are highly vulnerable to § 103 over Tencent US 2019/0204102 (or US 11,143,522) in view of He 2014, with the strongest single-combination variant adding Avazu/Youdao 2015 or Juan 2016 / Rendle 2010. Every substantive limitation has a clean, documented home: route-based ETA with a historical-data-trained model (Tencent); GBDT leaf-node 1-of-K encoding producing a high-dimensional sparse binary vector, combined with one-hot features (He); FM/FFM as the sparse-data classifier (Rendle, Juan, Avazu); and FTRL/online updating of that exact feature pipeline (Avazu, McMahan). The claimed "advance" is, on the record before me, an assembly of known techniques with a predictable benefit, and the patent's own specification supplies the admissions that make the § 103 case easier rather than harder. The main unknowns that could change this assessment are claims 2–6 and the applicant's own cited-reference list — neither of which I have.