Object image information generation method, device, equipment and storage medium

Abstract

The present disclosure provides an object portrait information generation method, device, equipment and storage medium, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation, auxiliary driving, financial risk control, network security and the like. The method comprises the following steps: analyzing historical operation information to obtain a business operation information sequence and a corresponding time interval sequence; calling a trained operation sequence encoding network to perform encoding processing on the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features; respectively performing identification processing on the sequence operation features, and based on obtaining a plurality of corresponding operation identification sub-results, generating object portrait information of an object. Thus, in the case that the object data is not comprehensive, the object portrait information can be accurately generated, and the dependence on object data and scene labels is reduced.

Description

Technical Field

[0001] This disclosure relates to the field of computers, and in particular to a method, apparatus, device and storage medium for generating object profile information. Background Technology

[0002] In the internet field, object profiling information is generally used to characterize the features of an object. Object profiling information can assist in performing downstream tasks such as information recognition and media recommendation.

[0003] In related technologies, object profile information is generally generated based on acquiring a large amount of object data. However, when there is little object data or it is not easily acquired, this method will greatly reduce the accuracy of object profile information, and may even fail to effectively generate reliable object profile information. Summary of the Invention

[0004] This disclosure provides a method, apparatus, device, and storage medium for generating object profile information to solve at least one technical problem in the prior art.

[0005] On the one hand, this disclosure provides a method for generating object profile information, including:

[0006] Obtain historical operation information of an object, wherein the historical operation information represents the operation timing information of the object requesting the target business from at least one business operation platform;

[0007] The historical operation information is parsed to obtain a business operation information sequence and a corresponding time interval sequence; the business operation information sequence includes multiple business sub-operation identifiers corresponding to the chronological order of the operation times corresponding to the request for the target business; the time interval sequence represents the operation time difference between each adjacent business sub-operation identifier in the business operation information sequence;

[0008] The trained operation sequence encoding network is invoked to encode the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features.

[0009] Multiple trained operation recognition sub-networks are invoked to identify the sequence operation features respectively, resulting in multiple corresponding operation recognition sub-results. Different operation recognition sub-networks are used to indicate the identification of business indicator types in different dimensions, and each operation recognition sub-result represents the probability that the object performs a business operation under the corresponding business indicator type.

[0010] Based on the business indicator types corresponding to each of the multiple operation recognition sub-networks and the multiple operation recognition sub-results, the object profile information of the object is obtained.

[0011] On the other hand, an object profile information generation device is also provided, the device comprising:

[0012] The acquisition module is used to acquire the historical operation information of an object, wherein the historical operation information represents the operation timing information of the object requesting the target business from at least one business operation platform;

[0013] The parsing module is used to parse the historical operation information to obtain a business operation information sequence and a corresponding time interval sequence; the business operation information sequence includes multiple business sub-operation identifiers corresponding to the chronological order of the operation times corresponding to the request for the target business; the time interval sequence represents the operation time difference between each adjacent business sub-operation identifier in the business operation information sequence;

[0014] The first processing module is used to call the trained operation sequence encoding network to encode the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features.

[0015] The second processing module is used to call multiple trained operation recognition sub-networks to recognize and process the sequence operation features respectively, and obtain multiple corresponding operation recognition sub-results; different operation recognition sub-networks are used to indicate the recognition of business indicator types of different dimensions, and each operation recognition sub-result represents the probability that the object performs a business operation under the corresponding business indicator type;

[0016] The profile determination module is used to obtain the object profile information of the object based on the business indicator type corresponding to each of the multiple operation recognition sub-networks and the multiple operation recognition sub-results.

[0017] On the other hand, an electronic device is also provided, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement any of the above-described object portrait information generation methods.

[0018] On the other hand, a computer-readable storage medium is also provided, wherein at least one instruction or at least one program is stored in the storage medium, and the at least one instruction or the at least one program is loaded and executed by a processor to implement any of the above-described object portrait information generation methods.

[0019] On the other hand, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the object profile information generation methods described above.

[0020] The present disclosure provides a method, apparatus, device, and storage medium for generating object profile information, which has the following technical effects:

[0021] This embodiment of the disclosure obtains historical operation information of an object, which represents the timing information of the operation requested by the object from at least one business operation platform for a target business; it parses the historical operation information to obtain a business operation information sequence and a corresponding time interval sequence; the business operation information sequence includes multiple business sub-operation identifiers corresponding to the chronological order of the operation times requested for the target business; the time interval sequence represents the operation time difference between adjacent business sub-operation identifiers in the business operation information sequence; it calls a trained operation sequence encoding network to encode the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features; it calls multiple trained operation recognition sub-networks to recognize the sequence operation features respectively to obtain multiple corresponding operation recognition sub-results; different operation recognition sub-networks are used to indicate the recognition of different dimensions of business indicator types, and each operation recognition sub-result represents the probability that the object performs a business operation under the corresponding business indicator type; based on the business indicator types corresponding to the multiple operation recognition sub-networks and the multiple operation recognition sub-results, it obtains object profile information of the object. By combining time interval sequences to enhance the expression of sequence operation features in historical operation information, and generating object profile information based on the operation recognition sub-results corresponding to multiple operation recognition sub-networks, object profile information can be accurately generated even when sample labels are lacking or object data is incomplete, reducing the dependence on object data and scene labels. Attached Figure Description

[0022] To more clearly illustrate the technical solutions and advantages in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the application environment of an object profile information generation method provided in this embodiment of the disclosure;

[0024] Figure 2 This is a flowchart illustrating a method for generating object profile information provided in an embodiment of this disclosure;

[0025] Figure 3 This is a partial flowchart illustrating a method for generating object profile information provided in an embodiment of this disclosure;

[0026] Figure 4 This is a partial flowchart illustrating a method for generating object profile information provided in an embodiment of this disclosure;

[0027] Figure 5 This is a schematic diagram of an attention calculation step provided in an embodiment of this disclosure;

[0028] Figure 6 This is a schematic diagram of the training process of an operational sequence coding network provided in an embodiment of this disclosure;

[0029] Figure 7 This is a schematic diagram illustrating the training process of an operational sequence coding network provided in an embodiment of this disclosure;

[0030] Figure 8 This is a schematic diagram of the supervised training process provided in an embodiment of this disclosure;

[0031] Figure 9 This is a structural block diagram of an object portrait information generation device provided in an embodiment of this disclosure;

[0032] Figure 10 This is a schematic diagram of the hardware structure of an electronic device for implementing the methods provided in the embodiments of this disclosure. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0034] To make the objectives, technical solutions, and advantages of this disclosure clearer, the embodiments of this disclosure will be described in further detail below with reference to the accompanying drawings.

[0035] To facilitate understanding of the technical solutions described above and their resulting technical effects in the embodiments of this disclosure, the terms used in the embodiments of this disclosure are briefly introduced as follows:

[0036] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0037] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0038] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0039] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0040] Autonomous driving technology typically includes high-precision maps, environmental perception, behavior decision-making, path planning, motion control, and other technologies, and autonomous driving technology has broad application prospects.

[0041] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0042] The solutions provided in this disclosure relate to technologies such as computer vision and machine learning in artificial intelligence, and are specifically illustrated through the following embodiments:

[0043] The object profile information generation method disclosed herein can be applied to, for example, Figure 1 The application environment shown. For example... Figure 1 As shown, the hardware environment may include at least a terminal 110 and a server 120.

[0044] The aforementioned terminals 110 include, but are not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc.

[0045] The aforementioned server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal and server can be connected directly or indirectly via wired or wireless communication, and this disclosure does not impose any limitations. It should be noted that the aforementioned server 120 can be implemented as a cloud server in the cloud.

[0046] In some embodiments, the server 120 can also be implemented as a node in a blockchain system. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and cryptographic algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0047] It should be noted that, in practical applications, the object profile information generation method provided in this disclosure embodiment can be implemented in a server or terminal, or jointly implemented by a terminal and a server.

[0048] Of course, the methods provided in this disclosure are not limited to... Figure 1 The hardware environment shown can also be used in other possible hardware environments, and this disclosure does not limit the scope of the embodiments. Figure 1The functions that each device in the hardware environment shown can perform will be described in subsequent method embodiments, and will not be elaborated on here.

[0049] It should be noted that this disclosure involves user information and other related data. When the following embodiments of this disclosure are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0050] Figure 2 This is a flowchart illustrating a method for generating object profile information according to an embodiment of this disclosure. This disclosure provides the operational steps of the method described in the embodiments or flowchart, but based on conventional or non-inventive methods, it may include more or fewer operational steps. The order of steps listed in the embodiments is merely one possible execution order among many steps and does not represent the only possible execution order. The executing entity of this object profile information generation method can be the object profile information generation device provided in the embodiments of this disclosure, or an electronic device integrating the object profile information generation device, wherein the object profile information generation device can be implemented in hardware or software. Taking the above-mentioned executing entity as an example... Figure 1 Taking the server in the example of this, such as Figure 2 As shown, the method may include:

[0051] S201: Obtain the object's historical operation information, which represents the object's request for the operation sequence information corresponding to the target business from at least one business operation platform.

[0052] Here, "object" refers to an entity that initiates a target business request from at least one business operation platform. For example, the object could be an application account, enterprise, organization, or user. The business operation platform can be an application platform that manages or runs the target business. The target business refers to the relevant business used to characterize the object's profile information.

[0053] Both the business operation platform and the target business are matched with the application scenarios corresponding to the object profile information that needs to be characterized. For example, if the application scenario is for order anomaly detection, the business operation platform would be various e-commerce platforms, such as shopping apps, ticketing apps, and live streaming apps, and the target business would be order generation. If the application scenario is for transaction anomaly detection, the business operation platform would be various payment platforms, and the target business would be payment services. If the application scenario is for financial risk detection, the business operation platform would be various banking apps, consumer finance apps, etc., and the target business would be loan services, lending services, etc.

[0054] Specifically, historical operation information represents the timing information of an object requesting the target service from at least one business operation platform. That is, this historical operation information may only contain timing information related to the historical requests for the target service.

[0055] Optionally, this historical operation information can be reflected based on the request time of the target service and the corresponding service operation platform. For example, firstly, at time t1, object A requested services m and n on service operation platform 1; then at time t2, object A requested services n and p on service operation platform 2; finally, at time t3, object A requested services k and m on service operation platform 3. If the target service is service n, then the corresponding historical operation information 1 can be {t1 + service operation platform 1, t2 + service operation platform 2}; if the target service is service m, then the corresponding historical operation information 2 can be {t1 + service operation platform 1, t3 + service operation platform 3}.

[0056] Optionally, the server can obtain the historical application operation logs of each object to be identified reported by each business operation platform, filter out the objects to be identified from the historical application operation logs, and request the historical operation information corresponding to the target business from at least one business operation platform.

[0057] S203: Parse historical operation information to obtain a sequence of business operation information and the corresponding time interval sequence.

[0058] The business operation information sequence includes multiple business sub-operation identifiers corresponding to the chronological order of operation times corresponding to the requested target business. These business sub-operation identifiers are related to the business operation platform. Optionally, the business sub-operation identifier can be represented by the platform identifier of the business operation platform, such as the platform name or platform ID.

[0059] For example, regarding the aforementioned historical operation information 1, if its business sub-operation identifiers are, in chronological order, business operation platform 1 and business operation platform 2, then the corresponding business operation information sequence can be {business operation platform 1, business operation platform 2}. Similarly, regarding the aforementioned historical operation information 2, if its business sub-operation identifiers are, in chronological order, business operation platform 1 and business operation platform 3, then the corresponding business operation information sequence can be {business operation platform 1, business operation platform 3}.

[0060] The time interval sequence represents the operation time difference between adjacent business sub-operation identifiers in the business operation information sequence. For example, for the aforementioned historical operation information 1, if the adjacent business sub-operation identifiers in its business operation information sequence are t1 and t2, then the corresponding time interval sequence can be {t0, Δt1, t2, t1, t2, t2}. t2-t1}, where t0 is the initialization time interval or the default time interval, Δ t2-t1 Let t2 be the time difference between t1 and t2. For the aforementioned historical operation information 2, the adjacent business sub-operation identifiers in its business operation information sequence are t1 and t3, respectively. Therefore, the corresponding time interval sequence can be {t0, Δt1, t2, t3}. t3-t1}, where t0 is the initialization time interval or the default time interval, Δ t3-t1 This represents the time difference between t3 and t1.

[0061] Optionally, the server can parse historical operation information to obtain a sequence of business operation information and an initial time sequence in chronological order. It can then calculate the difference between adjacent operation times in the initial time sequence to obtain multiple time intervals and form a corresponding time interval sequence.

[0062] S205: Call the trained operation sequence encoding network to encode the business operation information sequence and the corresponding time interval sequence to obtain the sequence operation features.

[0063] The operation sequence encoding network is used to encode the business operation information sequence and the corresponding time interval sequence to extract relevant feature information from them, thus obtaining sequence operation features. These sequence operation features reflect the hidden information of the business operation information sequence and the corresponding time interval sequence.

[0064] Optionally, the server can input the business operation information sequence and the corresponding time interval sequence into a trained operation sequence encoding network for encoding processing to obtain sequence operation features. This operation sequence encoding network can be an encoding network capable of learning hidden information about time-related sequences, such as a Transformer-based encoder network or a BERT network.

[0065] In an alternative implementation, such as Figure 3 As shown, a trained operation sequence encoding network is invoked to encode the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features, including:

[0066] S301: The business operation information sequence is vector-encoded using the first vector encoding module in the trained business sequence encoding network to obtain the business embedding vector.

[0067] Optionally, the server takes the business operation information sequence as one input and feeds it into the first vector encoding module of the trained business sequence encoding network for vector encoding to obtain a business embedding vector. This business embedding vector is an initial vector representation of the business operation information sequence. For example, the embedding representation m of each business sub-operation identifier in the business operation information sequence can be determined by looking up a table or vector initialization. i Then, according to the chronological order of the operation times corresponding to each business sub-operation identifier, the embedded representations of each business sub-operation identifier are concatenated as embedding vector components to obtain a business embedding vector. The length of this business embedding vector is consistent with the sequence length of the business operation information sequence. This business embedding vector E I It can be represented as:

[0068]

[0069] Where, m i It refers to the embedded representation of the i-th business sub-operation identifier in the business operation information sequence, where n is the sequence length of the business operation information sequence.

[0070] Specifically, each business sub-operation identifier in the business operation information sequence can be numbered, and the embedding vector component corresponding to each business sub-operation identifier can be obtained by lookup table, thereby obtaining the business embedding vector corresponding to the business operation information sequence.

[0071] S303: The time interval sequence is vector-encoded by the second vector encoding module in the service sequence encoding network to obtain the time embedding vector.

[0072] Optionally, the server uses the time interval sequence as another input, inputting it into the second vector encoding module of the trained business sequence encoding network for vector encoding to obtain a time embedding vector. This time embedding vector is an initial vector representation of the time interval sequence. Specifically, a lookup table can be performed on each time interval in the time interval sequence to obtain the embedding vector component corresponding to each time interval, thereby obtaining the time embedding vector corresponding to the time interval sequence. By introducing the interval time between operation sequences, two objects with similar operation sequences may obtain different time embedding vectors due to different operation frequencies, thus obtaining different object profile information based on different time embedding vectors, improving the expressive power of object profile information.

[0073] In an optional implementation, before vector encoding the time interval sequence through the second vector encoding module in the traffic sequence coding network to obtain the time embedding vector, the method includes:

[0074] S3031: Normalize the time interval sequence to obtain the initial time series matrix.

[0075] Optionally, before vector encoding the time interval sequence, the minimum time interval in the object's time interval sequence can be determined, and the time difference between the operation times corresponding to every two time intervals in the time interval sequence can be calculated. Then, each time difference is normalized based on the minimum time interval to obtain the adjusted time intervals. Finally, the initial time series matrix is ​​obtained based on the adjusted time intervals.

[0076] For example, the server can first determine the minimum time interval in the sequence of time intervals of the object. And calculate the time difference t between the operation times corresponding to every two time intervals in the time interval sequence. i -t j And take the absolute value of that time difference |t i -t j Next, the absolute value of the time difference |t i -t j Minimum time interval with this object The ratio is rounded down to normalize the time differences, resulting in adjusted time intervals. Each adjusted time interval is then used as a matrix element to form the initial time series matrix M. u The initial time series matrix has an n×n dimension, where n is the sequence length of the business operation information sequence. For example, the initial time series matrix M... u Adjusted time interval It can be represented as:

[0077]

[0078] Among them, t i and t j It is the time of the i-th operation and the time of the j-th operation. It is the minimum time interval. It rounds down to the nearest integer.

[0079] S3033: Numerical truncation is performed on the initial time series matrix based on a preset truncation threshold to obtain an intermediate time series matrix.

[0080] Optionally, after determining the initial time series matrix, the matrix elements in the initial time series matrix can be numerically truncated based on a preset truncation threshold to obtain an intermediate time series matrix. For example, only the initial time series matrix M... u Matrix elements exceeding the preset truncation threshold are numerically truncated and can be replaced by the preset truncation threshold; there is no need to modify the initial time series matrix M.u Matrix elements smaller than or equal to the preset truncation threshold are numerically truncated, while retaining their individual matrix elements. The truncated matrix elements are then combined with the remaining matrix elements to form an intermediate time series matrix. This intermediate time series matrix reflects the discretized time intervals between pairs of operations in the business operation information sequence. The truncated intermediate time series matrix is ​​shown below. It can be represented as:

[0081]

[0082] Here, `clip()` refers to the numeric truncation function, and its specific truncation operation can be... In this context, k is a preset truncation threshold, which is an adjustable hyperparameter. Specifically, truncation can be performed using the preset truncation threshold k, reducing the adjusted time intervals that exceed this threshold. The time interval is truncated and replaced with the preset truncation threshold k, retaining the adjusted time interval that is less than or equal to the preset truncation threshold k.

[0083] In one optional implementation, the time interval sequence is vector-encoded by the second vector encoding module in the traffic sequence coding network to obtain a time embedding vector, including:

[0084] The intermediate time series matrix is ​​vector-encoded by the second vector encoding module in the business sequence coding network to obtain the time embedding vector.

[0085] Optionally, after normalizing and truncating the time interval sequence, the resulting intermediate time series matrix is ​​used as input to the second vector encoding module. This module performs vector encoding on each matrix element of the intermediate time series matrix to obtain the embedded representation of the time interval between any two operation times. The embedded representations corresponding to each matrix element are then concatenated as embedded vector components to obtain a time embedding vector. This time embedding vector can be a matrix representation; for example, the time embedding vector... It can be represented as:

[0086]

[0087] in, The embedding representation represents the time interval between the i-th operation time and the j-th operation time after embedding processing, where n is the sequence length of the time interval sequence.

[0088] Specifically, the table can be looked up for each processed time interval in the intermediate time series matrix to obtain the embedding vector component corresponding to each processed time interval, and then the time embedding vector corresponding to the intermediate time series matrix can be obtained.

[0089] S305: Through the encoding conversion module in the service sequence coding network, feature extraction is performed on the service embedding vector and the time embedding vector to obtain the sequence operation features.

[0090] Optionally, the server extracts features by inputting the business embedding vector and the time embedding vector into the encoding conversion module to obtain sequence operation features.

[0091] In one optional implementation, the encoding conversion module includes an attention mechanism module and a feedforward module. For example... Figure 4 As shown, the above-mentioned sequence operation features are obtained by extracting features from the service embedding vector and the time embedding vector through the encoding conversion module in the service sequence coding network, including:

[0092] S401: Through the attention mechanism module, based on the correlation between the various business vector components in the business embedding vector of the fused temporal embedding vector, the business embedding vector and the temporal embedding vector are processed to obtain the initial sequence operation features.

[0093] Optionally, when the server performs attention calculation through the attention mechanism module, it can fuse the time vector component corresponding to the time embedding vector into each business vector component in the business embedding vector, and process the initial sequence operation features based on the correlation between the various business vector components with fused time vector components. The correlation between these various business vector components refers to the degree of correlation between each business vector component in the business embedding vector and other business vector components; the higher the correlation, the higher the attention coefficient.

[0094] In one optional implementation, an attention mechanism module processes the service embedding vector based on the correlation between the various service vector components in the fused temporal embedding vector to obtain initial sequence operation features, including:

[0095] The transformation matrix in the attention mechanism module is used to transform each business vector component in the business embedding vector to obtain the query matrix, key matrix and value matrix corresponding to the business embedding vector.

[0096] The key-value matrix is ​​obtained by fusing the corresponding key-value components in the key-value matrix based on each time vector component in the time embedding vector.

[0097] Based on the similarity between the query matrix and the fused key-value matrix corresponding to the business embedding vector, attention weights are obtained, and the corresponding value matrices are weighted based on the attention weights to obtain the initial sequence operation features.

[0098] Optionally, the attention mechanism module is a multi-head self-attention mechanism. It transforms each business vector component in the business embedding vector using a transformation matrix to form query, key, and value matrices, i.e., obtaining the query matrix Q, key-value matrix K, and value matrix V corresponding to the business embedding vector. The transformation matrix may include the query transformation matrix W. Q Key value transformation matrix W K Sum transformation matrix W V Next, based on each time vector component in the time embedding vector, the corresponding key components in the key value matrix are fused to obtain the fused key value matrix.

[0099] Specifically, the time-interval-based fusion mechanism here is as follows: Figure 5 As shown, the key-value matrix is ​​obtained by adding and fusing the corresponding key-value components in the key-value matrix with each time vector component in the time embedding vector. Then, the initial attention weight is obtained based on the similarity between the query matrix corresponding to the business embedding vector and the fused key-value matrix. That is, for each query, its similarity to each key in the sequence is calculated to form the initial attention weight; the larger the initial attention weight, the greater the similarity between similar operations. This initial attention weight e... ij It can be represented as:

[0100]

[0101] Where, m i and m j Let the i-th service embedding vector and the j-th service embedding vector be... W is the time interval vector between the i-th operation time and the j-th operation time. K and W Q This is the transformation matrix for each business embedding vector. i W Q For m i and W Q The dot product between them represents the query matrix Q of the i-th business embedding vector; m j W K For m j and W K The dot product between them represents the key matrix K of the j-th business embedding vector; For m j W K and The sum of the positions of the two elements represents the fused key-value matrix K', where d is the dimension of the key-value matrix K.

[0102] Next, the initial attention weights e are... ij Normalization is performed, and the initial attention weights e are... ijNormalized to the [0,1] interval, the attention weight α is obtained. ij The attention weight α ij The result can be obtained by normalization using the following softmax function:

[0103]

[0104] Next, the corresponding value matrix is ​​weighted based on the attention weights to obtain the initial sequence operation features z. i Its calculation expression can be:

[0105]

[0106] Where, m i Let α be the embedding vector for the i-th business. ij For attention weights, W V m is the transformation matrix used to transform the business embedding vector. i W v For m i and W V The dot product between them represents the value matrix V of the i-th business embedding vector, where n is the sequence length of the business operation information sequence.

[0107] The above embodiments integrate time interval sequence information into attention weights, comprehensively considering the potential differential operation information contained in the uneven time intervals between adjacent operation behaviors in the operation sequence. For example, the operation time between operation platform 1 and operation platform 2 differs by 5 days, while the operation time between operation platform 2 and operation platform 3 differs by one month. By capturing differential operation information, the accuracy of the object profile information can be improved.

[0108] Furthermore, the mechanism of fusion time interval adds a weight coefficient to the pooling weighted or attention mechanism in classic sequence fusion, which is more specific in characterizing the head profile of the object, thus making the expressive power of the embedded vector richer and more comprehensive, realizing a more accurate nonlinear description of temporal features, and further improving the risk characterization ability of the sub-models of the subsequent identification network sub-models.

[0109] S403: The initial sequence operation features are processed by the feedforward module to obtain the sequence operation features.

[0110] The feedforward module can include a normalization layer, a residual layer, and a forward feedback layer. The feedforward module processes the initial sequence operation features output from the previous step through the normalization and residual layers, then connects to a forward feedback layer, followed by another normalization and residual layer, and finally a DropOut layer. Through this forward computation, it extracts and transforms the features in the business operation information sequence, generating a hidden layer matrix of the same length as the business operation information sequence, thus obtaining the sequence operation features.

[0111] S207: Call multiple trained operation recognition sub-networks to recognize and process the sequence operation features respectively, and obtain the corresponding multiple operation recognition sub-results.

[0112] Different operation recognition subnetworks are used to indicate the types of business indicators in different dimensions. These different dimensions of business indicator types are related to the application scenarios corresponding to the object profile information. For example, in an application scenario targeting order anomaly identification, the different dimensions of business indicator types could include indicators for successful order placement, fraudulent orders, and abnormal orders. In an application scenario targeting transaction anomaly identification, the different dimensions of business indicator types could include indicators for normal payments and indicators for illegal payments. In an application scenario targeting financial risk identification, the different dimensions of business indicator types could include indicators for banking scenarios, consumer finance scenarios, anti-fraud scenarios, and overdue payments.

[0113] Each operation identification sub-result represents the probability that an object will perform a business operation under the corresponding business indicator type. These operation identification sub-results may be represented by, but are not limited to, identification probability values ​​and identification sub-scores.

[0114] In one optional implementation, multiple trained operation recognition sub-networks are invoked to recognize and process the sequence operation features respectively, resulting in multiple operation recognition sub-results. This includes: inputting the sequence operation features into the multiple operation recognition sub-networks for recognition and processing, and outputting the corresponding multiple operation recognition sub-scores; and using the multiple operation recognition sub-scores as multiple operation recognition sub-results.

[0115] Optionally, the server inputs the sequence operation features into multiple trained operation recognition sub-networks respectively, performs recognition processing of the corresponding business indicator types, outputs multiple corresponding operation recognition sub-scores, and uses these multiple operation recognition sub-scores as multiple operation recognition sub-results.

[0116] S209: Based on the business indicator types corresponding to each of the multiple operation recognition sub-networks and the results of multiple operation recognition sub-networks, obtain the object profile information of the object.

[0117] The object profile information is used to characterize the features of an object performing a target business operation. Optionally, multiple operation identification sub-results can be combined or fused with the corresponding business indicator types of multiple operation identification sub-networks to obtain the object profile information. For example, the object profile information can be represented as {(business indicator type 1, operation identification sub-result 1), (business indicator type 2, operation identification sub-result 2), ..., (business indicator type n, operation identification sub-result n)}.

[0118] In one optional implementation, object profile information of an object is obtained based on the business indicator types corresponding to each of the multiple operation recognition sub-networks and multiple operation recognition sub-results. This includes: performing vector transformation on the multiple operation recognition sub-results according to a preset type order of the business indicator types corresponding to each of the multiple operation recognition sub-networks to obtain the object profile information of the object. Specifically, the multiple operation recognition sub-results are vector transformed, and then concatenated according to a preset type order of the business indicator types corresponding to each of the multiple operation recognition sub-networks to obtain object profile information represented in vector form. For example, if n operation recognition sub-results are operation recognition sub-scores, and the operation recognition sub-scores corresponding to business indicator type 1, business indicator type 2, business indicator type 3, and business indicator type 4 are f1, f2, f3, and f4 respectively, and the preset type order is {business indicator type 2, business indicator type 4, business indicator type 3, and business indicator type 1}, then the object profile information of the object can be (f2, f4, f3, f1).

[0119] In practical applications, taking the application scenario of financial risk identification as an example, it includes four operation identification sub-networks, which correspond to four types of business indicators: banking scenario business indicators, consumer finance scenario business indicators, anti-fraud business indicators, and overdue business indicators. The historical operation information of an object is processed by the operation sequence encoding network and these four operation identification sub-networks in sequence for identification. The four operation identification sub-scores are 0.5, 0.1, 0.2, and 0.2, respectively. Then the object profile information of the object can be represented as (0.5, 0.1, 0.2, 0.2).

[0120] The above embodiments combine time interval sequences to enhance the expression of sequence operation features in historical operation information, and obtain object profile information based on the operation recognition sub-results corresponding to multiple operation recognition sub-networks. In the case of missing sample labels or incomplete object data, object profile information can be accurately generated, reducing the dependence on object data and scene labels.

[0121] Furthermore, this disclosure can provide downstream applications, such as business insurance model development, with stable and generalizable risk profile information to supplement risk capabilities. In an era of strict regulation of personal privacy information, it can fully characterize the risk of the target without outputting underlying sensitive information by effectively compressing the raw data.

[0122] In an alternative implementation, such as Figure 6 and Figure 7 As shown, the operational sequence coding network is trained through the following steps:

[0123] S601: Obtain the first training sample, which includes first historical sample operation information corresponding to multiple first sample objects; the first historical sample operation information represents the timing information of sample operation corresponding to the target business requested by the first sample object from at least one business operation platform.

[0124] The first sample object is the object used to train the operational sequence encoding network. This object refers to an entity that initiates a target business request from at least one business operation platform; for example, the object could be an application account, enterprise, organization, user, etc. The business operation platform can be an application platform that manages or runs the target business. The target business refers to the relevant business used to characterize the object's profile information.

[0125] Both the business operation platform and the target business are matched with the application scenarios corresponding to the object profile information that needs to be characterized. For example, if the application scenario is for order anomaly detection, the business operation platform would be various e-commerce platforms, such as shopping apps, ticketing apps, and live streaming apps, and the target business would be order generation. If the application scenario is for transaction anomaly detection, the business operation platform would be various payment platforms, and the target business would be payment services. If the application scenario is for financial risk detection, the business operation platform would be various banking apps, consumer finance apps, etc., and the target business would be loan services, lending services, etc.

[0126] The first historical operation information represents the timing information of sample operations requested by the first sample object from at least one business operation platform to the target business. That is, the first historical operation information may only be timing information related to historical requests for the target business. Optionally, the first historical operation information can reflect the request operation time of the target business and the corresponding business operation platform.

[0127] S603: Analyze the operation information of each first historical sample to obtain the business sample operation information sequence and the corresponding sample time interval sequence of each first sample object.

[0128] The business sample operation information sequence includes multiple business sample sub-operation identifiers corresponding to the chronological order of operation times for the requested target business. These sub-operation identifiers are related to the business operation platform; optionally, they can be represented by the platform identifier of the business operation platform, such as the platform name or platform ID. The sample time interval sequence represents the operation time difference between adjacent business sample sub-operation identifiers in the business sample operation information sequence.

[0129] S605: Mask the target business sample sub-operation identifier in each business sample operation information sequence to obtain the corresponding masked sample operation information sequence.

[0130] Specifically, the target business sample sub-operation identifier can be determined from the sequence of operation information of each business sample, and these target business sample sub-operation identifiers can be masked, for example, by replacing the target business sample sub-operation identifier with [mask], or by replacing the target business sample sub-operation identifier with other sub-operation identifiers.

[0131] S607: The operation information sequence of each masked sample and the corresponding sample time interval sequence are processed by the initial operation sequence encoding network to predict the predicted business sub-operation identifier results corresponding to the masked business sample sub-operations in each masked sample operation information sequence.

[0132] Among them, the predicted business sub-operation identifier result represents the probability that the masked business sample sub-operation in the masked sample operation information sequence belongs to various types of business sample sub-operation identifiers.

[0133] In one optional implementation, the operation information sequence of each masked sample and the corresponding sample time interval sequence are processed by an initial operation sequence encoding network to predict the predicted service sub-operation identifiers corresponding to the masked service sample sub-operations in each masked sample operation information sequence, including:

[0134] Each masked sample operation information sequence and its corresponding sample time interval sequence are input into the initial operation sequence encoding network for encoding processing to obtain the target sub-operation features corresponding to the masked business sample sub-operations in each masked sample operation information sequence; wherein, in the attention mechanism in the encoding processing, the processing is performed by fusing the time embedding vectors corresponding to the sample time interval sequences.

[0135] The features of each target sub-operation are classified and processed to predict the predicted business sub-operation identifiers corresponding to the masked business sample sub-operations in each masked sample operation information sequence.

[0136] S609: Based on the training loss determined by the predicted business sub-operation identifier result and the corresponding target business sample sub-operation identifier, the initial operation sequence coding network is trained to obtain the operation sequence coding network.

[0137] In the training framework of the Operation Sequence Encoding Network, the aim is to learn general information about behavioral sequences without introducing labeled samples from specific scenarios. However, business operations do exhibit certain patterns. Taking the application scenario of financial risk identification as an example, the interval between two lending transactions on a banking platform may be relatively long, while small-amount lending transactions due to tight funding needs may occur relatively frequently. This is why user risks are inherent in behavioral sequences. This disclosure addresses the training process of the Operation Sequence Encoding Network by randomly masking a portion of the business sub-operation identifiers (i.e., platform information) corresponding to the business operations. After transforming the input sequence into the entire network, it can reasonably infer the masked business sub-operation identifiers. The training loss is determined based on the predicted business sub-operation identifier results and the corresponding target business sample sub-operation identifiers.

[0138] Optionally, the loss function for this training loss can be the cross-entropy between the predicted business sub-operation identifier result in the business operation information sequence and the target business sample sub-operation identifier before masking. Optionally, the target sub-operation features output by the operation sequence encoding network. After classification using the σ function (softmax operation), the predicted business sub-operation identifier is obtained. This predicted business sub-operation identifier represents the probability that the masked business sample sub-operation in the masked sample operation information sequence belongs to each type of business sample sub-operation identifier. Next, the target business sample sub-operation identifier before masking is used as the true label, and its probability in the corresponding predicted business sub-operation identifier result belonging to each type of business sample sub-operation identifier is calculated to obtain the corresponding cross-entropy. For example, the target business sample sub-operation identifier before masking can be used as the true label, and the sign function corresponding to this true label and the probability in the corresponding predicted business sub-operation identifier result belonging to each type of business sample sub-operation identifier can be logarithmically transformed, negativeened, and then averaged to obtain the cross-entropy. For example, the calculation expression of the loss function in each batch is:

[0139]

[0140] Where batch_size is the number of samples in the batch. σ represents the output of a sample sequence after passing through the network, and the σ function represents the softmax operation. For each target business sample sub-operation identifier, i.e., the real label, the sign function (0 or 1) is used. If the real category of sample i is equal to j, it is set to 1; otherwise, it is set to 0. m represents the total category of the business sample sub-operation identifier.

[0141] The model is trained using this training loss, primarily optimizing the prediction accuracy of sub-operations of masked business samples. Backpropagation is employed to update the parameters in the model, including the input vector in the embedding layer and the transformation matrix in the encoding transformation module, until training convergence is achieved, resulting in the operation sequence encoding network. During training, the model no longer relies on prior pre-classification of business operation platforms. Instead, it learns the contextual relationships, including operation time intervals, from the operation sequences. This allows for some differences in the representation of business operation platforms in different combinations of sequences, ultimately enriching the representation of object profile information and enabling full utilization of a large amount of operation sequence data to obtain sequence operation features.

[0142] In an alternative implementation, multiple operation recognition subnetworks are trained through the following steps:

[0143] Obtain a second training sample, which includes second historical sample operation information corresponding to multiple second sample objects, and corresponding indicator identification labels; the second historical sample operation information represents the time sequence information of sample operations corresponding to the target business requested by the second sample object from at least one business operation platform.

[0144] Multiple initial operation recognition subnetworks are obtained, and different initial operation recognition subnetworks are used to indicate the types of business indicators in different dimensions.

[0145] Multiple initial operation recognition subnetworks are trained in a supervised manner using a second training sample until the training termination condition is met, resulting in multiple operation recognition subnetworks.

[0146] Optionally, after obtaining the operation sequence encoding network through self-supervised training, when performing supervised modeling using a small number of labeled samples from specific scenarios, the input needs to be processed into the input format corresponding to the operation sequence encoding network. After input, the object feature vector representation is obtained. Then, supervised learning is performed using the labeled samples to obtain multiple operation recognition sub-networks. These multiple operation recognition sub-networks can be simple MLP layers or decision tree models, etc., and the parameters can be iteratively optimized by combining a binary classification loss function. In this way, by obtaining the operation sequence encoding network through self-supervised training and training multiple operation recognition sub-networks in a supervised manner, a polysemous representation of the business operation platform is achieved, and it has scene label-independent generalization performance, which is very suitable for scenarios where object data is scarce or difficult to obtain, or where scene labels are few.

[0147] like Figure 8As shown, during the training of the operation sequence coding network, iterative optimization is performed according to the designed loss function. This primarily optimizes the prediction accuracy of behaviors / events in the randomly masked parts of the sequence. The learned operation sequence coding network can output corresponding sequence codes for behavior sequences not involved in training, i.e., sequence operation features. During the training of multiple operation recognition sub-networks, fine-tuning is performed using different scenarios and labeled samples. From the historically accumulated labeled dataset, the user's historical behavior sequence is first traced back according to the application time, and then input into the pre-trained model from the previous step to obtain the encoding of the user's corresponding time slice. Then, supervised modeling is performed according to different scenarios and label definitions to learn the fitting of the user sequence encoding to specific labeled data samples. For example, the KS effect of each model's output score, the gain in modeling existing risk features, and the PSI distribution of feature values ​​for each historical slice are examined. Then, the process continues as follows... Figure 8 As shown, in the backtracking inference and deployment phase, this step involves backtracking inference and scoring of historical segments across all users after the previous step has confirmed that the feature gain of the fine-tuned model is significant and the cross-cycle score is stable. This yields recognition sub-scores for different scenarios and labels. For example... Figure 8 The system can identify the sub-scores of scenario 1, scenario 2, label a, and label b, and save the feature scores by slicing them according to different periods. The latest period's output identification score can be deployed online, and the stability of the feature can be monitored and maintained on a daily basis over a long period of time.

[0148] The object profile information, generated by the dimensionality-reduced object features predicted by the self-supervised operation sequence encoding network and obtained by the supervised operation recognition subnetwork, can assist other profile information to enhance the multidimensional representation of the object, facilitating downstream application development. For example, in the application scenario of financial risk identification, this object profile information can be used to supplement general profile information to accurately characterize the operational risk level of the object, and can be used as needed in the development of some risk control models, providing the ability to identify the object's debt risk without exposing the object's original information.

[0149] For example, in application scenarios for order anomaly identification or transaction anomaly identification, the object profile information generated by this disclosure can be used to supplement general profile information to accurately characterize the risk level of fraudulent orders or abnormal transactions of an object. This can effectively identify fraudulent orders, abnormal transactions on payment platforms, etc., reduce transaction risks, and improve the security of network services.

[0150] Compared to end-to-end sequence learning tasks, the operation sequence encoding network disclosed in this paper can extract the underlying expression logic from large-scale data by constructing a self-supervised training objective. It has strong generalization performance and only needs to be adaptively adjusted according to a small number of different scene labels in the training operation recognition subnetwork. In the end, it provides downstream tasks (such as risk control models) with highly accurate, good generalization and relatively stable object profile information, effectively improving the operation recognition ability of downstream tasks, reducing the dependence on label samples, and solving the problem of overfitting due to scarce label samples.

[0151] To objectively verify the effectiveness of the method disclosed herein, an experiment was conducted using real risk control data from business operations, taking its application in a financial risk control scenario as an example. The experiment shows that the method disclosed herein exhibits good generalization performance without relying on labeled samples for training, and the fusion time interval mechanism further enhances the effect. Based on the original basic object features, it better uncovers the risk of shared debt behavior of objects, effectively improving risk control modeling for different financial scenarios.

[0152] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0153] Please refer to Figure 9 This diagram illustrates a structural block diagram of an object portrait information generation apparatus according to an embodiment of this disclosure. The apparatus has the functions described in the method example above; these functions can be implemented in hardware or by hardware executing corresponding software. The object portrait information generation apparatus may include:

[0154] The acquisition module 910 is used to acquire the historical operation information of an object, wherein the historical operation information represents the operation timing information of the object requesting the target business from at least one business operation platform;

[0155] The parsing module 920 is used to parse the historical operation information to obtain a business operation information sequence and a corresponding time interval sequence; the business operation information sequence includes multiple business sub-operation identifiers corresponding to the chronological order of the operation times corresponding to the request of the target business; the time interval sequence represents the operation time difference between each adjacent business sub-operation identifier in the business operation information sequence;

[0156] The first processing module 930 is used to call the trained operation sequence encoding network to encode the business operation information sequence and the corresponding time interval sequence to obtain sequence operation features.

[0157] The second processing module 940 is used to call multiple trained operation recognition sub-networks to perform recognition processing on the sequence operation features respectively, and obtain multiple corresponding operation recognition sub-results; different operation recognition sub-networks are used to indicate the recognition of business indicator types of different dimensions, and each operation recognition sub-result represents the probability that the object performs a business operation under the corresponding business indicator type;

[0158] The profile determination module 950 is used to obtain the object profile information of the object based on the business indicator type corresponding to each of the multiple operation recognition sub-networks and the multiple operation recognition sub-results.

[0159] In an optional implementation, the first processing module includes:

[0160] The first processing submodule is used to perform vector encoding on the service operation information sequence through the first vector encoding module in the trained service sequence encoding network to obtain the service embedding vector;

[0161] The second processing submodule is used to perform vector encoding on the time interval sequence through the second vector encoding module in the service sequence encoding network to obtain a time embedding vector;

[0162] The third processing submodule is used to extract features from the service embedding vector and the time embedding vector through the encoding conversion module in the service sequence coding network to obtain the sequence operation features.

[0163] In an alternative embodiment, the apparatus further includes:

[0164] The time interval sequence is normalized to obtain an initial time series matrix;

[0165] The initial time series matrix is ​​numerically truncated based on a preset truncation threshold to obtain an intermediate time series matrix.

[0166] The second processing submodule is specifically used for:

[0167] The intermediate time series matrix is ​​vector-encoded by the second vector encoding module in the business sequence encoding network to obtain the time embedding vector.

[0168] In one optional implementation, the encoding conversion module includes an attention mechanism module and a feedforward module; the third processing submodule includes:

[0169] The first processing unit is used to process the service embedding vector and the time embedding vector through the attention mechanism module based on the correlation between the service vector components in the service embedding vector that fuses the time embedding vector, to obtain initial sequence operation features;

[0170] The second processing unit is used to process the initial sequence operation features through the feedforward module to obtain the sequence operation features.

[0171] In an optional implementation, the first processing unit is specifically used for:

[0172] The transformation matrix in the attention mechanism module is used to transform each business vector component in the business embedding vector to obtain the query matrix, key matrix and value matrix corresponding to the business embedding vector.

[0173] Based on each time vector component in the time embedding vector, the corresponding key components in the key value matrix are fused to obtain a fused key value matrix.

[0174] Based on the similarity between the query matrix and the fused key-value matrix corresponding to the business embedding vector, attention weights are obtained, and the corresponding value matrices are weighted based on the attention weights to obtain the initial sequence operation features.

[0175] In an alternative implementation, the operational sequence coding network is trained through the following steps:

[0176] Obtain a first training sample, which includes first historical sample operation information corresponding to multiple first sample objects; the first historical sample operation information represents the timing information of sample operations corresponding to the target business requested by the first sample object from at least one business operation platform.

[0177] Parse the operation information of each first historical sample to obtain the business sample operation information sequence and the corresponding sample time interval sequence for each first sample object; the business sample operation information sequence includes multiple business sample sub-operation information corresponding to the chronological order of the operation time corresponding to the request for the target business; the sample time interval sequence represents the operation time difference between each adjacent business sample sub-operation identifier in the business sample operation information sequence;

[0178] The target business sample sub-operation identifiers in each of the aforementioned business sample operation information sequences are masked to obtain the corresponding masked sample operation information sequences.

[0179] The initial operation sequence encoding network processes each masked sample operation information sequence and the corresponding sample time interval sequence to predict the predicted business sub-operation identifier results corresponding to the masked business sample sub-operations in each masked sample operation information sequence.

[0180] The initial operation sequence encoding network is trained based on the training loss determined by the predicted business sub-operation identifier result and the corresponding target business sample sub-operation identifier to obtain the operation sequence encoding network.

[0181] In an optional implementation, the step of processing each masked sample operation information sequence and the corresponding sample time interval sequence through an initial operation sequence encoding network to predict the predicted service sub-operation identifier results corresponding to the masked service sample sub-operations in each masked sample operation information sequence includes:

[0182] Each of the masked sample operation information sequences and the corresponding sample time interval sequences are input into the initial operation sequence encoding network for encoding processing to obtain the target sub-operation features corresponding to the masked business sample sub-operations of each of the masked sample operation information sequences; wherein, in the attention mechanism in the encoding processing, the processing is performed by fusing the time embedding vectors corresponding to the sample time interval sequences;

[0183] The features of each target sub-operation are classified and processed to predict the predicted business sub-operation identifiers corresponding to the masked business sample sub-operations in each masked sample operation information sequence.

[0184] In an alternative implementation, the plurality of operation recognition subnetworks are trained through the following steps:

[0185] Obtain a second training sample, which includes second historical sample operation information corresponding to multiple second sample objects, and corresponding indicator identification labels; the second historical sample operation information represents the time sequence information of sample operations corresponding to the target business requested by the second sample object from at least one business operation platform.

[0186] Multiple initial operation recognition subnetworks are obtained, and different initial operation recognition subnetworks are used to indicate the types of business indicators in different dimensions.

[0187] The multiple initial operation recognition subnetworks are trained in a supervised manner using the second training samples until the training termination condition is met, thereby obtaining multiple operation recognition subnetworks.

[0188] In an optional implementation, the second processing module is specifically used for:

[0189] The sequence operation features are respectively input into the multiple operation recognition sub-networks for recognition processing, and the corresponding multiple operation recognition sub-scores are output.

[0190] The scores of the multiple operation identification sub-scores are used as the results of the multiple operation identification sub-scores;

[0191] The generation module is specifically used for:

[0192] According to the preset type order of the business indicator types corresponding to the multiple operation recognition sub-networks, the multiple operation recognition sub-results are vectorized to obtain the object profile information of the object.

[0193] The apparatus provided in the above embodiments can execute the corresponding methods in the embodiments of this disclosure, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in the above embodiments can be found in the methods provided in any embodiment of this application.

[0194] This disclosure provides an electronic device that may include a processor and a memory. The memory stores at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the method described in any of the above method embodiments.

[0195] Furthermore, Figure 10 This is a schematic diagram of the hardware structure of an electronic device for implementing the methods provided in the embodiments of this disclosure. (Refer to...) Figure 10 The electronic device includes a processor; a memory for storing processor-executable instructions; wherein, when the processor is configured to execute the instructions stored in the memory, it implements the steps of any of the methods in the above embodiments.

[0196] The electronic device can be a terminal, a server, or a similar computing device. Taking a server as an example... Figure 10 This is a block diagram illustrating an electronic device for generating object profile information according to an exemplary embodiment. The electronic device 1000 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 1010 (CPUs 1010 may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory 1030 for storing data, and one or more storage media 1020 (e.g., one or more mass storage devices) for storing application programs 1023 or data 1022. The memory 1030 and storage media 1020 may be temporary or persistent storage. The program stored in the storage media 1020 may include one or more modules, each module including a series of instruction operations on the electronic device. Furthermore, the CPU 1010 may be configured to communicate with the storage media 1020 and execute the series of instruction operations in the storage media 1020 on the electronic device 1000.

[0197] Electronic device 1000 may also include one or more power supplies 1060, one or more wired or wireless network interfaces 1050, one or more input / output interfaces 1040, and / or one or more operating systems 1021, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc.

[0198] The input / output interface 1040 can be used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the electronic device 1000. In one example, the input / output interface 1040 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In an exemplary embodiment, the input / output interface 1040 can be a radio frequency (RF) module for wireless communication with the Internet.

[0199] Those skilled in the art will understand that Figure 10 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the electronic device 1000 may also include components that are more... Figure 10 The more or fewer components shown, or having the same Figure 10 The different configurations shown.

[0200] This disclosure also provides a computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement any of the methods described above. For example, the storage medium includes a memory for instructions that can be executed by a processor of an electronic device 1000 to perform the methods described above. Optionally, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc.

[0201] This disclosure also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the methods described above.

[0202] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0203] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this disclosure have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0204] The various embodiments in this disclosure are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and electronic device embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0205] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0206] The above description is only a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. An object image information generation method characterized by, The method comprises: acquiring historical operation information of a subject, the historical operation information representing operation time sequence information of the subject when requesting a target service corresponding operation from at least one service operation platform; parsing the historical operation information to obtain a service operation information sequence and a corresponding time interval sequence; the service operation information sequence comprises a plurality of service sub-operation identifiers corresponding to the order of operation time of requesting the target service; and the time interval sequence represents the operation time difference between adjacent service sub-operation identifiers in the service operation information sequence; calling a trained operation sequence encoding network to encode the service operation information sequence and the corresponding time interval sequence to obtain sequence operation features; calling a plurality of trained operation recognition sub-networks to respectively recognize the sequence operation features to obtain a plurality of corresponding operation recognition sub-results; different operation recognition sub-networks are used to indicate recognition of different dimensions of service indicator types, and each operation recognition sub-result represents the probability of the subject performing service operation under the corresponding service indicator type; based on the service indicator types corresponding to the plurality of operation recognition sub-networks and the plurality of operation recognition sub-results, obtaining subject portrait information of the subject.

2. The method of claim 1, wherein, The method comprises: vector encoding the service operation information sequence by a first vector encoding module in the trained service sequence encoding network to obtain a service embedding vector; vector encoding the time interval sequence by a second vector encoding module in the service sequence encoding network to obtain a time embedding vector; extracting features of the service embedding vector and the time embedding vector by an encoding conversion module in the service sequence encoding network to obtain the sequence operation features.

3. The method of claim 2, wherein, Before the vector encoding of the time interval sequence by the second vector encoding module in the service sequence encoding network to obtain the time embedding vector, the method comprises: normalizing the time interval sequence to obtain an initial time sequence matrix; performing numerical truncation processing on the initial time sequence matrix based on a preset truncation threshold to obtain an intermediate time sequence matrix; the vector encoding of the time interval sequence by the second vector encoding module in the service sequence encoding network to obtain the time embedding vector, comprising: vector encoding the intermediate time sequence matrix by the second vector encoding module in the service sequence encoding network to obtain the time embedding vector.

4. The method of claim 2, wherein, The encoding conversion module comprises an attention mechanism module and a feedforward module; the feature extraction of the service embedding vector and the time embedding vector by the encoding conversion module in the service sequence encoding network to obtain the sequence operation features, comprising: processing the service embedding vector and the time embedding vector based on the correlation between each service vector component in the service embedding vector fused with the time embedding vector by the attention mechanism module to obtain initial sequence operation features; and performing feature extraction on the initial sequence operation features by the feedforward module to obtain the sequence operation features. The initial sequence operation feature is processed by the feedforward module to obtain the sequence operation feature.

5. The method of claim 4, wherein, The initial sequence operation feature is obtained by processing the business embedding vector based on the correlation between the business vector components in the business embedding vector fused with the time embedding vector through the attention mechanism module. The business embedding vector is transformed through the transformation matrix in the attention mechanism module to obtain a query matrix, a key-value matrix and a value matrix corresponding to the business embedding vector. The key-value components in the key-value matrix are fused based on the respective time vector components in the time embedding vector to obtain a fused key-value matrix. The attention weight is obtained based on the similarity between the query matrix and the fused key-value matrix corresponding to the business embedding vector, and the corresponding value matrix is weighted based on the attention weight to obtain the initial sequence operation feature.

6. The method according to any one of claims 1 to 5, characterized in that, The operation sequence encoding network is trained through the following steps: A first training sample is obtained, and the first training sample includes a plurality of first sample objects respectively corresponding to first historical sample operation information; the first historical sample operation information represents sample operation time sequence information corresponding to a target business requested by the first sample object to at least one business operation platform; Each first historical sample operation information is parsed to obtain a business sample operation information sequence and a corresponding sample time interval sequence of each first sample object; the business sample operation information sequence includes a plurality of business sample sub-operation information corresponding to the order of operation time of the target business; and the sample time interval sequence represents the operation time difference between adjacent business sample sub-operation identifiers in the business sample operation information sequence. The target business sample sub-operation identifier in each business sample operation information sequence is masked to obtain a corresponding masked sample operation information sequence. Each masked sample operation information sequence and the corresponding sample time interval sequence are processed through an initial operation sequence encoding network to predict the predicted business sub-operation identifier corresponding to each masked business sample sub-operation in the masked sample operation information sequence. The initial operation sequence encoding network is trained based on the training loss determined by the predicted business sub-operation identifier and the corresponding target business sample sub-operation identifier to obtain the operation sequence encoding network.

7. The method of claim 6, wherein, The initial operation sequence encoding network is trained based on the training loss determined by the predicted business sub-operation identifier and the corresponding target business sample sub-operation identifier to obtain the operation sequence encoding network. The mask sample operation information sequence and the corresponding sample time interval sequence are input into the initial operation sequence coding network for coding processing, and target sub-operation features corresponding to each masked business sample sub-operation in the mask sample operation information sequence are obtained; in the attention mechanism of the coding processing, the time embedding vector corresponding to the sample time interval sequence is processed by fusion; Each target sub-operation feature is classified and processed to obtain a predicted business sub-operation identification result corresponding to each masked business sample sub-operation in the mask sample operation information sequence.

8. The method of claim 1, wherein, The plurality of operation recognition sub-networks are trained by the following steps: Obtain a second training sample, the second training sample includes a second historical sample operation information corresponding to a plurality of second sample objects respectively, and a corresponding index identification label; The second historical sample operation information represents sample operation time sequence information corresponding to a target business requested by the second sample object to at least one business operation platform; Obtain a plurality of initial operation recognition sub-networks, and different initial operation recognition sub-networks are used to indicate recognition of different dimensions of business index types; The plurality of initial operation recognition sub-networks are supervised trained by the second training sample until a training end condition is reached, and the plurality of operation recognition sub-networks are obtained.

9. The method according to any of claims 1-5, 7 and 8, characterized by, The trained plurality of operation recognition sub-networks are called respectively to recognize the sequence operation features, and a plurality of operation recognition sub-results are obtained, including: The sequence operation features are input into the plurality of operation recognition sub-networks for recognition processing, and a plurality of operation recognition sub-scores corresponding to the sequence operation features are output; The plurality of operation recognition sub-scores are used as the plurality of operation recognition sub-results; Based on the plurality of operation recognition sub-networks corresponding to the business index types and the plurality of operation recognition sub-results, object portrait information of the object is obtained, including: The plurality of operation recognition sub-results are vector converted according to a preset type order of the plurality of operation recognition sub-networks corresponding to the business index types, and the object portrait information of the object is obtained.

10. An object image information generating apparatus characterized by comprising: The device includes: An acquisition module is configured to acquire historical operation information of an object, wherein the historical operation information represents operation time sequence information corresponding to a target business requested by the object to at least one business operation platform; An analysis module is configured to analyze the historical operation information to obtain a business operation information sequence and a corresponding time interval sequence; the business operation information sequence includes a plurality of business sub-operation identifications corresponding to an order of operation time of the target business; and the time interval sequence represents an operation time difference between each adjacent business sub-operation identification in the business operation information sequence; A first processing module is configured to call a trained operation sequence coding network to code process the business operation information sequence and the corresponding time interval sequence to obtain a sequence operation feature. The second processing module is configured to call the trained multiple operation recognition sub-networks to respectively perform identification processing on the sequence operation features to obtain corresponding multiple operation recognition sub-results; different operation recognition sub-networks are used to indicate identification of different dimensions of business index types, and each operation recognition sub-result represents a probability of performing a business operation by the object under a corresponding business index type. The portrait determination module is configured to obtain object portrait information of the object based on the multiple operation recognition sub-networks, the business index types corresponding to the multiple operation recognition sub-networks, and the multiple operation recognition sub-results.

11. An electronic device, comprising: The electronic device includes a processor and a memory, and the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the object portrait information generation method according to any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the object portrait information generation method according to any one of claims 1-9.

Images