Chapter 3: Deep Learning RecSys

Recommendation AI — WIA-AI-RECO-001

Wide and Deep per Heng-Tze Cheng + Levent Koc + Jeremiah Harmsen + Tal Shaked + Tushar Chandra + Hrishi Aradhye 2016, DeepFM, DLRM per Maxim Naumov, Two-Tower architecture, Transformer4Rec.

This chapter elaborates on the Deep Learning RecSys portion of Recommendation AI (WIA-AI-RECO-001). The recommendation ai canonical envelope shape, API surface, and protocol exchanges are designed to interoperate across implementations published under the WIA Standards family. Each Phase file in `spec/` documents one architectural layer (envelopes, REST surface, federation, ecosystem) and the present chapter focuses on the deep learning recsys subset of that layering.

Implementations of deep learning recsys need to reconcile the per-standard expectations of ISO/IEC 22989:2022 (AI concepts), ISO/IEC 23053:2022 (AI ML framework), ISO/IEC 24029-1/2 (Robustness), ISO/IEC 42001:2023 (AI management system), EU AI Act (Reg 2024/1689). The Phase 1 envelope class for Recommendation AI carries the per-host identifier and per-tenant identifier so that audit records can be reconstructed across hosts that federate through the Phase 3 protocol layer.

Operators stepping into deep learning recsys typically follow this trajectory: stand up the reference simulator at `simulator/index.html`, walk the canonical envelope per the CLI helper at `cli/recommendation-ai.sh envelope`, then exercise the Phase 2 endpoints with curl/Postman against a host running the reference container at `wia/recommendation-ai-host:1.0.0`. Once those steps complete, integrators wire the real backend one endpoint at a time per the Phase 1 §A.* envelope catalogue and the Phase 2 §B.* endpoint catalogue.

A common operational concern in the deep learning recsys space is observability. Every Phase 1 envelope SHOULD emit a structured log line at the host's audit transport with: timestamp per RFC 3339, host identifier, tenant identifier, envelope class, envelope identifier, operation outcome, and a W3C Trace Context `traceparent` propagated end-to-end. Phase 2 surfaces this trace identifier as the `X-WIA-Trace-Id` response header. Phase 3 protocol exchanges propagate the trace identifier inside the exchange envelope so that a federation crossing remains correlatable end-to-end.

The deep learning recsys portion of the standard composes with: WIA-OMNI-API for credential storage, WIA-AIR-SHIELD for runtime trust list, WIA-INTENT for workload intent declaration, and WIA Secure Enclave for sealed-data envelopes where personal data is processed. The composition lets one host running multiple WIA family standards reuse one identity, one signing-key chain, and one audit transport rather than maintaining N parallel per-standard implementations.

Normative references touched in this chapter

Implementation worksheet

  1. Read the corresponding Phase file in spec/ (Phase 1 for envelopes, Phase 2 for API, Phase 3 for protocol, Phase 4 for ecosystem).
  2. Walk the CLI helper: ./cli/recommendation-ai.sh envelope emits a sample Phase 1 envelope.
  3. Exercise the simulator at simulator/index.html.
  4. Cross-reference the standards listed above to align with the operator's per-jurisdiction obligations.
  5. Wire the conformance suite at https://github.com/WIA-Official/wia-recommendation-ai-conformance.

Cross-standard composition recap

This chapter, like every other Phase 1-4 chapter in the Recommendation AI eBook, composes with the wider WIA Standards family. Implementations that adopt the canonical envelope reuse the cross-standard audit transport (W3C Trace Context plus OpenTelemetry semantic conventions), the cross-standard identity (WIA-OMNI-API), and the cross-standard runtime trust list (WIA-AIR-SHIELD) without per-standard re-implementation. The federation handshake described in Phase 3 §3 follows the same challenge-response pattern as the rest of the WIA Standards family so that one operator can federate across multiple standards using one signing-key chain and one audit transport.