Is It Still the Same Person? Bidirectional Coherence Tests for Separating Account Takeover from Benign Preference Drift

When a fraud system flags a card or account, the hardest alerts are not the obvious ones. They are the cases where a real customer has changed what they do: a move to a new city, a new hobby, a first expensive purchase. To a model that asks "is this transaction unusual?", a genuine change in someone's habits and an account takeover look the same at the moment they happen. They look different afterwards. A person whose tastes change still behaves like themselves. They return to the same grocery store, keep the same subscriptions, and settle into the new pattern. A fraudster's activity does not connect to the owner's long-term habits. It arrives as a burst, often while the real owner keeps transacting in parallel. Industrial fraud teams already have time to look afterwards: alerts sit in a human review queue for hours or days, and the customer keeps generating events during that wait.

We propose to frame review after an alert as a test of whether the account's history reads as one continuous story. A bidirectional sequence model reads the behavior stream before and after the flagged event, meaning it processes the sequence both forward and backward in time. It is trained without fraud labels to tell whether that stream comes from one continuously evolving person or has been spliced together from two. The training negatives are built the way coherent-storytelling datasets build them, by corrupting real sequences. Here the corruption replaces or interleaves another user's behavior after a chosen point. Genuine change points inside a single user's history then serve as hard positives: the model must learn that change by itself is not a splice.

Core Claim

The claim is that a self-supervised "same author" coherence objective, read bidirectionally around the alert, separates benign preference drift from takeover better than supervised bidirectional fraud scores. It should also need far fewer of the slow-arriving fraud labels. The measured effect should appear mainly as fewer false alarms on accounts that are genuinely drifting, at equal fraud recall.

Three bodies of work motivate this. The first shows the gap. The Importance of Future Information in Credit Card Fraud Detection exploits the verification delay with a bidirectional LSTM (a recurrent network run in both time directions) over 4 past and 2 future transactions. Its target is supervised fraud labels on flat features, and it has no notion of the user's long-term habits. The second explains why labels are a bottleneck: Explainable Deep Behavioral Sequence Clustering for Transaction Fraud Detection notes chargeback labels lag by weeks and responds with unsupervised bidirectional embeddings plus clustering. That clustering groups sessions by similarity. It does not ask whether one account's history is internally consistent. The third frames drift as adversarial. Domain Knowledge-Enhanced LLMs for Fraud and Concept Drift Detection and Joint Detection of Fraud and Concept Drift in Online Conversations with LLM-Assisted Judgment pair a statistical drift tripwire with a forward-only language-model judge. Neither looks backward from later evidence, and neither models what benign drift should look like.

Two further lines supply the method. From narrative coherence, CoMM shows that negatives built by corrupting real sequences (step shuffling, frame mismatching) teach models sequential consistency. Character-Preserving Coherent Story Visualization treats the preservation of a character's identity across frames as the core coherence problem. We use "identity persists across the timeline" in the same sense, applied to an account holder. From preference drift, PISA separates stable habits from transient taste shifts using ACT-R memory activation, a cognitive-psychology model of how recall strength grows with frequency and decays as a power law over time. We reuse that activation as an explicit habit-memory signal: after the change point, do the user's high-activation merchants or items keep recurring?

Research Program

The model has three parts.

The first is a causal encoder that summarizes each account's history up to the alert. It also maintains PISA-style activation scores over merchants, categories, or items, so that a compact "habit profile" exists at alert time.

The second is a bidirectional encoder over a window of events around the alert. It produces two judgments. One estimates the probability that the post-alert segment was produced by the same person as the pre-alert segment. The other locates the most likely splice point.

The third is a habit-continuation channel. It measures how strongly post-alert events re-activate the pre-alert habit profile, and it is fed to the coherence head as an input.

Pretraining needs no fraud labels. For each account we take a window and choose a cut point. Positives are untouched windows, including windows centered on the user's own largest change points. These change points are detected by distance between the pre- and post-cut habit profiles, and we oversample them so that drift itself is not a shortcut cue. Negatives come in two kinds. Replacement splices substitute another account's events after the cut. Interleaved splices merge another account's events into the stream while the owner's events continue, mirroring the parallel streams that Interleaved Sequence RNNs for Fraud Detection models in production. Donor accounts are matched on coarse statistics such as spend level and geography, so the model cannot win by spotting crude mismatches.

After pretraining, the coherence score is fine-tuned with whatever fraud labels exist, and it can be read alongside the supervised score. The predicted splice point and the habit-continuation trace give investigators an account of the alert they can check. For example: "after the alert, none of the 20 most-activated merchants recur, and a second geographic stream appears." This is a grounded form of the narrative explanation that Skope-rules or free-text summaries only approximate.

The obvious risk is that splices of real users are not fraud. A real fraudster is not a random second customer, and a careful one may imitate the victim. We treat this as a question for the experiments rather than an assumption. We test transfer to real fraud labels, and we test against simulated fraudsters who deliberately preserve the victim's habits.

Baselines

The naive reference point is the alert score itself. This is the causal model at authorization time, applied unchanged to the review queue, which shows how much the delay window is worth at all.

The strongest existing method is the supervised bidirectional LSTM of The Importance of Future Information in Credit Card Fraud Detection. That paper reports AUCPR 0.403 for the 4-1-2 bidirectional model versus 0.344 for the unidirectional 4-1-0 model, and Early AUCPR 0.254 versus 0.165, on a private 30-million-transaction dataset. We will reimplement it on each of our datasets with the same past and future window sizes, the same features, and the same hyperparameter search budget as our model, and train it on all available labels. Beating it means higher AUCPR at full labels, or equal AUCPR with a significantly lower false-positive rate on benign change-point accounts at matched recall. A third win condition is matching its full-label AUCPR using a quarter or less of the labels.

The strongest classical competitor is a gradient-boosted tree model of the kind that Interleaved Sequence RNNs for Fraud Detection used as its production LightGBM baseline. For fairness we give it hand-built post-alert features, such as the recurrence rate of the top pre-alert merchants after the alert, the count of distinct concurrent locations, and spend ratios. That way it has access to the same "future" information and the same habit idea in tabular form.

For adaptive streams in simulation, the reference is the bandit-plus-trees method TBBTS from Adapting to Concept Drift in Credit Card Transaction Data Streams Using Contextual Bandits and Decision Trees. It reports Precision@K 0.724 against 0.455 for Random Forest and 0.376 for XGBoost on its benchmark.

Three ablations isolate the key ingredients. The first removes the backward pass, leaving coherence pretraining on a causal encoder. The second removes the habit-continuation channel. The third removes the hard positives, that is, pretraining without oversampled genuine change points. This last one tests directly whether teaching the model that drift is not a splice is what reduces false alarms.

Why not just…?

Why not just give the supervised Bi-LSTM a longer future window? This is the most serious threat. A longer window may let a supervised model learn habit continuation implicitly. The cited paper shows future context helps, but it does not show whether supervision alone learns to excuse benign drift. We will sweep future window length for both models. If the supervised model closes the gap on benign-drift false positives at full labels, the remaining claim is label efficiency, and we will report the result that way.

Why not just add post-alert habit features to a gradient-boosted model? This could plausibly capture much of the signal, which is why it is a full baseline with those features rather than a strawman. Our bet is that interleaved splices and partial habit preservation require sequence-level reasoning that fixed aggregates miss. The adaptive-fraudster experiment is where this will be decided.

Why not just send flagged accounts to a language-model judge? On simulated dialogues, the hybrid judge of Joint Detection of Fraud and Concept Drift in Online Conversations with LLM-Assisted Judgment reached 82% accuracy and took 77 seconds for 360 samples. Those dialogues are short, while transaction histories run to hundreds of events. The judge also reads forward only and has no habit memory. We will include a judge prompted with serialized windows on a sampled subset, as a cost-quality reference.

Why not just use a drift detector and suppress alerts where drift is detected? A drift detector fires on exactly the cases in question, both benign change and takeover. It therefore cannot separate them, and suppressing alerts where it fires would hide fraud.

Proposed Experiments

The first experiment is a public, controlled study on Last.fm listening sessions from PISA. These contain real taste drift but no fraud, so we inject replacement and interleaved splices. We measure splice-detection AUCPR and, most importantly, the false-alarm rate on the users' genuine largest drift episodes. The point is to show that the model detects the splices while ignoring real drift. No reported result was found on Lacuna for splice detection on Last.fm, so the bidirectional and gradient-boosted baselines establish the reference.

The second experiment uses the MultiMAuS simulator from Adapting to Concept Drift in Credit Card Transaction Data Streams Using Contextual Bandits and Decision Trees. We add scripted benign life-event drift for genuine cardholders, plus a fraudster that progressively imitates the victim's merchants. We report Precision@K against TBBTS and AUCPR against the other baselines as imitation strength rises. The decisive question is how fast our advantage decays as the fraudster copies more of the victim's habits.

The third experiment is real-label transfer. Neither the 30-million-transaction dataset nor the Alibaba clickstream data is public, so this requires a partner dataset with card identifiers and chargeback labels. On it we will run AUCPR, Early AUCPR, and label-fraction curves at 5%, 25%, and 100% of the labels. The bar to clear is the reimplemented supervised Bi-LSTM, with 0.403 AUCPR as the published reference on the original data.

The idea fails under either of two conditions. One is that coherence pretraining helps on synthetic splices but adds nothing on real fraud labels. The other is that the hard-positive ablation leaves benign-drift false positives unchanged. Either outcome would show that "same person" is not the right description of how fraud differs from benign drift.