TL;DR
This paper introduces a novel data-driven approach to constructing interindustry input-output tables by leveraging interfirm buyer-seller network data and firm textual descriptions. By extending the sparse block model with a two-dimensional Chinese restaurant process, the method can automatically discover latent industry structures and their interdependencies, offering a more granular and flexible alternative to traditional survey-based methods.
What Problem Does This Paper Address?
Traditional input-output (IO) tables are crucial for understanding the flow of goods and services within an economy, summarizing how industries buy from and sell to each other. However, these tables are typically built using extensive surveys and manual interpretation of economic data, a process that is time-consuming, costly, and can be slow to adapt to evolving economic structures. A key limitation is that these methods often pre-define industries and their relationships, potentially missing emerging or unconventional economic interactions. The authors aim to address this by developing a data-driven method that can automatically infer industry structures and their interdependencies directly from micro-level interfirm transaction data and textual information about firms' business lines.
Methods
The core of the proposed approach is an extension of the Sparse Block (SB) model, a graphical model that jointly analyzes network structure and node attributes. The authors introduce the Infinite Sparse Block with Node Textual Information (InfSBT) model.
Key methodological innovations include:
- Joint Modeling of Network and Text: The model integrates interfirm buyer-seller network data (who buys from whom) with textual descriptions of firms' main business lines. This allows for a richer understanding of industry relationships than using either data source alone.
- Two-Dimensional Chinese Restaurant Process (2DCRP): To handle an unbounded number of latent industries and their interactions, the authors extend the well-known Chinese restaurant process. This extension allows for the discovery of new industry "clusters" and new links between existing or new industries, mimicking both exogenous (new industries) and combinatorial (new links between existing industries) innovation processes.
- Sparse Block Model Foundation: The use of the SB model is motivated by the inherent sparsity in interfirm transaction data, where not all firms trade with all others. The SB model is efficient for such sparse networks.
- Inference: The model is trained using collapsed Gibbs sampling, a standard Bayesian inference technique.
The paper also discusses a "Reversed SBT" model, which prioritizes text generation before network generation, and compares its performance.
Datasets
The paper evaluates the proposed InfSBT model on two types of datasets:
- Synthetic Data: A network of 70 nodes and 249 edges, with 16 underlying industries and associated textual data. This dataset is used to assess the model's ability to recover known block structures and topic distributions.
- Real-World Interfirm Buyer-Seller Network: Data from a provider collecting interfirm buyer-seller information for 222 Japanese firms from the accounting year 2012. This dataset includes firm names, their buyers/sellers (up to five each), and textual descriptions of their main business lines (parsed using morphological analysis).
Key Insights and Ideas
- Data Science for Economic Structure: The paper demonstrates that established data science techniques, particularly in network analysis and topic modeling, can be powerful tools for understanding complex economic structures like interindustry relationships, offering a data-driven alternative to traditional methods.
- Latent Structure Discovery: The ability of the model to automatically discover latent industry structures (blocks) and their interdependencies is a significant advantage over traditional IO tables, which rely on pre-defined industry classifications. This is crucial for identifying emerging industries or nuanced economic relationships.
- Complementarity of Data Sources: Combining interfirm transaction data with textual descriptions of firms' business lines significantly enhances the accuracy of estimating industry interdependence and predictive performance. Textual data provides crucial context that network structure alone might miss.
- Flexibility of Infinite Models: The use of the 2DCRP allows the model to handle an open-ended number of industries, making it more adaptable to dynamic economies where new industries or business models emerge.
- Sparsity is Key: The authors highlight that interfirm networks are inherently sparse, and models like the SB model are well-suited to exploit this sparsity, leading to more efficient and accurate inference compared to models that assume denser connections.
Key Results
- Improved Predictive Accuracy: The InfSBT model consistently outperformed baseline models (SB, MMSB, RTM) in predicting unseen edge lists (link prediction) on both synthetic and real-world datasets, achieving higher AUC scores.
- Effective Block Structure Estimation: On synthetic data, the model successfully estimated the underlying block structure and topic distributions, with the SB model incorporating textual information showing superior performance.
- Meaningful Industry Clusters: For the real-world interfirm network, the model successfully identified meaningful industry clusters, such as a general construction business purchasing from various suppliers (hardware, concrete, machinery, etc.) and a wholesale business centered on concrete.
- Sparser Representation: The estimated industry structure from the proposed approach was sparser than traditional IO tables, which is attributed to the fixed-rank nomination scheme in the data (firms naming only their top 5 buyers/sellers) and the focus on major relationships.
- Word Prediction Performance: The model also showed strong performance in predicting unseen words, outperforming LDA and RTM when network information was incorporated.
Limitations and Future Work
- Data Granularity and Sparsity: The paper acknowledges that the fixed-rank nomination scheme in the interfirm data (limiting firms to naming only five buyers/sellers) leads to a sparser representation than ideal. To capture more minor relationships, data closer to the "ideal dataset" (listing all transactions with date, price, and volume) would be beneficial.
- Complementary Insights: The proposed approach, relying on buyer-seller networks, naturally omits sectors like public and finance, which are present in traditional IO tables. Future work could explore incorporating these sectors or other data sources (e.g., geographic information, multiplex relational information) to create a more comprehensive economic model.
- Computational Challenges: While inference is based on Gibbs sampling, the two-dimensional CRP extension introduces complexities that can make inference less straightforward than its one-dimensional counterpart.
- Interpretability of "New" Industries: The 2DCRP allows for the creation of new industries. Further work could focus on better interpreting what constitutes a "new" industry in this context and how to label or categorize them.
Summary updated