How should video platforms spot their next big stars

Amid the crowd of millions of content creators, digital platforms need to pick a few to support and promote, and the traditional selection process is not sophisticated enough

Featured faculty: Cao Xinyu and Wang Jingbo
Written by Putro Harnowo

Many may call it cliché or cheesy when watching a short movie about a bullied employee who turns out to be the CEO’s son or a poor husband revealing himself as a billionaire, but believe it or not, these microdramas have stolen the show. Their plot twists and cliff hangers have helped them find a way to Hollywood, as Fox Entertainment and Disney recently announced investments in producing bite-sized movies.

Microdramas have less than two-minute duration and a vertical format because they are meant to be watched on a smartphone. Their popularity can be traced back to 2018, when Chinese short-video platforms began featuring them. Fast forward to 2024, the country’s microdrama industry surpassed its box-office revenue with US$6.9 billion, according to the China Netcasting Services Association.

Nowadays, China’s microdrama industry has become highly competitive, powered by more than 100,000 enterprises producing around 3,000 series monthly. Video platforms recognise these miniseries as profitable content and not only host but also actively incentivise them through financial rewards, algorithm optimisation, and the like, to encourage more original and high-quality content.

However, the ever-increasing number of creators and limited financial resources have made selecting creators for such incentives more challenging. After analysing data from a major video platform, Wang Jingbo, Assistant Professor in the Department of Marketing, and Cao Xinyu, Vice-Chancellor Associate Professor of Marketing at the Chinese University of Hong Kong (CUHK) Business School, find that creators with a large number of followers tend to be selected for financial support.

This method may be counterproductive since the supports might not provide a significant boost to their already high performance. “The platform may simply select creators who already have the highest content quantity and quality, rather than those who were likely to experience the greatest improvement due to the incentives,” Professor Cao says.

Therefore, in a study titled A Deep-DiD method to estimate heterogeneous treatment effects: Application to content creator selection, Professors Cao and Wang, as well as Cheng Yan of Shanghai University of Finance and Economics, Shen Zuo-Jun (Max) of the University of Hong Kong, and independent researcher Zhang Yuhui, propose a new method to optimise a digital platform’s selection process.

Practical experiment and application in the digital world

The short video platform launched a signing programme in three countries in 2022. Selected creators were asked to sign a contract and receive monthly payments based on their performance.

Professor Cao and her team then try to measure how effective the signing programme was, using three key indicators: the number of video uploads per day, the time users spent on each video, and user engagement from likes, comments, shares, and follows. To create a fair comparison, they match 2,343 creators who signed the programme with the same number of other creators who did not sign up based on their performance trajectories.

Through a statistical method called difference-in-differences (DiD), which has been widely used in economics and social sciences to estimate the effect of specific interventions, the researchers find that the signed group shows a significant boost in all key indicators. After signing up, their average number of videos per day temporarily increases, while the positive impacts on user time and engagement last longer.

However, the researchers also find that the effects vary widely across different creators. “The DiD method calculates the overall average impact by blending all the individual effects,” says Professor Cao. “If an intervention affects different individuals in several ways, or if the impact changes over time, the analyses could lead to an inaccurate picture of the actual impact.”

“The platform may simply select creators who already have the highest content quantity and quality, rather than those who were likely to experience the greatest improvement due to the incentives.”

Professor Cao Xinyu

Therefore, Professor Cao and the team further examine the specific impact on each creator with an advanced method by leveraging a computer programme called a deep neural network. Dubbed as the Deep-DiD method, it possesses extra layers to process information and find hidden patterns that traditional techniques overlook.

How does Deep DiD work, and how good is it?

First, the researchers develop a Deep DiD model by integrating deep neural networks into a difference-in-differences framework to flexibly estimate individual-level heterogeneous treatment effects as nonparametric functions of high-dimensional pre-treatment features.

By inspecting rich data from the platform, the model can predict which creators would have improved the most if they joined the programme. “With this advanced predictive analysis, we can examine how a specific programme helped some individuals more than others,” Professor Cao adds.

Creators selected by the Deep DiD model as favourable candidates show 57 to 123 per cent higher performance than those selected by the platform. Overall, creators chosen by the model consistently show a 72 to 114 per cent higher performance.

In out-of-sample evaluations, creators selected by the Deep-DiD model exhibit substantially larger performance gains than those selected by the platform. Among signed creators, those also identified by the model experience 70 to 80 per cent higher realised performance jumps relative to the average signed creator, across both user time contributed and user engagement.

When comparing selection rules directly, creators ranked highest by the model have 57 to 123 per cent higher estimated treatment effects than platform-selected creators, indicating significant scope for improvement in targeting. Notably, nearly half of the creators identified by the model were not signed by the platform, reflecting systematic differences in selection criteria.

What makes the Deep DiD method remarkable is its flexibility. While the three key indicators above reflect outcomes that platforms are likely to consider, platforms can tailor any other metrics depending on their goals. This way, they can maximise the impact of their programmes by investing only in those who will gain the most.

“If a platform’s revenue depends on users’ watching time, then this metric can be prioritised. The goal can also be defined as a weighted combination of metrics or other customised outcomes,” Professor Cao says. “The process can be repeated multiple times to reduce randomness and improve the stability of the results. Whenever the platform plans to implement a new intervention, new rounds of estimation should be conducted using the corresponding data and inputs.”

Beyond the digital platform, the Deep DiD method may also be applied in other settings. For instance, a retail shop launching a rewards programme must choose its target wisely to ensure only the right customers contribute to revenue, a company deploying bonus systems needs to determine which employees would achieve the largest productivity boost, and a government introducing a subsidy scheme should predict which individuals would benefit most.

In the real world, figuring out how much an intervention truly impacts a beneficiary is never easy since many variables interact in complex and hidden ways. These intricate and unknown linkages are nearly impossible to figure out using traditional mathematical methods, but the deep neural network acts much more like a sophisticated brain and is exceptionally good at discovering complex and subtle patterns.