Service providing method and device based on multi-dimensional time sequence feature data matrix

Abstract

The application relates to a service providing method and device based on a multi-dimensional time sequence feature data matrix, an electronic device and a computer readable medium. The method comprises: obtaining operation data of a user; extracting time sequence data in the operation data according to a preset category to generate a multi-dimensional data matrix; determining position coding for time sequence data in each dimension of the multi-dimensional data matrix; inputting the multi-dimensional data matrix with position coding into a feature extraction model; the feature extraction model extracts the order and correlation between time sequence data in the multi-dimensional data matrix based on position coding and deep learning functions to generate a plurality of groups of feature data; generating time sequence feature data of the user through the plurality of groups of feature data; and determining a user portrait of the user according to the time sequence feature data to provide internet services for the user. The application can accurately, quickly and comprehensively generate a user portrait, thereby providing internet services for the user and improving system security.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer information processing, in particular, to a service providing method and device for generating user portrait based on multi-dimensional time sequence feature data matrix, electronic equipment and computer readable medium. BACKGROUND

[0002] User behavior data refers to data generated by a user performing various operations in a product. The user performs an operation at what time, in what place, and in what interactive manner (accessing / browsing / clicking, etc.). As a strong behavior representation of a user, depicting a user portrait through user behavior data has always been a technical focus. Based on user behavior data, a user portrait of the entire life cycle of a user can be sketched.

[0003] However, the current main way of mining user behavior data is manual, which requires a lot of manpower, and the features mined manually are highly dependent on experience in the business. Nowadays, data is often processed from interpretable rules. The processing method for time sequence data is often to select different time windows and aggregate the information in the time window, such as the number of times in the last 3 months, the average, the maximum, and the minimum. Further derivations can be made on this basis, such as the ratio in different time windows (e.g., the number of times in the last 3 months / the number of times in the last 6 months), the ratio of different indicators in the same time window (e.g., the number of non-official institutions in the last 3 months / the number of official institutions), etc. Although such a way can theoretically derive a large number of features, the feature screening work is heavy in actual application, and most of the features with weak correlation need to be removed, and it is limited by the experience and cognition of experts.

[0004] Therefore, a new service providing method and device for generating user portrait based on multi-dimensional time sequence feature data matrix, electronic equipment and computer readable medium are needed.

[0005] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present application, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] Therefore, the present application provides a service providing method and device for generating user portrait based on multi-dimensional time sequence feature data matrix, electronic equipment and computer readable medium, which can accurately and quickly generate a user portrait, and thus provide Internet services for users, improve system security, and ensure factor safety and improve service performance.

[0007] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0008] According to one aspect of this application, a service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix is ​​proposed. The method includes: acquiring user operation data; extracting temporal data from the operation data according to preset categories to generate a multi-dimensional data matrix; determining a positional code for the temporal data in each dimension of the multi-dimensional data matrix; inputting the multi-dimensional data matrix with positional codes into a feature extraction model; the feature extraction model extracting the order and correlation between the temporal data in the multi-dimensional data matrix based on the positional codes and deep learning functions to generate multiple sets of feature data; generating the user's temporal feature data through the multiple sets of feature data; and determining the user profile based on the temporal feature data to provide internet services to the user.

[0009] Optionally, the feature extraction model extracts the order and correlation between temporal data in the multi-dimensional data matrix based on positional encoding and deep learning functions, generating multiple sets of feature data, including: multiple machine translation layers in the feature extraction model, which cyclically pass the state of the previous neuron to the subsequent neuron through encoder and decoder, so as to extract the order and correlation between temporal data in multiple dimensions of the multi-dimensional data matrix and generate multiple sets of feature data.

[0010] Optionally, generating the user's temporal feature data through the multiple sets of feature data includes: multiple global average pooling layers in the feature extraction model, downsampling the multiple sets of feature data in the current layer and passing them to the previous layer to generate multiple sampled data; and concatenating the multiple sampled data to generate the user's temporal feature data.

[0011] Optionally, extracting time-series data from the operation data according to a preset category to generate a multi-dimensional data matrix includes: dividing the time-series data in the operation data into multiple data groups according to a preset category; one data group corresponds to one dimension of the multi-dimensional data matrix; and performing sequence transformation on the time-series data in each data group to generate the multi-dimensional data matrix.

[0012] Optionally, the time-series data in each data group is transformed into a sequence to generate the multi-dimensional data matrix, including: converting categorical variables in the time-series data into numerical codes for category sequence transformation; converting non-categorical variables in the time-series data into numerical sequences; and concatenating the sequence-transformed time-series data in each data group to generate the multi-dimensional data matrix.

[0013] Optionally, it further includes: extracting the sequence length threshold corresponding to each category; and filling or deleting time-series data in the operation data according to the length threshold.

[0014] Optionally, determining the user portrait of the user according to the time-series feature data to provide an Internet service for the user comprises: obtaining basic information of the user; generating basic feature data from the basic information; determining the user portrait of the user according to the basic feature data of the user and the time-series feature data; and providing an Internet service for the user according to the user portrait.

[0015] Optionally, the operation data of the user is obtained by: obtaining original data of the user from a third-party database, wherein the original data comprises time-series data and non-time-series data of multiple categories; determining a time range and a target category according to a business scenario; and filtering the original data according to the time range and the target category to generate the operation data.

[0016] Optionally, the position code is determined for the time-series data of each dimension in the multi-dimensional data matrix by: determining the position code for the time-series data of each dimension in the multi-dimensional time-series matrix according to a time difference between the time corresponding to the time-series data and a current time.

[0017] Optionally, the multi-dimensional data matrix with the position code is input into a feature extraction model by: performing sequence conversion on the non-time-series data in the operation data of the user to generate a non-time-series array; and inputting the multi-dimensional data matrix with the position code and the non-time-series array into the feature extraction model.

[0018] According to an aspect of the present application, a service providing device based on a multi-dimensional time-series feature data matrix to generate a user portrait is provided, which comprises: a data module configured to obtain operation data of a user; a matrix module configured to extract time-series data in the operation data according to a preset category to generate a multi-dimensional data matrix; a position module configured to determine a position code for the time-series data of each dimension in the multi-dimensional data matrix; an input module configured to input the multi-dimensional data matrix with the position code into a feature extraction model; an extraction module configured to extract an order and a correlation between the time-series data in the multi-dimensional data matrix based on the position code and a deep learning function of the feature extraction model to generate a plurality of groups of feature data; a feature module configured to generate time-series feature data of the user from the plurality of groups of feature data; and a service module configured to determine a user portrait of the user according to the time-series feature data to provide an Internet service for the user.

[0019] According to an aspect of the present application, an electronic device is provided, which comprises: one or more processors; a storage device configured to store one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0020] According to an aspect of the present application, a computer readable medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method as described above.

[0021] According to the service providing method, device, electronic equipment and computer readable medium for generating a user portrait based on a multi-dimensional time sequence feature data matrix, the operation data of a user is obtained; time sequence data in the operation data is extracted according to a preset category to generate a multi-dimensional data matrix; position encoding is determined for the time sequence data of each dimension in the multi-dimensional data matrix; the multi-dimensional data matrix with position encoding is input into a feature extraction model; the feature extraction model extracts the order and correlation between the time sequence data in the multi-dimensional data matrix based on the position encoding and deep learning function to generate a plurality of groups of feature data; the time sequence feature data of the user is generated through the plurality of groups of feature data; and the user portrait of the user is determined according to the time sequence feature data to provide an internet service for the user. In this way, the user portrait can be accurately, quickly and comprehensively generated, and the internet service for the user is provided, the system security is improved, and the service performance is improved.

[0022] It should be understood that the foregoing general description and the following detailed description are only examples and are not limiting the present application. BRIEF DESCRIPTION OF DRAWINGS

[0023] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0024] Figure 1 FIG. 1 is a flowchart illustrating a service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix according to an example embodiment.

[0025] Figure 2 FIG. 2 is a schematic diagram illustrating a service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix according to another example embodiment.

[0026] Figure 3 FIG. 3 is a schematic diagram illustrating a service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix according to another example embodiment.

[0027] Figure 4 FIG. 4 is a schematic diagram illustrating a service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix according to another example embodiment.

[0028] Figure 5This is a schematic diagram illustrating a service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix, according to another exemplary embodiment.

[0029] Figure 6 This is a flowchart illustrating a service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix, according to another exemplary embodiment.

[0030] Figure 7 This is a block diagram illustrating a service provision apparatus for generating user profiles based on a multi-dimensional temporal feature data matrix, according to an exemplary embodiment.

[0031] Figure 8 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0032] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this application will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0033] Figure 1 This is a flowchart illustrating a service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix, according to an exemplary embodiment. The service provision method 10 for generating user profiles based on a multi-dimensional temporal feature data matrix includes at least steps S102 to S114.

[0034] like Figure 1 As shown, in S102, user operation data is obtained. For example, the user's raw data can be obtained from a third-party database, which contains multiple categories of time-series and non-time-series data; the time range and target category are determined according to the business scenario; the raw data is filtered according to the time range and the target category to generate the operation data.

[0035] In this embodiment, the user can be an individual user or an enterprise user, and the user operation data can be the user's operational behavior data, such as page operation data, business access duration, business access frequency, etc., for individual users or enterprise users. For example, a short video APP needs to auction advertising slots. Before each short video is about to play an advertisement, each enterprise user needs to choose whether to bid for the advertising slot to place an advertisement. At this time, since the auction matching time is very short, the advertising slot auction can only be carried out through a set program. Obtaining the enterprise user's bidding information for the advertising slot is to obtain the enterprise user's user operation data. Similarly, business access duration and business access frequency can also be understood as business access information carried out by the program set by the enterprise user. The specific content of the user information can be determined according to the actual application scenario and is not limited here.

[0036] In a practical application scenario, taking enterprise users as an example, data can be obtained from their business access data within the platform, or data access or operation information within the system, serving as the raw data for subsequent processing. Raw data from different time ranges or categories can also be extracted depending on the application scenario. For example, when evaluating the security of individual users, raw data related to data services within the past 90 days can be extracted, such as whether the user has illegally downloaded or obtained data. When evaluating user preferences, raw data related to shopping within the past year can be extracted.

[0037] In S104, time-series data from the operation data is extracted according to preset categories to generate a multi-dimensional data matrix. The time-series data in the operation data can be divided into multiple data groups according to preset categories; one data group corresponds to one dimension of the multi-dimensional data matrix; the time-series data in each data group is sequence-transformed to generate the multi-dimensional data matrix.

[0038] In S106, a position code is determined for the time series data of each dimension in the multi-dimensional data matrix. The position code can be determined based on the time difference between the time corresponding to the time series data and the current time.

[0039] The time difference between the event occurrence date and the observation point cannot be described because the self attention mechanism architecture of the deep learning model Transformer does not use the time sequence information of the sequence. In the present application, the time sequence information is added to the model calculation through the positional encoding to improve the time sequence calculation function of the deep learning model Transformer.

[0040] If the time difference is counted in days, the number of categories that can be taken is too large, resulting in too many positional encoding parameters. Preferably, a scaling coefficient can be selected to reduce the range. In a specific application, the range of days is 0-800 days, which can be multiplied by 0.1 to reduce the range to 0-80. The positional encoding is added to the features (i.e. the result of concatenating the categorical embedding and the numerical features) to introduce the time sequence features in the deep learning model calculation.

[0041] The matrix calculation formula of the positional encoding is:

[0042]

[0043]

[0044] where d is the feature dimension. i is the row (time sequence dimension), which takes an integer value of 1-n, and n is the time sequence length. 2j represents an even column, and 2j+1 represents an odd column (feature dimension).

[0045] In S108, the multi-dimensional data matrix with positional encoding is input into the feature extraction model. The non-time sequence data in the user operation data is converted into a non-time sequence array; the multi-dimensional data matrix with positional encoding and the non-time sequence array are input into the feature extraction model.

[0046] The feature extraction function can be realized by stacking multiple machine translation layers (Transformer Layer) in the Transformer deep learning model. The structure of the Transformer Layer is as shown in Figure 2 Figure 3 ​The outputs of several groups of time series data are connected to a full connection layer and spliced, and then stacked with a Transformer Layer. Alternatively, a cross attention mechanism is used to extract the correlation between different time series data. Then, the several groups of time series data are respectively down-sampled by a global average pooling layer, and then spliced, connected to several full connection layers (MLP), and finally connected to an output layer, for example, a full connection layer with the number of neurons being the number of categories (softmax activation function) for a classification problem.

[0047] In S110, the feature extraction model extracts the order and correlation between time series data in the multidimensional data matrix based on position encoding and deep learning functions, and generates multiple groups of feature data. More specifically, multiple machine translation layers in the feature extraction model pass the state of previous neurons to subsequent neurons through an encoder-decoder cycle to extract the order and correlation between multiple dimensions of time series data in the multidimensional data matrix, and generate multiple groups of feature data.

[0048] In the embodiments of the present application, multiple groups of feature data can be generated by extracting the order and correlation between time series data in a multidimensional data matrix through a Transformer deep learning model. Figure 4 As shown, the Transformer attention-based architecture has a longest path (the distance from x1 to xn in the sequence) of O(1), that is, the entire sequence can be seen in one step, which is an advantage over the longest path of O(n) of RNN-type neural networks and the longest path of O(n / k) of CNN (k is the window size). Therefore, even time information that is far away in time can be captured through the attention mechanism, which is suitable for long time series. Therefore, it is also suitable for cases where the credit history is long, for example, if the recent credit behavior is suddenly changed compared to the early credit behavior, the corresponding features can be extracted through the attention mechanism.

[0049] In S112, the time series feature data of the user is generated through the multiple groups of feature data. Multiple global average pooling layers in the feature extraction model down-sample the multiple groups of feature data of the current layer and pass them to the previous layer to generate multiple sampling data; and the multiple sampling data are spliced to generate the time series feature data of the user.

[0050] In the embodiments of the present application, the overall network architecture of the Transformer deep learning model is as shown in Figure 5As shown. In a data service scenario, the network's underlying input is time-series data, which may include query sequences, data acquisition sequences, and data upload sequences. Other time-series and non-time-series data may also be added, including basic user information. The time-series data contains categorical variables, such as query reasons. These are converted to numerical codes (integers) and then categorical embedding is used to obtain relatively low-dimensional, dense features, which are then concatenated with the numerical features.

[0051] Furthermore, during model training, hyperparameter tuning is performed on parameters such as the number of transformer (translation module) blocks, the number of heads in the multi-head attention mechanism, the number of neurons in the fully connected layer, the elimination ratio, and the output dimension of the embedding layer. Appropriate ranges can be set based on experience, and then hyperparameters can be tuned using tools such as KerasTuner.

[0052] When training a neural network, the optimizer can use gradient descent methods such as Adam and RMSprop. The results of the output layer and several layers preceding it (such as the penultimate fully connected layer) are saved as features.

[0053] The features extracted by the aforementioned neural network can be used, along with features from other dimensions, to train models such as XGBoost and Logistic Regression. Alternatively, the score of the neural network's output layer can be used as a subset to train the model along with features from other dimensions.

[0054] In S114, a user profile is determined based on the time-series feature data to provide internet services to the user. This involves obtaining basic user information; generating basic feature data from the basic information; determining a user profile based on the user's basic feature data and the time-series feature data; and providing internet services to the user based on the user profile.

[0055] The service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix according to this application involves: acquiring user operation data; extracting temporal data from the operation data according to preset categories to generate a multi-dimensional data matrix; determining a position code for the temporal data in each dimension of the multi-dimensional data matrix; inputting the multi-dimensional data matrix with position codes into a feature extraction model; the feature extraction model extracting the order and correlation between the temporal data in the multi-dimensional data matrix based on position codes and deep learning functions to generate multiple sets of feature data; generating the user's temporal feature data through the multiple sets of feature data; and determining the user's user profile based on the temporal feature data to provide internet services to the user. This method can accurately, quickly, and comprehensively generate user profiles, thereby providing internet services to users, improving system security, ensuring user safety, and enhancing service performance.

[0056] According to the service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix provided by the present application, the encoder in the Transformer is used for time sequence feature extraction. The features learned from data are different from the way of processing features based on human experience, and some features that are difficult to summarize and discover from experience can be extracted. Moreover, the correlation with artificial features is low, and can be complementary to them. Since the parallelism of the Transformer is O(n), and the parallelism of the RNN is O(1) due to the time sequence, the parallelism of the Transformer used in the present application is high, and the speed can be easily and simply increased by using GPU and TPU. The feature distinguishing ability of the Transformer used in the present application is good. Since the features are obtained by training based on the y label of the sample performance of the neural network (supervised learning), the Transformer model selected in the present application has good distinguishing ability, and can simplify the feature screening work during modeling.

[0057] It should be clearly understood that the present application describes how to form and use specific examples, but the principles of the present application are not limited to any details of these examples. On the contrary, based on the teachings of the disclosure of the present application, these principles can be applied to many other embodiments.

[0058] Figure 6 is a flowchart of a service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix according to another exemplary embodiment. Figure 6 The flowchart 60 shown is for Figure 1 The detailed description of S104 "extracting time sequence data in the operation data according to the preset category to generate a multi-dimensional data matrix" in the flowchart shown.

[0059] As Figure 6 As shown in S602, the time sequence data in the operation data is divided into a plurality of data groups according to the preset category. The time sequence data is divided into a plurality of data groups according to different application scenarios and categories.

[0060] In the data service scenario, the time sequence features in the report can also be extracted, including the query records in the past 2 years, various data in the data acquisition information details, data usage records, and nearly 5 years of data download records under each detailed record. At the same time, necessary data cleaning and aggregation are performed, such as missing value filling, outlier truncation, numerical variable normalization, and classification variable conversion to numerical encoding, etc.

[0061] The data download records in the past 5 years can aggregate each detailed record, such as the data download amount of each month, to obtain time series data of nearly 60 months. The occurrence dates of events in the time series data, such as the query date, the account opening date, the data download month, and the data download amount, can be processed as the time difference between the observation point. In addition to the time series features, other static data, summary data, and the like, such as basic information, violation acquisition data, and the like, can be added as an optional input, and then combined with the intermediate output result of the time series module.

[0062] In S604, one data group corresponds to one dimension of the multi-dimensional data matrix.

[0063] In S606, the time series data in each data group is subjected to sequence transformation to generate the multi-dimensional data matrix.

[0064] In one embodiment, the categorical variables in the time series data can be converted into numerical encoding for category sequence transformation; the non-categorical variables in the time series data are subjected to numerical sequence transformation; and the sequence-transformed time series data in each group of data are spliced to generate the multi-dimensional data matrix.

[0065] Since there are categorical variables in the time series data, such as the query reason, the numerical encoding (integer) is converted into a relatively low-dimensional dense feature through a category transformation (Categorical embedding) operation, and then spliced with the numerical features.

[0066] Since the time series data of each sample is of different lengths, a maximum length max_length can be set, and data exceeding the maximum length is truncated. The maximum length can be set as the length of the longest sequence in the sample, and the truncation is minimized to avoid the loss of early information. The data less than the maximum length is padded with zeros to pad the sequence length to max_length.

[0067] In one embodiment, after zero padding, the padding mask operation can be used to eliminate the influence of zero padding on the overall data. For example, in model calculation, the attention score of the data after zero padding can be set to 0.

[0068] In S608, the time series data in the operation data is padded or deleted according to the length threshold. The sequence length threshold corresponding to each category can be extracted to perform padding

[0069] In actual application scenarios, the Transformer has high computational complexity, which is proportional to the square of the sequence length n. The Transformer has various variants, mainly optimizing the multi-head attention mechanism to reduce computational complexity, and other attention mechanisms can be optimized, such as local attention, truncated attention, big bird attention, linformer attention, etc.

[0070] In actual application scenarios, the timing data of the user safety report can be very long, up to thousands. Since the transformer has high computational complexity and good parallelism, it can generally be trained by GPU, TPU. If limited by the device, the training time is too long, and bidirectional RNN + attention can also be considered, using the final layer hidden state of RNN at all time steps as the key and value of attention. Here, RNN refers to LSTM, GRU and the like that are good at memorizing.

[0071] In actual application scenarios, different timing sequences can be used as needed when extracting timing data, such as separating data download information and data upload information into different timing data, adding timing data of postpaid payment records, etc. The input in addition to the timing data can also be increased.

[0072] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiments are implemented as computer programs executed by the CPU. When the computer program is executed by the CPU, the above-mentioned functions defined by the above-mentioned method provided by the present application are executed. The program can be stored in a computer readable storage medium, which can be a read-only memory, a magnetic disk or an optical disk, etc.

[0073] In addition, it should be noted that the above-mentioned figures are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present application, and are not for limiting purposes. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be executed synchronously or asynchronously, for example, in multiple modules.

[0074] The following is a device embodiment of the present application, which can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0075] Figure 7 is a block diagram of a service providing apparatus for generating a user portrait based on a multi-dimensional timing feature data matrix according to an exemplary embodiment. As shown in Figure 7As shown, the service providing apparatus 70 for generating user portrait based on multi-dimensional time sequence feature data matrix comprises a data module 702, a matrix module 704, a position module 706, an input module 708, an extraction module 710, a feature module 712 and a service module 714.

[0076] The data module 702 is configured to acquire operation data of a user.

[0077] The matrix module 704 is configured to extract time sequence data in the operation data according to a preset category to generate a multi-dimensional data matrix.

[0078] The position module 706 is configured to determine position encoding for time sequence data of each dimension in the multi-dimensional data matrix.

[0079] The input module 708 is configured to input the multi-dimensional data matrix with position encoding into a feature extraction model.

[0080] The extraction module 710 is configured to extract order and correlation between time sequence data in the multi-dimensional data matrix by the feature extraction model based on position encoding and deep learning function to generate a plurality of groups of feature data.

[0081] The feature module 712 is configured to generate time sequence feature data of the user by the plurality of groups of feature data.

[0082] The service module 714 is configured to determine user portrait of the user according to the time sequence feature data to provide internet service for the user.

[0083] The extraction module 710 is specifically configured to be a plurality of machine translation layers in the feature extraction model, and the state of previous neurons is transmitted to subsequent neurons by an encoder-decoder cycle to extract order and correlation between time sequence data of a plurality of dimensions in the multi-dimensional data matrix to generate a plurality of groups of feature data.

[0084] The feature module 712 is specifically configured to be a plurality of global average pooling layers in the feature extraction model, and the plurality of groups of feature data of a current layer are down-sampled and transmitted to a previous layer to generate a plurality of sampling data; and the plurality of sampling data are spliced to generate the time sequence feature data of the user.

[0085] The matrix module 704 is specifically configured to divide time sequence data in the operation data into a plurality of data groups according to a preset category; one data group corresponds to one dimension of the multi-dimensional data matrix; and time sequence data in each data group is subjected to sequence conversion to generate the multi-dimensional data matrix.

[0086] The matrix module 704 is specifically configured to convert a classification variable in the time series data into a numerical type code for category sequence conversion; perform numerical sequence conversion on a non-classification variable in the time series data; and splice the sequence-converted time series data in each group of data to generate the multi-dimensional data matrix.

[0087] The matrix module 704 is further configured to extract a sequence length threshold corresponding to each category; and fill or delete the time series data in the operation data according to the length threshold.

[0088] The service module 714 is specifically configured to obtain basic information of a user; generate basic feature data based on the basic information; determine a user portrait for the user based on the basic feature data and the time series feature data of the user; and provide an Internet service for the user based on the user portrait.

[0089] The data module 702 is specifically configured to obtain original data of a user from a third-party database, wherein the original data includes time series data and non-time series data of multiple categories; determine a time range and a target category based on a business scenario; and filter the original data based on the time range and the target category to generate operation data.

[0090] The position module 706 is specifically configured to determine a position code for time series data of each dimension in the multi-dimensional time series matrix based on a time difference between a time corresponding to the time series data and a current time.

[0091] The input module 708 is specifically configured to perform sequence conversion on non-time series data in the user operation data to generate a non-time series array; and input the multi-dimensional data matrix with the position code and the non-time series array into a feature extraction model.

[0092] The service providing device for generating a user portrait based on a multi-dimensional time series feature data matrix according to the present application obtains operation data of a user; extracts time series data in the operation data according to a preset category to generate a multi-dimensional data matrix; determines a position code for time series data of each dimension in the multi-dimensional data matrix; inputs the multi-dimensional data matrix with the position code into a feature extraction model; the feature extraction model extracts an order and a correlation between the time series data in the multi-dimensional data matrix based on the position code and a deep learning function to generate multiple groups of feature data; generates time series feature data of the user based on the multiple groups of feature data; and determines a user portrait for the user based on the time series feature data to provide an Internet service for the user. In this way, the user portrait can be accurately, quickly and comprehensively generated, and the Internet service for the user can be provided, thereby improving system security and ensuring factor security and service performance.

[0093] As Figure 8As shown, the electronic device provided by the embodiment of the present application comprises a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 complete mutual communication through the communication bus 840.

[0094] The memory 830 is used for storing a computer program.

[0095] The processor 810 is used for executing the program stored in the memory 830 to realize the video expression action-based data allocation permission adjustment method of any of the above-mentioned embodiments.

[0096] The electronic device provided by the embodiment of the present application is used for the processor 810 to acquire the initial permission of data allocation and access information of a target by executing the program stored in the memory 830; to determine the video text content through the access information; to establish a real-time video link with the target and display the video text content to generate video data according to the video connection; to identify the expression action of a user in the video data to determine the corresponding permission adjustment coefficient; and to adjust the data allocation permission of the user according to the initial permission and the permission adjustment coefficient.

[0097] The communication interface 820 is used for communication between the above-mentioned electronic device and other devices.

[0098] The memory 830 can comprise a random access memory 830 (RAM) and can also comprise a non-volatile memory 830, for example at least one disk memory 830.

[0099] The embodiment of the present application provides a computer readable storage medium, the computer readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the service providing method for generating a user portrait based on a multi-dimensional time sequence feature data matrix of any of the above-mentioned embodiments. For example, operation data of a user can be acquired; time sequence data in the operation data is extracted according to a preset category to generate a multi-dimensional data matrix; position coding is determined for time sequence data of each dimension in the multi-dimensional data matrix; the multi-dimensional data matrix with position coding is input into a feature extraction model; the feature extraction model extracts the order and correlation between time sequence data in the multi-dimensional data matrix based on position coding and deep learning function to generate a plurality of groups of feature data; time sequence feature data of the user is generated through the plurality of groups of feature data; and a user portrait of the user is determined according to the time sequence feature data to provide an internet service for the user.

[0100] The exemplary embodiments of this application are specifically illustrated and described herein. But, it is to be understood that the application is not limited to the details of the illustrated implementations; rather, this application is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A service provision method for generating user profiles based on a multi-dimensional temporal feature data matrix, characterized in that, include: Obtain user operation data; The time-series data in the operation data is divided into multiple data groups according to a preset category; A data set corresponds to one dimension of a multi-dimensional data matrix; The time-series data in each data group are transformed into a multi-dimensional data matrix. The location encoding of the time series data in each dimension of the multi-dimensional time series matrix is ​​determined based on the time difference between the time corresponding to the time series data and the current time. Input the multi-dimensional data matrix with location encoding into the feature extraction model; The feature extraction model has multiple machine translation layers that pass the state of previous neurons to subsequent neurons through encoders and decoders in a loop, so as to extract the order and correlation between time-series data of multiple dimensions in the multi-dimensional data matrix and generate multiple sets of feature data. The user's temporal feature data is generated using the multiple sets of feature data; Based on the time-series feature data, a user profile is determined to provide internet services to the user.

2. The method as described in claim 1, characterized in that, The user's time-series feature data is generated using the multiple sets of feature data, including: The feature extraction model has multiple global average pooling layers that downsample the multiple sets of feature data in the current layer and pass them to the previous layer to generate multiple sampled data. The multiple sampled data are concatenated to generate the user's time-series feature data.

3. The method as described in claim 1, characterized in that, The time-series data in each data group are transformed into sequences to generate the multi-dimensional data matrix, including: Convert categorical variables in time series data into numerical codes for category sequence transformation; Numerical sequence transformation of non-categorical variables in time series data; The time-series data after sequence transformation in each data set are concatenated to generate the multi-dimensional data matrix.

4. The method as described in claim 1, characterized in that, Also includes: Extract the sequence length threshold corresponding to each category; The timing data in the operation data is filled or deleted according to the length threshold.

5. The method as described in claim 1, characterized in that, Determining a user profile based on the time-series feature data to provide internet services to the user includes: Obtain basic user information; Basic feature data is generated from the aforementioned basic information; A user profile is determined for the user based on the user's basic characteristic data and the time-series characteristic data; Internet services are provided to the user based on the user profile.

6. The method as described in claim 1, characterized in that, Obtain user action data, including: The user's raw data is obtained from a third-party database, and the raw data contains multiple categories of time-series data and non-time-series data; Determine the time frame and target categories based on the business scenario; The raw data is filtered according to the time range and the target category to generate the operation data.

7. The method as described in claim 1, characterized in that, Determining the position encoding for the time-series data in each dimension of the multi-dimensional data matrix includes: The positional encoding of the time series data in each dimension of the multi-dimensional time series matrix is ​​determined based on the time difference between the time corresponding to the time series data and the current time.

8. The method as described in claim 1, characterized in that, Inputting a multi-dimensional data matrix with positional encoding into the feature extraction model includes: Transform the non-time-series data in the user operation data into a sequence to generate a non-time-series array; The multi-dimensional data matrix with positional encoding and the non-temporal array are input into the feature extraction model.

9. A service provisioning apparatus for generating user profiles based on a multi-dimensional temporal feature data matrix, characterized in that, include: The data module is used to acquire user operation data; The matrix module is used to divide the time-series data in the operation data into multiple data groups according to a preset category; A data set corresponds to one dimension of a multi-dimensional data matrix; The time-series data in each data group are transformed into a multi-dimensional data matrix. The location module is used to determine the location encoding of the time series data in each dimension of the multi-dimensional time series matrix based on the time difference between the time corresponding to the time series data and the current time. The input module is used to input a multi-dimensional data matrix with positional encoding into the feature extraction model; The extraction module is used in the multiple machine translation layers of the feature extraction model. It passes the state of the previous neuron to the subsequent neuron through the encoder and decoder in a loop to extract the order and correlation between the time series data of multiple dimensions in the multi-dimensional data matrix and generate multiple sets of feature data. The feature module is used to generate the user's time-series feature data from the multiple sets of feature data; The service module is used to determine the user profile of the user based on the time-series feature data in order to provide Internet services to the user.

10. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 8.

11. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images