ReconGraph: Probabilistic Consistency Inference over Heterogeneous Temporal Graphs for Enterprise Financial Reconciliation
Chandrakanth Thadkapally
Manuscript complete. Not submitted; no venue or submission date set.
Enterprise reconciliation is not a record-pair problem. A single economic event can produce an order, an invoice, a shipment, a receipt, a payment, a settlement, a commission, an adjustment and a ledger posting, written by different systems on different clocks. ReconGraph represents these artefacts as a heterogeneous temporal graph and infers a globally consistent reconciliation subgraph over it, rather than accepting or rejecting candidate pairs independently. Relation- and time-specific message passing combines attribute, amount, temporal, neighbourhood and business-rule evidence; a factorised consistency layer penalises violations of quantity conservation, temporal ordering, debit–credit balance and link cardinality. On RECONBENCH — a synthetic benchmark constructed from the public IBM AMLSim / AML-Data ecosystem — ReconGraph reaches 0.949 link F1, 0.944 graph-consistency accuracy and 0.892 anomaly recall, while reducing the false-reconciliation rate to 0.033, a 32.7% reduction against the strongest graph-matching baseline. Ablation, calibration, robustness and scaling studies isolate the contribution of the accounting constraints. The result is a reconciliation decision that is auditable as a subgraph with residuals attached, rather than an isolated similarity score.
Keywordsenterprise reconciliation · heterogeneous temporal graph · probabilistic inference · accounting constraints · entity matching · anomaly detection · calibration
- 01Poses reconciliation as inference of a globally consistent subgraph over a heterogeneous temporal graph, so that one-to-many, many-to-one and partial allocations are ordinary cases rather than exceptions to a bijection.
- 02Combines relation-aware and time-aware message passing over typed financial records with a learned local edge probability, so neighbourhood evidence can override a locally attractive but globally impossible link.
- 03Adds a factorised consistency layer that scores quantity-conservation, temporal-order, debit–credit balance and cardinality residuals, making the reason a subgraph was rejected inspectable per constraint family.
- 04Releases RECONBENCH, a synthetic order-to-cash and procure-to-pay benchmark derived from public AML simulation data, with ground-truth allocation edges and controlled corruption applied only after ground truth is frozen.
- 05Evaluates on link quality, false reconciliation, graph consistency, anomaly recall, calibration, runtime and review workload, rather than on pairwise accuracy alone.
- Financial artefacts become typed nodes — invoices, shipments, receipts, payments, settlements, adjustments, ledger postings — and candidate correspondences become typed edges (fulfils, settles, allocates, reverses, posts-to), each carrying event time and ingestion time.
- Candidate generation runs three cascaded lanes: exact normalised identifiers for precision, phonetic and edit-distance keys for recovery, and amount–time windows for the identifier-free case. Every retained edge records why it was blocked in, so candidate-generation misses can be audited separately from classifier errors.
- A logistic local model scores each candidate edge from attribute similarity, normalised amount agreement, temporal evidence, graph context and business-rule features, with missing fields masked so absence is not read as agreement.
- Two layers of relation- and time-specific attention propagate neighbourhood evidence without smoothing local reconciliation chains away.
- A factorised consistency layer converts accounting structure into residuals over candidate subgraphs; posterior edge beliefs are temperature-calibrated so that a confidence threshold means something operationally.
| Method | Prec. | Recall | F1 | FRR | GCA | Anom. | ECE |
|---|---|---|---|---|---|---|---|
| Exact rules | .962 | .681 | .798 | .038 | .742 | .611 | .118 |
| Fuzzy probabilistic | .931 | .814 | .869 | .069 | .801 | .702 | .083 |
| Neural pairwise | .944 | .873 | .907 | .056 | .834 | .748 | .061 |
| Graph matching | .951 | .895 | .922 | .049 | .897 | .821 | .048 |
| ReconGraph | .967 | .931 | .949 | .033 | .944 | .892 | .026 |
FRR — false-reconciliation rate; GCA — graph-consistency accuracy; ECE — expected calibration error. Lower is better for FRR and ECE.
| Topology | Method | Prec. | Recall | F1 | GCA |
|---|---|---|---|---|---|
| One-to-one | Graph matching | .962 | .929 | .945 | .928 |
| One-to-one | ReconGraph | .973 | .951 | .962 | .966 |
| One-to-many | Graph matching | .941 | .872 | .905 | .881 |
| One-to-many | ReconGraph | .963 | .929 | .946 | .936 |
| Many-to-one | Graph matching | .927 | .841 | .882 | .854 |
| Many-to-one | ReconGraph | .951 | .887 | .918 | .910 |
| Partial payment | Graph matching | .938 | .858 | .896 | .867 |
| Partial payment | ReconGraph | .958 | .914 | .936 | .925 |
The gap widens as the topology moves away from one-to-one — which is the population pairwise matching handles worst.
- Pairwise systems improve F1 from rules to neural matching, but their graph-consistency accuracy stays below their F1 throughout: they get more links right without getting the explanation right. Global inference closes that gap.
- Under 50% joint corruption of identifiers, fields, timestamps and amounts, F1 holds at 0.864, because weight moves off damaged identifiers and onto amount conservation, event ordering and neighbourhood structure.
- Ablation separates the two mechanisms: removing graph context costs the most F1 (0.949 → 0.918), while removing accounting constraints costs the most consistency (GCA 0.944 → 0.872). They are not redundant.
- Temperature calibration cuts ECE by 63.4%. At an auto-accept threshold of 0.92 the simulated auto-reconciliation rate is 71.6%, which is the number an operations team actually plans against.
- Scaling from 0.1M to 5M records is near-linear: peak GPU memory 1.1 → 16.8 GB, inference 8 → 226 seconds, because blocking keeps connected components independent and average degree near-constant.
- RECONBENCH is synthetic. It is built to be reproducible and ethically shareable, not to be a substitute for a live enterprise trial, and no claim here transfers to production without one.
- Candidate blocking bounds recall. A true allocation that never became a candidate cannot be recovered by any downstream stage, and the blocking-reason audit exists precisely because that failure is otherwise invisible.
- The constraint set is the one this work formalised. Foreign exchange, tax treatment and late adjustment practice vary by jurisdiction and organisation, and an over-weighted constraint suppresses genuine exceptions rather than surfacing them.