User behavior prediction system and method based on historical data

Abstract

This application discloses a user behavior prediction system and method based on historical data, which uses deep learning-based artificial intelligence technology to deeply mine the user's historical behavior data, capture the time-series evolution characteristics of the user's behavior pattern, and then predict the user's purchase behavior in advance based on the dynamic interactive correlation between the user's behavior pattern and commodity data. In this way, it can provide enterprises with more accurate marketing strategies and personalized recommendation services, enhance user experience, and promote sales conversion.

Description

Technical Field

[0001] The present application relates to the field of intelligent prediction, and more specifically, to a user behavior prediction system and method based on historical data. Background Art

[0002] With the rapid development of the Internet and mobile Internet, various applications and platforms have emerged one after another, and the number of users has increased dramatically. These platforms have accumulated a large amount of user behavior data during their operations, such as browsing history, search history, purchase history, social interaction, etc. These data have become an important resource for companies to analyze user behavior, optimize products and services, and formulate market strategies.

[0003] Through in-depth mining and analysis of user behavior data, companies can understand user needs more accurately, thereby providing users with personalized products and services and improving user satisfaction and loyalty. At the same time, user behavior data also provides strong support for the company's product development, marketing and customer service, and can help companies optimize resource allocation and enhance market competitiveness.

[0004] Traditional user behavior prediction methods are mainly based on simple statistical models, such as collaborative filtering and Markov chains. These methods have great limitations when processing large-scale user behavior data. They find it difficult to capture the complexity and temporal dynamics of user behavior, which often leads to inaccurate prediction results.

[0005] Therefore, an optimized user behavior prediction system and method based on historical data is expected. Summary of the invention

[0006] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides a user behavior prediction system and method based on historical data, which uses artificial intelligence technology based on deep learning to deeply mine the user's historical behavior data, capture the time-series evolution characteristics of the user's behavior pattern, and then based on the dynamic interactive correlation between the user behavior pattern and the commodity data, realize the early prediction of the user's purchase behavior. In this way, it can provide enterprises with more accurate marketing strategies and personalized recommendation services, enhance user experience, and promote sales conversion.

[0007] According to one aspect of the present application, a user behavior prediction system based on historical data is provided, which includes:

[0008] User behavior record acquisition module, used to obtain the historical records of user behavior data;

[0009] A user behavior pattern analysis module, configured to perform semantic understanding on each user behavior data in the historical records of the user behavior data and perform time series propagation aggregation of semantic information to obtain a time series aggregation representation vector of the user behavior pattern;

[0010] A commodity data acquisition module, used to acquire commodity data;

[0011] A commodity data semantic coding module, used for semantically coding the commodity data to obtain a commodity data semantic coding vector;

[0012] The user purchase behavior prediction module is used to predict the probability level of the user purchasing the product based on the dynamic interactive fusion information of the temporal aggregation representation vector of the user behavior pattern and the semantic encoding vector of the product data.

[0013] According to another aspect of the present application, a method for predicting user behavior based on historical data is provided, which includes:

[0014] Get historical records of user behavior data;

[0015] Performing semantic understanding on each user behavior data in the historical records of the user behavior data and performing time series propagation aggregation of semantic information to obtain a time series aggregation representation vector of the user behavior pattern;

[0016] Get product data;

[0017] Performing semantic encoding on the commodity data to obtain a semantic encoding vector of the commodity data;

[0018] Based on the dynamic interactive fusion information of the user behavior pattern time series aggregation representation vector and the commodity data semantic encoding vector, the probability level of the user purchasing the commodity is predicted.

[0019] Compared with the prior art, the user behavior prediction system and method based on historical data provided by this application uses deep learning-based artificial intelligence technology to deeply mine the user's historical behavior data, capture the temporal evolution characteristics of the user's behavior pattern, and then based on the dynamic interactive correlation between the user's behavior pattern and commodity data, realize the early prediction of the user's purchase behavior. In this way, it can provide enterprises with more accurate marketing strategies and personalized recommendation services, enhance user experience, and promote sales conversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0021] Figure 1 is a block diagram of a user behavior prediction system based on historical data according to an embodiment of the present application;

[0022] Figure 2 A data flow diagram of a user behavior prediction system based on historical data according to an embodiment of the present application;

[0023] Figure 3 A block diagram of a user behavior pattern analysis module in a user behavior prediction system based on historical data according to an embodiment of the present application;

[0024] Figure 4 A block diagram of a user purchase behavior prediction module in a user behavior prediction system based on historical data according to an embodiment of the present application;

[0025] Figure 5 Flow chart of a method for predicting user behavior based on historical data according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0027] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0028] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0029] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0030] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0031] Traditional user behavior prediction methods are mainly based on simple statistical models, such as collaborative filtering and Markov chains. These methods have great limitations when processing large-scale user behavior data. They are difficult to capture the complexity and temporal dynamics of user behavior, which often leads to inaccurate prediction results. Therefore, an optimized user behavior prediction system and method based on historical data is expected.

[0032] In the technical solution of the present application, a user behavior prediction system based on historical data is proposed. Figure 1 4 is a block diagram of a user behavior prediction system based on historical data according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a user behavior prediction system based on historical data according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to an embodiment of the present application, a user behavior prediction system 300 based on historical data includes: a user behavior record acquisition module 310, used to acquire historical records of user behavior data; a user behavior pattern analysis module 320, used to perform semantic understanding on each user behavior data in the historical records of the user behavior data and to perform time-series propagation aggregation of semantic information to obtain a user behavior pattern time-series aggregation representation vector; a commodity data acquisition module 330, used to acquire commodity data; a commodity data semantic encoding module 340, used to perform semantic encoding on the commodity data to obtain a commodity data semantic encoding vector; a user purchase behavior prediction module 350, used to predict the probability level of a user purchasing a commodity based on the dynamic interactive fusion information of the user behavior pattern time-series aggregation representation vector and the commodity data semantic encoding vector.

[0033] In particular, the user behavior record acquisition module 310 is used to acquire the historical records of user behavior data. It should be understood that the user behavior data covers various interactive behaviors of users on the platform, including but not limited to clicks, browsing, comments, sharing, favorites, adding to shopping carts and final purchases, etc., which is the key to understanding and predicting users' future purchase intentions.

[0034] In particular, the user behavior pattern analysis module 320 is used to perform semantic understanding on each user behavior data in the historical record of the user behavior data and to perform semantic information time series propagation aggregation to obtain a user behavior pattern time series aggregation representation vector. In particular, in a specific example of the present application, Figure 3As shown, the user behavior pattern analysis module 320 includes: a user behavior data embedding coding unit 321, which is used to embed and code each user behavior data in the historical record of the user behavior data to obtain a time series of user behavior data embedded coding vectors; a user behavior data context coding unit 322, which is used to input the time series of the user behavior data embedded coding vectors into a behavior pattern feature extractor based on a converter structure to obtain a time series of context user behavior data semantic coding vectors; a user behavior feature time series aggregation unit 323, which is used to perform time series aggregation on the time series of the context user behavior data semantic coding vectors to obtain the user behavior pattern time series aggregation representation vector.

[0035] Specifically, the user behavior data embedding coding unit 321 is used to embed and code each user behavior data in the historical record of the user behavior data to obtain a time series of user behavior data embedded coding vectors. Since user behavior data has the characteristics of high dimensionality and sparsity, in order to effectively achieve dimensionality reduction and semantic feature representation of the data, the present application adopts text embedding coding technology to convert each user behavior data in the historical record of the user behavior data into a low-dimensional dense continuous vector representation, while retaining the semantic meaning of the data, understanding the user's behavior characteristics at different time points, so as to obtain a time series of user behavior data embedded coding vectors. In an embodiment of the present application, a pre-trained word embedding model, such as Word2Vec or BERT model, can be used to implement the embedding coding process.

[0036] Specifically, the user behavior data context encoding unit 322 is used to input the time series of the user behavior data embedded in the encoding vector into a behavior pattern feature extractor based on a transformer structure to obtain a time series of context user behavior data semantic encoding vectors. Considering that the user's behavior characteristics are usually continuous and correlated, in order to capture the contextual semantic dependencies of the user behavior characteristics, the present application further uses a behavior pattern feature extractor based on a transformer structure to perform contextual association encoding on the time series of the user behavior data embedded in the encoding vector, capture the complex associations between user behaviors at different time points through a self-attention mechanism, and integrate the global context information of the user behavior into the user behavior feature representation at each time step to obtain a time series of context user behavior data semantic encoding vectors.

[0037] Specifically, the user behavior feature time series aggregation unit 323 is used to perform time series aggregation on the time series of the context user behavior data semantic encoding vector to obtain the user behavior pattern time series aggregation representation vector. In order to understand the temporal evolution of the user behavior pattern, it is necessary to further perform time series aggregation operations on the time series of the context user behavior data semantic encoding vector. At the same time, considering that for user behavior prediction, the contribution weights of the behavior features at different time steps in the time series may decay over time, that is, the behavior patterns at more distant time steps may gradually weaken the impact on the current behavior. Therefore, the present application introduces a node feature propagation network based on a node energy decay mechanism, which performs information aggregation on the time series of the context user behavior data semantic coding vector in the time dimension, and utilizes the node energy decay mechanism to dynamically adjust the feature fusion weight according to the feature energy intensity of each context user behavior data semantic coding vector and the feature space span between it and the current (i.e., the end of the time series) context user behavior data semantic coding vector, so as to perform weighted aggregation on the time series of the context user behavior data semantic coding vector, emphasize the influence of recent behavior patterns, and weaken the influence of long-term behavior patterns, thereby effectively propagating and aggregating key information in the user behavior data, and obtaining a weighted aggregation representation of the user behavior pattern evolving over time, i.e., a user behavior pattern time series aggregation representation vector.

[0038] In an embodiment of the present application, the time series of the context user behavior data semantic coding vector is time-series aggregated to obtain the user behavior pattern time-series aggregated representation vector, including: calculating the node energy statistical paradigm value of each context user behavior data semantic coding vector in the time series of the context user behavior data semantic coding vector to obtain a time series of node energy statistical paradigm values, wherein the node energy statistical paradigm value corresponding to the current context user behavior data semantic coding vector in the time series of the node energy statistical paradigm value is used as the current node energy statistical paradigm value, and other node energy statistical paradigm values ​​are used as historical node energy statistical paradigm values ​​to obtain a time series of current node energy statistical paradigm values ​​and historical node energy statistical paradigm values; counting the node propagation space span value between each other context user behavior data semantic coding vector in the time series of the context user behavior data semantic coding vector and the current context user behavior data semantic coding vector to obtain the node propagation space span value. The method comprises the following steps: determining a time series of node propagation space span values ​​and a time series of historical node energy statistical paradigm values ​​based on the time series of node propagation space span values ​​and the time series of historical node energy statistical paradigm values, determining the node energy propagation attenuation coefficient values ​​of other context user behavior data semantic coding vectors in the time series of context user behavior data semantic coding vectors to obtain a time series of node energy propagation attenuation coefficient values, wherein the node energy propagation attenuation coefficient values ​​are inversely correlated with the historical node energy statistical paradigm values; using the time series of node energy propagation attenuation coefficient values ​​as a weight sequence, calculating the weighted sum of all other context user behavior data semantic coding vectors in the time series of context user behavior data semantic coding vectors to obtain a historical node energy attenuation time series aggregation feature vector; and fusing the current node energy statistical paradigm value to calculate the weighted sum of the historical node energy attenuation time series aggregation feature vector and the current context user behavior data semantic coding vector to obtain the user behavior pattern time series aggregation representation vector.

[0039] Among them, the process of calculating the node energy statistical normal value of each context user behavior data semantic coding vector in the time series of the context user behavior data semantic coding vector to obtain the time series of the node energy statistical normal value includes: calculating the weighted sum between the variance and the bias term of the context user behavior data semantic coding vector to obtain the first intermediate term, and calculating the square of the difference between the maximum value and the mean of the context user behavior data semantic coding vector, the variance of the context user behavior data semantic coding vector and the weighted sum between the bias term to obtain the second intermediate term; calculating the first intermediate term divided by the second intermediate term to obtain the node energy statistical normal value.

[0040] More specifically, based on the time series of the node propagation space span values ​​and the time series of the historical node energy statistical paradigm values, the process of determining the node energy propagation attenuation coefficient values ​​of other context user behavior data semantic coding vectors in the time series of the context user behavior data semantic coding vector to obtain the time series of node energy propagation attenuation coefficient values ​​includes: using a preset attenuation constant as a weight, calculating the weighted sum of each node propagation space span value in the time series of the node propagation space span value and its own exponential function value with the natural constant as the base to obtain the time series of spatial span attenuation factors; calculating the corresponding division results of each historical node energy statistical paradigm value in the time series of the historical node energy statistical paradigm values ​​and each spatial span attenuation factor in the time series of the spatial span attenuation factors to obtain the time series of the node energy propagation attenuation coefficient values.

[0041] In summary, in the above embodiment, the time series of the context user behavior data semantic encoding vector is input into the node feature propagation network based on the node energy decay mechanism to obtain the user behavior pattern time series aggregation representation vector, including: processing the time series of the context user behavior data semantic encoding vector with the following time series propagation fusion formula to obtain the user behavior pattern time series aggregation representation vector, wherein the time series propagation fusion formula is:

[0042] X={x1,x2,...x k-1 ,x k}

[0043]

[0044] Wherein, X represents the time series of the semantic encoding vector of the contextual user behavior data, the value of k is the number of feature vectors in the time series of the semantic encoding vector of the contextual user behavior data, and x1, x2, x i 、x k-1 and x k respectively represent the first, second, ith, k-1th and current contextual user behavior data semantic encoding vectors in the time series of the contextual user behavior data semantic encoding vectors, represents the mth eigenvalue of the semantic encoding vector of the i-th context user behavior data, the value of L is the length of the semantic encoding vector of the i-th context user behavior data, μ i is the mean of the eigenvalues ​​of the semantic encoding vector of the i-th context user behavior data, σ 2 i is the variance of each eigenvalue of the semantic encoding vector of the i-th context user behavior data, max(x i) represents the maximum value of each feature value of the semantic encoding vector of the i-th context user behavior data, ∈ is the bias term, is the node energy statistical paradigm value of the semantic encoding vector of the i-th context user behavior data, Count(x k →x i ) represents the node propagation space span value between the i-th context user behavior data semantic encoding vector and the current context user behavior data semantic encoding vector, γ and δ are different preset attenuation constants, a and b are different weight hyperparameters, is the statistical paradigm value of the current node energy, and v1 represents the temporal aggregation representation vector of the user behavior pattern.

[0045] It is worth mentioning that in other specific examples of the present application, semantic understanding of each user behavior data in the historical record of the user behavior data and semantic information time series propagation aggregation can be performed in other ways to obtain a user behavior pattern time series aggregation representation vector, for example: input each user behavior data in the historical record of the user behavior data; for each user behavior data, map it to a semantic concept; use natural language processing technology, such as word embedding, to extract the semantic meaning of the user behavior data; represent the user behavior data sequence as a semantic information time series graph; use a time series propagation model such as a graph neural network or a recursive neural network to propagate semantic information on the semantic information time series graph; through time series propagation, each user behavior data can aggregate semantic information from previous and next behavior data; use pooling operations (such as maximum pooling or average pooling) to aggregate the semantic information in the semantic information time series graph to each time step; to obtain the user behavior pattern time series aggregation representation vector.

[0046] In particular, the commodity data acquisition module 330 is used to acquire commodity data. It should be understood that the commodity data is a description text of commodity attributes, including but not limited to the name, category, brand, price and other multi-dimensional information of the commodity. In the technical solution of the present application, by acquiring commodity data and analyzing the user's purchase intention for the commodity, commodities that better meet the user's potential needs can be recommended to the user.

[0047] In particular, the commodity data semantic encoding module 340 is used to semantically encode the commodity data to obtain a commodity data semantic encoding vector. In order to realize the combined analysis of user behavior pattern characteristics and commodity attribute characteristics, in the technical solution of the present application, first, the commodity data is semantically encoded to fully understand the semantic content of the commodity data, capture the intrinsic attribute meaning of the commodity, and convert the commodity data from a text structure into a low-dimensional dense vector representation compatible with the user behavior pattern temporal aggregation representation vector space, so as to generate a commodity data semantic encoding vector.

[0048] In particular, the user purchase behavior prediction module 350 is used to predict the probability level of the user purchasing the product based on the dynamic interactive fusion information of the user behavior pattern time series aggregation representation vector and the product data semantic encoding vector. In particular, in a specific example of the present application, Figure 4 As shown, the user purchase behavior prediction module 350 includes: an information interaction fusion unit 351, which is used to input the user behavior pattern time series aggregation representation vector and the commodity data semantic encoding vector into a feature vector dynamic interaction fusion module based on gated response to obtain a commodity-user behavior joint representation vector; a prediction result generation unit 352, which is used to input the commodity-user behavior joint representation vector into a classifier-based purchase behavior predictor to obtain a prediction result, and the prediction result is used to represent the probability level label of the user purchasing the commodity.

[0049] Specifically, the information interaction fusion unit 351 is used to input the user behavior pattern temporal aggregation representation vector and the commodity data semantic coding vector into the feature vector dynamic interaction fusion module based on the gated response to obtain the commodity-user behavior joint representation vector. That is, the commodity data semantic coding vector and the user behavior pattern temporal aggregation representation vector are fused and interactively modeled to capture the dynamic interactive association between the user behavior pattern and the commodity attribute. In the technical solution of the present application, a feature vector dynamic interaction fusion module based on the gated response is introduced to realize the dynamic interactive fusion of user behavior pattern features and commodity attribute features. Specifically, the feature vector dynamic interaction fusion module is based on the gating mechanism, and by learning the correlation between the commodity data semantic coding vector and the user behavior pattern temporal aggregation representation vector, the fusion weight of the two is dynamically adjusted to realize the adaptive fusion of the user behavior pattern and commodity attribute information, thereby obtaining the commodity-user behavior joint representation vector to comprehensively reflect the feature association between user behavior and commodity attributes, and provide more accurate and rich information basis for subsequent recommendation decisions.

[0050] In an embodiment of the present application, the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector are input into a feature vector dynamic interaction fusion module based on gated response to obtain a commodity-user behavior joint representation vector, including: performing feature cascading on the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector to obtain a commodity-user behavior cascade feature vector; inputting the commodity-user behavior cascade feature vector into a gated response function to obtain an information fusion response gate; calculating a difference with the information fusion response gate, and using the information fusion response gate and the difference as weights to calculate the position-weighted sum of the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector to obtain the commodity-user behavior joint representation vector.

[0051] Among them, the process of inputting the commodity-user behavior cascade feature vector into the gated response function to obtain the response gate of information fusion includes: using a predetermined weight vector to multiply the commodity-user behavior cascade feature vector to obtain an information interaction fusion correlation coefficient; adding the information interaction fusion correlation coefficient and a predetermined bias parameter and activating it through a sigmoid function to obtain the response gate of the information fusion.

[0052] In summary, in the above embodiment, the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector are input into the feature vector dynamic interactive fusion module based on gated response to obtain the commodity-user behavior joint representation vector, including: fusing the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector with the following interactive fusion formula to obtain the commodity-user behavior joint representation vector, wherein the interactive fusion formula is:

[0053]

[0054]

[0055] Wherein, v1 is the temporal aggregation representation vector of the user behavior pattern, v2 is the semantic encoding vector of the commodity data, σ is the sigmoid function, W is the predetermined weight vector, b is the predetermined bias parameter, is the response gate of information fusion, v a It is the joint representation vector of the product and user behavior.

[0056] Specifically, the prediction result generating unit 352 is used to input the product-user behavior joint representation vector into the classifier-based purchase behavior predictor to obtain a prediction result, and the prediction result is used to represent the probability level label of the user purchasing the product. That is, in the technical solution of the present application, a purchase behavior predictor is constructed using a pre-trained classifier model, and the possibility of the user's purchase behavior for the product is predicted by performing feature analysis and probability distribution calculation on the product-user behavior joint representation vector, and a probability level label of the user's purchase of the product is generated, such as high level, medium level and low level, so as to provide an accurate decision-making basis for personalized recommendation of products and guide the formulation of marketing strategies of enterprises.

[0057] Preferably, the applicant of the present application takes into account that the user behavior pattern temporal aggregation representation vector and the commodity data semantic coding vector respectively express the temporal node propagation association characteristics of the embedded coding temporal context of the user behavior data and the encoded semantic characteristics of the commodity data. In this way, when the user behavior pattern temporal aggregation representation vector and the commodity data semantic coding vector are dynamically interactively fused with the feature vectors based on gated responses, the obtained commodity-user behavior joint representation vector will also lead to outlier fusion class reasoning mapping deviation due to insufficient multi-dimensional coding response interaction coverage of heterogeneous data, thereby affecting the accuracy of the prediction results obtained by the commodity-user behavior joint representation vector through the classifier-based purchase behavior predictor.

[0058] Therefore, in a preferred example, the prediction results obtained by inputting the commodity-user behavior joint representation vector into a classifier-based purchase behavior predictor include:

[0059] Calculating a feature mean of the product-user behavior joint representation vector, and dividing the feature mean by a difference between a maximum feature value and a minimum feature value of the product-user behavior joint representation vector to obtain a product-user behavior joint representation distribution representation value;

[0060] Subtract the commodity-user behavior joint representation distribution representation value from one and divide the result by the commodity-user behavior joint representation distribution representation value to obtain a commodity-user behavior joint representation distribution modulation value;

[0061] Activating the product-user behavior joint representation vector through a probabilistic function to obtain a probabilistic product-user behavior joint representation vector;

[0062] Subtract the probabilistic product-user behavior joint representation vector from the product-user behavior joint representation distribution modulation value, take the absolute value and calculate the negative of the logarithmic value with base 2 to obtain the probabilistic product-user behavior joint representation distribution modulation information vector;

[0063] After dividing the distribution representation value of the joint representation of the product and user behavior by one minus the difference of each eigenvalue of the probabilistic joint representation vector of the product and user behavior, summing all eigenvalues ​​of the probabilistic joint representation vector of the product and user behavior, and dividing by the length of the joint representation vector of the product and user behavior to obtain a modulation bias value of the distribution of the probabilistic joint representation of the product and user behavior;

[0064] The optimized product-user behavior joint representation vector is obtained by multiplying the probabilistic product-user behavior joint representation distribution modulation information vector and the probabilistic product-user behavior joint representation distribution modulation bias value by a weight as a hyperparameter; and

[0065] The optimized commodity-user behavior joint representation vector is input into a classifier-based purchase behavior predictor to obtain a prediction result.

[0066] The optimized product-user behavior joint representation vector is expressed as:

[0067]

[0068] And among them:

[0069]

[0070] in, represents the feature mean of the product-user behavior joint representation vector, v max and v min denote the maximum eigenvalue and the minimum eigenvalue in the product-user behavior joint representation vector, respectively; p denotes the distribution representation value of the product-user behavior joint representation; V denotes the probabilistic product-user behavior joint representation vector obtained by activating the product-user behavior joint representation vector through a probabilistic function; V i represents the i-th eigenvalue of the probabilistic product-user behavior joint representation vector, log represents a logarithmic function with base 2, ε is a weight as a hyperparameter, the value of L is the length of the probabilistic product-user behavior joint representation vector, and V' represents the optimized product-user behavior joint representation vector.

[0071] That is, the probability information distribution planning based on the eigenvalue of the joint representation vector of commodity-user behavior is performed through the Bernoulli probability modulation distribution of the joint representation vector of commodity-user behavior relative to the eigenvalue distribution, and the probability reverse mapping of the probability characteristics of the joint representation vector of commodity-user behavior is used as the extended coverage of the set mapping space of the joint representation vector of commodity-user behavior, so as to independently understand the interaction path between the intuitive probability information distribution and the abstract probability space mapping of the joint representation vector of commodity-user behavior, so as to improve the accuracy of the prediction result obtained by inputting the optimized joint representation vector of commodity-user behavior into the purchase behavior predictor based on the classifier by avoiding the counterfactual reasoning mapping of the outlier feature distribution of the joint representation vector of commodity-user behavior.

[0072] It is worth mentioning that in other specific examples of the present application, the probability level of a user purchasing a product can be predicted based on the dynamic interaction fusion information of the user behavior pattern time series aggregation representation vector and the product data semantic encoding vector in other ways, for example: inputting the user behavior pattern time series aggregation representation vector and the product data semantic encoding vector into a dynamic interaction model; the dynamic interaction model can be a multi-layer perceptron, an attention mechanism or a graph neural network; learning the interaction information of the user behavior pattern and the product data through the dynamic interaction model, and outputting a fused representation vector; inputting the fused representation vector into a classifier; the classifier outputs the probability level of the user purchasing the product.

[0073] As described above, the user behavior prediction system 300 based on historical data according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a user behavior prediction algorithm based on historical data. In a possible implementation, the user behavior prediction system 300 based on historical data according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the user behavior prediction system 300 based on historical data can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the user behavior prediction system 300 based on historical data can also be one of the many hardware modules of the wireless terminal.

[0074] Alternatively, in another example, the user behavior prediction system 300 based on historical data and the wireless terminal may also be separate devices, and the user behavior prediction system 300 based on historical data may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0075] Furthermore, a user behavior prediction method based on historical data is also provided.

[0076] Figure 5 Flow chart of a method for predicting user behavior based on historical data according to an embodiment of the present application. Figure 5 As shown, according to the user behavior prediction method based on historical data of the embodiment of the present application, the steps include: S1, obtaining the historical records of user behavior data; S2, performing semantic understanding and semantic information time series propagation aggregation on each user behavior data in the historical records of the user behavior data to obtain a user behavior pattern time series aggregation representation vector; S3, obtaining product data; S4, performing semantic encoding on the product data to obtain a product data semantic encoding vector; S5, predicting the probability level of the user purchasing the product based on the dynamic interactive fusion information of the user behavior pattern time series aggregation representation vector and the product data semantic encoding vector.

[0077] In summary, the user behavior prediction method based on historical data according to the embodiment of the present application is explained, which uses deep learning-based artificial intelligence technology to deeply mine the user's historical behavior data, captures the time-series evolution characteristics of the user's behavior pattern, and then realizes the early prediction of the user's purchase behavior based on the dynamic interactive correlation between the user's behavior pattern and the commodity data. In this way, it can provide enterprises with more accurate marketing strategies and personalized recommendation services, enhance user experience, and promote sales conversion.

[0078] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A user behavior prediction system based on historical data, characterized in that: include: User behavior record acquisition module, used to obtain the historical records of user behavior data; A user behavior pattern analysis module, configured to perform semantic understanding on each user behavior data in the historical records of the user behavior data and perform time series propagation aggregation of semantic information to obtain a time series aggregation representation vector of the user behavior pattern; A commodity data acquisition module, used to acquire commodity data; A commodity data semantic coding module, used for semantically coding the commodity data to obtain a commodity data semantic coding vector; A user purchase behavior prediction module, used to predict the probability level of a user purchasing a product based on the dynamic interactive fusion information of the user behavior pattern time series aggregation representation vector and the product data semantic encoding vector; Wherein, the user behavior pattern analysis module includes: A user behavior data embedding coding unit, used for embedding and coding each user behavior data in the historical records of the user behavior data to obtain a time series of user behavior data embedding coding vectors; A user behavior data context encoding unit, used to input the time series of the user behavior data embedded in the encoding vector into a behavior pattern feature extractor based on a transformer structure to obtain a time series of context user behavior data semantic encoding vectors; A user behavior feature time series aggregation unit, used for performing time series aggregation on the time series of the context user behavior data semantic encoding vector to obtain the user behavior pattern time series aggregation representation vector; Wherein, the user behavior feature time series aggregation unit includes: A node energy calculation subunit, used for calculating the node energy statistical paradigm value of each context user behavior data semantic coding vector in the time series of the context user behavior data semantic coding vector to obtain a time series of node energy statistical paradigm values, wherein the node energy statistical paradigm value corresponding to the current context user behavior data semantic coding vector in the time series of the node energy statistical paradigm value is used as the current node energy statistical paradigm value, and the other node energy statistical paradigm values ​​are used as historical node energy statistical paradigm values ​​to obtain a time series of the current node energy statistical paradigm value and the historical node energy statistical paradigm values; A node propagation space span statistics subunit, used to count the node propagation space span values ​​between each other context user behavior data semantic coding vector in the time series of the context user behavior data semantic coding vector and the current context user behavior data semantic coding vector to obtain a time series of node propagation space span values; An energy propagation attenuation coefficient calculation subunit is used to determine the node energy propagation attenuation coefficient values ​​of other context user behavior data semantic encoding vectors in the time series of the context user behavior data semantic encoding vector based on the time series of the node propagation space span value and the time series of the historical node energy statistical paradigm value to obtain the time series of the node energy propagation attenuation coefficient values, wherein the node energy propagation attenuation coefficient value is inversely correlated with the historical node energy statistical paradigm value; A historical node information aggregation subunit, used to calculate the weighted sum of all other context user behavior data semantic encoding vectors in the time series of the context user behavior data semantic encoding vector using the time series of the node energy propagation attenuation coefficient value as a weight sequence to obtain a historical node energy attenuation time series aggregation feature vector; The global feature propagation aggregation subunit is used to fuse the current node energy statistical paradigm value to calculate the weighted sum of the historical node energy decay time series aggregation feature vector and the current context user behavior data semantic encoding vector to obtain the user behavior pattern time series aggregation representation vector.

2. The user behavior prediction system based on historical data according to claim 1, characterized in that: The user behavior feature time series aggregation unit is used to: The time series of the semantic encoding vector of the contextual user behavior data is input into a node feature propagation network based on a node energy decay mechanism to obtain a temporal aggregation representation vector of the user behavior pattern.

3. The user behavior prediction system based on historical data according to claim 2, characterized in that: The node energy calculation subunit is used to: Calculate the weighted sum of the variance and the bias term of the contextual user behavior data semantic encoding vector to obtain a first intermediate term, and calculate the square of the difference between the maximum value and the mean of the contextual user behavior data semantic encoding vector, the variance of the contextual user behavior data semantic encoding vector, and the bias term to obtain a second intermediate term; The first intermediate term is divided by the second intermediate term to obtain the node energy statistical normal form value.

4. The user behavior prediction system based on historical data according to claim 3, characterized in that: The energy propagation attenuation coefficient calculation subunit is used to: Using a preset attenuation constant as a weight, calculating the weighted sum of each node propagation space span value in the time series of the node propagation space span value and its own exponential function value with the natural constant as the base to obtain a time series of the space span attenuation factor; The corresponding division results of each historical node energy statistical paradigm value in the time series of the historical node energy statistical paradigm value and each spatial span attenuation factor in the time series of the spatial span attenuation factor are calculated to obtain the time series of the node energy propagation attenuation coefficient value.

5. The user behavior prediction system based on historical data according to claim 4, characterized in that: The user purchase behavior prediction module includes: An information interaction fusion unit, used for inputting the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector into a feature vector dynamic interaction fusion module based on gated response to obtain a commodity-user behavior joint representation vector; The prediction result generating unit is used to input the commodity-user behavior joint representation vector into a classifier-based purchase behavior predictor to obtain a prediction result, wherein the prediction result is used to represent a probability level label of the user purchasing the commodity.

6. The user behavior prediction system based on historical data according to claim 5, characterized in that: The information interaction and fusion unit includes: A feature joint representation subunit, used for performing feature cascade on the user behavior pattern temporal aggregation representation vector and the commodity data semantic encoding vector to obtain a commodity-user behavior cascade feature vector; A gated response calculation subunit, used for inputting the commodity-user behavior cascade feature vector into a gated response function to obtain a response gate for information fusion; The feature fusion subunit is used to calculate a difference with the response gate of the information fusion, and use the response gate of the information fusion and the difference as weights to calculate the position-weighted sum of the temporal aggregation representation vector of the user behavior pattern and the semantic encoding vector of the commodity data to obtain the commodity-user behavior joint representation vector.

7. The user behavior prediction system based on historical data according to claim 6, characterized in that: The gated response calculation subunit is used to: Multiplying the product-user behavior cascade feature vector by a predetermined weight vector to obtain an information interaction fusion correlation coefficient; After the information interaction fusion correlation coefficient and the predetermined bias parameter are added, activation processing is performed through a sigmoid function to obtain a response gate for the information fusion.

8. A method for predicting user behavior based on historical data, using the user behavior prediction system based on historical data according to claim 1, characterized in that: include: Get historical records of user behavior data; Performing semantic understanding on each user behavior data in the historical records of the user behavior data and performing time series propagation aggregation of semantic information to obtain a time series aggregation representation vector of the user behavior pattern; Get product data; Performing semantic encoding on the commodity data to obtain a semantic encoding vector of the commodity data; Based on the dynamic interactive fusion information of the user behavior pattern time series aggregation representation vector and the commodity data semantic encoding vector, the probability level of the user purchasing the commodity is predicted.

Images