AutoCDSR: Automation in C-V2X and Recommendations

• AutoCDSR is a dual-purpose automation technique enabling optimized candidate CSR allocation in C-V2X Mode 4 and automated cross-domain knowledge transfer in sequential recommendation.
• It employs distributed algorithms and Markov chain models to assign candidate single-subframe resources dynamically, minimizing delays and collision probabilities in vehicular networks.
• Empirical results demonstrate substantial performance gains, with delay reductions of up to 80.7% in C-V2X and recommendation accuracy improvements of up to 16.7% in cross-domain sequential recommendation.

AutoCDSR denotes an automated approach for optimizing candidate single-subframe resource allocation in cellular vehicle-to-everything (C-V2X) Mode 4 and, independently, an adaptive method for automating cross-domain knowledge transfer in sequential recommendation. While these two domains share automation and self-configuration as underlying principles, their technical approaches and implementation contexts are distinct, covering distributed vehicular communication systems and recommender systems, respectively.

1. Automated Candidate CSR Allocation in C-V2X Mode 4

In C-V2X Mode 4, AutoCDSR refers to systems that dynamically determine the optimal number of candidate single-subframe resources (CSRs) per vehicle and assign multiple prioritized data streams across these resources for parallel transmission, directly enhancing medium access control (MAC) performance. Candidate CSR allocation is inversely proportional to vehicle density (N): as N increases, the number of CSRs per vehicle ($n_\mathrm{CSR}$) must decrease to maintain system capacity. The number of vehicles supported is described by $$N_\mathrm{max} = 0.8 \times CSR_\mathrm{tot} / n_\mathrm{CSR}$$, where $CSR_\mathrm{tot}$ is the total number of available CSRs in a selection window ($\Gamma$), accordance with 3GPP standards.

$n_\mathrm{CSR}$	$\Gamma$ = 20ms	$\Gamma$ = 50ms	$\Gamma$ = 100ms
1	400	1000	2000
2	200	500	1000
3	134	334	667
4	100	250	500

AutoCDSR systems utilize an algorithmic ruleset to select $(n_\mathrm{CSR}, \Gamma)$ based on vehicle count, minimizing packet delays for lower-priority traffic without degrading collision probability (maximum observed increase: 0.6% at $N=99$) (A et al., 2022).

2. Optimal Assignment and Grouping of Data Streams

Vehicles often transmit up to four prioritized streams: HPD, DENM, CAM, and MHD. AutoCDSR partitions these streams across multiple CSRs, selecting the grouping that minimizes the weighted sum of expected average delays: $\Delta D_{n_G} = \sum_{l} w_l D_{n_G, l}$, where $w_l$ is the priority weight, and $D_{n_G, l}$ the DTMC-modeled delay for stream $l$ using grouping $n_G$. The algorithm enumerates all stream-to-CSR groupings (for $n_\mathrm{CSR}=2,3$ per Stirling numbers), computes delays, and chooses the group with minimal $\Delta D_{n_G}$. This automatically mitigates indefinite preemption of low-priority streams, a flaw in the single-CSR paradigm (A et al., 2022).

3. Performance Impact and Implementation Considerations

Empirical results demonstrate substantial delay reductions for lower-priority streams (e.g., CAM up to 77.4%, MHD up to 80.7%), with negligible gained packet collision rates and improved channel utilization. AutoCDSR is lightweight, employing simple lookups and Markov chain analyses, and is well suited to real-time implementation, requiring modifications to semi-persistent scheduling (SPS) and potentially demanding parallel PHY for simultaneous transmissions. This suggests practical deployability in future distributed, multi-service vehicular networks (A et al., 2022).

4. AutoCDSR in Automated Certification Workflows

The term AutoCDSR also refers to automated continuous design and safety review workflows in safety-critical software certification (Shankar et al., 2022), notably in projects such as DesCert (Design for Certification) within the DARPA ARCOS program. Here, AutoCDSR encompasses continuous, automated assurance-driven development—integrating requirements capture (via CLEAR), architectural decomposition (RADL), assertional contracts, and evidence generation using formal tools (Sally, PVS, Text2Test, Randoop, Daikon, SeaHorn). All evidence is structured and curated within an assurance ontology and the Rapid Assurance Curation Kit (RACK), supporting traceability, composability, and continuous certification.

Key formulae express compositional rigor:

$$\text{ArchitectureProperties} \land \text{ComponentContracts} \implies \text{ControllerOutput}$$

The workflow thereby supports automation, formal rigor, traceable evidence integration, and efficient assurance for complex systems, aligning entirely with AutoCDSR objectives (Shankar et al., 2022).

5. AutoCDSR in Cross-domain Sequential Recommendation

In the context of cross-domain sequential recommendation (CDSR), AutoCDSR is a self-configuring method to automate knowledge transfer between behavioral sequences across domains (Ju et al., 27 May 2025). The approach constructs a multi-objective optimization to simultaneously maximize recommendation accuracy and minimize harmful cross-domain self-attention. The total cross-domain attention is:

$$a_{\text{cd}} = \sum_{i=1}^M \sum_{j=1}^M \text{softmax}(\mathbf{A})_{i,j} \cdot \mathbb{I}(d(x_i) \neq d(x_j))$$

The training objective is:

$$\min_{\boldsymbol{\theta}} \big( \mathcal{L}_\text{rec}(\boldsymbol{\theta}), \mathcal{L}_\text{cd-attn}(\boldsymbol{\theta}) \big)$$

Update directions use the Multiple Gradient Descent Algorithm (MGDA) constrained to Pareto-optimal tradeoffs, with a preference for recommendation performance. This fully automates regularization per mini-batch, mitigating negative transfer and enabling adaptive knowledge exchange.

AutoCDSR$^+$ uses Information Bottleneck (IB) tokens to further structure cross-domain transfer, confining interaction only through these tokens, providing additional control. Empirical results show Recall@10 improvements of 9.8% (SASRec) and 16.0% (Bert4Rec), NDCG@10 improvements up to 16.7%, with computational overheads far lower than state-of-the-art alternatives (Ju et al., 27 May 2025).

Model	Recall@10	NDCG@10	Speed (iter/s)	Overhead vs base	SOTA Perf. Match?
SASRec (base)	0.291–0.514	0.203–0.514	10.27	0%	No
Bert4Rec (base)	0.264–0.483	0.181–0.501	10.27	0%	No
+AutoCDSR	+9.8–16%	+12–16.7%	9.31	+9.3%	Yes
+AutoCDSR$^+$	further	further	8.22	+19.9%	Yes/Better

6. Common Themes and Plausible Implications

AutoCDSR implementations in both vehicular communications and data-driven recommender systems share a focus on dynamic self-configuration, performance enhancement, and mitigation of systemic inefficiencies (e.g., stale packets, negative transfer). A plausible implication is that similar automation and multi-objective optimization principles will permeate other domains where resource allocation, assurance, and adaptive transfer are critical. Both contexts rely on formal models to guide configuration (DTMC for networking, multi-objective optimization for neural attention), and both yield measurable gains in latency, throughput, and robustness. Future research may generalize AutoCDSR automation strategies for broader application in multi-resource, multi-priority, and cross-domain environments.

7. References to Key Papers and Implementations

• Performance Enhancement of C-V2X Mode 4 Utilizing Multiple Candidate Single-subframe Resources (A et al., 2022)
• DesCert: Design for Certification (Shankar et al., 2022)
• Revisiting Self-attention for Cross-domain Sequential Recommendation (Ju et al., 27 May 2025)

Source code and further implementation details regarding the sequential recommendation AutoCDSR method are publicly available at https://github.com/snap-research/AutoCDSR.