M3C: a Multi-Domain Multi-Objective, Mixed-Modality Framework for Cost-Effective, Industry Scale Recommendation

1. The novelty is somewhat limited. It primarily focuses on addressing specific engineering challenges within a particular domain by integrating multiple existing models and techniques. This work appears more appropriate for submission to the industrial track of a data mining conference such as SIGIR or KDD.
2. The novel aspects of this framework primarily reside in the design of the Zipper, Filter, and Sketch modules. However, the explanations and illustrations regarding these components are sometimes unclear. I found it challenging to grasp certain details of the Zipper algorithm, such as the role of "workers" and the rationale behind extracting the set of features Fe. Figure 3 was also confusing to me. Similarly, understanding the pseudocode in Algorithm 1 was difficult. Furthermore, Section 3.3.2 glosses over many details of the M3R Sketch, while Appendix A2.2 remains difficult to comprehend and follow. The authors don’t disclose plans to release the code and data, making it more difficult to understand the details and reproduce the model.
3. A recent published work, “M3oE: Multi-Domain Multi-Task Mixture-of Experts Recommendation Framework”, shares quite similar theme and technical approach. While it is acknowledged in the related works, further detailed discussion and comparison appear necessary.

W1: The structure and writing of this paper require significant improvement. The absence of a framework overview makes it difficult to understand the overall structure of M3C, hindering the ability to follow each component's function throughout the recommendation system pipeline.

W2: The paper does not adequately address the challenges outlined in Section 2. Issues such as the ‘non-overlapping ID spaces’ and the ‘optimization-validation cycle’ are not explicitly discussed.

W3: In this paper, does the author acknowledge that "MIXED-MODALITY" refers to various types of features, such as sparse, dense, and sequential features, rather than the commonly recognized inputs like images, videos, or audio?

W4: Some important details are either too brief or missing, such as the Knowledge Distillation of M3C Networks and RoPE embedding.

1. The proposed M3C consists of many components but without ablation studies, making it hard to judge the benefits each componenet brings.
2. The choices of design and strategies are heavily depended on user scenarios. Readers without the corresponding background can hardly understand the paper. E.g., I doubt few academia folks have learnt the term "attribution window".

• The introduction is well-written and engaging; however, the description of the framework lacks depth in certain areas. For instance, the domain-objective pairs partitioner and the mechanism for grouping these pairs are not clearly articulated. This omission makes it challenging to assess the flexibility of the proposed method and to understand how it balances commonalities and distinctions among pairs. The same problem applies to ECS. It is highly recommended that the authors provide detailed mathematical formulations or pseudo-code for this section.

• There seems to be an omission regarding the multi-domain setting. It would be helpful to know on which domains the proposed method learns and effectively transfers knowledge.

• The choice of baselines, except for Wukong, are somewhat outdated. It would be beneficial to include comparisons with more recent state-of-the-art multi-domain and multi-task baselines to position the proposed method.

• Considering that Wukong is utilized as component of the model, there seems trivial performance gap between the proposed method and Wukong in Table 1, even the proposed method is tuned by M3C Sketch. It deserves further ablation study to evaluate the contribution of each component.