Technical Implementation of Ownership Penetration: Graph Data Modeling and Traversal Algorithms

1. Why Ownership Penetration is a Technical Problem

Ownership penetration refers to tracing a company’s ownership chain starting from a target entity along investment relationships. Upward penetration continuously asks “who are the shareholders of this company, and who are the shareholders of those shareholders?” until it reaches an ultimate natural person or a recognized actual controller. Downward penetration maps out the investment chain to identify which enterprises this company ultimately controls and where the benefits flow.

This capability is widely used in due diligence, credit approval, Anti-Money Laundering (AML), beneficial owner identification, related party transaction screening, and bidding qualification reviews. Take beneficial owner identification as an example: standard practices require penetrating to the natural person who ultimately holds the company’s equity, often using a specific ownership threshold (e.g., 25%) as a determination criterion, cross-verified with actual controller declarations.

Building this into a stable system capability immediately presents four technical challenges:

1. Multi-source heterogeneous data: Business registration, public information, listed company announcements, and legal documents are scattered across different channels. Field definitions, update frequencies, and data quality vary significantly.
2. Severe entity ambiguity: Companies and natural persons with identical names are common. Corporate renames and re-registrations after cancellation further increase confusion.
3. Complex structures: Under conglomerate operations, ownership chains can easily reach seven or eight layers. Cross-holdings and circular shareholdings create cyclic structures where naive traversal can fall into infinite loops.
4. Dynamic evolution: Capital increases, equity transfers, and liquidations occur constantly. The system must support “point-in-time” snapshot semantics rather than just maintaining a single “current state” table.

Therefore, ownership penetration is fundamentally a comprehensive technical problem combining “data engineering + graph computing + intelligence.” The following sections expand on these four layers: data, algorithms, engineering, and intelligence.

2. Data Foundation and Graph Modeling: The Prerequisite for Penetration

1. Data Sources

Available public data generally includes enterprise registration and publicity information (shareholder lists, investment ratios, subscribed and paid-in capital, change records), listed company announcements and periodic reports, legal documents, and administrative penalty information. The engineering key is not “whether data exists,” but how to clean, align, and fuse data from different channels into a consistent ownership relationship network.

2. Property Graph Modeling

Ownership relationships are naturally graph-structured: nodes are companies or natural persons, edges are shareholding relationships, and attributes like investment ratio, subscribed/paid-in capital, and change date are attached to the edges. Adopting a property graph model is highly recommended:

{
  "nodes": [
    { "id": "C1", "type": "company", "name": "Target Company" },
    { "id": "C2", "type": "company", "name": "Intermediate Holding Company" },
    { "id": "P1", "type": "person",  "name": "Natural Person" }
  ],
  "edges": [
    { "src": "P1", "dst": "C1", "rel": "invest", "ratio": 0.40, "effective_date": "2024-06-30" },
    { "src": "C2", "dst": "C1", "rel": "invest", "ratio": 0.35, "effective_date": "2024-06-30" }
  ]
}

Two modeling recommendations:

• Edges must carry time attributes. Equity is time-variant. The penetration result for “June 30, 2024” and “today” can be completely different. The query interface should support fetching snapshots based on a specific point in time.
• Unify the ratio scale. Distinguish between direct and indirect shareholding via intermediaries. Normalize ratios uniformly as decimals to avoid calculation errors caused by mixing “percentages” with “capital contribution amounts.”

3. Storage Selection

Relational databases are suitable for detailed records and transactions, but multi-layer variable-length path queries (e.g., “find all natural persons upwards”) require either recursive SQL or repeated query concatenation at the application layer, which is costly and hard to optimize. Graph databases naturally fit multi-hop traversals and path retrievals. Knowledge graph platforms further provide entity alignment, attribute fusion, and version management capabilities. If you prefer not to build from scratch, you can utilize general-purpose graph database or knowledge graph services provided by cloud vendors (such as Baidu AI Cloud), scaling on demand to lower initial costs.

3. Core Algorithms: Multi-hop Traversal, Effective Shareholding, and Cycle Handling

1. Multi-hop Upward Traversal and Effective Shareholding Calculation

The standard algorithm for effective shareholding is: along a single path, multiply the shareholding ratios segment by segment to get that path’s contribution; if the same shareholder has multiple paths to the target company, sum the contributions of each path. Below is a runnable schematic code:

from collections import defaultdict
# Adjacency list: Company -> [(Shareholder, Shareholding Ratio)], ratio normalized to 0~1
# Naming convention: C prefix for company nodes, P prefix for person nodes; actual systems should use node type fields
graph = {
    "C1": [("P1", 0.40), ("C2", 0.35), ("C3", 0.25)],
    "C2": [("P2", 0.60), ("C3", 0.40)],
    "C3": [("P1", 0.50), ("P3", 0.50)],
}
def trace_up(graph, node, path=(), eff=1.0, out=None, depth=0, max_depth=8):
    """Bottom-up traversal: accumulate effective shareholding contribution on each shareholder chain"""
    if out is None:
        out = []
    if depth >= max_depth:          # Defense: limit abnormal deep chains
        return out
    for holder, ratio in graph.get(node, []):
        if holder in path:          # Cycle pruning: do not re-enter the same path
            continue
        cur_eff = eff * ratio
        if holder.startswith("P"):  # Natural person: path endpoint, record result
            out.append((holder, cur_eff, path + (node, holder)))
        else:                       # Corporate shareholder: continue upwards
            trace_up(graph, holder, path + (node,), cur_eff, out, depth + 1, max_depth)
    return out
def aggregate(paths):
    """Sum contributions of multiple paths for the same shareholder to get the aggregated effective shareholding"""
    total = defaultdict(float)
    for holder, eff, _ in paths:
        total[holder] += eff
    return dict(sorted(total.items(), key=lambda x: -x[1]))
paths = trace_up(graph, "C1")
for holder, eff, path in paths:
    print(f"{holder:<4} Effective Shareholding={eff:.2%}  Path: {' -> '.join(path)}")
print("Aggregated Effective Shareholding: ", aggregate(paths))

Using the graph data above as an example: Natural person P1’s direct shareholding in C1 is 40%, with indirect paths via C2 and C3 contributing 7% and 12.5% respectively, resulting in an aggregated effective shareholding of approximately 59.5%, making P1 the primary candidate for the ultimate beneficial owner. P2 contributes 21%, and P3 contributes 19.5%. The value of penetration analysis lies precisely here—making the actual equity relationships hidden behind multi-layer structures explicit and computable.

2. Ultimate Beneficial Owner Determination

Calculating aggregated effective shareholding is only the first step. Determining the “ultimate beneficial owner / actual controller” typically requires integrating multiple signals:

• Threshold rules: If the aggregated effective shareholding reaches a set threshold (e.g., 25%), the entity is included as a beneficial owner candidate. Specific thresholds depend on regulatory and business standards.
• Maximum shareholding rule: If no one reaches the threshold, the individual with the highest aggregated effective shareholding is selected.
• Cross-validation: Compare with filed actual controller declarations, board composition, and concert party arrangements. Output warnings instead of making arbitrary conclusions when contradictions arise.
• Special arrangements: Situations where control is obtained through voting rights entrustment or nominee shareholders cannot be covered by shareholding ratios alone and require manual review of textual materials.

3. Circular Shareholding Handling

Cross-holdings create cycles in the graph, for example, Company A holds shares in Company B, and B反过来 holds shares in A. There are three handling principles: First, at the traversal level, deduplicate the “current path” (like the cycle pruning in the code) to ensure algorithm termination. Second, at the calculation level, clarify the criteria—a common practice is not to count back-edges in circular paths towards effective shareholding to avoid duplicate amplification of equity. Third, at the analysis level, output the cycle itself as a risk signal, as circular shareholding structures warrant due diligence attention. Engineering-wise, regularly running strongly connected component detection on the entire graph can proactively isolate subgraphs containing cycles.

4. Engineering Implementation Key Points

1. Entity Alignment and Disambiguation. This determines the upper limit of data quality. Corporate entities should primarily use unified registration numbers (or equivalent identifiers like Unified Social Credit Code) as primary keys, with names serving only as auxiliary references. Natural persons require cross-referencing multi-source information for disambiguation, handling both “same name, different person” and “different name, same person” scenarios, preserving confidence scores for downstream decision-making when necessary.

2. Temporal and Version Management. Each equity change generates a new version rather than overwriting old records. The query interface must support point-in-time historical lookups. Due diligence conclusions exported externally must be able to answer “as of [specific date]” scope questions.

3. Performance Optimization. Full-graph layer-by-layer traversal is costly for massive conglomerates. Common techniques include: pruning by shareholding ratio threshold (contributions below the threshold have limited impact on conclusions and can be down-weighted or truncated, provided the criteria are noted in the report), limiting maximum penetration depth, caching penetration results for hot companies, and pre-computing beneficial owner conclusions offline for frequently queried entities.

4. Explainability. Penetration results must be accompanied by a complete path evidence chain—every contributing path, every segment ratio, data source, and update time must be traceable. This is a hard requirement for due diligence and compliance scenarios, and the prerequisite for business stakeholders to trust algorithmic results.

5. Data Compliance. Penetration involves natural person and business operational information. Only data obtained through legal and compliant channels should be used, adhering to personal information protection laws and regulations. When outputting results, adhere to the principle of minimum necessity, avoiding using personal information for purposes beyond authorized scopes.

5. How Large Language Models Help: Reading Documents, Not Calculating Numbers

The rapid evolution of Large Language Model (LLM) capabilities in recent years has compensated for the long-standing shortcoming of “unstructured data processing” in ownership penetration. Mature auxiliary methods include:

• Material Extraction: Extracting equity change events, shareholder names, and ratios from announcements, legal documents, and news. After rule validation, these are written into the graph, significantly reducing manual entry costs.
• Entity Disambiguation Assistance: Utilizing contextual semantics to determine whether two mentions of a “company” refer to the same entity, providing candidates and confidence references for entity alignment.
• Graph Q&A and Report Generation: Business users ask questions in natural language, such as “How many paths are there from this company penetrating to natural persons?” The system translates the question into a graph query, and the LLM organizes the structured results into readable report paragraphs.

However, the boundaries must be recognized: LLMs are not adept at precise numerical calculations and carry hallucination risks. Directly letting an LLM “calculate” effective shareholding or “draw” actual controller conclusions is unreliable. A rational system division of labor is: LLMs handle document reading, extraction, and interaction; the graph computing engine handles precise calculations; and critical conclusions are reviewed by a rule engine with a manual audit step retained. Only by performing their respective duties can the system achieve both “intelligence” and “rigor.”

6. Conclusion

Ownership penetration appears to be a simple query feature, but it is essentially a complete technical pipeline: organizing multi-source equity data via property graphs, calculating effective shareholding through multi-hop traversals and path aggregation, determining ultimate beneficial owners via thresholds and cross-validation, and ensuring engineering viability through entity alignment, temporal management, explainability, and data compliance. LLMs provide effective gains in unstructured material extraction and human-computer interaction. For enterprises, whether building in-house or adopting general services like cloud-based graph databases and knowledge graphs, the prudent path is to first solidify the data model and calculation criteria, and then incrementally add intelligent capabilities.