Big data push processing method and system based on Internet finance

Abstract

According to the big data push processing method and system based on Internet finance, the financial product interaction data flow of the target user on the Internet finance platform is obtained, the discrete interaction trajectory vector of each interaction data unit is extracted, and a fine-grained interaction model between the user and the financial product is effectively constructed. Furthermore, a global interaction mode of financial product interaction data flow is captured by generating a transmission interaction trajectory vector, and a target interaction trajectory vector is generated by combining a discrete interaction trajectory vector of an xth interaction data unit, so that dynamic and global understanding of user interaction behaviors is realized; the interest point prediction can be accurately carried out on the xth interaction data unit, the corresponding interest point prediction label is generated, and accurate information pushing is carried out on the target user according to the interest point prediction label. Therefore, the pertinence and effectiveness of information pushing are improved, the user experience is improved, and meanwhile, a more intelligent user service strategy is provided for an internet financial platform.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a big data push processing method and system based on Internet finance. Background Art

[0002] With the rapid development of Internet finance, the types and number of financial products are increasing, and the interactive behaviors of users on Internet financial platforms are becoming more complex and diverse. In order to improve user experience and enhance user stickiness, Internet financial platforms need to understand users' interactive behaviors more accurately and push personalized information based on users' interests.

[0003] Traditional information push methods are often based on users' static attributes or historical behavior data, which makes it difficult to capture users' real-time interactive behaviors and interest changes at the current moment. Especially in the field of Internet finance, users' interactive behaviors often have the characteristics of time series and periodic updates, and traditional static analysis methods are difficult to accurately reflect users' dynamic interactive patterns. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a big data push processing method based on Internet finance, the method comprising:

[0005] Obtaining a financial product interaction data stream of a target user on an Internet financial platform, wherein the financial product interaction data stream includes x interactive data units that are updated in stages in time series, where x is a positive integer;

[0006] Obtaining discrete interaction trajectory vectors corresponding to x interaction data units in the financial product interaction data stream, wherein the discrete interaction trajectory vector is an independent interaction trajectory vector of each interaction data unit;

[0007] Generate a transfer interaction trajectory vector based on the discrete interaction trajectory vectors corresponding to the first y interaction data units in the financial product interaction data stream, where the transfer interaction trajectory vector is a global interaction trajectory vector of the financial product interaction data stream, and y is a positive integer less than x;

[0008] generating a target interaction trajectory vector corresponding to the xth interaction data unit based on the transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the xth interaction data unit;

[0009] Based on the target interaction trajectory vector corresponding to the x-th interaction data unit, perform interest point prediction on the x-th interaction data unit, generate an interest point prediction tag corresponding to the x-th interaction data unit, and push information to the target user based on the interest point prediction tag corresponding to the x-th interaction data unit.

[0010] On the other hand, an embodiment of the present invention also provides a big data push processing system based on Internet finance, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application effectively constructs a fine-grained interaction model between users and financial products by acquiring the target user's financial product interaction data stream on the Internet financial platform and extracting the discrete interaction trajectory vector of each interaction data unit. Furthermore, by generating a transfer interaction trajectory vector to capture the global interaction mode of the financial product interaction data stream, and generating a target interaction trajectory vector in combination with the discrete interaction trajectory vector of the xth interaction data unit, a dynamic and global understanding of user interaction behavior is achieved, and the points of interest of the xth interaction data unit can be accurately predicted, and the corresponding point of interest prediction label can be generated, and accurate information can be pushed to the target user accordingly. Thereby, the pertinence and effectiveness of information push are improved, the user experience is enhanced, and at the same time, a more intelligent user service strategy is provided for the Internet financial platform, which enhances the user stickiness and competitiveness of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the big data push processing method based on Internet finance provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of the hardware architecture of a big data push processing system based on Internet finance provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The present invention is a flowchart of a method for pushing and processing big data based on Internet finance provided by an embodiment of the present invention. The method for pushing and processing big data based on Internet finance is introduced in detail below.

[0015] Step S110, obtaining a financial product interaction data stream of a target user on an Internet financial platform, wherein the financial product interaction data stream includes x interaction data units that are updated periodically in time series, where x is a positive integer.

[0016] In this embodiment, it is assumed that there is a target user named Mr. G on an Internet financial platform called "Investment Platform A". Mr. G participates in interactive activities of various financial products on the platform. These financial products include fund investment, fixed deposits, and purchase of financial products.

[0017] When Mr. G logged into the platform for the first time and viewed the details page of various fund products, this operation was recorded as an interaction data unit. This interaction data unit includes information such as the time of the operation (e.g., 10:00 on January 1, 2023), the type of operation (viewing fund details), the name of the fund product involved (e.g., "Growth Fund A"), and the result of the operation (no purchase or investment operation was performed).

[0018] As time goes by, Mr. G may perform more operations. For example, he deposited a sum of money into his time deposit account at 14:00 on January 5, 2023. This operation is also recorded as an interaction data unit, which contains information such as operation time, operation type (time deposit), deposit amount (for example, 5,000 yuan), etc.

[0019] Assume that within a period of time, Mr. G has performed 10 different types of financial product interaction operations on the platform, then x here is equal to 10. These interaction data units are arranged in chronological order, forming Mr. G's financial product interaction data flow on the "Investment Platform A" Internet financial platform. Each interaction data unit reflects a state of interaction between Mr. G and a financial product at a specific point in time, and the ordered set of these interaction data units fully describes the trajectory of Mr. G's financial product interaction behavior on the platform.

[0020] Step S120, obtaining discrete interaction trajectory vectors corresponding to each of the x interaction data units in the financial product interaction data stream, wherein the discrete interaction trajectory vector is an independent interaction trajectory vector of each of the interaction data units.

[0021] Let's continue with Mr. G's interaction on the "Investment Platform A" platform as an example. For the first interactive data unit, that is, the operation of viewing fund details at 10:00 on January 1, 2023. First, the entire financial product interaction data stream is cleaned of interference features. In this example, there may be some interference features, such as the platform's advertising push records, some system prompt records that are not related to the financial product interaction itself, etc. After removing these interference feature data, the financial product interaction data stream after the interference feature cleaning is obtained.

[0022] Then, based on the cleaned data stream, the interactive state transfer feature corresponding to the interactive data unit is obtained. For the operation of viewing fund details, the interactive transfer nodes may include entering the fund details page, browsing the fund's historical return chart, viewing the fund's management team information, etc. Assuming that in the data stream after the interference feature is cleaned, it is found that the interactive transfer event corresponding to the interactive transfer node of browsing the fund's historical return chart is the largest (for example, the longest stay time, the highest operation frequency, etc.), then this node is determined as the target interactive transfer node.

[0023] Taking this target interactive transfer node as the reference transfer node, set A = 2 interactive transfer nodes in the forward direction (for example, enter the fund details page, click on the fund risk level to view), and B = 1 interactive transfer node in the backward direction (for example, view the fund's investment strategy). In this way, the interactive transfer node set corresponding to this interactive data unit is generated.

[0024] Next, the interactive transfer events corresponding to this interactive transfer node set are decomposed to generate multiple interactive data sub-units. Assume that three interactive data sub-units are obtained after decomposition, and each sub-unit contains interactive transfer events corresponding to C = 2 interactive transfer nodes. For each interactive data sub-unit, the interactive state transition features are obtained based on its corresponding interactive transfer events. For example, for an interactive data sub-unit containing two interactive transfer events, namely, entering the fund details page and clicking to view the fund risk level, its interactive state transition features may include the time interval from entering the page to clicking to view the risk level, the changes in page elements before and after clicking to view the risk level, etc.

[0025] For each interaction data sub-unit in the interaction data unit, obtain the corresponding gain coefficient. Assume that the gain coefficient of the first interaction data sub-unit is 0.3, the second is 0.4, and the third is 0.3. Then, based on these gain coefficients, weighted calculation is performed on the interaction state transition features corresponding to each interaction data sub-unit to generate a first weighted feature.

[0026] For each interactive data subunit, the time memory flow characteristics between each target interactive data subunit (here refers to the subsequent interactive data subunit) corresponding to the interactive data subunit are obtained. Taking the first interactive data subunit as an example, its first boundary information is determined, including the starting and ending ranges of the interactive transfer events it contains (for example, from entering the fund details page to clicking on the fund risk level to view the end), and the definition information of the relevant attributes involved in these interactive transfer events (for example, the page elements involved, operation permissions, etc.). For the target interactive data subunit, its second boundary information is determined, including the interactive transfer event range, the definition information of the relevant attributes, and the relative relationship information with the first interactive data subunit. Based on these boundary information, the associated range information is determined, and then the interactive transfer events in the interactive data subunit are analyzed, the position and importance of each interactive transfer event in the entire interactive data subunit are identified, and the first event memory structure is constructed. At the same time, based on the associated range information, the position and importance of each interactive transfer event in the target interactive data subunit are identified, and the second event memory structure is constructed. Through a series of calculations and analyses, the quantified time memory flow parameters are obtained, the time memory flow feature model is constructed, and the time memory flow features are extracted.

[0027] Finally, the second weighted features corresponding to each interaction data sub-unit in the interaction data unit are fused to generate the third weighted feature. The first weighted feature and the third weighted feature are fused to generate the discrete interaction trajectory vector corresponding to this interaction data unit. In the same way, the discrete interaction trajectory vectors corresponding to all 10 interaction data units of Mr. G on the platform can be obtained.

[0028] Step S130, based on the discrete interaction trajectory vectors corresponding to the first y interaction data units in the financial product interaction data stream, generate a transfer interaction trajectory vector, the transfer interaction trajectory vector is the global interaction trajectory vector of the financial product interaction data stream, and y is a positive integer less than x.

[0029] Assume that in Mr. G's 10 interaction data units, y = 5. For the discrete interaction trajectory vectors corresponding to the first 5 interaction data units, first obtain the gain coefficient corresponding to each discrete interaction trajectory vector. For example, the gain coefficient corresponding to the discrete interaction trajectory vector of the first interaction data unit is 0.15, the second is 0.18, the third is 0.2, the fourth is 0.22, and the fifth is 0.25.

[0030] Then, the five discrete interaction trajectory vectors are transformed into feature space to generate their corresponding mapping interaction trajectory vectors. This feature space transformation may involve converting the original discrete interaction trajectory vector from one representation space to another space that is more suitable for fusion calculation. For example, the original discrete interaction trajectory vector based on multiple dimensions such as time and operation type is converted into a vector space representation based on a certain mathematical model.

[0031] Based on the gain coefficients obtained previously, the mapping interaction trajectory vectors corresponding to each of the five interaction data units are fused. The specific fusion method may be weighted summation or the like. Assume that after calculation, a new vector is generated. This vector is the transmission interaction trajectory vector, which integrates the information of the first five interaction data units and reflects the characteristics of Mr. G's financial product interaction trajectory at this stage from a global perspective.

[0032] Step S140: generating a target interaction trajectory vector corresponding to the xth interaction data unit based on the transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the xth interaction data unit.

[0033] For example, in Mr. G's example, x = 10, and the discrete interaction trajectory vector corresponding to the 10th interaction data unit has been obtained in step S120. Now integrate the transfer interaction trajectory vector generated in step S130 and the discrete interaction trajectory vector corresponding to the 10th interaction data unit. The integration method may be to combine or operate the two vectors on the corresponding dimensions. For example, addition operations are performed on certain dimensions, weighted average operations are performed on other dimensions, and so on. After such an integration operation, the new vector generated is the target interaction trajectory vector corresponding to the 10th interaction data unit. This target interaction trajectory vector contains both the global interaction trajectory information of the first 5 interaction data units and the independent interaction trajectory information of the 10th interaction data unit itself.

[0034] Step S150, based on the target interaction trajectory vector corresponding to the xth interaction data unit, predict the point of interest for the xth interaction data unit, generate a point of interest prediction tag corresponding to the xth interaction data unit, and push information to the target user based on the point of interest prediction tag corresponding to the xth interaction data unit.

[0035] Taking Mr. G’s 10th interaction data unit as an example, the corresponding target interaction trajectory vector contains rich information. The point of interest prediction label may be predicted based on multiple factors. Assume that on this Internet financial platform, the logic of the point of interest prediction label is constructed as follows:

[0036] If some dimensions in the target interaction trajectory vector indicate frequent viewing of high-risk and high-yield financial products, and a tendency to invest a small amount of money in high-risk products in recent operations, and more consultation interactions with platform customer service about high-risk products, then the point of interest prediction label may be "interested in high-risk and high-yield financial products and in the risk assessment and learning stage."

[0037] Assume that if the target interaction trajectory vector shows that the user frequently switches between time deposits and low-risk financial products and is more sensitive to interest rate fluctuations, then the predicted label of the point of interest may be "pay attention to low-risk stable return financial products and seek the optimal interest rate."

[0038] According to the target interaction trajectory vector corresponding to the 10th interaction data unit, after complex calculations and logical judgments, the corresponding point of interest prediction label is predicted to be "interested in high-risk and high-yield financial products and in the risk assessment and learning stage."

[0039] Based on this interest point prediction tag, the platform can push information to Mr. G. For example, the platform can push some articles about high-risk financial product investment strategies, high-risk fund recommendations, and some risk control training courses to Mr. G. These information pushes are designed to meet the interests shown by Mr. G during the interaction with financial products, improve user experience, and also help the platform better serve users, promote the sales of financial products and users' financial management.

[0040] Based on the above steps, the embodiment of the present application effectively constructs a fine-grained interaction model between users and financial products by acquiring the target user's financial product interaction data stream on the Internet financial platform and extracting the discrete interaction trajectory vector of each interaction data unit. Furthermore, by generating a transmission interaction trajectory vector to capture the global interaction mode of the financial product interaction data stream, and generating a target interaction trajectory vector in combination with the discrete interaction trajectory vector of the xth interaction data unit, a dynamic and global understanding of user interaction behavior is achieved, and the points of interest of the xth interaction data unit can be accurately predicted, and the corresponding point of interest prediction label can be generated, and accurate information can be pushed to the target user accordingly. Thereby, the pertinence and effectiveness of information push are improved, the user experience is enhanced, and at the same time, a more intelligent user service strategy is provided for the Internet financial platform, which enhances the user stickiness and competitiveness of the platform.

[0041] In a possible implementation, step S120 may include:

[0042] Step S121 , performing interference feature cleaning on the financial product interaction data stream to generate an interference feature cleaned financial product interaction data stream, wherein the interference feature data is removed from the interference feature cleaned financial product interaction data stream.

[0043] In this embodiment, the interaction of Mr. G on the Internet financial platform "Investment Platform A" is taken as an example to explain the relevant steps in detail.

[0044] First, the interference features of the financial product interaction data stream are cleaned. In Mr. G's interaction data, interference features may include pop-up click records that are implanted in the page by the platform to promote its own business and are irrelevant to Mr. G's actual financial product interaction, or records of some maintenance prompts automatically performed by the system. After removing these interference feature data from the financial product interaction data stream, the financial product interaction data stream after interference feature cleaning is obtained. This cleaned data stream more purely reflects the interaction between Mr. G and financial products.

[0045] Step S122: based on the financial product interaction data stream cleaned of the interference features, obtaining the interaction state transfer features corresponding to each of the x interaction data units.

[0046] Step S123: acquiring discrete interaction trajectory vectors corresponding to each of the x interaction data units based on the interaction state transition features corresponding to each of the x interaction data units.

[0047] In a possible implementation manner, the interaction data unit includes interaction transfer events corresponding to multiple interaction transfer nodes.

[0048] Step S122 includes:

[0049] Step S1221, based on the financial product interaction data stream after the interference feature cleaning, obtain the target interaction transmission node corresponding to each of the x interaction data units, and the target interaction transmission node is used to represent the interaction transmission node with the largest interaction transmission event among the multiple interaction transmission nodes corresponding to the interaction data unit.

[0050] Step S1222, for each of the interaction data units, take the target interaction transmission node corresponding to the interaction data unit as the reference transmission node, obtain A interaction transmission nodes in the forward direction, and B interaction transmission nodes in the backward direction, and generate an interaction transmission node set corresponding to the interaction data unit, wherein A and B are preset positive integers.

[0051] Step S1223: acquiring the interaction state transfer feature corresponding to the interaction data unit based on the interaction transfer event corresponding to each interaction transfer node in the interaction transfer node set corresponding to the interaction data unit.

[0052] In a possible implementation, step S1223 may include:

[0053] Step S1223-1, decomposing the interactive transfer events corresponding to the interactive transfer node set to generate a plurality of interactive data sub-units, wherein the interactive data sub-units include interactive transfer events corresponding to C interactive transfer nodes.

[0054] Step S1223-2: for each of the interaction data sub-units, based on the interaction transfer event corresponding to the interaction data sub-unit, obtain the interaction state transfer feature corresponding to the interaction data sub-unit.

[0055] Step S1223-3: based on the interaction state transition features corresponding to each of the plurality of interaction data sub-units, obtain the interaction state transition feature corresponding to the interaction data unit.

[0056] In this embodiment, for each interaction data unit, after obtaining the interaction state transition feature, it is necessary to further determine the gain coefficients corresponding to each interaction data sub-unit. The determination of these gain coefficients may be based on factors such as the importance of the interaction transfer event in the interaction data sub-unit to the entire interaction data unit. For example, for an interaction data sub-unit containing an important operation (such as confirming a purchase), its gain coefficient may be higher. Based on the gain coefficients corresponding to each interaction data sub-unit in the interaction data unit, the interaction state transition features corresponding to each interaction data sub-unit in the interaction data unit are weighted and calculated to generate a first weighted feature.

[0057] For each interactive data sub-unit in the interactive data unit, obtain the time memory flow characteristics between each target interactive data sub-unit (here refers to the subsequent interactive data sub-unit) corresponding to the interactive data sub-unit. Taking the interactive data sub-unit that includes clicking to enter the financial product page and conducting risk assessment as an example, determine its first boundary information, including the starting and ending ranges of these two interactive transfer events (such as from clicking to enter the financial product page to completing the risk assessment), and the definition information of the relevant attributes involved in these interactive transfer events (such as the risk level options involved in the risk assessment process, etc.). For the target interactive data sub-unit (such as the interactive data sub-unit that includes entering the purchase amount and confirming the purchase), determine its second boundary information, including the interactive transfer event range, the definition information of the relevant attributes, and the relative relationship information with the previous interactive data sub-unit (such as the order relationship, etc.). Based on these boundary information, determine the associated range information, then analyze the interactive transfer events in the interactive data sub-unit, identify the position and importance of each interactive transfer event in the entire interactive data sub-unit, and construct the first event memory structure. At the same time, based on the associated range information, identify the position and importance of each interactive transfer event in the target interactive data sub-unit, and construct the second event memory structure. Through a series of calculations and analyses, the quantified time memory flow parameters are obtained, the time memory flow feature model is constructed and the time memory flow features are extracted.

[0058] The second weighted features corresponding to each interaction data sub-unit in the interaction data unit are fused to generate the third weighted feature. Finally, the first weighted feature and the third weighted feature are fused to generate the discrete interaction trajectory vector corresponding to this interaction data unit. According to this process, the discrete interaction trajectory vectors corresponding to all x interaction data units of Mr. G on the platform can be obtained.

[0059] Take an interactive data unit of Mr. G as an example, for example, a financial product purchase operation he performed on February 1, 2023. This interactive data unit contains interactive transmission events corresponding to multiple interactive transmission nodes, which may include operations such as logging into the platform, searching for financial products, viewing financial product details, conducting risk assessments, entering purchase amounts, and confirming purchases. Each operation is an interactive transmission event. First, the target interactive transmission node corresponding to this interactive data unit must be obtained. In this example, assuming that the interactive transmission event corresponding to the interactive transmission node for viewing the details of financial products is the largest, the "interactive transmission event is the largest" here can mean that in this interactive data unit, the data flow involved in the operation of viewing the details of financial products is the largest (for example, the amount of data loaded when the page stays is the largest) or the operation duration is the longest. Then the interactive transmission node for viewing the details of financial products is determined as the target interactive transmission node.

[0060] For this interactive data unit, the target interactive transfer node of viewing the details of the financial product is used as the reference transfer node, assuming that the preset A = 3, B = 2. Then obtain 3 interactive transfer nodes in the forward direction, namely, logging into the platform, searching for financial products, and clicking to enter the financial product page; and 2 interactive transfer nodes in the backward direction, namely, conducting risk assessment and entering the purchase amount, so as to generate the interactive transfer node set corresponding to this interactive data unit.

[0061] Based on the interactive transfer events corresponding to each interactive transfer node in this interactive transfer node set, the interactive state transfer features corresponding to this interactive data unit are further obtained. First, the interactive transfer events corresponding to the interactive transfer node set are decomposed to generate multiple interactive data sub-units. Assume that C = 2, that is, each interactive data sub-unit includes interactive transfer events corresponding to 2 interactive transfer nodes. For example, an interactive data sub-unit includes two interactive transfer events, namely, logging into the platform and searching for financial products, and another interactive data sub-unit includes two interactive transfer events, namely, clicking to enter the financial product page and conducting risk assessment. For each interactive data sub-unit, the interactive state transfer features are obtained based on the corresponding interactive transfer events. For the interactive data sub-unit that includes logging into the platform and searching for financial products, the interactive state transfer features may include the time interval from logging into the platform to starting to search for financial products, the matching degree between the keywords used when searching for financial products and the financial products recommended by the platform, etc.

[0062] Then, based on the interaction state transfer features corresponding to each of the multiple interaction data sub-units, the interaction state transfer features corresponding to the entire interaction data unit are obtained. For example, the interaction state transfer features corresponding to each interaction data sub-unit are weighted and summarized, and the weights may be determined based on factors such as the importance of each interaction data sub-unit in the entire interaction data unit. In this way, the interaction state transfer features corresponding to this interaction data unit are obtained. In the same way, all x interaction data units of Mr. G on the platform can be processed to obtain the interaction state transfer features corresponding to each interaction data unit.

[0063] In a possible implementation, step S123 includes:

[0064] Step S1231: for each of the interaction data units, obtain a gain coefficient corresponding to each of the interaction data sub-units in the interaction data unit.

[0065] In this embodiment, for example, Mr. G has an interactive data unit on the platform about the operation process of purchasing a certain fund product. This interactive data unit is decomposed into multiple interactive data sub-units, one of which contains two interactive transmission events: logging on to the platform and searching for fund products, and another interactive data sub-unit contains two interactive transmission events: viewing fund details and assessing risks. For the interactive data sub-unit containing logging on to the platform and searching for fund products, since the logging on to the platform is the starting operation of the entire interaction, it is relatively basic and important, and may be assigned a relatively low gain coefficient, such as 0.2; while the operation of searching for fund products is more critical for locating the target fund, and may be assigned a gain coefficient of 0.3. For the interactive data sub-unit containing viewing fund details and assessing risks, viewing fund details can obtain key information about the fund, and may be assigned a gain coefficient of 0.35. Assessing risks is very important for investment decisions, and may be assigned a gain coefficient of 0.45. The determination of these gain coefficients is derived by comprehensively considering multiple factors such as the importance of the interactive transmission event in the entire interactive data unit and the impact on subsequent operations.

[0066] Step S1232: Based on the gain coefficients corresponding to the interaction data sub-units in the interaction data unit, weighted calculation is performed on the interaction state transition features corresponding to the interaction data sub-units in the interaction data unit to generate a first weighted feature.

[0067] Continuing with the interactive data unit for purchasing fund products as an example, it is assumed that the interactive state transition features of the interactive data subunits including logging into the platform and searching for fund products include the time interval from logging in to searching (assuming 5 seconds), the matching degree between the search keywords and the platform recommended funds (assuming 80%), etc. These interactive state transition features are weighted according to the gain coefficient, such as the weighted value of the time interval is 5×0.2 = 1, and the weighted value of the matching degree is 80%×0.3 = 0.24. Similar weighted calculations are performed on the interactive state transition features of the interactive data subunits including viewing fund details and assessing risks, and then these weighted results are combined to generate the first weighted feature.

[0068] Step S1233, for each interaction data sub-unit in the interaction data unit, obtaining the time memory flow characteristics between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit, wherein the target interaction data sub-unit is used to represent the interaction data sub-unit after the interaction data sub-unit in the interaction data unit.

[0069] Step S1234 fuses the time memory flow features between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit to generate a second weighted feature corresponding to the interaction data sub-unit.

[0070] Step S1235: fusing the second weighted features corresponding to the respective interaction data sub-units in the interaction data unit to generate a third weighted feature.

[0071] Step S1236: Fusing the first weighted feature and the third weighted feature to generate a discrete interaction trajectory vector corresponding to the interaction data unit.

[0072] In this embodiment, it is assumed that there are three interactive data sub-units in the interactive data unit for purchasing fund products. After calculating their respective second weighted features, the weights are determined according to factors such as the importance of each interactive data sub-unit in the entire interactive data unit, and the third weighted feature is obtained by fusion calculation. Finally, the first weighted feature and the third weighted feature are fused to generate a discrete interaction trajectory vector corresponding to the interactive data unit. This discrete interaction trajectory vector integrates various information such as the interactive state transfer characteristics and time memory flow characteristics of each interactive data sub-unit in the interactive data unit, and comprehensively describes the characteristics of this interactive data unit in the entire financial product interaction process. In the same way, Mr. G's other interactive data units on the platform can be processed to obtain the discrete interaction trajectory vectors corresponding to each interactive data unit.

[0073] In a possible implementation, step S1233 includes:

[0074] Step S1233-1, determining first boundary information of the interactive data subunit, wherein the first boundary information includes the start and end ranges of the interactive transfer events contained in the interactive data subunit, and definition information of relevant attributes involved in the interactive transfer events.

[0075] Step S1233-2, determine the second boundary information of the target interaction data sub-unit, the second boundary information includes the definition information of the interaction transfer event range and related attributes of the target interaction data sub-unit, and the relative relationship information between the target interaction data sub-unit and the interaction data sub-unit.

[0076] Step S1233-3, based on the first boundary information and the second boundary information, determine the association scope information between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit, the association scope information covers various data contact information existing from the interaction data sub-unit to the target interaction data sub-unit, and the data contact information includes direct interaction transfer event association information and indirect contact information constructed by related attributes.

[0077] Step S1233-4: Analyze the interactive transfer events in the interactive data sub-unit according to the associated scope information, and identify the position and importance of each interactive transfer event in the entire interactive data sub-unit.

[0078] Step S1233-5, constructing a first event memory structure within the interactive data sub-unit according to the logical relationship between interactive transfer events, wherein the first event memory structure is presented in the form of a logical hierarchy diagram, wherein nodes represent interactive transfer events, and edges represent logical connections between interactive transfer events. For each interactive transfer event, the position of the interactive transfer event on the timeline is determined, and the time intervals between different interactive transfer events are analyzed to obtain structural information of the first event memory structure.

[0079] Step S1233-6, based on the associated scope information, identify the position and importance of each interactive transfer event in the target interactive data sub-unit, construct a second event memory structure by analyzing the logical relationship between each interactive transfer event, and determine the position of each interactive transfer event on the timeline, analyze the time interval between the interactive transfer events, and the time sequence relationship with the events in the interactive data sub-unit, and generate the structural information of the second event memory structure.

[0080] Step S1233-7, based on the structural information of the first event memory structure, the structural information of the second event memory structure and the associated range information, find the nodes and edges connected to the event memory structure of the target interactive data sub-unit, obtain the basic connection part, identify the memory transfer path along the basic connection part, and organize and classify the identified memory transfer paths according to the length, complexity, total time span, and time interval distribution of the memory transfer paths, and divide the memory transfer paths into different categories.

[0081] Step S1233-8, for each memory transfer path, determine the key nodes on the memory transfer path, and for each key node, calculate the time delay of the key node in the memory transfer process, where the time delay refers to the time difference from a starting interaction transfer event in the interaction data sub-unit to the key node.

[0082] Step S1233-9, analyzing the time interval distribution between different key nodes on the memory transfer path, and determining the time interval distribution characteristics of the memory transfer time interval between different key nodes.

[0083] Step S1233-10, calculate the overall time span of the memory transfer path, the overall time span is the total time from the starting interaction transfer event of the interaction data sub-unit to the final interaction transfer event of the target interaction data sub-unit, and the overall time span reflects the overall time cost of memorizing from the interaction data sub-unit to the target interaction data sub-unit.

[0084] Step S1233-11, quantize the calculated time delay, time interval distribution characteristics and overall time span to form quantized time memory flow parameters.

[0085] Step S1233-12, determining the basic structure of the target model according to the quantized time memory flow parameters, wherein the basic structure is a model based on a vector space, in which each dimension represents a defined time memory flow parameter.

[0086] Step S1233-13, normalize each time memory flow parameter, and fill the normalized time memory flow parameter into the corresponding dimension of the target model to construct a time memory flow feature model.

[0087] Step S1233-14, extracting the time memory flow feature from the constructed time memory flow feature model according to the predefined feature extraction rules.

[0088] In detail, taking the interactive data subunit containing viewing fund details and assessing risks as an example, its first boundary information is determined. The interactive transmission event contained in this interactive data subunit starts with clicking the button to view fund details and ends with completing the risk assessment submission. The definition information of the relevant attributes involved includes various types of information contained in the viewed fund details page (such as the fund's investment portfolio, historical returns, etc.), the type of risk assessment (such as different levels of assessment such as conservative and stable), etc. For the target interactive data subunit (assuming it is an interactive data subunit for entering the purchase amount and confirming the purchase), its second boundary information is determined. The scope of its interactive transmission event is from entering the purchase amount to confirming the purchase. The definition information of the relevant attributes includes the numerical range of the purchase amount, the payment method when confirming the purchase, etc. The relative relationship information between the interactive data subunit is a sequential relationship, that is, the purchase operation can only be performed after the risk assessment is completed.

[0089] Based on the first boundary information and the second boundary information, the association range information between the interactive data subunit and each target interactive data subunit corresponding to the interactive data subunit is determined. In this example, the direct interactive transmission event association information is that the purchase operation can only be performed after the risk assessment is passed, and the indirect contact information constructed by the relevant attributes includes the potential association between the risk assessment result and the purchase amount (such as a high risk tolerance may result in a large purchase amount). Based on the association range information, the interactive transmission events in the interactive data subunit are analyzed. For the event of viewing fund details, it is a basic operation for obtaining information in the entire interactive data subunit, which is more important and is at a lower level of the logical hierarchy diagram; assessing risk is a decision-making operation based on viewing details, which is more important and is at a higher level of the logical hierarchy diagram. According to the logical relationship between interactive transmission events, the first event memory structure in the interactive data subunit is constructed. For the event of viewing fund details, it occurs earlier on the timeline, and the time interval between the risk assessment is assumed to be 30 seconds. The time interval between different interactive transmission events and other information constitute the structural information of the first event memory structure.

[0090] Based on the associated scope information, identify the location and importance of each interactive transfer event in the target interactive data subunit. Entering the purchase amount is the pre-operation of the purchase operation, and confirming the purchase is the final operation, which is more critical. The second event memory structure is constructed by analyzing the logical relationship between each interactive transfer event, and the position of each interactive transfer event on the timeline is determined. The time interval between the interactive transfer events and the time sequence relationship with the events in the interactive data subunit are analyzed to generate the structural information of the second event memory structure. For example, the time interval between entering the purchase amount and confirming the purchase is assumed to be 10 seconds.

[0091] Based on the structural information of the first event memory structure, the structural information of the second event memory structure, and the associated range information, the nodes and edges connected to the event memory structure of the target interactive data subunit are found to obtain the basic connection part. The memory transfer path is identified along the basic connection part. For example, from viewing the fund details to entering the purchase amount to confirming the purchase may be a memory transfer path. According to the length of the memory transfer path (here are 3 interactive transfer events), complexity (due to different types of operations such as risk assessment and amount input, the complexity is high), total time span (from viewing details to confirming the purchase, assuming a total of 60 seconds), time interval distribution (viewing details to assessing risks 30 seconds, entering the amount to confirming the purchase 10 seconds), the identified memory transfer paths are sorted and classified, and the memory transfer paths are divided into different categories.

[0092] For each memory transfer path, determine the key nodes on the memory transfer path. In the above memory transfer path, viewing fund details, assessing risks, and confirming purchases may be key nodes. For each key node, calculate the time delay of the key node during the memory transfer process. For the starting key node of viewing fund details, the time delay is 0 seconds; for the key node of assessing risks, the time difference from viewing fund details, the starting interactive transfer event, to assessing risks is 30 seconds; for the key node of confirming purchases, the time difference from viewing fund details to confirming purchases is 60 seconds. Analyze the time interval distribution between different key nodes on the memory transfer path, and determine the time interval distribution characteristics of the time intervals of memory transfer between different key nodes. For example, the time interval from viewing details to assessing risks is longer, and the time interval from entering amount to confirming purchase is shorter. Calculate the overall time span of the memory transfer path, that is, the total time from the starting interactive transfer event (viewing fund details) of the interactive data subunit to the final interactive transfer event (confirming purchase) of the target interactive data subunit is 60 seconds. This overall time span reflects the overall time cost of memory from the interactive data subunit to the target interactive data subunit.

[0093] The calculated time delay, time interval distribution characteristics and overall time span are quantified to form quantified time memory flow parameters. For example, the time delay can be directly expressed in seconds, the time interval distribution characteristics can be expressed in the proportional relationship of the time intervals, and the overall time span is expressed in total seconds. According to the quantified time memory flow parameters, the basic structure of the target model is determined. This basic structure is a model based on vector space, in which each dimension represents a defined time memory flow parameter. For example, one dimension represents the time delay from viewing fund details to assessing risks, and one dimension represents the time interval from inputting an amount to confirming a purchase. Each time memory flow parameter is normalized, and the normalized time memory flow parameters are filled into the corresponding dimensions of the target model to construct a time memory flow feature model. From the constructed time memory flow feature model, the time memory flow features are extracted according to the pre-defined feature extraction rules.

[0094] In a possible implementation, step S130 includes:

[0095] Step S131: acquiring a gain coefficient corresponding to each of the first y interaction data units based on the discrete interaction trajectory vectors corresponding to each of the first y interaction data units.

[0096] For example, suppose Mr. G has 10 interaction data units in the financial product interaction data stream on the platform, where y = 5. For the first interaction data unit, its corresponding discrete interaction trajectory vector contains comprehensive feature information about Mr. G's browsing of various financial products (such as funds, bonds, etc.) after logging into the platform for the first time. The gain coefficient of this interaction data unit needs to consider multiple factors, such as the fundamental contribution of the operations in this interaction data unit to the overall financial product interaction behavior. If the operations in this interaction data unit are the first comprehensive browsing of the platform's financial products, this lays the foundation for subsequent in-depth interactions, so it may be assigned a relatively high gain coefficient, such as 0.2.

[0097] For the second interactive data unit, let’s assume that Mr. G conducts a detailed review and preliminary evaluation of a specific fund product. This operation has a certain guiding role in the entire interactive process and has a greater impact on subsequent operations. After comprehensive consideration, it may be assigned a gain coefficient of 0.22. The third interactive data unit is that Mr. G further consults the platform customer service about some questions about the fund product. This interactive data unit plays an important role in obtaining accurate information and may be assigned a gain coefficient of 0.23. The fourth interactive data unit is that Mr. G transfers a small amount of funds as a fund preparation for the subsequent purchase of financial products. Considering its contribution to fund preparation, it may be assigned a gain coefficient of 0.18. The fifth interactive data unit is that Mr. G once again checks the recent returns of the fund product. This operation helps the final investment decision and may be assigned a gain coefficient of 0.17.

[0098] Step S132 , performing feature space conversion on the discrete interaction trajectory vectors corresponding to the first y interaction data units to generate mapped interaction trajectory vectors corresponding to the first y interaction data units.

[0099] Taking the first interactive data unit as an example, its discrete interactive trajectory vector is a feature representation formed based on a series of previous operations (such as the time and order of browsing different financial product pages, the frequency of clicking on different sections, etc.). When performing feature space conversion, the original representation based on these operation features may be converted into a representation in another mathematical space. For example, the original discrete interactive trajectory vector is in a space with operation type, operation time, etc. as coordinate axes, and is converted to a space with new features as coordinate axes through a certain mathematical transformation (such as linear transformation or nonlinear transformation, which may be based on the conversion algorithm preset by the platform, and this algorithm may take into account the logical relationship between different operations and the distribution characteristics of the overall interactive data, etc.), thereby generating a corresponding mapping interactive trajectory vector. For the second interactive data unit, its discrete interactive trajectory vector contains the operational features of viewing different sections (such as investment portfolio, management team, etc.) of the specific fund product details page. When performing feature space conversion, it also follows the algorithm preset by the platform to convert these features into a new space and generate a new mapping interactive trajectory vector. The third, fourth and fifth interactive data units also generate corresponding mapping interactive trajectory vectors according to the feature content of their respective discrete interactive trajectory vectors through the same feature space conversion mechanism.

[0100] Step S134: Based on the gain coefficients corresponding to the first y interaction data units, the mapping interaction trajectory vectors corresponding to the first y interaction data units are merged to generate the transfer interaction trajectory vector.

[0101] Finally, according to the gain coefficient determined above, the mapping interaction trajectory vector corresponding to the first interaction data unit is multiplied by its gain coefficient 0.2, the mapping interaction trajectory vector corresponding to the second interaction data unit is multiplied by 0.22, the mapping interaction trajectory vector corresponding to the third interaction data unit is multiplied by 0.23, the mapping interaction trajectory vector corresponding to the fourth interaction data unit is multiplied by 0.18, and the mapping interaction trajectory vector corresponding to the fifth interaction data unit is multiplied by 0.17. Then the five multiplication results are fused, and the fusion method may be in the form of vector addition (of course, it may also be other fusion methods defined based on the platform algorithm, and this algorithm will take into account factors such as the status and mutual relationship of different mapping interaction trajectory vectors in the overall interaction). Through such a fusion operation, the final generated transmission interaction trajectory vector integrates the information of the first y interaction data units, and reflects the interaction trajectory characteristics of Mr. G's financial products at this stage (the interaction stage covered by the first y interaction data units) from a more macro perspective. This transmission interaction trajectory vector will serve as an important basic data for subsequent processing steps and is used for correlation analysis and other operations with other interaction data units.

[0102] In a possible implementation, step S140 includes: integrating the transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the x-th interaction data unit to generate a target interaction trajectory vector corresponding to the x-th interaction data unit.

[0103] In this embodiment, during Mr. G's interaction with financial products on the platform, the transmission interaction trajectory vector was obtained by processing the first y interaction data units. This vector integrates the information of the first y interaction data units after conversion and weighted fusion of different operation characteristics, and reflects the interaction trajectory characteristics of a certain stage as a whole. Now consider the xth interaction data unit, and its corresponding discrete interaction trajectory vector is an independent interaction trajectory representation formed based on the operation characteristics of the interaction data unit itself. For example, the xth interaction data unit is a discrete interaction trajectory vector composed of the relevant interaction information of Mr. G's final purchase operation on a specific financial product, which includes detailed features of operations such as purchase amount input, purchase confirmation, and payment method selection.

[0104] In order to generate the target interaction trajectory vector corresponding to the xth interaction data unit, the transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the xth interaction data unit need to be integrated. This integration process is an operation that organically combines the information of the two vectors. Specifically, during integration, operations may be performed according to various dimensions of the vectors based on a pre-set algorithm. For example, for some dimensions representing the operation type, if the transfer interaction trajectory vector represents the previous operation tendency of viewing and evaluating various financial products on this dimension, and the discrete interaction trajectory vector corresponding to the xth interaction data unit represents the current operation of purchasing financial products on this dimension, then during the integration process, the information of the two on this dimension may be merged according to certain rules, which may be weighted merging or integration based on logical relationships. For the dimension representing the operation time, the time information in the transfer interaction trajectory vector reflects the time distribution of the previous interaction operation, and the time information in the discrete interaction trajectory vector corresponding to the xth interaction data unit is the time feature of the current purchase operation. During integration, factors such as the time sequence and time interval of the two will be considered to merge these information together. By integrating the transmitted interaction trajectory vector and the discrete interaction trajectory vector corresponding to the x-th interaction data unit in various dimensions, a target interaction trajectory vector corresponding to the x-th interaction data unit is finally generated. This target interaction trajectory vector contains both the information of the previous overall interaction trajectory and the specific interaction information of the current x-th interaction data unit.

[0105] In a possible implementation, the financial product interaction data stream is extracted from initial interaction big data. The method further includes:

[0106] Step A110, removing the first y interaction data units in the initial interaction big data from the initial interaction big data to generate updated interaction big data.

[0107] Step A120, obtaining the interest point prediction labels corresponding to each of the remaining interaction data units in the updated interaction big data.

[0108] Step A130, based on the interest point prediction labels corresponding to the respective remaining interaction data units, the respective remaining interaction data units are divided to generate a plurality of interaction data unit clusters, each of the interaction data unit clusters corresponding to one of the interest point prediction labels.

[0109] Step A140 , extracting one of the remaining interaction data units from the plurality of interaction data unit clusters for multiple times, and combining them to generate a plurality of combination unit blocks.

[0110] Step A150, for each interest point prediction process description model, obtain the influence weight of the interest point prediction process description model under the limitation of each of the combined unit blocks, wherein the influence weight represents the positive contribution of the first remaining interactive data unit in the combined unit block, and the proportion of the positive contribution of the combined unit block.

[0111] Step A160, based on the influence weights of the interest point prediction process description model respectively defined in each of the combination unit blocks, obtain the target influence weight corresponding to the interest point prediction process description model.

[0112] Step A170, based on the target influence weights corresponding to each of the interest point prediction process description models, determine the interest point prediction process description model that matches the initial interaction big data.

[0113] Wherein, step A150 includes:

[0114] Step A151, for each remaining interaction data unit in the combination unit block, obtain the cumulative value of positive contribution between the remaining interaction data unit and the interest point prediction process description model, the cumulative value of positive contribution is used to represent the sum of positive contributions between each interaction transmission event in the remaining interaction data unit and the interest point prediction process description model under the limitation of the interest point prediction label corresponding to the remaining interaction data unit.

[0115] Step A152, adding the cumulative values ​​of positive contributions between each remaining interaction data unit in the combination unit block and the interest point prediction process description model respectively to generate a first addition result.

[0116] Step A153, dividing the accumulated value of the positive contribution between the first remaining interaction data unit and the interest point prediction process description model by the first addition result, to obtain the influence weight of the interest point prediction process description model under the limitation of the combination unit block.

[0117] First, the financial product interaction data stream is extracted from the initial interaction big data. In Mr. G’s case, the initial interaction big data contains the original data related to all his financial product interactions on the “Investment Platform A” platform. These data come from a wide range of sources and cover his various operations after logging into the platform, such as browsing different financial product pages, checking interest rates, viewing product details, performing fund operations (depositing, withdrawing, transferring, etc.), and interacting with the platform’s customer service.

[0118] Next, the first y interaction data units in the initial interaction big data are removed from the initial interaction big data to generate the updated interaction big data. Assuming y = 5, for Mr. G, these first five interaction data units may be some basic operations after he initially logged into the platform, such as browsing the platform welcome page when logging in for the first time, checking basic account information, and initially browsing the financial product categories in the platform navigation bar, as well as simply checking some popular recommended financial products. After removing these five interaction data units, the remaining interaction data units constitute the updated interaction big data.

[0119] Then, the interest point prediction labels corresponding to each of the remaining interaction data units in the updated interaction big data are obtained. Taking one of Mr. G's remaining interaction data units as an example, it is assumed that this interaction data unit is his operation of conducting a detailed study on a specific high-risk fund product, including viewing the fund's historical return trend, investment portfolio, management team and other detailed information, and continuing to pay attention to the fund's dynamics for a period of time thereafter. Based on these operations, the interest point prediction label may be determined as "having an in-depth research interest in high-risk fund products and paying attention to long-term return potential." For other remaining interaction data units, such as Mr. G's re-evaluation operation of the fixed deposit product interest rate adjustment, the corresponding interest point prediction label may be "paying attention to the change and stability of the fixed deposit interest rate" and so on.

[0120] Based on the interest point prediction labels corresponding to each of the remaining interaction data units, each of the remaining interaction data units is divided to generate multiple interaction data unit clusters, each of which corresponds to an interest point prediction label. Continuing with Mr. G's example, all the remaining interaction data units marked as "interested in in-depth research on high-risk fund products and concerned about long-term return potential" will be divided into one interaction data unit cluster, and those interaction data units marked as "concerned about changes and stability in time deposit interest rates" will be divided into another interaction data unit cluster. In this way, the remaining interaction data units in the updated interaction big data are classified according to the interest point prediction labels.

[0121] After that, extract one remaining interactive data unit from each of the multiple interactive data unit clusters for multiple times, and combine to generate multiple combined unit blocks. For example, extract an interactive data unit from the interactive data unit cluster of "having an in-depth research interest in high-risk fund products and paying attention to the long-term return potential", and then extract an interactive data unit from the interactive data unit cluster of "paying attention to the change and stability of fixed deposit interest rates", and combine these two interactive data units together to form a combined unit block. By performing multiple operations in this way, multiple different combined unit blocks can be generated.

[0122] For each point of interest prediction process description model, obtain the influence weight of the point of interest prediction process description model under the limitation of each combination unit block. The point of interest prediction process description model here is a model structure used to predict points of interest for interactive data units, which contains a series of algorithms and rules to analyze various features in interactive data units to determine the point of interest prediction label. Taking a specific point of interest prediction process description model as an example, it is assumed that this model will focus on factors such as the frequency of operations, the amount involved in the operations, and the risk level of the operation object when analyzing interactive data units. For a combination unit block, it contains interactive data units extracted from different interactive data unit clusters.

[0123] For each remaining interactive data unit in the combined unit block, the cumulative value of the positive contribution between the remaining interactive data units and the interest point prediction process description model is obtained. Taking an interactive data unit from the interactive data unit cluster of "having in-depth research interest in high-risk fund products and paying attention to long-term return potential" in the combined unit block as an example, the interactive transfer events in this interactive data unit include operations such as viewing the historical return trend of high-risk funds and investment portfolios. Under the limitation of the interest point prediction label "having in-depth research interest in high-risk fund products and paying attention to long-term return potential" corresponding to the interactive data unit, the positive contribution between the interactive data unit and the interest point prediction process description model is calculated as follows: the interactive transfer event of viewing the historical return trend, because it is an important basis for analyzing the long-term return potential of high-risk fund products, according to the evaluation criteria for the importance of operations in the interest point prediction process description model, may be given a higher positive contribution, such as 0.3; the operation of viewing the investment portfolio is also critical for evaluating fund risks and returns, and may be given a positive contribution of 0.25. By adding up the positive contributions of each interaction transfer event in this interaction data unit, we can obtain the cumulative positive contribution value between the remaining interaction data unit and the interest point prediction process description model, which is assumed to be 0.55.

[0124] The cumulative positive contribution values ​​of the remaining interaction data units in the combined unit block and the interest point prediction process description model are added to generate the first addition result. Assuming that there is another interaction data unit from the "pay attention to the change and stability of fixed deposit interest rates" interaction data unit cluster in the combined unit block, and its cumulative positive contribution value is 0.4, then the first addition result is 0.55 + 0.4 = 0.95.

[0125] The cumulative value of the positive contribution between the first remaining interaction data unit and the interest point prediction process description model is divided by the first addition result to obtain the influence weight of the interest point prediction process description model under the combination unit block limitation. In the above example, the first remaining interaction data unit is an interaction data unit from the "in-depth research interest in high-risk fund products and attention to long-term return potential" interaction data unit cluster, and its positive contribution cumulative value is 0.55, so the influence weight of the interest point prediction process description model under this combination unit block limitation is 0.55 / 0.95 ≈ 0.579.

[0126] Based on the influence weights of the interest point prediction process description model under the constraints of each combination unit block, the target influence weight corresponding to the interest point prediction process description model is obtained. By performing the above calculations on multiple combination unit blocks, the influence weights of the interest point prediction process description model under different combination unit blocks are obtained, and then these influence weights are combined according to a certain algorithm (such as weighted average, etc.) to obtain the target influence weight.

[0127] Finally, based on the target influence weights corresponding to each point of interest prediction process description model, the point of interest prediction process description model that matches the initial interaction big data is determined. Assume that there are multiple point of interest prediction process description models, each model has its corresponding target influence weight. By comparing these target influence weights, the point of interest prediction process description model with the highest target influence weight is selected. This model is the point of interest prediction process description model that best matches the initial interaction big data. It can most accurately predict points of interest based on the interaction data units in the initial interaction big data, thereby providing the platform with better user interest analysis and service decision-making basis.

[0128] In a possible implementation, the method further includes:

[0129] Step S101, obtaining sample data to be learned, wherein the sample data to be learned includes x sample interaction data units that are updated periodically in time series, each of the sample interaction data units carries a marked interest point prediction label, and x is a positive integer.

[0130] In this embodiment, in the learning process of building a prediction network for points of interest in financial products, it is necessary to obtain sufficient sample data to be learned. For the "Investment Platform A" platform, these sample data to be learned come from the financial product interaction records of many users (including Mr. G) on the platform. Assume that the sample data to be learned here includes x sample interaction data units that are updated in stages in time series, and x is a positive integer. For example, for Mr. G's sample interaction data unit, it may contain his operation records on different financial products over a period of time. Each sample interaction data unit carries a tagged point of interest prediction label, which is determined in advance based on the user's actual behavior and clear business logic.

[0131] Take a sample interaction data unit of Mr. G as an example, assuming that this sample interaction data unit is data related to his purchase operation of a certain financial product. This sample interaction data unit contains detailed operation information, such as operation time, operation type (purchase), name of the financial product involved, purchase amount, etc. The corresponding annotated point of interest prediction label may be "tendency to purchase stable income financial products". This label is obtained by analyzing Mr. G's historical behavior, his other related operations on the platform (such as previous browsing of similar products, attention to income stability, etc.) and his basic account information (such as risk tolerance assessment results, etc.).

[0132] Step S102: using the financial product interest point prediction network, obtain the sample discrete interaction trajectory vectors corresponding to the x sample interaction data units in the sample data to be learned.

[0133] Next, the financial product interest point prediction network is a network model specifically used to analyze financial product interaction data and predict interest points. For each sample interaction data unit of Mr. G, the network generates a sample discrete interaction trajectory vector based on the various operation information contained therein.

[0134] Taking Mr. G's sample interaction data unit about the purchase of financial products as an example, the network will analyze the order of the operation time and other operations, the relationship between the operation type and his previous operation type (such as whether he has viewed similar financial products before), the purchase amount and the proportion of his account funds and historical investment amount, etc. After complex calculations and feature extraction, these factors are integrated into a sample discrete interaction trajectory vector. This vector comprehensively describes the characteristics of this sample interaction data unit in the entire financial product interaction process in a specific mathematical representation form.

[0135] Step S103, using the financial product interest point prediction network to generate a sample transfer interaction trajectory vector based on the sample discrete interaction trajectory vectors corresponding to the first y sample interaction data units in the sample data to be learned.

[0136] Then, assume that in Mr. G's sample data to be learned, y = 5. For the first five sample interaction data units, each unit has its corresponding sample discrete interaction trajectory vector. The financial product interest point prediction network will comprehensively consider the information of these five vectors to generate the sample transfer interaction trajectory vector.

[0137] For example, the first sample interaction data unit may be Mr. G's basic browsing operation after logging into the platform for the first time. Its sample discrete interaction trajectory vector contains information such as the sections he browsed and the duration of his stay; the second sample interaction data unit is his initial viewing operation of a certain type of fund product, including the type of fund viewed, viewing duration and other information. The network will process these sample discrete interaction trajectory vectors according to the preset algorithm. This algorithm may take into account factors such as the temporal sequence of each sample interaction data unit, the importance weight of different operations (such as operations that have a more direct impact on purchase decisions have a higher weight). By performing feature fusion, weighted calculation and other operations on these five sample discrete interaction trajectory vectors, a sample transfer interaction trajectory vector is finally generated. This sample transfer interaction trajectory vector reflects the overall characteristics of the interaction stage covered by the first five sample interaction data units from a more macro perspective.

[0138] Step S104: generating a target sample interaction trajectory vector corresponding to the x-th sample interaction data unit based on the sample transfer interaction trajectory vector and the sample discrete interaction trajectory vector corresponding to the x-th sample interaction data unit.

[0139] Assume that x = 10. For the 10th sample interaction data unit (for example, the sample interaction data unit related to Mr. G's large investment operation in a high-risk financial product), the corresponding sample discrete interaction trajectory vector contains unique information about this investment operation, such as a large investment amount and a high-risk investment product.

[0140] The financial product interest point prediction network will integrate the previously generated sample transfer interaction trajectory vector (combining the information of the first five sample interaction data units) and the sample discrete interaction trajectory vector corresponding to the 10th sample interaction data unit. This integration process may involve operations such as vector addition and multiplication, as well as weighted processing according to the importance of information in different dimensions. For example, for the dimension representing the operation type, if the sample transfer interaction trajectory vector reflects the comprehensive trend of multiple previous operation types in this dimension, and the sample discrete interaction trajectory vector corresponding to the 10th sample interaction data unit highlights the operation type of this high-risk investment in this dimension, the network will fuse these two pieces of information according to the preset rules. For the dimension representing the operation amount, the amount values ​​of the two and their relative importance in the entire interaction process will also be considered in a similar way for fusion. Through such operations, the target sample interaction trajectory vector corresponding to the 10th sample interaction data unit is finally generated.

[0141] Step S105, using the financial product interest point prediction network to predict interest points for the xth sample interaction data unit based on the target sample interaction trajectory vector corresponding to the xth sample interaction data unit, and generating a training interest point prediction label corresponding to the xth interaction data unit.

[0142] For the 10th sample interaction data unit (Mr. G's large investment operation in high-risk financial products), the financial product interest point prediction network will predict the interest point based on its corresponding target sample interaction trajectory vector.

[0143] The financial product interest point prediction network will analyze various information in the target sample interaction trajectory vector, such as a large amount of operation may indicate that Mr. G has high expectations for high returns, and the type of high-risk products invested indicates that he may have a high risk tolerance and is interested in high-risk and high-return financial products. Based on these analysis results, combined with the pre-built interest point prediction logic within the network (this logic is based on a large amount of historical data and financial business knowledge), the training interest point prediction label corresponding to the 10th sample interaction data unit is generated. Assume that this training interest point prediction label is "strong investment interest in high-risk and high-return financial products and high risk tolerance".

[0144] Step S106, based on the training interest point prediction label and the annotated interest point prediction label corresponding to the x-th sample interaction data unit, the parameters of the financial product interest point prediction network are optimized to generate an optimized financial product interest point prediction network.

[0145] For the 10th sample interaction data unit, its annotated POI prediction label is “strong investment interest in high-risk and high-return financial products and high risk tolerance” (assumed earlier), and the training POI prediction label generated by the financial product POI prediction network is also “strong investment interest in high-risk and high-return financial products and high risk tolerance”.

[0146] If the two are exactly the same, it means that the network is accurate in predicting this sample interaction data unit. However, if there is a difference, for example, the annotated interest point prediction label is "some interest in high-risk and high-return financial products but medium risk tolerance", while the training interest point prediction label is "strong investment interest in high-risk and high-return financial products and high risk tolerance", it means that the network has a deviation in prediction. At this time, the parameters of the financial product interest point prediction network will be adjusted according to the preset optimization algorithm. This optimization algorithm may be based on the principle of error back propagation, and calculate the degree of influence of each parameter on the prediction result according to the difference between the annotated interest point prediction label and the training interest point prediction label, and then adjust the parameters in the network accordingly. By performing such operations on a large number of sample interaction data units, the parameters of the financial product interest point prediction network are continuously optimized, and finally an optimized financial product interest point prediction network is generated. This optimized network can more accurately predict interest points based on the characteristics of the sample interaction data unit, thereby improving the prediction accuracy in practical applications.

[0147] Figure 2 The hardware structure of the big data push processing system 100 based on Internet finance for implementing the above-mentioned big data push processing method based on Internet finance provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the big data push processing system 100 based on Internet finance may include a processor 110 , a machine-readable storage medium 120 , a bus 130 and a communication unit 140 .

[0148] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the big data push processing system 100 based on Internet finance uses to execute or use to complete the exemplary method described in the present invention.

[0149] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the big data push processing method based on Internet finance in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0150] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned Internet finance-based big data push processing system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0151] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned big data push processing method based on Internet finance is implemented.

[0152] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A big data push processing method based on Internet finance, characterized in that: The method comprises: Obtaining a financial product interaction data stream of a target user on an Internet financial platform, wherein the financial product interaction data stream includes x interactive data units that are updated in stages in time series, where x is a positive integer; Obtaining discrete interaction trajectory vectors corresponding to x interaction data units in the financial product interaction data stream, wherein the discrete interaction trajectory vector is an independent interaction trajectory vector of each interaction data unit; Generate a transfer interaction trajectory vector based on the discrete interaction trajectory vectors corresponding to the first y interaction data units in the financial product interaction data stream, where the transfer interaction trajectory vector is a global interaction trajectory vector of the financial product interaction data stream, and y is a positive integer less than x; generating a target interaction trajectory vector corresponding to the xth interaction data unit based on the transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the xth interaction data unit; Based on the target interaction trajectory vector corresponding to the x-th interaction data unit, perform interest point prediction on the x-th interaction data unit, generate an interest point prediction tag corresponding to the x-th interaction data unit, and push information to the target user based on the interest point prediction tag corresponding to the x-th interaction data unit.

2. The big data push processing method based on Internet finance according to claim 1 is characterized in that: The obtaining of discrete interaction trajectory vectors corresponding to each of the x interaction data units in the financial product interaction data stream includes: Cleaning the financial product interaction data stream for interference features to generate a financial product interaction data stream cleaned for interference features, wherein the interference feature data is removed from the financial product interaction data stream cleaned for interference features; Based on the financial product interaction data stream cleaned from the interference features, obtaining the interaction state transfer features corresponding to each of the x interaction data units; Based on the interaction state transfer features corresponding to each of the x interaction data units, a discrete interaction trajectory vector corresponding to each of the x interaction data units is obtained.

3. The big data push processing method based on Internet finance according to claim 2 is characterized in that: The interactive data unit includes interactive transfer events corresponding to multiple interactive transfer nodes respectively; The obtaining of the interaction state transfer features corresponding to each of the x interaction data units based on the financial product interaction data stream cleaned from the interference features includes: Based on the financial product interaction data stream cleaned of interference features, obtaining a target interaction transfer node corresponding to each of the x interaction data units, wherein the target interaction transfer node is used to represent the interaction transfer node with the largest interaction transfer event among multiple interaction transfer nodes corresponding to the interaction data unit; For each of the interactive data units, taking the target interactive transfer node corresponding to the interactive data unit as the reference transfer node, obtaining A interactive transfer nodes in the forward direction and B interactive transfer nodes in the backward direction, and generating an interactive transfer node set corresponding to the interactive data unit, wherein A and B are preset positive integers; Based on the interaction transfer events corresponding to the respective interaction transfer nodes in the interaction transfer node set corresponding to the interaction data unit, the interaction state transfer feature corresponding to the interaction data unit is acquired.

4. The big data push processing method based on Internet finance according to claim 3 is characterized in that: The acquiring the interaction state transition feature corresponding to the interaction data unit based on the interaction transfer event corresponding to each interaction transfer node in the interaction transfer node set corresponding to the interaction data unit includes: Decomposing the interactive transfer events corresponding to the interactive transfer node set to generate a plurality of interactive data sub-units, wherein the interactive data sub-units include the interactive transfer events corresponding to C interactive transfer nodes; For each of the interaction data sub-units, based on the interaction transfer event corresponding to the interaction data sub-unit, acquiring the interaction state transfer feature corresponding to the interaction data sub-unit; Based on the interaction state transition features corresponding to each of the plurality of interaction data sub-units, the interaction state transition feature corresponding to the interaction data unit is acquired.

5. The big data push processing method based on Internet finance according to claim 4 is characterized in that: The acquiring, based on the interaction state transfer features corresponding to the x interaction data units, the discrete interaction trajectory vectors corresponding to the x interaction data units respectively, comprises: For each of the interaction data units, obtaining a gain coefficient corresponding to each of the interaction data sub-units in the interaction data unit; Based on the gain coefficients corresponding to the interaction data sub-units in the interaction data unit, weighted calculation is performed on the interaction state transition features corresponding to the interaction data sub-units in the interaction data unit to generate a first weighted feature; For each interaction data sub-unit in the interaction data unit, obtaining the time memory flow characteristics between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit; wherein the target interaction data sub-unit is used to represent the interaction data sub-unit after the interaction data sub-unit in the interaction data unit; fusing the time memory flow features between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit to generate a second weighted feature corresponding to the interaction data sub-unit; Fusing the second weighted features corresponding to each interaction data sub-unit in the interaction data unit to generate a third weighted feature; fusing the first weighted feature and the third weighted feature to generate a discrete interaction trajectory vector corresponding to the interaction data unit; The step of obtaining the time memory flow characteristics between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit includes: Determine first boundary information of the interactive data subunit, wherein the first boundary information includes a start range and an end range of an interactive transfer event included in the interactive data subunit, and definition information of relevant attributes involved in the interactive transfer event; Determine second boundary information of the target interactive data subunit, wherein the second boundary information includes definition information of an interactive transfer event range and related attributes of the target interactive data subunit, and relative relationship information between the target interactive data subunit and the interactive data subunit; Based on the first boundary information and the second boundary information, determining association scope information between the interaction data sub-unit and each target interaction data sub-unit corresponding to the interaction data sub-unit, wherein the association scope information covers various data contact information existing from the interaction data sub-unit to the target interaction data sub-unit, and the data contact information includes direct interaction transfer event association information and indirect contact information constructed by related attributes; Analyzing the interactive transfer events in the interactive data subunit according to the associated scope information, and identifying the position and importance of each interactive transfer event in the entire interactive data subunit; According to the logical relationship between the interactive transfer events, a first event memory structure in the interactive data subunit is constructed, wherein the first event memory structure is presented in the form of a logical hierarchical graph, wherein nodes represent interactive transfer events, and edges represent logical connections between interactive transfer events; For each interactive transfer event, determine the position of the interactive transfer event on the time axis, and analyze the time intervals between different interactive transfer events to obtain the structural information of the first event memory structure; And, based on the association scope information, identify the position and importance of each interactive transfer event in the target interactive data subunit, construct a second event memory structure by analyzing the logical relationship between each interactive transfer event, and determine the position of each interactive transfer event on the time axis, analyze the time interval between the interactive transfer events, and the time sequence relationship with the events in the interactive data subunit, and generate the structural information of the second event memory structure; Based on the structural information of the first event memory structure, the structural information of the second event memory structure and the associated range information, searching for nodes and edges connected to the event memory structure of the target interactive data subunit, obtaining a basic connection part, identifying a memory transfer path along the basic connection part, and sorting and classifying the identified memory transfer paths according to the length, complexity, total time span and time interval distribution of the memory transfer paths, and classifying the memory transfer paths into different categories; For each memory transfer path, determine the key nodes on the memory transfer path, and for each key node, calculate the time delay of the key node during the memory transfer process, where the time delay refers to the time difference from a starting interaction transfer event in the interaction data subunit to the key node; Analyze the time interval distribution between different key nodes on the memory transfer path, and determine the time interval distribution characteristics of the time interval of memory transfer between different key nodes; Calculate the overall time span of the memory transfer path, where the overall time span is the total time elapsed from the starting interaction transfer event of the interaction data subunit to the final interaction transfer event of the target interaction data subunit, and the overall time span reflects the overall time cost of memorizing from the interaction data subunit to the target interaction data subunit; The calculated time delay, time interval distribution characteristics and overall time span are quantified to form quantized time memory flow parameters; Determining a basic structure of a target model according to the quantized time memory flow parameters, wherein the basic structure is a model based on a vector space, wherein each dimension represents a defined time memory flow parameter; Normalize each time memory flow parameter, and fill the normalized time memory flow parameter into the corresponding dimension of the target model to construct a time memory flow feature model; From the constructed time memory flow feature model, the time memory flow features are extracted according to the predefined feature extraction rules.

6. The big data push processing method based on Internet finance according to claim 1 is characterized in that: The generating of the transfer interaction trajectory vector based on the discrete interaction trajectory vectors corresponding to the first y interaction data units in the financial product interaction data stream includes: Based on the discrete interaction trajectory vectors corresponding to the first y interaction data units, respectively, obtaining the gain coefficients corresponding to the first y interaction data units; Performing feature space conversion on the discrete interaction trajectory vectors corresponding to the first y interaction data units to generate mapped interaction trajectory vectors corresponding to the first y interaction data units; Based on the gain coefficients corresponding to the first y interaction data units, the mapping interaction trajectory vectors corresponding to the first y interaction data units are fused to generate the transfer interaction trajectory vector.

7. The big data push processing method based on Internet finance according to claim 1 is characterized in that: The generating a target interaction trajectory vector corresponding to the xth interaction data unit based on the transferred interaction trajectory vector and the discrete interaction trajectory vector corresponding to the xth interaction data unit includes: The transfer interaction trajectory vector and the discrete interaction trajectory vector corresponding to the x-th interaction data unit are integrated to generate a target interaction trajectory vector corresponding to the x-th interaction data unit.

8. The big data push processing method based on Internet finance according to claim 1 is characterized in that: The financial product interaction data stream is extracted from the initial interaction big data; the method further comprises: removing the first y interaction data units in the initial interaction big data from the initial interaction big data to generate updated interaction big data; Obtaining the interest point prediction labels corresponding to the remaining interaction data units in the updated interaction big data; Based on the interest point prediction labels corresponding to the remaining interaction data units, the remaining interaction data units are divided to generate a plurality of interaction data unit clusters, each interaction data unit cluster corresponding to one interest point prediction label; extracting one of the remaining interaction data units from the plurality of interaction data unit clusters for multiple times respectively, and combining them to generate a plurality of combination unit blocks; For each interest point prediction process description model, obtain the influence weight of the interest point prediction process description model under the limitation of each combination unit block, wherein the influence weight represents the positive contribution of the first remaining interaction data unit in the combination unit block, and accounts for the proportion of the positive contribution of the combination unit block; Based on the influence weights of the interest point prediction process description model respectively defined by each of the combination unit blocks, obtaining the target influence weight corresponding to the interest point prediction process description model; Determining a point of interest prediction process description model that matches the initial interaction big data based on the target influence weights corresponding to each of the point of interest prediction process description models; The step of obtaining the influence weights of the interest point prediction process description model under the limitation of each of the combined unit blocks includes: For each remaining interaction data unit in the combined unit block, an accumulated positive contribution value between the remaining interaction data unit and the interest point prediction process description model is obtained, wherein the accumulated positive contribution value is used to represent the sum of the positive contributions between each interaction transfer event in the remaining interaction data unit and the interest point prediction process description model under the limitation of the interest point prediction label corresponding to the remaining interaction data unit; Adding the cumulative values ​​of positive contributions between each remaining interaction data unit in the combined unit block and the interest point prediction process description model to generate a first addition result; The accumulated value of the positive contribution between the first remaining interaction data unit and the interest point prediction process description model is divided by the first addition result to obtain the influence weight of the interest point prediction process description model under the limitation of the combination unit block.

9. The big data push processing method based on Internet finance according to claim 1 is characterized in that: The method further comprises: Acquire sample data to be learned, wherein the sample data to be learned includes x sample interaction data units that are updated in stages in time series, each of the sample interaction data units carries a marked interest point prediction label, and x is a positive integer; Using the financial product interest point prediction network, obtaining sample discrete interaction trajectory vectors corresponding to each of the x sample interaction data units in the sample data to be learned; Generate a sample transfer interaction trajectory vector using the financial product interest point prediction network based on the sample discrete interaction trajectory vectors corresponding to the first y sample interaction data units in the sample data to be learned; Generate a target sample interaction trajectory vector corresponding to the x-th sample interaction data unit based on the sample transfer interaction trajectory vector and the sample discrete interaction trajectory vector corresponding to the x-th sample interaction data unit; Using the financial product interest point prediction network to predict interest points for the x-th sample interaction data unit based on the target sample interaction trajectory vector corresponding to the x-th sample interaction data unit, and generating a training interest point prediction label corresponding to the x-th interaction data unit; Based on the training interest point prediction label and the annotated interest point prediction label corresponding to the x-th sample interaction data unit, the financial product interest point prediction network is optimized to generate an optimized financial product interest point prediction network.

10. A big data push processing system based on Internet finance, characterized in that: The big data push processing system based on Internet finance includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the big data push processing method based on Internet finance as described in any one of claims 1 to 9 above.

Images