Patent Citation Networks Anticipate Technology Lock-In Risk: Multimodal Evidence from China’s Semiconductor Industry

In the high-stakes world of technology, success is often measured by how much a company produces and how often others copy its ideas. When a firm files many patents and those patents are frequently cited by competitors, it is usually seen as a sign of strength and leadership. However, there is a hidden danger lurking beneath these visible signs of prosperity. Just as a traveler might keep walking down a familiar path because it feels safe, even when better routes are closing off, a company can become trapped in its own way of thinking. This phenomenon, known as technological lock-in, happens when an organization becomes so dependent on a specific set of knowledge and connections that it loses the ability to adapt when the world changes. The risk is that a company can look incredibly productive on paper while its actual capacity to innovate and pivot quietly disappears.

Researchers Jilong Zhang and Lei Chen set out to find a way to spot this danger before it is too late. They focused on China's semiconductor industry, a sector where the ability to reorient quickly is vital for national security and economic stability. Instead of waiting for a company's patent output to drop or its citations to fade—signals that often arrive only after the damage is done—they asked if the structure of a company's connections could reveal the trouble earlier. They treated the web of patent citations not just as a record of past achievements, but as a map of future possibilities. By analyzing how companies link their new ideas to old ones, and how those ideas spread through different groups of inventors, the team hoped to build a warning system that could see the narrowing of a company's horizons while it was still busy celebrating its current success.

To test this idea, the researchers gathered a massive collection of 541,592 invention patents filed in China between 2001 and 2023. They did not simply count how many patents each company had or how many times they were cited. Instead, they built a complex digital model that looked at three different types of information simultaneously. First, they examined the obvious numbers: the volume of patents, the size of the teams creating them, and the sheer number of citations. Second, they read the text of the patent summaries to understand what specific problems the companies were trying to solve and whether their focus was widening or narrowing. Third, and most importantly, they mapped the hidden architecture of the citation network. They used a method that traces how a patent connects to others, looking for signs that a company was only talking to itself or a small, closed group of peers, rather than reaching out to diverse communities of knowledge.

The team trained a computer model to predict which companies would eventually become deeply embedded in a single, dominant technological path—a state they defined as a "lock-in" risk. They found that the most accurate way to spot this risk was to combine all three types of information. The model that worked best was a type of artificial intelligence known as a random forest, which excels at finding patterns in messy, complex data. When tested on data it had never seen before, this model achieved an AUC of 0.54 while keeping false alarms relatively low. This performance was significantly better than simpler statistical methods, proving that the hidden structure of connections holds vital clues that raw numbers alone cannot provide.

One of the most striking discoveries was that the structure of the citation network provided information that could not be replaced by patent counts or text analysis. Even when a company had a high volume of patents and a strong reputation, the model could detect if its connections were becoming too narrow. The network data acted like a structural X-ray, revealing whether a company's knowledge was flowing freely across different fields or if it was circling back on itself. The researchers found that removing this network data from their model caused a sharp drop in its ability to predict risk, confirming that the way ideas are connected is just as important as the ideas themselves.

However, the study also revealed a crucial limitation: this warning system works very differently for large companies compared to small ones. For big entities with long histories of patenting and dense networks of citations, the model was highly effective. It could reliably spot when their search for new knowledge was becoming stagnant. But for smaller companies with fewer patents and sparser connections, the model struggled to make accurate predictions. The data for these smaller players was often too thin to support a confident diagnosis. The researchers concluded that while the tool is powerful for monitoring established giants, it cannot be used as a standalone rule for small startups or early-stage research teams. For them, the numbers need to be combined with human judgment and other forms of evidence.

To make these findings useful for real-world decision-making, the authors created a "Technology Lock-in Risk Index." This is a score from zero to one hundred that estimates the likelihood of a company entering a locked-in state in the coming year. The score is not just a single number; it breaks down the risk into three parts, showing whether the danger comes from the company's visible output, the topics it is discussing, or the hidden structure of its connections. A high score does not mean a company has failed; rather, it serves as an early warning signal that prompts a deeper look. It suggests that while the company may still be producing results, its path forward might be narrowing, and it may need to seek out new partnerships or explore different technological avenues to stay resilient.

The study suggests that public agencies and industry leaders should not rely solely on traditional metrics like patent counts when assessing the health of their innovation systems. By adding network-based indicators to their monitoring dashboards, they can detect structural weaknesses before they lead to a crisis. The goal is not to punish companies for being successful, but to help them recognize when their success is making them rigid. For large corporations, this means regularly checking if their knowledge sources are diverse enough. For smaller innovators, it means understanding that their potential might not be fully visible in the patent data yet, requiring a more nuanced approach to evaluation. Ultimately, the research offers a new way to see the invisible currents of knowledge flow, helping societies navigate the complex terrain of technological change with greater foresight.