Analyzing transactions in payment processing system
By using a transaction analysis unit to calculate probability scores and quality vectors in the payment processing system, and combining these with decision rules to select processing methods, the fraud risk and efficiency issues in transaction decision-making are resolved, resulting in more accurate and efficient transaction processing.
Patent Information
- Application Number
- CN202380026996.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-03-18
- Filing Date
- 2023-03-16
- Publication Date
- 2024-11-05
AI Technical Summary
In existing payment processing systems, transaction decision models struggle to balance fraud risk management with transaction efficiency, leading to erroneous transaction processing decisions that impact merchant revenue and costs.
The transaction analysis unit calculates probability score vectors and probability quality vectors, combines them with decision rules, selects the optimal processing method, applies it to transactions, and then outputs and trains the model to improve decision accuracy.
It improves the accuracy and efficiency of transaction processing, reduces the risk of transaction cancellations and fraud, and optimizes merchants' revenue and cost control.
Smart Images

Figure CN118901074A_ABST
Abstract
Description
Background Art
[0001] The present invention relates generally to processing financial transactions and, more particularly, to a system and method for analyzing transactions in a payment processing system to determine whether a process should be applied to the transaction.
[0002] Figure 1 1 is a flow chart of a payment processing system 100. The payment processing system 100 includes a buyer 102, a merchant 104, a payment processor 106, a card network 108, and an issuing bank 110. Note that the payment processing system 100 may include more than Figure 1 More or fewer entities may be shown. For purposes of discussion, it is assumed that the entities in the payment processing system are in electronic communication with each other via known electronic communication means.
[0003] Buyer 102 makes a purchase with an electronic payment card at merchant 104 (operation 120), and merchant 104 creates a transaction in response. Merchant 104 submits the transaction to payment processor 106 (operation 122). Payment processor 106 submits the transaction to card network 108 (operation 124). Card network 108 requests authorization for the transaction from issuing bank 110 (operation 126). Issuing bank 110 is the entity that issues the electronic payment card to buyer 102.
[0004] The issuing bank 110 determines whether to approve or deny the transaction and sends a response to the card network 108 (operation 128). The issuing bank 110 may consider any number of factors when determining whether to approve or deny the transaction, such as whether the buyer has sufficient credit to be able to complete the transaction. The card network 108 sends the response to the payment processor 106 (operation 130), and the payment processor 106 sends the response to the merchant 104 (operation 132).
[0005] For a better shopping experience, it is desirable to complete the transaction approval process as quickly as possible. At certain points in the process (e.g., at the payment processor 106, the card network 108, or the issuing bank 110), certain actions may be automated and may include the use of artificial intelligence (AI) algorithms.
[0006] For example, a classification model may be implemented to determine whether buyer 102 needs to be authenticated (e.g., by password authentication or by fingerprint authentication if the buyer is using a mobile device). The classification model may be an automated risk assessment tool that decides at the transaction level whether to authenticate a buyer based on suspicion that the transaction will later become fraudulent. The decision whether to authenticate a user may be referred to as a "soft intervention," meaning that the decision is limited to whether the buyer should be authenticated, rather than blocking the transaction if the risk of fraud is high.
[0007] In order to improve the operation of the classification model, feedback can be provided. This feedback is used to train the classification model to help classify situations that were predicted to occur as a result of requesting buyer authentication. This feedback is typically limited to whether the transaction ultimately turned out to be fraudulent or whether the transaction was authorized (and remained a genuine sale).
[0008] By limiting feedback to whether a transaction was fraudulent (meaning the transaction was completed), it may not include all possible outcomes, such as if the transaction was not completed (thus, it cannot be determined whether the transaction was fraudulent). If the classification model determines that a buyer authentication was requested and the buyer cancels the transaction because they do not want to complete the authentication (called a "drop-off"), this feedback is generally not considered because the transaction was not completed. Therefore, in some cases, requesting buyer authentication may result in a loss of sales. It may be beneficial to train the classification model to incorporate this additional feedback.
[0009] Large-scale transaction processing systems benefit from intelligent optimization measures that increase conversion rates. These measures or "treatments" can range from secure customer authentication checks to in-flight adjustments to transaction data fields. While the definition of a "favorable outcome" may vary, the common element is often the application of data-driven machine learning models to make automated decisions about whether an incoming transaction should be processed or better yet, not processed.
[0010] The decision can be suboptimal when the wrong transaction receives processing, or when the wrong type of processing is received. This results in missed revenue and unnecessary costs for the merchant due to canceled transactions or fraud that could have been prevented. Summary of the invention
[0011] A method for analyzing transactions in a payment processing system, the method comprising: receiving a transaction; classifying the transaction; analyzing the transaction; selecting a treatment to be applied to the transaction; applying the selected treatment to the transaction; and outputting the transaction after the selected treatment is applied to the transaction. Classifying the transaction comprises calculating a probability score vector for the transaction, the probability score vector indicating the probability of each of one or more possible outcomes of the transaction. Analyzing the transaction comprises calculating one or more probability mass vectors for the transaction, the probability mass vector indicating the impact value and associated probability of one or more possible treatments to be applied to the transaction. Selecting the treatment to be applied comprises applying a set of decision rules to the probability score vector and the one or more probability mass vectors. The transaction output is based on the selected treatment to be applied to the transaction, and the transaction then continues through the payment processing system before reaching an outcome that is subsequently used to train a classification and analysis unit.
[0012] A system for analyzing transactions in a payment processing system, the system comprising at least one processor and a non-transitory computer-readable medium containing instructions, the instructions, when executed by the at least one processor, causing the at least one processor to perform the following operations: receiving a transaction; classifying the transaction; analyzing the transaction; selecting a process to be applied to the transaction; applying the selected process to the transaction; and after the selected process is applied to the transaction, outputting the transaction. Classifying the transaction comprises calculating a probability score vector for the transaction, the probability score vector indicating a probability of each of one or more possible outcomes of the transaction. Analyzing the transaction comprises calculating one or more probability mass vectors for the transaction, the probability mass vector indicating an impact value and an associated probability of one or more possible processes to be applied to the transaction. Selecting the process to be applied comprises applying a set of decision rules to the probability score vector and the one or more probability mass vectors.
[0013] A transaction analysis unit for analyzing transactions in a payment processing system, the transaction analysis unit comprising at least one processor and a non-transitory computer-readable medium containing instructions, the instructions, when executed by the at least one processor, causing the at least one processor to perform the following operations: receiving a transaction; classifying the transaction; analyzing the transaction; selecting a process to be applied to the transaction; applying the selected process to the transaction; and after the selected process is applied to the transaction, outputting the transaction. Classifying the transaction comprises calculating a probability score vector for the transaction, the probability score vector indicating a probability of each of one or more possible outcomes of the transaction. Analyzing the transaction comprises calculating one or more probability mass vectors for the transaction, the probability mass vector indicating an impact value and an associated probability of one or more possible processes to be applied to the transaction. Selecting the process to be applied comprises applying a set of decision rules to the probability score vector and the one or more probability mass vectors. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 is a flow chart of a system that may implement the present disclosure according to the disclosed embodiments.
[0015] Figure 2 is a flow chart of a system for analyzing financial transactions of processing applications according to the disclosed embodiments.
[0016] Figure 3 is a flow chart of example decision logic used by a process decision unit according to the disclosed embodiments.
[0017] Figure 4 is a flow chart of a method for analyzing financial transactions of a processing application according to a disclosed embodiment.
[0018] Figure 5 is a flow chart of a method for classifying transactions according to a disclosed embodiment.
[0019] Figure 6 is a flow chart of a method for determining a process to apply to a transaction in accordance with a disclosed embodiment. DETAILED DESCRIPTION
[0020] The disclosed embodiments include systems and methods for analyzing transactions in a payment processing system. Before explaining certain embodiments of the present disclosure in detail, it should be understood that the present disclosure is not limited in its application to the details of the construction and the arrangement of components set forth in the following description or shown in the accompanying drawings. In addition to the described embodiments, the present disclosure is capable of practicing and executing the embodiments in various ways. Moreover, it should be understood that the words and terms used herein and in the accompanying drawings are for descriptive purposes and should not be considered limiting.
[0021] Thus, those skilled in the art will appreciate that the concepts upon which this disclosure is based may readily be utilized as a basis for the designing of other structures, methods and systems for carrying out the several purposes of the present disclosure.
[0022] Reference will now be made in detail to the present exemplary embodiments of the present disclosure, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numerals are used throughout the drawings to refer to the same or like parts.
[0023] A method for analyzing transactions in a payment processing system may include receiving a transaction. The transaction may be a purchase made by a buyer from a merchant, or other types of financial transactions using an electronic payment card that require approval prior to authorization. The transaction may be received in various ways from an internal or external source (e.g., through a customer-facing application programming interface or connected to an adjacent internal upstream processing system). The transaction may be a streaming unit of bound data points relating to various aspects of the transaction, such as a buyer identifier, a merchant identifier, a transaction identifier, a transaction amount, a transaction date, and other data points required to process the transaction to determine whether the transaction should be approved or denied.
[0024] The transaction may be classified by a processor, for example, by calculating a probability score vector for the transaction, the probability score vector indicating the probability of each of one or more possible outcomes of the transaction. The probability score vector may include a score for each possible transaction outcome. For example, each score in the probability score vector may be a floating point number between 0 and 1, and the sum of all scores in the probability score vector may equal 1.
[0025] The transaction can be analyzed by a processor, for example, by calculating one or more probability mass vectors for the transaction, the probability mass vector indicating the impact value and associated probability of one or more possible treatments to be applied to the transaction. Each impact quality probability vector represents a discrete probability distribution of the impact value of a possible treatment calculated. The impact value can be expressed in financial terms and can represent the possible loss or gain of the transaction for each of one or more possible treatments. The impact quality probability vector of each treatment can capture a series of impact values and associated probabilities, the probability indicating the calculated improvement value of the treatment compared to not applying the treatment to the transaction. The impact value can reflect the estimated effect of each treatment on various possible outcomes of the transaction. Similar to the probability score vector, each probability in the probability mass vector can be a floating point number between 0 and 1, and the sum of all probabilities in the probability mass vector can be equal to 1.
[0026] The selection of the processing to be applied to the transaction can be based on the probability score vector and one or more probability mass vectors. The decision logic can apply a series of decision rules (for example) to examine the probability score vector and the probability mass vector to select the processing to be applied to the transaction. The rules can include comparing the probability score vector and the probability mass vector with various thresholds and selecting the processing to be performed based on the thresholds. For example, the thresholds can include a first threshold associated with the probability score vector, a second threshold associated with the probability in the probability mass vector, and a third threshold associated with the impact value in the probability mass vector. In one embodiment, the processor (for example) can calculate the expected value of the transaction for all possible processing of the transaction and across all possible results, and can select the processing that results in the highest expected value of the transaction as the processing to be applied to the transaction. The impact of each possible processing can be captured in a single expected value, which can be calculated as the inner product of the impact value vector and the associated probability vector.
[0027] Transaction outcomes are affected by the selected processing applied to the transaction and result from interactions with other entities in the payment processing system. Transaction outcomes may be recorded for use in training machine learning models to aid in future predictions of transaction outcomes in one or both of classifying transactions and analyzing transactions.
[0028] Figure 2 2 is a flow chart of a system 200 for analyzing financial transactions of a processing application. The system 200 may be implemented as Figure 1 The payment processing system 100 is shown as a single unit in a payment processing system. In one embodiment, the system 200 may be implemented in more than one location, such as in a payment processor 106, a card network 108, and / or an issuing bank 110.
[0029] The system 200 includes a transaction analysis unit 202 and a database 204 including stored historical transactions. The transaction analysis unit 202 can be implemented as software, hardware, or a combination of software and hardware. For example, the transaction analysis unit 202 can be implemented as software running on a processor. The processor may include a central processing unit (CPU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other processing devices configured to receive and process data and instructions.
[0030] The database 204 can be implemented in different formats, so that the database 204 can store a large amount of semi-structured data (e.g., data in JavaScript Object Notation (JSON)). In one embodiment, the scalable storage provided by the cloud platform (e.g., Amazon Web Services, Google Cloud, or Microsoft Azure) is sufficient to support the database 204. The transaction analysis unit 202 and the database 204 can be located in a single server or can be located in separate servers. The operation of the system 200 does not change based on the relative location of the transaction analysis unit 202 and the database 204.
[0031] The transaction analysis unit 202 analyzes new transactions 206 using a machine learning model based on historical transactions stored in the database 204 to predict transaction outcomes 208. The new transactions 206 can arrive at the transaction analysis unit 202 in a variety of ways from internal sources or external sources (e.g., through a customer-facing application programming interface (API) or connected to an adjacent internal upstream processing system). For example, if the transaction analysis unit 202 is located at a card network (e.g., Figure 1 If the card network 108 is shown, the new transaction 206 can be processed from a payment processor (e.g., Figure 1 Payment processor 106) is shown arriving.
[0032] A new transaction 206 is a streaming unit of bundled data points arriving from one or more sources. It is assumed that the content and format of the new transaction 206 are consistent over time. This does not mean that each incoming new transaction 206 needs to have exactly the same data structure. For example, a semi-structured data format (such as a JSON data type) may contain additional branches depending on the origin of the new transaction 206. The interpretation of the data points in each transaction should not change rapidly over time. The use of hashes and tokens in new transactions 206 is allowed as long as the generation mechanism behind the hashes or tokens does not change frequently (as this may hinder the correct interpretation of the model).
[0033] In the transaction analysis unit 202, the new transaction 206 is sent to the classification model 210 and the processing decision unit 212. In one embodiment, the classification model 210 and the processing decision unit 212 can be implemented as software, hardware, or a combination of software and hardware, as an independent unit or as part of the transaction analysis unit 202. For example, the classification model 210 and the processing decision unit 212 can be implemented as software running on a processor. The processor may include a CPU, a GPU, an ASIC, an FPGA, or other processing devices configured to receive and process data and instructions.
[0034] Classification model 210 takes new transaction 206 as input and generates predictions 220 of the probabilities of possible transaction outcomes. Classification predictions 220 are sent to treatment decision unit 212. As described below, treatment decision unit 212 uses causal inference model 232 discussed below to determine whether to apply treatment 214a, 214b, or 214c to new transaction 206 (treatment decision 222). After one or more treatments 214a-214c are applied to new transaction 206, it continues through a payment processing system (e.g., such as Figure 1 The new transaction 206 is processed by the payment processing system 100 shown in the figure, and finally realizes the transaction result 208. The processing 214a-214c can include one or more possible processings for the new transaction 206, including not applying any processing to the new transaction (shown as processing 214c).
[0035] Process 214a-214c in Figure 2214a, 214b) plus a "no treatment" option (treatment 214c). There may be more or less than two treatment types, and it is assumed that there is always a "zero option" in which no treatment is applied, but without loss of generality, a "default treatment" may also be considered. In some embodiments, each treatment 214a-214c may represent a "soft interaction" with the new transaction 206, meaning that the treatment 214a-214c cannot completely block the new transaction 206 (which means that the new transaction 206 cannot be rejected by the treatment decision 222). There are many possible operations on the new transaction 206 (which may be considered as treatments): (i) data points in the transaction may be removed, edited, added, or shuffled; (ii) data security checks; (iii) customer security checks; (iv) dedicated AML (anti-money laundering) checks; (v) KYC (know your customer) checks; (vi) KYB (know your business) checks; or (vii) credit risk checks. Note that this list of treatments is non-limiting and other treatments are possible. Processing can be performed internally or externally (e.g., through a customer-facing application programming interface (API)). Implementation of a process can include queries to internal or external databases, computations using any number of data points from the transaction as input (including computations through statistical models, machine learning models, or artificial intelligence models (AI)), and adding the output of those computations back into the transaction as new data points. A process can consist of any number or combination of such operations.
[0036] In payment processing systems (e.g. Figure 1 In the context of the payment processing system 100 shown, specific examples of processing and associated desired results may include the following.
[0037] Strong Customer Authentication (SCA) checks (including asking for a password, mailing address, personal security questions, or biometric checks (e.g., fingerprint or self-portrait on a mobile device) are examples of "transaction processing" because SCA is a soft intervention that may or may not be applied and is intended to influence the outcome of the transaction 208. SCA is particularly intended to prevent fraudulent outcomes in e-commerce payments.
[0038] Deleting data fields in new transaction 206 that have bad or missing content may be a process that reduces the number of declines by the issuing bank.
[0039] Applying tokenization to a sub-selection of data fields in the new transaction 206 can be used to increase card network acceptance rates. Different card networks may have different preferences, and it may not be the same type of tokenization that is preferred for all transaction segments, which means that several processes 214a-214c may need to be considered simultaneously.
[0040] Processing new transactions 206 to determine predicted transaction outcomes 208 is the main synchronous prediction process and is Figure 2 This portion of the flow represents the relevant steps that a new transaction 206 goes through. For the sake of customer experience, this processing typically occurs in "real time" (i.e., virtually instantaneous or without discernible delay; in most payment systems, this may be a split second. Note that for the purposes of this discussion, Figure 2 The interaction with other payment processing infrastructure and other payment service providers is omitted. The final step from processing to transaction result usually involves one or more interactions with external parties (such as the issuing bank or card network).
[0041] Transaction results 208 are observable, obtainable, and quantifiable. Each new transaction 206 can have only one transaction result 208. Examples of transaction results in the context of a payment processing system include: authorized "real" sale, fraud (card networks distinguish between various types of fraud), chargeback (card networks distinguish between various types of chargebacks), declined (new transaction 206 is not accepted by the issuing bank and there are various types), canceled, or refunded. Note that the list of transaction results is not exhaustive and other transaction result types are possible.
[0042] In some embodiments, it may be assumed that there is one desired outcome that is more favorable than other possible outcomes. In this sense, a "favorable outcome" may be determined by the entity that is operating the system 200. The system 200 may be operated by various entities in the payment processing system 100, and each entity may have different goals when operating the system 200. For example, a payment processor (e.g., payment processor 106) may be primarily concerned with maximizing its processing capacity and facilitating payments by buyers, but may also be concerned with compliance with fraud rates due to rules set by a card network (e.g., card network 108).
[0043] In some embodiments, reaching a "favorable outcome" may occur implicitly as a result of how the entity operating the system 200 uses the transaction analysis unit 202 and how the entity chooses to configure the processing decision unit 212 in conjunction with the selected selection of the processing 214a-214c. The causal reasoning model training unit 232 may also know what outcome is "favorable" by way of model building. The distinction between how "favorable" an outcome is plays out in the decision logic in the processing decision unit 212. This may be reflected in the thresholds applied and the resulting decisions to apply certain processing 214a-214c, which may be encapsulated in the decision logic selected by the entity operating the transaction analysis unit 202.
[0044] There may be a delay between the time when the synchronization process has completed and the observation of the final transaction result 208. For example, for fraud and chargebacks in e-commerce payments, it is not uncommon for fraud to be reported up to 100 days after the transaction originally occurred. This delay, called "maturation," can be taken into account when establishing the correct result. In this case, the delayed feedback can still be included in the database 204 and used to train the classification model 210 and the processing decision unit 212.
[0045] The remaining actions and flows indicated by other arrow types are asynchronous and do not need to occur in real time. Figure 2 The dashed lines in represent the asynchronous flow of data points stored in the database 204. In some implementations, the new transactions 206, the classification predictions 220, the processing decisions 222, the processing results 224, and the transaction results 208 are all stored in the database 204.
[0046] Figure 2 The dashed lines from database 204 to classification model 210 and processing decision unit 212 in FIG. 2 represent the two models based on past transactions and their results ( Figure 2 230 and 232). After (re)training, the trained model is ready to make predictions for new transactions and will be installed in the synchronous streaming environment, replacing the previous model version. How often retraining occurs depends on the nature of the data stream and how it may change over time: throughput, the proportion of observed outcome classes and the amount of available processing all contribute to this.
[0047] The transaction results 208 become the “labels” used as target variables in the supervised training of the two models. To help train the models, new transactions 206 are linked to their result labels before being submitted to the database 204.
[0048] Transactions canceled by buyers before the transaction is completed, referred to as "backouts," are sometimes difficult to capture and store. Importantly, all transactions after the processing stage are concluded in the database 204 and assigned a transaction result 208. Cancellations after the processing stage may be assigned a result label of "canceled," while transactions backed out earlier in the process may be ignored. In some embodiments, to be able to capture the results of transactions that may be canceled at a later point in time, a processing result field may be added to the feedback sent to the database 204 (operation 224). As an example, the processing result field may include a binary flag (e.g., a "yes / no" flag) indicating whether the processing was previously applied to the transaction. As another example, the processing result field may include a detailed processing result (e.g., "partially passed," "fully passed," "high risk," or other processing result indications). In some embodiments, the training units 230 and 232 may treat any transaction that does not have an associated processing result field as a canceled transaction. In these embodiments, the "canceled" label may be viewed as an implicit result from the analysis of historical transactions that do not have a valid (or empty) processing result field.
[0049] The classification model 210 quantifies the propensity of a new transaction 206 to reach a certain (discrete) outcome state. For example, in a payment processing system where there are two possible outcomes (a real sale and a fraudulent attempt), a binary classification model can be used to predict the probability that a transaction proves to be fraudulent. The output of the classification model can be viewed as a risk score. As another example, if there are more than two possible transaction outcomes, a multi-class classification model can be used for the classification model 210.
[0050] In one embodiment, the classification model 210 receives the new transaction 206 as input, selects data fields from the new transaction 206, performs transformations of the selected fields (including any parsing or preprocessing defined during model training) to form a numerical feature vector, and calculates predictions in the form of a probability score vector. When the classification model 210 is initially trained, the data fields to be selected from the new transaction 206 are determined. The probability score vector contains a score for each possible transaction outcome. In one embodiment, each score can be a floating point number between 0 and 1, and the sum of all scores in the vector should equal 1. For example, if there are three different possible transaction outcomes O1, O2, O3, an example of a valid probability score vector (classification model prediction 220) can be a vector P (O1, O2, O3) = (0.35, 0.10, 0.55). Note that other embodiments of the probability score vector are possible, and the score for each possible transaction outcome can have different values; for example, each possible transaction outcome can be represented as an integer value between 0 and 100, and the sum of all scores in the vector should equal 100.
[0051] Fraud detection situations often require complex, so-called "stateful" feature vectors. This means that multiple entities (or "identifier variables") are tracked over time and become features of the input vector to the classifier. An example is the number of transactions using the same credit card in the past hour. In one embodiment, the classification model 210 and its associated training and validation models 230 utilize stateful feature vectors.
[0052] In one embodiment, the classification model 210 can be a supervised statistical model with a discrete target variable and a multidimensional numerical input vector. Many machine learning (ML) and artificial intelligence (AI) models can be used as the classification model 210. One requirement of the classification model 210 is that the model can be deployed in a streaming system and quickly generate predictions. For example, when the distinction between high-risk and low-risk result categories is known or relatively straightforward to make, a decision tree algorithm can be a good algorithm choice. As another example, when there are many (e.g., more than ten) features that can be extracted from transaction data, and the contribution of these features to the tendency to reach a given transaction result is not easy to analyze from the context, a more advanced classification learner such as a random forest, gradient enhancement, or an artificial neural network can be used. As another example, an unimportant classification model based on a single feature can be used, which directly links the result prediction to a specific market segment, merchant, or industry sector. Note that other ML or AI models can be used in the context of the present disclosure, and the selection of a specific ML or AI model as the classification model 210 does not change the overall operation of the system 200.
[0053] The treatment decision unit 212 uses the causal reasoning model 232 to quantify the effectiveness of each available treatment 214a-214c with respect to achieving the most favorable outcome, which can be expressed in monetary terms as a transaction-level value. For the purpose of explanation, the remainder of this discussion will be based on monetary terms in U.S. dollars. Note that the transaction-level impact (or "uplift") value of each outcome can be expressed in any currency or other monetary value without affecting the overall operation of the system 200.
[0054] The transaction-level impact value may be interpreted as the net “improvement” of selecting process T1 rather than letting the transaction pass without the process (e.g., process 214c). In one embodiment, the improvement value may include a confidence limit (e.g., an uncertainty range), or be calculated for each process T1. i The output of the causal inference model 232 for each incoming transaction 206 is a probability mass function of the lift value of each available transaction T with a specified probability. i. For example, Uplift(T1) = [(-$1.00, 0.25); ($0.00, 0.40); ($1.00, 0.35)]. As shown in this example, each probability score can be a floating point number between 0 and 1, and the sum of all probability scores in the matrix should equal 1. This example indicates that the causal inference model 232 predicts a negative uplift of $1.00 with a probability of 0.25, a zero uplift with a probability of 0.40, and a positive uplift of $1.00 with a probability of 0.35.
[0055] Historical transaction data from database 204 is used to train causal inference model 232. In some embodiments, causal inference model 232 can be retrained at regular intervals, and the retraining can be achieved automatically. In general, in order to successfully train and retrain a causal inference model, there should be observations for all treatments and across all outcomes, and preferably covering various "transaction segments". As used herein, a "transaction segment" is defined as an important subset in the feature space of pre-processing variables that are related to the outcome variable. For example, a simple segmentation can be based on which merchant the transaction belongs to. As another example, more advanced segmentation can use data-driven unsupervised methods to identify transaction clusters with observed similar behaviors.
[0056] For example, Blocked Randomized Trials can be used as the causal inference model 232. Dividing transactions into segments or "blocks" and randomly applying treatments within each segment can be a good way to quickly gain insight into what is the best treatment option for each segment.
[0057] As another example of a causal inference model 232, linear regression can be used with an indicator variable (0 / 1) for capturing the treatment applied and a continuous vector variable for all pre-treatment variables to account for suspected confounders. The advantage of using linear regression is that it is a straightforward approach, but given the nonlinearity it is computationally difficult to have linear regression in higher dimensions.
[0058] As another example of a causal inference model 232, a matched pairs method or nearest neighbor method may be used. With this type of model, a transaction is compared to similar past transactions. The idea behind this type of model is that the more similar two incoming transactions are to each other, the more likely their processing results are to be the same.
[0059] As another example of the causal inference model 232, a regression discontinuity design (RDD) can be used. In a more customized modeling approach, an RDD is a regression (linear or other) with a sharp transition at the decision point between applying a treatment and not applying a treatment. As used herein, the classification model prediction 220 can be used as a continuous independent variable ("assignment variable"), and the net result value (estimate) in monetary form at the transaction level is used as a dependent variable. For n available treatments (n>1), n-1 separate RDD models will be trained. The benefit of using RDD as a causal inference model 232 is that all confounding variables that are believed to have an impact on the final transaction outcome have been captured in the classification model 210. In addition, the use of a score threshold in the decision to apply treatments 214a-214c provides a natural, sharp transition between untreated and treated observations required in the RDD.
[0060] The treatment decision unit 212 is configured to decide which treatment 214a-214c should be applied, if any, for all new transactions 206. By combining the predictions from the classification model 210 and the causal inference model 232, the treatment decision unit 212 acts as a high-level decision lever intended to control the transaction outcome 208. The treatment decision unit 212 proceeds in two steps: (1) it runs the causal inference model 232 to predict the transaction-specific benefit (e.g., net uplift) of applying each of the treatments 214a, 214b relative to not applying any treatment (treatment 214c); (2) it combines the treatment predictions with the classification predictions 220 to make a decision based on a set of logical rules as to what (if any) treatment to apply. At run time, there is no "human-in-the-loop" involved in the decision making process made by the treatment decision unit 212.
[0061] The decision logic used by the treatment decision unit 212 depends on the system application and user preferences. In one embodiment, the decision logic can be a fixed, finite set of rules that take inputs from the classification model 210 and the causal inference model 232 and produce a single, clear, automatic decision about which treatment 214a, 214b to apply (including the option of no treatment; treatment 214c). The decision logic may include any one or more of the following: comparing the classification score with a set threshold; comparing the treatment improvement value with a set threshold; calculating the expected value of each treatment 214a-214c; comparing the uncertainty value band of each treatment improvement with a set threshold; calculating the expected value of all outcomes; selecting the treatment 214a-214c with the highest expected improvement value; or selecting the treatment 214a-214c randomly or with a fixed probability. It should be noted that the decision logic may include fewer, more, or different rules and achieve similar results as described herein.
[0062] In one embodiment, these rules can be combined in a decision tree to form an implementable decision function consisting of multiple "if / then" statements and suitable for computing real-time decisions. The decision function in the treatment decision unit 212 controls the way in which the system 200 makes decisions between treatment types 214a-214c. The thresholds and logic in the decision function are determined (and modified with appropriate frequency) in an offline analysis based on the past performance of both the classification model 210 and the causal reasoning model 232, as well as user preferences regarding how many transactions receive which treatment(s). A "user" can be, for example, an acquirer hosting the transaction analysis unit 202 and / or a merchant submitting a new transaction 206.
[0063] Considerations in establishing decision rules may involve how long the system 200 has been running, how much history has been established in the database 204, and the observed prevalence of various treatments and results. For example, for a binary result, the decision tree may initially decide between a single treatment and no treatment. In this example, if the classification score prediction for the result O1 exceeds 0.80, the decision logic may always apply treatment T1, and apply treatment T1 to all other transactions with a probability of 5%; and apply treatment T2 in all other cases. In this example, the result O1 may represent a high-risk transaction result that can be avoided with treatment T1, while treatment T2 may represent a low-impact, low-cost treatment. This example rule allows the collection of treatment data without allowing high-risk transactions to pass. At a later point in time, additional rules may be added, such as "if the expected value of the treatment improvement is greater than $0.10, only apply treatment T1". By setting up the decision logic in this way, the system 200 is flexible in how the model output is used and produces transparent results.
[0064] Figure 3 is a flow chart of an example decision logic 300 used by the treatment decision unit 212. In this example, there are two possible transaction outcomes (O1 and O2), and there are three available treatments (T1, T2, and T3, where T3 is a "no treatment" option). A decision tree can be used as the decision logic 300. P(O1) represents the predicted probability of the outcome O1 from the classification model 210. E(Uplift(T2)) represents the expected value and associated probability of the calculated uplift value of treatment T2, and E(Uplift(T3)) represents the expected value and associated probability of the calculated uplift value of treatment T3.
[0065] Decision logic 300 represents a series of steps that lead to the selection of a treatment to be applied. Once decision logic 300 is established (in some embodiments, in an offline design step), it is automatically executed by treatment decision unit 212. A determination is made as to whether classification model 210 predicts a probability of outcome O1 greater than 0.8 (this is the "score threshold"; "operation 302). Note that the score threshold of 0.8 is an exemplary value, and the score threshold may be set to any value by the entity operating treatment decision unit 212. If the probability of outcome O1 is greater than 0.8 (operation 302, "yes" branch), then treatment T1 is applied (operation 304). If the predicted probability of outcome O1 is less than 0.8 (operation 302, "no" branch), then based on the prediction generated by causal inference model 232, treatment T1 is applied. Further checks are performed on the process lift values and associated probability vectors. In this example, the expected values of lift for processes T2 and T3 are compared (operation 306). If the expected value of lift for process T2 is greater than the expected value of lift for process T3 (operation 306, "yes" branch), process T2 is applied (operation 308). If the expected value of lift for process T2 is less than the expected value of lift for process T3 (operation 306, "no" branch), process T3 is applied (operation 310). Therefore, the process with the higher expected value of lift will be applied.
[0066] A more detailed example of the application of the decision logic 300 is as follows. A new transaction 206 arrives and the classification model 210 produces an outcome probability prediction vector P(O1, O2) = (0.3, 0.7). The causal inference model 232 running in the treatment decision unit 212 defaults to producing lift predictions Uplift(T1) = [(-$1.00, 0.4), (-$0.50, 0.6)], Uplift(T2) = [($0.50, 0.35), ($0.75, 0.65)], and Uplift(T3) = [($0.00, 1.0)], because T3 represents "no treatment". The decision logic 300 evaluates these values as follows. Because P(O1) = 0.3 (i.e., P(O1) < 0.8; operation 302), treatment T1 is not applied. E(Uplift(T2))=($0.50×0.35)+($0.75×0.65)=$0.66 and E(Uplift(T3))=$0.00×1.0=$0.00, so E(Uplift(T2))>E(Uplift(T3)) (operation 306) and so process T2 is applied to new transaction 206 (operation 308).
[0067] Return to reference Figure 2 In some embodiments, it may be beneficial to separate the classification model 210 from the causal inference model 232. The first benefit is that predicting the most likely outcome and the treatment impact separately enables the user to define a more fine-grained decision function. The relevance of this is that within a segment of high-risk transactions, different treatments may be appropriate. There may be high-risk transactions for which no effective treatment has yet been performed. Conversely, where treatment may be very effective for certain clusters of transactions, it may be the case that those transactions only pose a low to medium risk of becoming a harmful outcome such as fraud. All of this can be controlled in the treatment decision unit 212 because it receives and combines separate predictions of outcome probabilities and treatment impacts. For example, the direct treatment decision logic in the treatment decision unit 212 will only apply treatment when the outcome risk score exceeds a certain threshold and the treatment efficiency expressed in the lift prediction is above a certain threshold level. Alternatively, it may be a user's preference to have a rule that always applies treatment to outcome scores above a selected threshold, regardless of the estimated treatment impact.
[0068] A second benefit of having separate outcome and treatment impact scores is the ability to provide greater transparency and explainability of decisions to apply treatments. In cases where unexpected outcomes are observed, the framework will be able to show why and how the treatment decision was made.
[0069] A third benefit of having separate classification models 210 and causal inference models 232 is that the proposed modeling configuration can be simpler to build than an integrated model. For example, in an existing payment processing system that already has specific risk scoring rules, the causal inference model 232 and the processing decision unit 212 can be added, while the existing risk rules serve as the classification model 210.
[0070] Figure 4 400 is a flow chart of a method 400 for analyzing financial transactions of a processing application. In one embodiment and for the purpose of discussion, the method 400 may be Figure 2 Receive a new transaction for processing by the transaction analysis unit 202 (operation 402). The new transaction can be received in a variety of ways from an internal source or an external source (e.g., through a customer-facing application programming interface (API) or connected to an adjacent internal upstream processing system). For example, if the transaction analysis unit 202 is located at a card network (e.g., such as Figure 1 If a card network 108 is shown, the new transaction can be processed from a payment processor (e.g., Figure 1 Payment processor 106) is shown arriving.
[0071] The transaction is classified by the classification model 210 (operation 404). The classification model 210 takes the transaction as input and generates a prediction of the possible transaction outcomes.
[0072] The new transaction and the classification model prediction are sent to the treatment decision unit 212 (operation 406). The treatment decision unit 212 uses the causal reasoning model to analyze the new transaction and the classification model prediction to determine whether transaction treatment is required (operation 408). As described above, one or more of a plurality of possible transaction treatments may be applied to the new transaction, including the option of not applying any treatment to the new transaction.
[0073] The transaction receives processing (or no processing, depending on the decision in 408), and is then output back to the payment processing system (operation 410). This can be done in a variety of ways, and the output connection can be internal or external (e.g., through an application programming interface (API)). Depending on the type of processing that was experienced, the output transaction may contain more or fewer data points or modified data content than when the output transaction was received (operation 402). Downstream interactions of the transaction with other entities in the payment processing system (such as checks performed by the issuing bank) result in the transaction outcome.
[0074] Figure 5 is a flow chart of a method 500 for classifying transactions, which in some embodiments may be Figure 4 In one embodiment, method 500 may be performed by Figure 2 The classification model 210 is shown executed.
[0075] The classification model 210 receives a new transaction (operation 502). The new transaction can be received from an internal source or an external source (e.g., through a customer-facing application programming interface (API) or connected to an adjacent internal upstream processing system) in a variety of ways. For example, if the transaction analysis unit 202 is located at a card network (e.g., such as Figure 1 If a card network 108 is shown, the new transaction can be processed from a payment processor (e.g., Figure 1 Payment processor 106) is shown arriving.
[0076] Data fields are selected from the new transaction (operation 504) and transformed to form a numerical feature vector (operation 506). Because the data fields may vary between transactions (i.e., may not always be in a fixed format for all transactions), different data fields may be selected from different transaction types. Offline model preparation based on historical transaction samples defines which data fields are selected by the classification model 210. In a real-time system, the classification model 210 simply selects the appropriate fields and performs the necessary analytical or numerical transformations (operation 506).
[0077] Based on the digital feature vector, the classification model 210 calculates a probability score vector (operation 508). The probability score vector contains a score for each possible transaction outcome. In one embodiment, each score can be a floating point number between 0 and 1, and the sum of all scores in the vector should be equal to 1. For example, if there are three different possible transaction outcomes (O1, O2, O3), an example of a valid probability score vector can be a vector P(O1, O2, O3) = (0.35, 0.10, 0.55). Note that other embodiments of the probability score vector are possible, and the score for each possible transaction outcome can have different values; for example, each possible transaction outcome can be represented as an integer value between 0 and 100, and the sum of all scores in the vector should be equal to 100.
[0078] The probability score vector is output from the classification model 210 as the classification model prediction 220 (operation 510). The classification model prediction is sent to the processing decision unit 212 for further processing and is stored in the database 204.
[0079] Figure 6 is a flow chart of a method 600 for determining a process to be applied to a transaction, which in some embodiments may be Figure 4 In one embodiment, method 600 may be performed by Figure 2 The processing decision unit 212 is shown to execute.
[0080] A new transaction and a probability score vector for the new transaction are received at the processing decision unit 212 (operation 602). The new transaction may be received in a variety of ways from an internal source or an external source (e.g., via a customer-facing application programming interface (API) or connected to an adjacent internal upstream processing system). For example, if the transaction analysis unit 202 is located at a card network (e.g., such as Figure 1 If a card network 108 is shown, the new transaction can be processed from a payment processor (e.g., Figure 1 The probability score vector may be received from the classification model 210 (eg, the classification model prediction 220).
[0081] The treatment decision unit 212 analyzes the new transaction and the probability score vector to determine the net lift value of applying each of the plurality of treatments to the transaction (operation 604). The predicted transaction-specific value can be interpreted as the net "lift" of selecting treatment T1 versus allowing the transaction to pass without processing. In one embodiment, the lift value can include a confidence bound (e.g., uncertainty range) or probability mass function for the lift value, which is a probability mass function for each treatment T1. i Computational. Each available processing T for each input transaction can be i The net lift predictions are stored as t for each available process i A matrix of assigned probabilities. For example, Uplift(T1) = [(-$1.00, 0.25); ($0.00, 0.40); ($1.00, 0.35)]. As shown in this example, each probability score can be a floating point number between 0 and 1, and the sum of all probability scores in the matrix should equal 1.
[0082] The treatment to be applied to the transaction is determined by the treatment decision unit 212 based on the net lift and the probability score vector (operation 606). The treatment decision unit 212 combines the treatment prediction with the classification model prediction based on a set of logic rules to reach a decision as to what treatment (if any) to apply. The logic rules can be a fixed, finite set of rules that take the classification model prediction and the net lift prediction and produce a single, unambiguous, automatic decision as to which treatment to apply (including the option of no treatment). The treatment decision is applied to the transaction and output (operation 608), resulting in a transaction with a treatment decision ( Figure 2 222).
[0083] Processes 214a-214c apply the reasoning to the transaction, producing a transaction with a processing result 224. This output transaction may or may not contain new data fields, changed data fields, or removed data fields compared to the original transaction 206. The transaction is then forwarded to an adjacent internal system of the same entity or an external entity system. After the transaction has passed through all the processing steps of the payment processing network, the transaction reaches its final state known as transaction result 208. Transaction result 208 becomes a "label" used as a target variable in the supervised training of machine learning models (classification model 210 and causal inference model used by processing decision unit 212). To help train the model, the "new transaction" 206 is linked to its result label 208 before being submitted to the database 204.
[0084] According to an embodiment of the present disclosure, a non-transitory computer readable medium storing instructions for a processor may be provided, wherein the processor is configured to analyze a Figure 1 and Figure 2 The example system and above Figures 3 to 6 A flowchart of a transaction in a payment processing system. For example, instructions stored in a non-transitory computer-readable medium can be executed by a processor to partially or fully perform a process for analyzing transactions in a payment processing system. Common forms of non-transitory media include, for example, floppy disks, flexible disks, hard disks, solid-state drives, magnetic tapes or any other magnetic data storage media, compact disk read-only memories (CD-ROMs), any other optical data storage media, any physical media with a pattern of holes, random access memories (RAMs), programmable read-only memories (PROMs), erasable programmable read-only memories (EPROMs), FLASH-EPROMs or any other flash memory, non-volatile random access memories (NVRAMs), caches, registers, any other memory chips or boxes, and networked versions thereof.
[0085] Although the present disclosure has been shown and described with reference to specific embodiments, it will be appreciated that the present disclosure may be practiced in other environments without modification. The foregoing description is given for illustrative purposes. It is not exhaustive and is not limited to the precise forms or embodiments disclosed. Modifications and adaptations will be apparent to those skilled in the art in view of the specification and practice of the disclosed embodiments.
[0086] Computer programs based on written descriptions and disclosed methods are within the skill range of experienced developers. Various programs or program modules can be created using any technology known to those skilled in the art, or various programs or program modules can be designed in conjunction with existing software. For example, they can be designed in or with the aid of: .Net Framework, .Net Compact Framework (and related languages, such as Visual Basic, C, etc.), Java, C++, Objective-C, Python, R, Scala, Hypertext Markup Language (HTML), HTML / AJAX combinations, XMLX combinations, XML, or HTML with included Java applets.
[0087] In addition, although illustrative embodiments have been described herein, those skilled in the art will understand, based on this disclosure, that the scope of any and all embodiments has equivalent elements, modifications, omissions, combinations (e.g., combinations of aspects of various embodiments), adaptations and / or changes. The limitations in the claims should be broadly interpreted based on the language used in the claims and are not limited to the examples described in this specification or described during the application process. The embodiments should be understood to be non-exclusive. In addition, the steps of the disclosed method or portions of the steps of the disclosed method may be modified in any way, including by reordering steps, inserting steps, repeating steps, and / or deleting steps (including between steps of different exemplary methods). Therefore, the description and embodiments are to be considered illustrative only, with the true scope being represented by the full scope of the appended claims and their equivalents.
Claims
1. A method for analyzing transactions in a payment processing system, the method comprising: Receive transactions; classifying the transaction, including calculating a probability score vector for the transaction, wherein the probability score vector indicates a probability of each of one or more possible outcomes of the transaction; analyzing the transaction, including calculating one or more probability mass vectors for the transaction, wherein the one or more probability mass vectors indicate impact values and associated probabilities of one or more possible treatments to be applied to the transaction; selecting a process to be applied to the transaction, comprising applying a set of decision rules to the probability score vector and the one or more probability mass vectors; applying the selected process to the transaction; and After the selected processing is applied to the transaction, the transaction is output.
2. The method according to claim 1, wherein: The categories include: selecting one or more data fields from the transaction; transforming the selected one or more data fields to form a numerical feature vector; and Based on the numerical feature vector, the probability score vector is calculated.
3. The method according to claim 1, wherein: The one or more possible outcomes of the transaction include: the transaction is authorized; the transaction indicates fraud; the transaction indicates a chargeback; the transaction is not authorized; the transaction is canceled; or the transaction is a refund.
4. The method according to claim 1, wherein: The impact value includes a monetary boost for applying the selected treatment to the transaction.
5. The method according to claim 1, wherein: The one or more possible processing to be applied to the transaction include any one or more of the following: no processing; adjusting data fields in the transaction; performing a data security check; performing a customer security check; performing an anti-money laundering check; or performing a credit risk check.
6. The method according to claim 5, wherein: Adjusting the data fields in the transaction includes any one or more of the following: deleting a data field; editing a data field; adding a data field; or reordering the data fields in the transaction.
7. The method according to claim 1, wherein: The set of decision rules includes a set of logic rules.
8. The method according to claim 1, further comprising: The transaction outcome is fed back to a classification model, wherein the step of classifying the transaction includes applying the classification model to predict a probability of each of the one or more possible outcomes of the transaction.
9. The method according to claim 1, further comprising: The transaction results are fed back to a treatment impact training model, wherein the step of analyzing the transaction includes applying the treatment impact training model to predict the probability and impact value of applying each of the one or more possible treatments to the transaction.
10. A system for analyzing transactions in a payment processing system, the system comprising: at least one processor; as well as A non-transitory computer-readable medium containing instructions that, when executed by the at least one processor, cause the at least one processor to: Receive transactions; classifying the transaction, including calculating a probability score vector for the transaction, wherein the probability score vector indicates a probability of each of one or more possible outcomes of the transaction; analyzing the transaction, including calculating one or more probability mass vectors for the transaction, wherein the one or more probability mass vectors indicate impact values and associated probabilities of one or more possible treatments to be applied to the transaction; selecting a process to be applied to the transaction, comprising applying a set of decision rules to the probability score vector and the one or more probability mass vectors; applying the selected process to the transaction; and After the selected processing is applied to the transaction, the transaction is output.
11. The system according to claim 10, wherein: The non-transitory computer readable medium contains further instructions for classifying, which when executed by the at least one processor cause the at least one processor to: selecting one or more data fields from the transaction; transforming the selected one or more data fields to form a numerical feature vector; and The probability score vector is calculated based on the numerical feature vector.
12. The system according to claim 10, wherein: The one or more possible outcomes of the transaction include: the transaction is authorized; the transaction indicates fraud; the transaction indicates a chargeback; the transaction is not authorized; the transaction is canceled; or the transaction is a refund.
13. The system according to claim 10, wherein: The one or more possible processing to be applied to the transaction include any one or more of the following: no processing; adjusting data fields in the transaction; performing a data security check; performing a customer security check; performing an anti-money laundering check; or performing a credit risk check.
14. The system according to claim 13, wherein: The non-transitory computer-readable medium contains additional instructions for adjusting data fields in the transaction, which, when executed by the at least one processor, cause the at least one processor to perform any one or more of the following: deleting data fields; editing data fields; adding data fields; or reordering data fields in the transaction.
15. The system according to claim 10, wherein: The non-transitory computer readable medium contains further instructions that, when executed by the at least one processor, cause the at least one processor to: The transaction outcome is fed back to a classification model, wherein classifying the transaction includes applying the classification model to predict a probability of each of the one or more possible outcomes of the transaction.
16. The system of claim 10, wherein: The non-transitory computer readable medium contains further instructions that, when executed by the at least one processor, cause the at least one processor to: The transaction results are fed back to a treatment impact training model, wherein the operation of analyzing the transaction includes applying the treatment impact training model to predict a series of impact values and associated probabilities of applying each of the one or more possible treatments to the transaction.
17. A transaction analysis unit for analyzing transactions in a payment processing system, the transaction analysis unit comprising: at least one processor; as well as A non-transitory computer-readable medium containing instructions that, when executed by the at least one processor, cause the at least one processor to: Receive transactions; classifying the transaction, including calculating a probability score vector for the transaction, wherein the probability score vector indicates a probability of each of one or more possible outcomes of the transaction; analyzing the transaction, including calculating one or more probability mass vectors for the transaction, wherein the one or more probability mass vectors indicate impact values and associated probabilities of one or more possible treatments to be applied to the transaction; selecting a process to be applied to the transaction, comprising applying a set of decision rules to the probability score vector and the one or more probability mass vectors; applying the selected process to the transaction; and After the selected processing is applied to the transaction, the transaction is output.
18. The transaction analysis unit according to claim 17, wherein: The transaction analysis unit is located at any one or more of the following: a payment processor; a card network; or a card-issuing bank.
19. The transaction analysis unit according to claim 17, wherein: The non-transitory computer readable medium contains further instructions that, when executed by the at least one processor, cause the at least one processor to: The transaction outcome is fed back to a classification model, wherein classifying the transaction includes applying the classification model to predict a probability of each of the one or more possible outcomes of the transaction.
20. The transaction analysis unit according to claim 17, wherein: The non-transitory computer readable medium contains further instructions that, when executed by the at least one processor, cause the at least one processor to: The transaction result is fed back to a treatment impact training model, wherein the operation of analyzing the transaction includes applying the treatment impact training model to predict the probability and impact value of applying each of the one or more possible treatments to the transaction.
