Unveiling User Satisfaction and Creator Productivity Trade-Offs in Recommendation Platforms

We greatly appreciate the reviewer's suggestion for additional justification of our problem setting. Below, we respond to the questions raised.

Evidence of volume competition Major UGC platforms, such as YouTube, TikTok, and Instagram, primarily generate profits through ad impressions, which are directly correlated with content impressions. Given this revenue model, it is logical that platform incentives are designed to encourage creators to compete for more impressions. By fostering an environment that motivates creators to increase their content visibility, platforms can maximize their ad revenue. This setup inherently triggers Cournot competition-like dynamics in UGC platforms, where creators compete for audience attention by producing more content.

Our work is done in collaboration with a world-leading content recommendation and sharing platform with billions of users. We do observe Cournot competition-like dynamics from offline data analysis, and we will include such evidence in the revision once authorized by the company to disclose this information. Generally, we observed two phenomena:

• Content embeddings from the same creators are clustered around a small region, indicating that most creators consistently produce content within the same topic.
• Creator productivity is significantly influenced by the platform’s incentivizing plan; when the platform allocates more user traffic to a group of creators, the average production frequency of this group increases.

These observations motivate the formulation of the $C^4$ framework.

How our model handles heterogeneity We do allow for heterogeneity in budgets, as the cost function can be creator-dependent. For the discussion of heterogeneity in terms of content creation, please refer to our common response above.

How general is our model to different platforms Our $C^4$ framework is applicable to mature content recommendation and sharing platforms where creators are aware of their expertise, consistently produce within their niche, and strategically adjust their production frequency to balance gain and cost. As discussed in lines 36-41, many existing leading platforms, such as YouTube and Instagram, exhibit these characteristics. On these platforms, most influential creators have clear and consistent branding and target specific user interest groups. Consequently, the platform’s recommendation strategy primarily impacts their willingness to engage, i.e., their content production frequency. However, our framework is less applicable to platforms like Netflix, where a different platform-as-creator dynamics prevail. The dynamics under such an environment is also interesting but is beyond the scope of our work. We have clarified this scope in the first paragraph of the introduction (see line 18), where we exclude non-UGC platforms like Netflix from our discussion.

We thank the reviewer for pointing out the concerns and raising the clarification questions and we respond to them below accordingly.

Response to weakness 1: unit user traffic First, we have to clarify that our model does not assume each user contributes only one unit of traffic. Since we do not impose any assumptions on the user distribution, we can easily represent active users by incorporating multiple instances of such a user in the formulation of creator utility function. For concerns regarding the assumption of homogeneous creation, please refer to our common response above.

Response to weakness 2 and question 1: computational efficiency Thank you for these great suggestions. We discussed computational efficiency in detail in Section 6.2, providing a computational complexity upper bound of $O(n^2m + nm^2 + m^3)$, where one can cluster users and creators to share parameters within the same cluster ($n$ and $m$ represent the number of clusters for creators and users, respectively). By appropriately setting $n$ and $m$, we can tune the resolution of clustering and balance between the accuracy and efficiency of the optimization process described in Algorithm 1. Moreover, this optimization process does not need to be run frequently (e.g., on a daily basis), as every time Algorithm 1 identifies the optimal treatment for each user group based on the current user and creator preference distribution. It is only necessary to rerun the algorithm if there is a significant change in the global user or creator preference distribution. Therefore, the intervention strategy computed by Algorithm 1 does have strong potential for practical applications.

Response to weakness 3 and question 2: Difference from $C^3$ game We highlighted the main differences between $C^4$ and $C^3$ in lines 32-41. First of all, we clarify that $C^4$ is NOT an extension of $C^3$, despite the similarity between their names. In fact, these two models correspond to entirely different competitive environments and assumptions about creator behavior. In $C^3$, creators have fixed creation frequencies and their strategy is to choose topics for producing content. In contrast, in $C^4$, creators' content has a stationary topic distribution, and their strategy is to determine the creation frequency. These two settings represent different stages of online content sharing platforms: $C^3$ models the early stage when creators are new to the platform and explore various possibilities to find their identity, while $C^4$ models a mature stage where creators are aware of their expertise and consistently produce within their niche, strategically adjusting their production frequency to balance gain and cost. Therefore, the formulation of $C^4$ addresses a very different practical problem and should be considered a complementary setting to $C^3$, and the results obtained in the previous $C^3$ setting do not provide any analytical insight for $C^4$.

Response to limitation 1: experiment scale As explained in Section 6.2, the number of users is not the bottleneck for the scalability of our algorithm because we can cluster the user population and update $\beta$ at the group level (this is also a practical treatment in large-scale real-world systems). Therefore, we believe that our experiment with 1,000 users is representative and should be sufficient for larger-scale experiments.

Dear Reviewer 1ynk, as the discussion phase is coming to a close, we wanted to gently remind you to review our response and consider reevaluating our paper in light of the additional information provided. If you have any further suggestions or concerns, we would be more than happy to engage in further discussion to improve our work.

We thank the reviewer for appreciating the insights of our problem setting and results. For our response to the suggested main weakness and question 2, please refer to our common response above. Below, we address the remaining questions.

Why total content volume affects long term user engagement It is not a surprise that long-term user engagement on a platform is crucially dependent on the total content creation volume due to the following reasons.

• First, Regular and abundant content updates provide users with a better experience of personalization, encouraging them to return to the platform more often or new users to join the platform, which increases engagement time and user satisfaction, and thus long-term success of the platform.

• Even though active creators and users are usually large in number on leading UGC platforms, the creator-to-user ratio is typically small. For example, TikTok has billions of viewers but only around 1 million active creators. As a result, most users are highly loyal to specific creators and eagerly consume whatever they create. This user stickiness means that even a small increase in content creation frequency by these creators can significantly boost total user engagement. Therefore, while minor fluctuations in overall content creation frequency might not drastically affect user churn rates, a consistent decrease in content creation from active creators will have a significant negative impact on user engagement.

In addition, we acknowledge that emphasizing the contrast between short-term and long-term goals might be a bit misleading. Our key point is to highlight the tradeoff between recommendation relevance and overall content creation volume in optimizing user engagement. Both factors are clearly important to the engagement target but we show that they are inherently conflicting when considering content creators' strategic responses. We will clarify this point in revision.

Alternative model that accounts for volume in utility This is a very constructive and interesting idea, however, to justify such a user utility model that correlates with the volume of available content on the platform (even if they do not consume them), we need to introduce an ad-hoc user behavioral assumption describing how users' satisfaction or return to the platform relates to the amount of available content. And oftentimes, ordinary users do not even know the amount of available content on a particular platform (think about YouTube and Instagram), although a positive correlation might empirically exist. We recognize this as an intriguing direction and intend to investigate it in future work.

Collusion among creators Indeed, many creators today are not working as individuals but as representatives of a team (e.g., belonging to the same studio or society). In our model, a player can be regarded as a team rather than an individual for such situations. However, if the reviewer is referring to collusion among different players to game the system, this is beyond the scope of our current framework and we might need a new model for studying such an intriguing phenomenon.