Information pushing method and device, equipment, medium and product

Abstract

The invention discloses an information pushing method and device, equipment, a medium and a product, and relates to the technical field of artificial intelligence, and the method comprises the steps: determining a target object having a target business product, obtaining a target behavior feature of the target object in a target time period, and obtaining a target behavior feature of the target object; obtaining candidate behavior characteristics of the candidate push object in the target time period; generating a target object behavior portrait of the target object in the target time period according to the target behavior characteristics, and generating candidate object behavior portraits of the candidate push objects in the target time period according to the candidate behavior characteristics; and according to the target object behavior portrait and the candidate object behavior portraits, determining a to-be-pushed object from the candidate pushing objects, and performing information pushing on the to-be-pushed object according to the to-be-pushed information corresponding to the target business product. According to the information pushing method and device, the product universality of to-be-pushed object recognition is improved, the timeliness and effectiveness of information pushing are further improved, and the information pushing effect is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium, and product for pushing information. Background Technology

[0002] In the context of digital transformation, bank channel departments face the challenge of a growing online customer base and the need to push relevant information about different business products to attract customers, in order to meet the performance evaluation of business products by their superiors.

[0003] Currently, information push methods for business products typically require training customized customer group identification models for different business products. The poor universality of customer group identification models across different business products necessitates separate training of customer group identification models for each business product to achieve information push, resulting in poor timeliness and effectiveness of information push and affecting the push effect. Summary of the Invention

[0004] This invention provides an information push method, apparatus, device, medium, and product to solve the problem that existing information push methods for business products have poor timeliness and effectiveness, which affects the information push effect.

[0005] According to one aspect of the present invention, an information push method is provided, the method comprising:

[0006] Identify the target objects that already possess the target business products, and obtain the target behavioral characteristics of the target objects within the target time period, as well as obtain the candidate behavioral characteristics of the candidate push objects within the target time period;

[0007] Generate a target object behavior profile of the target object within the target time period based on the target behavior features, and generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior features;

[0008] Based on the target object behavior profile and the candidate object behavior profile, the target object to be pushed is determined from the candidate push objects, and information is pushed to the target object according to the push information corresponding to the target business product.

[0009] According to another aspect of the present invention, an information push device is provided, the device comprising:

[0010] The behavior feature determination module is used to determine the target object that already holds the target business product, and to obtain the target behavior features of the target object within the target time period, as well as to obtain the candidate behavior features of the candidate push object within the target time period.

[0011] The object behavior profile generation module is used to generate a target object behavior profile of the target object within the target time period based on the target behavior features, and to generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior features.

[0012] The module for determining the target object is used to determine the target object from the candidate push objects based on the target object behavior profile and the candidate object behavior profile, and to push information to the target object according to the push information corresponding to the target business product.

[0013] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0014] At least one processor; and

[0015] A memory communicatively connected to the at least one processor; wherein,

[0016] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the information push method according to any one of the present invention.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the information push method according to any one of the present invention.

[0018] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the information push method described in any one of the present invention.

[0019] This invention identifies target objects that already possess the target business product, obtains the target object's target behavioral characteristics within a target time period, and obtains candidate behavioral characteristics of candidate push objects within the target time period; generates a target object behavioral profile based on the target behavioral characteristics within the target time period, and generates candidate push object behavioral profiles based on the candidate behavioral characteristics within the target time period; and determines the push object from the candidate push objects based on the target object behavioral profile and the candidate object behavioral profiles, and pushes information to the candidate push objects according to the push information corresponding to the target business product. The beneficial effects are:

[0020] By analyzing the behavioral profiles of candidate push targets and target users who already own the target business products within the same target time period, the system identifies the target users from the candidate push targets. This achieves the effect of identifying the target users based on their behavioral profiles, eliminating the need for separate customer group identification model training for the target business products. This improves the product universality of the target user identification and further enhances the timeliness and effectiveness of information push, ensuring the effectiveness of information push.

[0021] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

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

[0023] Figure 1 A flowchart of an information push method provided in Embodiment 1 of the present invention;

[0024] Figure 2 This is a flowchart of an object behavior profiling method provided in Embodiment 2 of the present invention;

[0025] Figure 3A This is a flowchart of an object behavior profiling determination method provided in Embodiment 3 of the present invention;

[0026] Figure 3B This is a schematic diagram of the structure of a behavior profile generation model provided in Embodiment 3 of the present invention;

[0027] Figure 4 This is a flowchart of a method for determining a target to be pushed, provided in Embodiment 4 of the present invention;

[0028] Figure 5 This is a schematic diagram of the structure of an information push device provided in Embodiment 5 of the present invention;

[0029] Figure 6 This is a schematic diagram of the structure of an electronic device that implements the information push method of this invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "candidate," "target," "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] Example 1

[0033] Figure 1 This is a flowchart of an information push method provided in Embodiment 1 of the present invention. This embodiment is applicable to situations where a target object is selected from candidate push objects, and the target business product information is pushed to the target object. This method can be executed by an information push device, which can be implemented in hardware and / or software. Figure 1 As shown, the method includes:

[0034] S101. Identify the target objects that already possess the target business products, and obtain the target behavioral characteristics of the target objects within the target time period, as well as obtain the candidate behavioral characteristics of the candidate push objects within the target time period.

[0035] In this context, the target business product refers to a specific product or service that is the core of the promotion in a particular business scenario; it is the core carrier around which the entire information push method revolves. For example, the types of target business products include, but are not limited to, physical products, digital services, financial products, or application functions. This embodiment does not limit the specific type of target business product.

[0036] The target audience for "already owning" a target business product refers to the group of users who currently own, purchase, or subscribe to the target business product. For example, if the target business product is a financial product, the target audience for "already owning" a target business product refers to the group of users who have already purchased that financial product.

[0037] Candidate push targets refer to user groups who do not yet own, purchase, or subscribe to the target business product, but have the potential to convert. For example, if the target business product is a financial product, the candidate push targets refer to user groups who are likely to purchase that financial product.

[0038] The target time period refers to a pre-defined time range with specific business significance in behavioral feature analysis. This time period is used to uniformly capture the behavioral characteristics of the target object and the candidate push object, ensuring the consistency of the analysis benchmark. The target time period can be flexibly set according to needs, such as a natural week, natural month, natural year, or promotional activity cycle, etc. This embodiment does not limit the specific duration of the target time period.

[0039] Target behavioral features refer to quantifiable behavioral indicators strongly correlated with business objectives, extracted from the behavioral data of target objects within a target time period. Candidate behavioral features refer to quantifiable behavioral indicators strongly correlated with business objectives, extracted from the behavioral data of candidate push objects within a target time period. It is understandable that target behavioral features and candidate behavioral features are fully aligned in feature dimensions. For example, if target behavioral features include "login frequency" and "spending amount," then candidate behavioral features will also include "login frequency" and "spending amount."

[0040] In one implementation, a list of target objects, consisting of object IDs, transmitted by the demand side is obtained. Based on the object IDs of each target object in the list, the target objects that already possess the target business product are identified. Further, based on the object IDs of each target object in the list and a preset target time period, target behavioral features are extracted from a behavioral feature data warehouse. For example, assuming the object ID of any target object is "00001" and the target time period is "November," the behavioral feature data for object ID "00001" in November is extracted from the behavioral feature data warehouse as the target behavioral feature.

[0041] Furthermore, the system acquires the candidate push object filtering criteria transmitted by the demand side, and filters candidate push objects from all object sets according to these criteria. Then, based on the object ID of the candidate push object and the preset target time period, it extracts candidate behavioral features from the behavioral feature data warehouse. For example, assuming the object ID of any candidate push object is "00002" and the target time period is "November," the system extracts the behavioral feature data of object ID "00002" in November from the behavioral feature data warehouse as the candidate behavioral features corresponding to that candidate push object.

[0042] Optionally, the target behavioral features and candidate behavioral features can be behavioral features of all dimensions by default, or they can be behavioral features of some dimensions selected from behavioral features of all dimensions based on behavioral feature filtering conditions.

[0043] In another implementation, target objects that already possess the target business product are selected based on the unique product identifier of the target business product, and candidate push objects that do not possess the target business product but may potentially possess it are selected. Further, target behavior data of the target objects and candidate behavior data of the candidate push objects are extracted from the object behavior logs according to a preset target time period. Feature extraction is then performed on the target behavior data and candidate behavior data respectively to obtain target behavior features and candidate behavior features.

[0044] It is worth noting that the behavioral characteristics collected in this embodiment are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, and necessary confidentiality measures have been taken. This process does not violate public order and good morals, and corresponding operation entry points are provided for users to choose to authorize or refuse. If the user chooses to refuse, the process proceeds to the expert decision-making stage. Moreover, the collection and application of behavioral characteristics in this embodiment are only used for information push notifications for the target business product and will not be applied to other scenarios.

[0045] S102. Generate a target object behavior profile of the target object within the target time period based on the target behavior characteristics, and generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior characteristics.

[0046] Among them, the target audience behavior profile refers to a group behavior model constructed through feature analysis and pattern extraction based on the target audience's target behavioral characteristics within a target time period. Its essence is to define the behavioral paradigm of the product's core users. The candidate audience behavior profile refers to a demand potential assessment model generated based on the candidate push audience's candidate behavioral characteristics within a target time period. Its essence is to identify potential groups with behaviors similar to those of the target audience.

[0047] In one implementation, based on the target behavior characteristics, behavioral tags representing the commonalities of the target object are extracted through aggregation analysis to form a structured target object behavior profile; at the same time, based on the candidate behavior characteristics, behavioral tags of each candidate push object are extracted through aggregation analysis to form a structured candidate object behavior profile.

[0048] In another implementation, the target behavior features and candidate behavior features are respectively input into a pre-trained behavior profile generation model, wherein the behavior profile generation model includes at least one encoder and at least one decoder. The encoder sequentially encodes the target behavior features and candidate behavior features to obtain a first encoded feature vector and a second encoded feature vector. Further, the decoder sequentially decodes the first encoded feature vector and the second encoded feature vector to obtain a first decoded feature vector and a second decoded feature vector. A target object behavior profile is then generated based on the first decoded feature vector, and a candidate object behavior profile is generated based on the second decoded feature vector.

[0049] S103. Based on the target object behavior profile and the candidate object behavior profile, determine the target object to be pushed from the candidate push objects, and push information to the target object according to the push information corresponding to the target business product.

[0050] The target audience for push notifications refers to a group of users with high conversion potential selected from the candidate audience through behavioral profiling; essentially, they are the ultimate target audience for information push notifications. The push notification information refers to customized push content designed for the target business product and adapted to the characteristics of the target audience.

[0051] In one implementation, a preset similarity calculation model is used to compare the behavioral profiles of each candidate object with the behavioral profiles of each target object in multiple dimensions to quantify the behavioral fit between each candidate push object and the target object; then, based on the business objectives, a dynamic threshold is set to automatically select candidate push objects with high behavioral fit as the objects to be pushed.

[0052] In another implementation, candidate similarity vectors are generated for each candidate object behavior profile based on the cosine similarity between each candidate object behavior profile and each target object behavior profile. Each candidate object behavior profile is then clustered based on the candidate similarity vectors. Based on the clustering results, the target object behavior profile is determined from the candidate object behavior profiles. Finally, the target object is determined based on the target object behavior profile.

[0053] Furthermore, it retrieves pre-configured push notification information corresponding to the target business product, or dynamically generates personalized push notification information by matching a pre-set push notification template library with the target object's behavioral profile. Then, it pushes the information to the target object based on the push notification information, such as pushing the push notification information to the client terminal held by the target object.

[0054] This invention, through its embodiments, identifies target objects that already possess the target business product, obtains the target object's target behavioral characteristics within a target time period, and obtains candidate behavioral characteristics of candidate push objects within the target time period; generates a target object behavioral profile based on the target behavioral characteristics within the target time period, and generates candidate push object behavioral profiles based on the candidate behavioral characteristics within the target time period; and determines the objects to be pushed from the candidate push objects based on the target object behavioral profiles and candidate object behavioral profiles, and pushes information to the candidate push objects according to the push information corresponding to the target business product. The beneficial effects are:

[0055] By analyzing the behavioral profiles of candidate push targets and target users who already own the target business products within the same target time period, the system identifies the target users from the candidate push targets. This achieves the effect of identifying the target users based on their behavioral profiles, eliminating the need for separate customer group identification model training for the target business products. This improves the product universality of the target user identification and further enhances the timeliness and effectiveness of information push, ensuring the effectiveness of information push.

[0056] Example 2

[0057] Figure 2 This is a flowchart of a method for determining object behavior profiles according to Embodiment 2 of the present invention. This embodiment further optimizes and expands the technical solution of the above embodiment, "generating a candidate object behavior profile of a candidate push object within a target time period based on candidate behavior characteristics," and can be combined with the above-mentioned optional implementation methods. For example... Figure 2 As shown, the method includes:

[0058] S201. Determine the candidate sub-time periods included in the target time period, and determine the sum of the behavioral feature values ​​corresponding to the candidate behavioral features in each candidate sub-time period.

[0059] Candidate sub-time periods refer to smaller time units with continuous or fixed intervals that further subdivide the target time period. They are used to analyze the temporal changes of behavioral characteristics in a more refined manner. The core of this approach is to decompose the overall time period to reveal the dynamic patterns of the object's behavior. For example, candidate sub-time periods can be divided into target time periods according to the natural week.

[0060] The sum of behavioral feature values ​​refers to the cumulative sum of all observations of a certain candidate behavioral feature within a single candidate sub-time period, used to quantify the intensity of the object's behavior within that candidate sub-time period.

[0061] In one implementation, the target time period is evenly divided into sub-time periods to obtain at least one candidate sub-time period, such as dividing the target time period into natural weeks as candidate sub-time periods. Further, date tags corresponding to each candidate behavioral feature are determined, and the candidate behavioral features are sorted in ascending order according to the date tags. Based on the sorting result, it is determined whether there is a date difference greater than "1 day" between the date tags. If so, a first date tag and a second date tag with a date difference greater than "1 day" are determined, and a new log tag is inserted between the first and second date tags to avoid the date difference being greater than "1 day". The feature value of the candidate behavioral feature corresponding to the new log tag is then set to zero. For example, assuming the date tag of candidate behavioral feature 1 is "November 1st", and the date tag of its adjacent candidate behavioral feature 2 is "November 3rd", then a candidate behavioral feature 3 with a feature value of 0 is inserted between candidate behavioral feature 1 and candidate behavioral feature 2, and its date tag is set to "November 2nd".

[0062] Furthermore, the feature values ​​of the candidate behavioral features contained in each candidate sub-time period are summed to determine the sum of the behavioral feature values ​​corresponding to each candidate behavioral feature in each candidate sub-time period.

[0063] S202. Based on the temporal relationship between candidate sub-time periods and the sum of behavioral feature values, construct an initial behavioral feature tensor and normalize the initial behavioral feature tensor to obtain an optimized behavioral feature tensor.

[0064] The initial behavioral feature tensor refers to the original data structure constructed based on the temporal relationship of candidate sub-time periods and the sum of behavioral feature values. Its essence is the quantitative expression of user behavior in the time-feature dimension. The optimized behavioral feature tensor is a standardized data structure after normalizing the initial behavioral feature tensor, aiming to eliminate differences in units and highlight behavioral patterns.

[0065] In one implementation, the initial behavioral feature tensor is obtained by horizontally concatenating the sums of the behavioral feature values ​​according to the temporal relationship between the candidate sub-time periods. For example, if the temporal relationship between the candidate sub-time periods is "candidate sub-time period 1 - candidate sub-time period 2 - candidate sub-time period 3 - candidate sub-time period 4", and the sums of the behavioral feature values ​​corresponding to each candidate sub-time period are "18, 22, 30, 42", then the initial behavioral feature tensor [18, 22, 30, 42] is generated.

[0066] Furthermore, the initial behavioral feature tensor is normalized to obtain the optimized behavioral feature tensor.

[0067] Optionally, the initial behavior feature tensor is normalized to obtain an optimized behavior feature tensor, including:

[0068] A1. When the feature type of the candidate behavior feature is continuous, generate an equidistant element sequence based on the maximum initial element value in the initial behavior feature tensor and the preset element spacing value, and use the kernel density estimation algorithm to determine the probability density corresponding to each element value in the equidistant element sequence.

[0069] Among them, the feature type is continuous, which means that the value of the candidate behavioral feature belongs to the continuous real number field in mathematics and has infinite divisibility. For example, behavioral features such as consumption amount and page dwell time are continuous.

[0070] The maximum initial element value, also known as the maximum sum of behavioral feature values, refers to the highest sum of values ​​for a certain behavioral feature across all candidate sub-time periods in the initial behavioral feature tensor. The preset element spacing value is a fixed difference between adjacent elements in a pre-defined equidistant element sequence, used to control the balance between accuracy and computational efficiency in probability density estimation. The equidistant element sequence, generated as an arithmetic progression from the minimum value to the maximum initial element value according to the preset element spacing value, serves as the coordinate axis for probability density calculation.

[0071] Kernel density estimation is a nonparametric probability density estimation method that obtains a continuous probability distribution by placing a smoothing kernel function at each data point and superimposing the outputs of all kernels. Probability density describes the likelihood of a continuous random variable around a given value.

[0072] In one implementation, when the candidate behavioral feature is of continuous type, the maximum initial element value and a preset element spacing value are determined in the initial behavioral feature tensor. Starting with "0" as the sequence start point and ending with twice the maximum initial element value, an equidistant element sequence is generated according to the preset element spacing value. For example, ,in This represents the maximum initial element value.

[0073] Furthermore, the kernel density estimation algorithm is used to determine the probability density corresponding to each element value in the equidistant element sequence, as shown in the following formula:

[0074] ;

[0075] in, This represents the r-th element in an equidistant sequence. This represents the probability density corresponding to the r-th element. This represents the u-th element in the initial behavioral feature tensor. This represents the number of elements in the initial behavioral feature tensor. This represents the preset bandwidth parameter of the kernel density estimation algorithm. This represents the kernel density function expressed as a standard normal distribution.

[0076] A2. Based on the numerical difference between the sum of the values ​​of each behavioral feature and the values ​​of each element, determine the target element value corresponding to the sum of the values ​​of each behavioral feature from the element values, and sum the first probability density corresponding to the target element value and the second probability density corresponding to the preceding element value to generate the first cumulative probability density.

[0077] The target element value refers to the element value in the equally spaced element sequence that is closest to the sum of the numerical values ​​of a certain behavioral feature. The preceding element value is the element value that precedes the target element value in the sequence.

[0078] In one implementation, the element value with the smallest difference between the sum of the numerical values ​​of each behavioral feature and the value of each element is determined as the target element value. A first cumulative probability density is generated by summing the first probability density corresponding to the target element value and the second probability density corresponding to the preceding element values.

[0079] A3. Based on the first cumulative probability density corresponding to each target element value, determine the first normalized value of the sum of behavioral feature values ​​corresponding to each target element value, and obtain the optimized behavioral feature tensor based on each first normalized value.

[0080] In one implementation, the first cumulative probability density corresponding to each target element value is used as the first normalized value of the sum of the behavioral feature values ​​corresponding to each target element value, and then the optimized behavioral feature tensor is obtained based on each first normalized value.

[0081] When the candidate behavioral features are continuous, an equidistant element sequence is generated based on the maximum initial element value and a preset element spacing value in the initial behavioral feature tensor. A kernel density estimation algorithm is then used to determine the probability density corresponding to each element value in the equidistant element sequence. Based on the difference between the sum of all behavioral feature values ​​and the values ​​of each element, target element values ​​corresponding to the sum of all behavioral feature values ​​are determined from the element values. The first cumulative probability density is generated by summing the first probability density corresponding to the target element value and the second probability density corresponding to the preceding element value. Based on the first cumulative probability density corresponding to each target element value, a first normalized value of the sum of behavioral feature values ​​corresponding to each target element value is determined. Finally, an optimized behavioral feature tensor is obtained based on each first normalized value. The beneficial effects are:

[0082] Firstly, the sum of the behavioral feature values ​​of candidate behavioral features in candidate sub-time periods is preserved by using the initial behavioral feature tensor, and a continuous probability distribution is generated by using the kernel density estimation algorithm, so that the normalization result has both dynamic evolution characteristics and population distribution location awareness.

[0083] Secondly, the probability density-based normalization mechanism has a natural adaptability to complex forms such as long-tailed distributions and multi-peaked distributions, and solves the problems of feature distribution skewness and scale conflict.

[0084] Thirdly, optimizing the behavioral feature tensor is essentially an embedded representation of spatiotemporal semantics, thereby providing high-quality spatiotemporal fusion input for subsequent behavioral profile generation models and improving the generation accuracy of candidate object behavioral profiles.

[0085] Optionally, the initial behavior feature tensor is normalized to obtain an optimized behavior feature tensor, including:

[0086] B1. When the feature type of the candidate behavioral feature is discrete, determine the number of times the sum of the values ​​of each behavioral feature appears in the initial behavioral feature tensor, and generate the mapping elements corresponding to the sum of the values ​​of each behavioral feature according to the proportion of the number of times each value appears.

[0087] In this context, "discrete feature type" refers to candidate behavioral features whose values ​​belong to a finite countable set mathematically, with clear intervals between values ​​that cannot be infinitely subdivided, such as coupon types. "Number of occurrences" refers to the total frequency of a particular behavioral feature's sum across all candidate sub-time periods in the initial behavioral feature tensor. "Percentage of occurrences" is determined based on the number of occurrences and the total number of occurrences in the initial behavioral feature tensor. "Mapping element" refers to converting the discrete sum of behavioral feature values ​​into a probability proxy value based on the percentage of occurrences, which serves as input for subsequent distribution fitting.

[0088] In one implementation, when the feature type of the candidate behavioral feature is discrete, the number of times each behavioral feature's numerical sum appears in the initial behavioral feature tensor is determined, and the percentage of each behavioral feature's numerical sum is determined based on the numerical sum's occurrence count and the total number of numerical sums in the initial behavioral feature tensor. Further, based on the percentage of each behavioral feature's numerical sum's occurrence count, a logarithmic operation is performed to generate the mapping elements corresponding to each behavioral feature's numerical sum.

[0089] For example, ;in, Represents the sum of behavioral characteristics. The number of times the value appears, This indicates the total number of times the value appears. Represents the sum of behavioral characteristics. The corresponding mapping element.

[0090] B2. The probability density function is obtained by estimating the exponential distribution parameters of each mapping element, and the second cumulative probability density of each mapping element with respect to the probability density function is determined.

[0091] The probability density function is a continuous probability distribution model, describing the probability density of a random variable at any given value. In discrete feature scenarios, it is used to model the distribution pattern of the mapped elements. The second cumulative probability density represents the cumulative distribution function value of the mapped elements.

[0092] For example, the second cumulative probability density is determined as follows:

[0093] ;

[0094] in, Represents the sum of behavioral characteristics. The corresponding mapping element, Represents the probability density function. Represents mapping elements Regarding probability density function The second cumulative probability density.

[0095] B3. Based on the second cumulative probability density corresponding to each mapping element, determine the second normalized value of the sum of behavioral feature values ​​corresponding to each mapping element, and obtain the optimized behavioral feature tensor based on each second normalized value.

[0096] When the candidate behavioral features are discrete, the number of occurrences of each behavioral feature's total numerical value in the initial behavioral feature tensor is determined, and mapping elements corresponding to the total numerical values ​​of each behavioral feature are generated based on the proportion of each value's occurrence. The probability density function is obtained by estimating the exponential distribution parameters of each mapping element, and the second cumulative probability density of each mapping element with respect to the probability density function is determined. Based on the second cumulative probability density of each mapping element, the second normalized value of the total numerical value of the behavioral features corresponding to each mapping element is determined, and the optimized behavioral feature tensor is obtained based on each second normalized value. The beneficial effects are:

[0097] Traditional discrete features lack continuous dimensions, making it impossible to directly compare the importance of different behavioral features. This scheme generates mapping elements by the proportion of numerical occurrences, transforming discrete values ​​into probabilistic indicators representing the scarcity of behaviors. Furthermore, it calculates the second cumulative probability density by fitting an exponential distribution, thus normalizing all discrete behavioral features to the [0,1] probability space and achieving comparability of discrete behavioral features.

[0098] S203. Generate a candidate object behavior profile for the candidate push object within the target time period based on the optimized behavior feature tensor.

[0099] By identifying candidate sub-time periods within the target time period and determining the sum of behavioral feature values ​​corresponding to each candidate sub-time period; constructing an initial behavioral feature tensor based on the temporal relationship between candidate sub-time periods and the sum of behavioral feature values; and normalizing the initial behavioral feature tensor to obtain an optimized behavioral feature tensor; and generating a candidate object behavior profile of the candidate push object within the target time period based on the optimized behavioral feature tensor, the beneficial effects are:

[0100] Firstly, by capturing the periodic rhythm of behavior through candidate sub-time periods, and using the initial behavior feature tensor to encode spatiotemporal relationships, a high-quality spatiotemporal fusion input is provided for the subsequent behavior profile generation model, thereby improving the generation accuracy of candidate object behavior profiles.

[0101] Secondly, by normalizing the initial behavioral feature tensor, an optimized behavioral feature tensor is obtained, eliminating dimensional interference and further improving the accuracy of subsequent candidate object behavioral profile generation.

[0102] Example 3

[0103] Figure 3A This is a flowchart of a method for determining object behavior profiles provided in Embodiment 3 of the present invention. This embodiment further optimizes and expands the technical solution of the above embodiments, "generating a candidate object behavior profile of a candidate push object within a target time period based on candidate behavior characteristics," and can be combined with the above-mentioned optional implementation methods. For example... Figure 3A As shown, the method includes:

[0104] S301. Input the candidate behavioral features into the behavioral profile generation model.

[0105] The behavior profile generation model includes an embedding layer, a location encoding layer, an encoder, a decoder, a first linear layer, and a second linear layer.

[0106] S302. The candidate behavioral features are processed in a low-dimensional and dense manner through the embedding layer to generate an initial feature vector.

[0107] The embedding layer, also known as the embedding layer, is a neural network layer that maps high-dimensional discrete features or structured continuous features to a low-dimensional continuous vector space. Low-dimensional dense processing refers to the transformation process of converting sparse high-dimensional features into dense low-dimensional vectors through the embedding layer. The initial feature vector is a low-dimensional dense vector output by the embedding layer, representing a unified form of candidate behavioral features.

[0108] In one implementation, a predefined embedding matrix is ​​constructed in the embedding layer to compress high-dimensional sparse features into a dense representation. When candidate behavioral features are input into the embedding layer, the row vector corresponding to each feature value index in the embedding matrix is ​​found based on the feature value of the candidate behavioral feature to obtain its dense embedding vector. Then, all dense embedding vectors are averaged or concatenated to integrate them into a unified initial feature vector.

[0109] S303. The initial feature vector is positionally encoded by the positional encoding layer to generate a positionally encoded vector. The positionally encoded vector and the initial feature vector are then added together to generate a summed feature vector.

[0110] The positional encoding layer is a dedicated component in the behavior profiling generation model. It injects sequential information of elements into the model, ensuring the continuity of the learned behavioral features. Positional encoding is a mathematical representation that maps discrete positional indices to continuous vectors. The positional encoding vector is the output of positional encoding, representing the mathematical vectorized expression of a specific position. The summative feature vector is a fused vector generated by element-wise addition of the initial feature vector (original behavioral semantics) and the positional encoding vector (behavioral sequence information).

[0111] In one implementation, based on the sequence length and dimension of the initial feature vector, a position encoding vector is generated by a position encoding layer using predefined sine and cosine function formulas, thereby encoding the absolute position information of each behavioral feature. Then, the generated position encoding vector is added element-wise to the initial feature vector, that is, directly summed under the same dimension and the same position index, to form an additive feature vector that integrates behavioral semantics and temporal order, which is used as the input of the subsequent encoder to achieve sequence modeling.

[0112] S304. The encoder encodes the summed feature vectors to generate an encoded feature vector.

[0113] The encoder is the core component of the behavior profiling generation model, responsible for fusing contextual information and extracting high-order features. The behavior profiling generation model includes at least one encoder, and for example, it may include four interconnected encoders. Encoding refers to the process by which the encoder compresses and semantically abstracts the summed feature vectors. The encoded feature vector is the high-order semantic representation output by the encoder, carrying the global statistical characteristics and latent patterns of the sequence behavior.

[0114] Optionally, the encoder includes a first multi-head self-attention module and a first feedforward neural network module.

[0115] The first multi-head self-attention module is a group of multi-view attention mechanisms running in parallel within the encoder, used to mine dependencies between behavioral sequences from different subspaces. The first feedforward neural network module is a fully connected nonlinear transformation layer located after the first multi-head self-attention module, acting independently at each position of the sequence.

[0116] The encoder encodes the summed feature vectors to generate encoded feature vectors, including:

[0117] The first self-attention feature vector is generated by performing multi-head self-attention calculation based on the summed feature vector through the multi-head self-attention module; the encoded feature vector is generated by performing nonlinear transformation based on the first self-attention feature vector through the first feedforward neural network module.

[0118] Multi-head self-attention computation involves calculating the correlation between sequence elements in parallel using multiple sets of weight matrices. Nonlinear transformation uses activation functions to perform nonlinear mapping on the results of linear operations, overcoming the expressive limitations of linear models.

[0119] In one implementation, the encoder receives the summed feature vector output by the positional encoding layer as input. Then, a multi-head self-attention module calculates the relevance weights of each vector in the vector sequence to all other vectors, thereby generating a context-aware weighted sum representation that captures dependencies between behaviors and generates a first self-attention feature vector. Next, residual connections and layer normalization are applied to the first self-attention feature vector to ensure effective gradient propagation. Subsequently, a first feedforward neural network module performs further nonlinear transformations to enhance expressive power, and residual connections and layer normalization are applied again. Finally, if the encoder has a multi-layer stacked structure, the above multi-head self-attention calculation and nonlinear transformation steps are repeated multiple times to gradually abstract higher-order semantics and output the encoded feature vector.

[0120] By using a multi-head self-attention module to perform multi-head self-attention calculations based on the summed feature vectors to generate the first self-attention feature vector, the beneficial effects are: enabling the model to dynamically capture the complex dependencies of behavior in the summed feature vector sequence from different semantic subspaces, thereby generating a context-aware first self-attention feature vector; this design avoids the bias of a single attention mechanism, improves the robustness and expressive power of the model, and at the same time alleviates the gradient vanishing problem through residual connections and layer normalization, ensuring the stability of feature fusion.

[0121] By performing a nonlinear transformation on the first self-attention feature vector using the first feedforward neural network module, an encoded feature vector is generated. The beneficial effects are: the vector at each sequence position is transformed nonlinearly independently, which breaks through the limitations of the linear model, deepens the contextual information of the first self-attention feature vector, and enhances the model's ability to generalize to noise and sparse data.

[0122] S305. The encoded feature vector is linearly mapped through the first linear layer to generate the mean vector and variance vector, and Gaussian sampling is performed based on the mean vector and variance vector to generate the sampled feature vector.

[0123] The first linear layer is a fully connected neural network layer that plays a core role in generating probability distribution parameters. Linear mapping refers to the affine transformation of the encoded feature vector through the weight matrix and bias vector. The mean vector describes the central location of the latent variable probability distribution, while the variance vector describes the degree of dispersion of the latent variable probability distribution. The sampled feature vector is a latent variable sample randomly drawn from a Gaussian distribution defined by the mean and variance vectors.

[0124] In one implementation, a dual-linear mapping is performed on the encoded feature vector through a first linear layer: a set of independent weight matrices and bias vectors maps the input to a mean vector, and another set maps it to a variance vector in parallel; finally, Gaussian sampling is performed based on the reparameterization technique to generate sampled feature vectors.

[0125] S306. The decoder performs decoding based on the summed feature vector and the sampled feature vector to generate a decoded feature vector.

[0126] The decoder is the core component of the behavior profile generation model, responsible for fusing latent variables with contextual information to reconstruct or generate the target output. Decoding refers to the inverse transformation process by which the decoder combines latent space samples with the original context to generate a semantic representation of the target. The decoded feature vector is a high-order behavioral semantic representation output by the decoder and is the direct precursor to the candidate object's behavior profile.

[0127] Optionally, the decoder includes a masked multi-head self-attention module, a second multi-head self-attention module, and a second feedforward neural network module.

[0128] The masked multi-head self-attention module is a self-attention mechanism in the decoder that processes the original behavior sequence. It forces the model to only acquire current and preceding information through masking, preventing it from knowing subsequent information at any given moment. This embodiment provides a unique masking mechanism to help the model adaptively learn sparse candidate behavior features and effectively improve these features, significantly increasing the model's marketing recommendation scale. Specifically, for the additive feature vector, the original feature mask data is first calculated. Following the time dimension, if the values ​​for the past N weeks are all zero (e.g., 52 weeks), the mask flag is set to 1; otherwise, it is 0.

[0129] The second multi-head self-attention module is a group of multi-view attention mechanisms running in parallel within the decoder, used to mine dependencies between behavioral sequences from different subspaces. The second feedforward neural network module is a fully connected nonlinear transformation layer following the second multi-head self-attention module, acting independently at each position of the sequence.

[0130] The decoder performs decoding based on the summed feature vector and the sampled feature vector to generate a decoded feature vector, including:

[0131] The masked multi-head self-attention module performs masked multi-head self-attention calculation based on the summed feature vector to generate a second self-attention feature vector; the second multi-head self-attention module uses the sampled feature vector as the key vector and value vector respectively, and the summed feature vector as the query vector, to perform multi-head self-attention calculation to generate a third self-attention feature vector; the second feedforward neural network module performs nonlinear transformation based on the third self-attention feature vector to generate a decoded feature vector.

[0132] In one implementation, the summed feature vector is received by a masked multi-head self-attention module. A masking mechanism ensures temporal compliance. For example, for the summed feature vector, the original feature mask data is first calculated. If the values ​​for the past N weeks are all zero (e.g., week 52), the mask flag is set to 1; otherwise, it is 0. Simultaneously, the summed feature vector is split into multiple subspaces. For each subspace, the dot product similarity of the query, key, and value matrices is independently calculated to generate attention weights to capture dependencies within the behavioral sequence. After weighted aggregation, the multi-head outputs are concatenated, and residual connections and layer normalization are applied to generate a second self-attention feature vector containing only the historical behavioral context.

[0133] Furthermore, multi-head self-attention calculation is performed through the second multi-head self-attention module: the summed feature vector is used as the query vector, and the sampled feature vector is used as the key vector and value vector. The similarity between the query vector and the key vector is calculated in parallel by multiple attention heads, and the value vector is dynamically weighted and aggregated to generate a third self-attention feature vector that integrates individual behavioral details and generalization pattern statistics.

[0134] Furthermore, the third self-attention feature vector is nonlinearly transformed through the second feedforward neural network module: a high-order feature transformation is achieved through two fully connected layers, which maps the third self-attention feature vector into interpretable semantics, and then the decoded feature vector is output through residual connection and layer normalization.

[0135] By using a masked multi-head self-attention module to perform masked multi-head self-attention calculation based on the summed feature vectors, a second self-attention feature vector is generated. The beneficial effect is that the masking mechanism forces the model to only obtain current and previous information, and cannot know the subsequent information at this moment, ensuring that the portrait generation strictly conforms to the temporal causality and avoiding logical errors that could lead to future information leakage.

[0136] By using the second multi-head self-attention module, the sampled feature vectors are used as key vectors and value vectors respectively, and the summed feature vector is used as the query vector. Multi-head self-attention calculation is performed to generate the third self-attention feature vector. The beneficial effects are: to realize the cross-attention mechanism, to provide a generalized knowledge base with the sampled feature vectors as key vectors and value vectors, and to use the summed feature vector as the query vector. Through multi-head parallel calculation of similarity weights, the generalization mode and individual evidence are dynamically fused to generate the third self-attention feature vector, which effectively balances personalization and generalization and solves the overfitting risk in data sparse scenarios.

[0137] By performing a nonlinear transformation on the third self-attention feature vector through the second feedforward neural network module, a decoding feature vector is generated. The beneficial effect is that it provides a highly discriminative and interpretable feature foundation for the subsequent generation of candidate object behavior profiles by the second linear layer, which significantly improves the accuracy and adaptability of profile generation.

[0138] S307. The decoded feature vector is linearly mapped through the second linear layer to generate a candidate object behavior profile.

[0139] The second linear layer, located at the output of the decoder, is a fully connected neural network layer responsible for the final probabilistic generation of candidate object behavior profiles. Linear mapping is a direct transformation from input features to the output space achieved through affine transformation.

[0140] By performing low-dimensional dense processing on candidate behavioral features through an embedding layer to generate an initial feature vector, the beneficial effect is that it compresses high-dimensional sparse candidate behavioral features into a low-dimensional dense representation, effectively solving the curse of dimensionality problem caused by excessively high dimensionality in the discrete feature space.

[0141] The initial feature vector is positionally encoded by the positional encoding layer to generate a positionally encoded vector. The positionally encoded vector and the initial feature vector are then added together to generate a summed feature vector. The benefits are that the model can perceive the sequential dependency of behavior while preserving the semantics of the behavior, avoid the problem of loss of temporal information in traditional sequence models, and provide the encoder with a complete input that integrates semantics and structure.

[0142] By encoding the summed feature vectors using an encoder, an encoded feature vector is generated. The benefit is that it improves the model's ability to model complex behavioral chains.

[0143] The first linear layer performs a linear mapping on the encoded feature vector to generate a mean vector and a variance vector. Gaussian sampling is then performed on the mean vector and the variance vector to generate a sampled feature vector. The beneficial effects are: transforming deterministic features into probability distribution parameters, realizing the continuity and diversity modeling of the latent space, avoiding the risk of overfitting caused by deterministic encoding, and enhancing the model's ability to generalize to sparse data.

[0144] The decoder decodes the summed feature vector and the sampled feature vector to generate a decoded feature vector. The beneficial effect is that it achieves the effect of fusing latent variables with contextual information and reconstructing the decoded feature vector.

[0145] By performing a linear mapping on the decoded feature vector through the second linear layer, a behavioral profile of the candidate object is generated. The beneficial effect is that it realizes the efficient conversion from latent space features to a behavioral profile that can be implemented in business, and finally outputs a user profile that is personalized, generalizable and interpretable.

[0146] Optionally, the behavioral profile generation model is generated in the following way:

[0147] 1) Using a preset batch size as the standard, normalize the customer behavior tensor. Divide into batch tensors of multiple sizes.

[0148] 2) A Transformer-based deep neural network model is used, with batch tensors as training data, and the network is trained according to preset training epochs and learning rate parameters. Each batch tensor generates a reconstructed tensor of the same size after passing through the decoder. The decoder generates two vectors: the expected value and the variance of the object behavior profile representation vector. The loss function used for model training consists of variational autoencoder loss, negative logarithmic loss, and contrastive loss.

[0149] 3) Select the network parameters that perform best on the validation set as the final network parameters to obtain the behavior profile generation model.

[0150] Figure 3B This is a schematic diagram of the structure of a behavior profile generation model provided in Embodiment 3 of the present invention, as shown below. Figure 3B As shown, the behavior profile generation model includes an embedding layer 300, a location encoding layer 301, an encoder 302, a first linear layer 303, a decoder 304, and a second linear layer 305. The encoder 302 includes a multi-head self-attention module 3021, a residual connection and layer normalization module 3022, and a first feedforward neural network 3023. The decoder 304 includes a masked multi-head self-attention module 3041, a residual connection and layer normalization module 3022, a second multi-head self-attention module 3042, and a second feedforward neural network module 3043.

[0151] The specific functions of each functional layer and module in the above-mentioned behavior profile generation model are detailed in the above-mentioned embodiments and will not be repeated here.

[0152] Example 4

[0153] Figure 4 This is a flowchart of a method for determining a target to be pushed to, provided in Embodiment 4 of the present invention. This embodiment further optimizes and expands the technical solution of the above embodiments, "determining the target to be pushed from candidate push objects based on the target object's behavioral profile and the candidate object's behavioral profile," and can be combined with the above-mentioned optional implementation methods. For example... Figure 4 As shown, the method includes:

[0154] S401. Based on the cosine similarity between each candidate object's behavioral profile and each target object's behavioral profile, generate candidate similarity vectors corresponding to each candidate object's behavioral profile.

[0155] Among them, the candidate similarity vector refers to the set of similarities between a single candidate object behavior profile and all target object behavior profiles.

[0156] S402. Cluster the behavioral profiles of each candidate object according to the similarity vectors of each candidate object to obtain at least one candidate cluster, and determine the target cluster from each candidate cluster according to the modulus of each candidate cluster.

[0157] Here, candidate clusters refer to groups of candidate object behavior profiles grouped using clustering algorithms, and the modulus of a candidate cluster is defined as the average norm of all vectors within the cluster. The target cluster refers to the cluster with the highest modulus selected from the candidate clusters as the push target.

[0158] In one implementation, a clustering model is trained with a preset number of clusters (2), and the candidate cluster with the largest modulus value is selected as the target cluster to obtain the object recognition model to be pushed.

[0159] S403. Take the behavioral profiles of the candidate objects contained in the target cluster as the behavioral profiles of the objects to be pushed, and determine the objects to be pushed from the candidate objects based on the behavioral profiles of the objects to be pushed.

[0160] Among them, the behavioral profile of the target object to be pushed refers to the set of behavioral profiles of each candidate target object within the target cluster.

[0161] In one implementation, the target object identification model is used to extract the target object behavior profile contained in the target cluster, and the target object is determined according to the object number corresponding to the target object behavior profile.

[0162] By generating candidate similarity vectors for each candidate object's behavior profile based on the cosine similarity between each candidate object's behavior profile and each target object's behavior profile, and then clustering each candidate object's behavior profile using these candidate similarity vectors to obtain at least one candidate cluster, and determining the target cluster from each candidate cluster based on the modulus of each candidate cluster, the candidate object's behavior profile contained in the target cluster is used as the behavior profile of the object to be pushed to, and the object to be pushed to is determined from the candidate objects based on the behavior profile of the object to be pushed to. The beneficial effects are:

[0163] Firstly, replacing candidate object behavior profiles with candidate similarity vectors reduces computational complexity and improves the efficiency of determining the objects to be pushed.

[0164] Secondly, the target cluster is selected based on the modulus of the candidate clusters to improve the correlation between the object to be pushed and the target object.

[0165] Example 5

[0166] Figure 5 This is a schematic diagram of an information push device provided in Embodiment 5 of the present invention. It can be applied to situations where a target object is selected from candidate push objects and pushes the target business product information to the target object. Figure 5 As shown, the device includes:

[0167] The behavior feature determination module 51 is used to determine the target object that has already held the target business product, and to obtain the target behavior features of the target object in the target time period, and to obtain the candidate behavior features of the candidate push object in the target time period.

[0168] The object behavior profile generation module 52 is used to generate a target object behavior profile of the target object within the target time period based on the target behavior features, and to generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior features.

[0169] The target object determination module 53 is used to determine the target object from the candidate push objects based on the target object behavior profile and the candidate object behavior profile, and push information to the target object according to the push information corresponding to the target business product.

[0170] Optionally, the object behavior profile generation module 52 is specifically used for:

[0171] Determine the candidate sub-time periods included in the target time period, and determine the sum of the behavioral feature values ​​corresponding to the candidate behavioral features in each of the candidate sub-time periods;

[0172] Based on the temporal relationship between the candidate sub-time periods and the sum of the behavioral feature values, an initial behavioral feature tensor is constructed, and the initial behavioral feature tensor is normalized to obtain an optimized behavioral feature tensor.

[0173] Generate a candidate object behavior profile for the candidate push object within the target time period based on the optimized behavior feature tensor.

[0174] Optionally, the object behavior profile generation module 52 is further used for:

[0175] When the feature type of the candidate behavior feature is continuous, an equidistant element sequence is generated based on the maximum initial element value in the initial behavior feature tensor and the preset element spacing value, and the probability density corresponding to each element value in the equidistant element sequence is determined by the kernel density estimation algorithm.

[0176] Based on the numerical difference between the sum of the behavioral feature values ​​and each element value, a target element value corresponding to the sum of the behavioral feature values ​​is determined from each element value. Then, a first cumulative probability density is generated by summing the first probability density corresponding to the target element value and the second probability density corresponding to the preceding element value. The preceding element value is the element value whose sequence order is before the target element value.

[0177] Based on the first cumulative probability density corresponding to each of the target element values, a first normalized value of the sum of the behavioral feature values ​​corresponding to each of the target element values ​​is determined, and the optimized behavioral feature tensor is obtained based on each of the first normalized values.

[0178] Optionally, the object behavior profile generation module 52 is further used for:

[0179] When the feature type of the candidate behavioral feature is discrete, the number of times each of the behavioral feature values ​​appears in the initial behavioral feature tensor is determined, and mapping elements corresponding to the sum of each behavioral feature value are generated according to the proportion of each value's occurrence; wherein, the proportion of the number of times the value appears is determined based on the number of times the value appears and the total number of times the value appears in the initial behavioral feature tensor.

[0180] The probability density function is obtained by estimating the exponential distribution parameters of each of the mapping elements, and the second cumulative probability density of each of the mapping elements with respect to the probability density function is determined.

[0181] Based on the second cumulative probability density corresponding to each of the mapping elements, a second normalized value of the sum of the behavioral feature values ​​corresponding to each of the mapping elements is determined, and the optimized behavioral feature tensor is obtained based on each of the second normalized values.

[0182] Optionally, the object behavior profile generation module 52 is further used for:

[0183] The candidate behavioral features are input into the behavioral profile generation model; wherein, the behavioral profile generation model includes an embedding layer, a location encoding layer, an encoder, a decoder, a first linear layer, and a second linear layer;

[0184] The candidate behavior features are processed in a low-dimensional, dense manner through the embedding layer to generate an initial feature vector.

[0185] The initial feature vector is positionally encoded by the positional encoding layer to generate a positional encoding vector, and the positional encoding vector and the initial feature vector are added together to generate a summed feature vector.

[0186] The encoder encodes the summed feature vectors to generate an encoded feature vector.

[0187] The encoded feature vector is linearly mapped through the first linear layer to generate a mean vector and a variance vector, and Gaussian sampling is performed based on the mean vector and the variance vector to generate a sampled feature vector;

[0188] The decoder performs decoding based on the summed feature vector and the sampled feature vector to generate a decoded feature vector;

[0189] The second linear layer performs a linear mapping on the decoded feature vector to generate the candidate object behavior profile.

[0190] Optionally, the encoder includes a first multi-head self-attention module and a first feedforward neural network module;

[0191] The object behavior profile generation module 52 is further used for:

[0192] The multi-head self-attention module performs multi-head self-attention calculation based on the summed feature vector to generate a first self-attention feature vector;

[0193] The first feedforward neural network module performs a nonlinear transformation on the first self-attention feature vector to generate the encoded feature vector.

[0194] Optionally, the decoder includes a masked multi-head self-attention module, a second multi-head self-attention module, and a second feedforward neural network module;

[0195] The object behavior profile generation module 52 is further used for:

[0196] The masked multi-head self-attention module performs masked multi-head self-attention calculation based on the summed feature vector to generate a second self-attention feature vector.

[0197] The second multi-head self-attention module uses the sampled feature vectors as key vectors and value vectors respectively, and the summed feature vectors as query vectors, to perform multi-head self-attention calculations to generate a third self-attention feature vector.

[0198] The second feedforward neural network module performs a nonlinear transformation based on the third self-attention feature vector to generate the decoded feature vector.

[0199] Optionally, the module 53 for determining the object to be pushed is specifically used for:

[0200] Based on the cosine similarity between each candidate object behavior profile and each target object behavior profile, a candidate similarity vector is generated for each candidate object behavior profile.

[0201] Clustering of the behavioral profiles of each candidate object based on the candidate similarity vectors yields at least one candidate cluster, and the target cluster is determined from each candidate cluster based on the modulus of each candidate cluster.

[0202] The behavioral profiles of the candidate objects contained in the target cluster are used as the behavioral profiles of the objects to be pushed, and the objects to be pushed are determined from the candidate objects based on the behavioral profiles of the objects to be pushed.

[0203] The information push device provided in the embodiments of the present invention can execute the information push method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0204] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0205] Example 6

[0206] Figure 6A schematic diagram of an electronic device 60 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0207] like Figure 6 As shown, the electronic device 60 includes at least one processor 61 and a memory, such as a read-only memory (ROM) 62 and a random access memory (RAM) 63, communicatively connected to the at least one processor 61. The memory stores computer programs executable by the at least one processor. The processor 61 can perform various appropriate actions and processes based on the computer program stored in the ROM 62 or loaded from storage unit 68 into the RAM 63. The RAM 63 can also store various programs and data required for the operation of the electronic device 60. The processor 61, ROM 62, and RAM 63 are interconnected via a bus 64. An input / output (I / O) interface 65 is also connected to the bus 64.

[0208] Multiple components in electronic device 60 are connected to I / O interface 65, including: input unit 66, such as keyboard, mouse, etc.; output unit 67, such as various types of monitors, speakers, etc.; storage unit 68, such as disk, optical disk, etc.; and communication unit 69, such as network card, modem, wireless transceiver, etc. Communication unit 69 allows electronic device 60 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0209] Processor 61 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 61 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 61 performs the various methods and processes described above, such as information push methods.

[0210] In some embodiments, the information pushing method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 68. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 60 via ROM 62 and / or communication unit 69. When the computer program is loaded into RAM 63 and executed by processor 61, one or more steps of the information pushing method described above may be performed. Alternatively, in other embodiments, processor 61 may be configured to perform the information pushing method by any other suitable means (e.g., by means of firmware).

[0211] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0212] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0213] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0214] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0215] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0216] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product within the cloud computing service system to address the shortcomings of traditional physical hosts and virtual private servers, such as high management difficulty and weak business scalability.

[0217] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0218] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for pushing information, characterized in that, The method includes: Identify the target objects that already possess the target business products, and obtain the target behavioral characteristics of the target objects within the target time period, as well as obtain the candidate behavioral characteristics of the candidate push objects within the target time period; Generate a target object behavior profile of the target object within the target time period based on the target behavior features, and generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior features; Based on the target object behavior profile and the candidate object behavior profile, the target object to be pushed is determined from the candidate push objects, and information is pushed to the target object according to the push information corresponding to the target business product.

2. The method according to claim 1, characterized in that, The step of generating a candidate object behavior profile for the candidate push object within the target time period based on the candidate behavior features includes: Determine the candidate sub-time periods included in the target time period, and determine the sum of the behavioral feature values ​​corresponding to the candidate behavioral features in each of the candidate sub-time periods; Based on the temporal relationship between the candidate sub-time periods and the sum of the behavioral feature values, an initial behavioral feature tensor is constructed, and the initial behavioral feature tensor is normalized to obtain an optimized behavioral feature tensor. Generate a candidate object behavior profile for the candidate push object within the target time period based on the optimized behavior feature tensor.

3. The method according to claim 2, characterized in that, The step of normalizing the initial behavior feature tensor to obtain the optimized behavior feature tensor includes: When the feature type of the candidate behavior feature is continuous, an equidistant element sequence is generated based on the maximum initial element value in the initial behavior feature tensor and the preset element spacing value, and the probability density corresponding to each element value in the equidistant element sequence is determined by the kernel density estimation algorithm. Based on the numerical difference between the sum of the behavioral feature values ​​and each element value, a target element value corresponding to the sum of the behavioral feature values ​​is determined from each element value. Then, a first cumulative probability density is generated by summing the first probability density corresponding to the target element value and the second probability density corresponding to the preceding element value. The preceding element value is the element value whose sequence order is before the target element value. Based on the first cumulative probability density corresponding to each of the target element values, a first normalized value of the sum of the behavioral feature values ​​corresponding to each of the target element values ​​is determined, and the optimized behavioral feature tensor is obtained based on each of the first normalized values.

4. The method according to claim 2, characterized in that, The step of normalizing the initial behavior feature tensor to obtain the optimized behavior feature tensor includes: When the feature type of the candidate behavioral feature is discrete, the number of times each of the behavioral feature values ​​appears in the initial behavioral feature tensor is determined, and mapping elements corresponding to the sum of each behavioral feature value are generated according to the proportion of each value's occurrence; wherein, the proportion of the number of times the value appears is determined based on the number of times the value appears and the total number of times the value appears in the initial behavioral feature tensor. The probability density function is obtained by estimating the exponential distribution parameters of each of the mapping elements, and the second cumulative probability density of each of the mapping elements with respect to the probability density function is determined. Based on the second cumulative probability density corresponding to each of the mapping elements, a second normalized value of the sum of the behavioral feature values ​​corresponding to each of the mapping elements is determined, and the optimized behavioral feature tensor is obtained based on each of the second normalized values.

5. The method according to claim 1, characterized in that, The step of generating a candidate object behavior profile for the candidate push object within the target time period based on the candidate behavior features includes: The candidate behavioral features are input into the behavioral profile generation model; wherein, the behavioral profile generation model includes an embedding layer, a location encoding layer, an encoder, a decoder, a first linear layer, and a second linear layer; The candidate behavior features are processed in a low-dimensional, dense manner through the embedding layer to generate an initial feature vector. The initial feature vector is positionally encoded by the positional encoding layer to generate a positional encoding vector, and the positional encoding vector and the initial feature vector are added together to generate a summed feature vector. The encoder encodes the summed feature vectors to generate an encoded feature vector. The encoded feature vector is linearly mapped through the first linear layer to generate a mean vector and a variance vector, and Gaussian sampling is performed based on the mean vector and the variance vector to generate a sampled feature vector; The decoder performs decoding based on the summed feature vector and the sampled feature vector to generate a decoded feature vector; The second linear layer performs a linear mapping on the decoded feature vector to generate the candidate object behavior profile.

6. The method according to claim 5, characterized in that, The encoder includes a first multi-head self-attention module and a first feedforward neural network module; The step of encoding the summed feature vectors using the encoder to generate encoded feature vectors includes: The multi-head self-attention module performs multi-head self-attention calculation based on the summed feature vector to generate a first self-attention feature vector; The first feedforward neural network module performs a nonlinear transformation on the first self-attention feature vector to generate the encoded feature vector.

7. The method according to claim 5, characterized in that, The decoder includes a masked multi-head self-attention module, a second multi-head self-attention module, and a second feedforward neural network module; The step of decoding using the decoder based on the summed feature vector and the sampled feature vector to generate a decoded feature vector includes: The masked multi-head self-attention module performs masked multi-head self-attention calculation based on the summed feature vector to generate a second self-attention feature vector. The second multi-head self-attention module uses the sampled feature vectors as key vectors and value vectors respectively, and the summed feature vectors as query vectors, to perform multi-head self-attention calculations to generate a third self-attention feature vector. The second feedforward neural network module performs a nonlinear transformation based on the third self-attention feature vector to generate the decoded feature vector.

8. The method according to claim 1, characterized in that, The step of determining the target object to be pushed from the candidate push objects based on the target object behavior profile and the candidate object behavior profile includes: Based on the cosine similarity between each candidate object behavior profile and each target object behavior profile, a candidate similarity vector is generated for each candidate object behavior profile. Clustering of the behavioral profiles of each candidate object based on the candidate similarity vectors yields at least one candidate cluster, and the target cluster is determined from each candidate cluster based on the modulus of each candidate cluster. The behavioral profiles of the candidate objects contained in the target cluster are used as the behavioral profiles of the objects to be pushed, and the objects to be pushed are determined from the candidate objects based on the behavioral profiles of the objects to be pushed.

9. An information push device, characterized in that, The device includes: The behavior feature determination module is used to determine the target object that already holds the target business product, and to obtain the target behavior features of the target object within the target time period, as well as to obtain the candidate behavior features of the candidate push object within the target time period. The object behavior profile generation module is used to generate a target object behavior profile of the target object within the target time period based on the target behavior features, and to generate a candidate object behavior profile of the candidate push object within the target time period based on the candidate behavior features. The module for determining the target object is used to determine the target object from the candidate push objects based on the target object behavior profile and the candidate object behavior profile, and to push information to the target object according to the push information corresponding to the target business product.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the information push method according to any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to execute the information push method according to any one of claims 1-8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the information push method according to any one of claims 1-8.

Images