Information pushing method and device, electronic equipment, computer readable storage medium and computer program product

Abstract

The invention provides an information pushing method and device, electronic equipment, a computer readable storage medium and a computer program product. The method comprises the steps of determining second behavior information of a first object at first time based on attribute information and first behavior information of the first object; the occurrence time corresponding to the first behavior information is earlier than the first time; determining state information of the first object at the first time based on the attribute information of the first object; and based on the behavior information and the state information, carrying out information pushing on the first object. According to the invention, the accuracy of information pushing can be improved.

Description

Technical Field

[0001] The present application relates to data processing technology, and in particular to an information push method, device, electronic device, computer-readable storage medium, and computer program product. Background Art

[0002] When it comes to information push, user interest in the information varies from person to person, so it's important to factor in user preferences when pushing information. Related technologies typically predict user behavior at specific points in time based on user characteristics, and then push targeted information based on the predicted results. However, behavioral information plays a limited role in information push, which impacts its accuracy. Summary of the Invention

[0003] The embodiments of the present application provide an information push method, device, electronic device, computer-readable storage medium, and computer program product, which can improve the efficiency of information push.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] This embodiment of the present application provides an information push method, the method comprising:

[0006] Determining, based on attribute information and first behavior information of a first object, second behavior information of the first object at a first time; the occurrence time corresponding to the first behavior information is earlier than the first time;

[0007] determining, based on the attribute information of the first object, state information of the first object at the first time;

[0008] Based on the second behavior information and the state information, information is pushed to the first object.

[0009] The present invention provides an information push device, including:

[0010] a determination module, configured to determine, based on attribute information and first behavior information of a first object, second behavior information of the first object at a first time, wherein the occurrence time corresponding to the first behavior information is earlier than the first time; and determine, based on the attribute information of the first object, state information of the first object at the first time;

[0011] A push module is used to push information to the first object based on the behavior information and the state information.

[0012] An embodiment of the present application provides an electronic device, comprising:

[0013] a memory for storing computer-executable instructions or computer programs;

[0014] The processor is used to implement the information push method provided in the embodiment of the present application when executing the computer executable instructions or computer program stored in the memory.

[0015] An embodiment of the present application provides a computer-readable storage medium storing a computer program or computer-executable instructions for implementing the information push method provided in the embodiment of the present application when executed by a processor.

[0016] An embodiment of the present application provides a computer program product, including a computer program or computer-executable instructions. When the computer program or computer-executable instructions are executed by a processor, the information push method provided in the embodiment of the present application is implemented.

[0017] The embodiments of the present application have the following beneficial effects: first, by combining the first behavior information that can reflect the first object, that is, the behavior taken by the first object in response to the pushed information in the past period of time, and the attribute information, the behavior taken by the first object in response to the information in the first time is predicted; second, because in the information push scenario, the first object's tendency to receive information is closely related to the state of the first object, the first object is usually more likely to accept information push when it is idle. Therefore, in the embodiments of the present application, the state of the first object in the first time is also predicted based on the attribute information of the first object. Finally, by combining the two types of information, the behavior taken by the first object in response to the information and the state of the first object, information is pushed to the first object. In this way, it is possible to simultaneously combine the two different factors of the first object's behavior and state to push information, so that the information push is more in line with the personalized preferences of the first object, and ultimately the accuracy of the information push is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1 is a schematic diagram of the architecture of the information push system 100 provided in an embodiment of the present application;

[0019] Figure 2A This is a flow chart of the information push method provided in the embodiment of the present application;

[0020] Figure 2B is a schematic diagram of a process for determining second behavior information provided in an embodiment of the present application;

[0021] Figure 2C This is a schematic diagram of a process for determining status information provided by an embodiment of the present application;

[0022] Figure 2D This is another flowchart of the information push method provided by an embodiment of the present application;

[0023] Figure 2E This is a flow chart of the model training method provided in the embodiment of the present application;

[0024] Figure 3 This is another flowchart of the model training method provided in the embodiment of the present application;

[0025] Figure 4 It is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0027] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0028] In the following description, the terms "first" and "second" are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understandable that "first" and "second" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0029] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0030] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meanings as those commonly understood by those skilled in the art. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0031] The relevant data collection and processing in the embodiments of this application should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0032] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0033] 1) Information push: A general term for technologies that proactively send information from a server to a user in an internet environment. This mechanism allows service providers to directly send content or notifications to a user's device, such as a smartphone, tablet, or other networked device. In the embodiments of this application, the server can implement information push via SMS, phone calls, and other methods.

[0034] 2) Personalized Preferences: Providing users with customized products, services, or experiences based on their specific needs, interests, behavioral habits, and historical interactions. In this embodiment, personalized preferences are used to push information to each user based on their behavior, needs, preferences, purchase history, and other characteristics. For example, the user's availability is determined based on their behavioral history and other characteristics to achieve personalized information push.

[0035] 3) Dual-Tower Model: A machine learning architecture widely used in fields such as recommender systems and natural language processing. It consists of two independent neural network "towers," each responsible for learning different feature representations of the dataset. The two towers are typically trained separately, but their outputs can be used for collaborative tasks such as rating prediction, recommendations, or text similarity comparison.

[0036] 4) Embedding layer: A network layer in a deep learning network, mainly used to perform embedding operations, converting discrete data (such as words, item IDs, category labels, etc.) into continuous, fixed-size vector representations, which can effectively solve the problem of feature sparsity.

[0037] 5) Multilayer Perceptron (MLP): A basic form of neural network used for supervised learning tasks, including classification and regression.

[0038] In marketing scenarios, users' interests in marketing information vary widely. Related technologies typically predict marketing timing based on a single user's historical marketing information consumption, and then use the predictions to recommend the next marketing opportunity to the user. However, when user feature data is limited, relying solely on a single user's historical marketing information often fails to accurately capture the user's individual preferences, hindering accurate marketing timing predictions and impacting the efficiency of marketing information delivery.

[0039] The embodiments of the present application provide an information push method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can improve the accuracy of information push. The following describes exemplary applications of the electronic device provided by the embodiments of the present application. The electronic device provided by the embodiments of the present application can be implemented as various types of terminals, such as laptops, tablet computers, desktop computers, set-top boxes, smartphones, smart speakers, smart watches, smart TVs, and in-vehicle terminals, and can also be implemented as a server. The following describes exemplary applications when the device is implemented as a terminal.

[0040] See also Figure 1 , Figure 1 This is a schematic diagram of the architecture of the information push system 100 provided in an embodiment of the present application. The terminal 400 is connected to the server 200 via the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0041] Among them, the terminal 400 responds to the staff's input operation on the image interface 401, determines the first object and the first time, and determines the second behavior information of the first object at the first time based on the attribute information and the first behavior information of the first object; determines the status information of the first object at the first time based on the attribute information of the first object; determines the information interest of the first object at the first time based on the second behavior information and the status information, and pushes information to the first object at an appropriate time according to the information interest.

[0042] Alternatively, a marketing personnel enters the first object and first time for which information push is desired through the graphical interface 401 of the terminal 400. The terminal 400 then receives the first object and first time and sends them to the server 200. Based on the first object's attribute information and the first behavior information, the server 200 determines the first object's second behavior information at the first time; based on the first object's attribute information, it determines the first object's status information at the first time; and based on the second behavior information and the status information, it determines the first object's information interest level at the first time and sends this information to the terminal 400. After receiving the information interest level, the terminal 400 pushes information to the first object at an appropriate time based on the information interest level.

[0043] Below, the information push method provided by the embodiment of the present application will be described in conjunction with the exemplary application and implementation of the terminal provided in the embodiment of the present application. As mentioned above, the electronic device that implements the information push method of the embodiment of the present application can be a terminal, a server, or a combination of the two. Therefore, the execution entity of each step will not be repeated below.

[0044] It should be noted that the information push examples below are illustrated using the information push scenario in the Internet environment as an example. Based on their understanding of the following, those skilled in the art can apply the information push method provided in the embodiments of the present application to information processing in other scenarios.

[0045] See also Figure 2A , Figure 2A This is a flow chart of the information push method provided by the embodiment of the present application, which will be combined with Figure 2A The steps shown are explained.

[0046] In step 101, second behavior information of the first object at a first time is determined based on attribute information and first behavior information of the first object.

[0047] Here, the first object refers to a user who needs to push information, and the first time is the time when the information is planned to be pushed to the first object, such as 8 o'clock, 19 o'clock, 7 o'clock in the morning, 4 o'clock in the afternoon, etc. Attribute information is used to describe the basic characteristics of the first object, including user-related information such as the age, gender, and occupation of the first object. The occurrence time corresponding to the first behavior information is earlier than the first time. The first behavior information is used to describe the user's behavior information related to the information to be pushed in the past period of time, such as answering the phone for pushing information, collecting the product corresponding to the push information, logging in to the software related to the push information, etc. The second behavior information can be a prediction result of the first object's reception of information at the first time, including whether the first object will receive the pushed information in the first time, or the probability of the first object receiving the pushed information in the first time.

[0048] For example, in a telemarketing scenario, the push information could be promotional calls. The first behavior information could include historical telemarketing information and historical login information. Statistics would include the first subject's connection rate, number of connections, number of calls, and call duration for marketing calls during various time periods over a month. This would serve as historical telemarketing information. Statistics would also include the number of logins to marketing-related software and the duration of logins during various time periods over a month. These historical telemarketing and login information would be combined to form the first behavior information. The second behavior information could include: a probability of receiving the message, a probability of not receiving the message, or a 40% probability of receiving the message.

[0049] Taking the product recommendation scenario as an example, push information refers to recommended products. The first behavior information can include historical purchase information and historical login information. The first subject's purchase rate and number of purchases during various time periods within a month are counted as historical purchase information. The number of times the first subject logged into software related to product recommendation services during various time periods within a month, as well as the duration of logins, are counted as historical login information. The historical purchase and login information are combined to form the first behavior information. The second behavior information can be any of the following: will purchase the product, will not purchase the product, or have a 55% probability of purchasing the product.

[0050] In some embodiments, see Figure 2B , the above step 101 can be implemented by steps 1011 and 1012, that is, based on the attribute information and the first behavior information of the first object, the second behavior information of the first object at the first time is determined. Figure 2B To explain:

[0051] In step 1011 , based on the attribute information of the first object, the first behavior information, and the time information of the first time, the behavior pattern characteristics of the first object at the first time are determined.

[0052] Here, the first time may be expressed in multiple ways, for example, 7:30 pm and 19:30 express the same time, and the time information is used to express the first time in a specified format specification, for example, using a 24-hour system, the first time 7:30 pm is expressed as 19:30. The behavioral pattern characteristics are used to describe the first object's preference for receiving information. The behavioral pattern characteristics can be used to predict whether the first object will receive the information at the first time. The behavioral pattern characteristics can be determined based on the combination of attribute information, first behavior information and time information, or the behavioral pattern characteristics can be determined based on at least one feature of the attribute information, first behavior information and time information. Feature extraction can be performed on the attribute information, first behavior information and time information first, and then cluster analysis can be performed on the extracted features to obtain the behavioral pattern characteristics.

[0053] In some embodiments, a first feature sequence and a second feature sequence are extracted from the attribute information, the first behavior information and the time information; and based on the first feature sequence and the second feature sequence, the behavior pattern characteristics of the first object at the first time are determined.

[0054] Here, first, the attribute information, the first behavior information, and the time information are subjected to a first feature extraction, respectively, to obtain attribute features, behavior features, and time features. Then, the attribute features, behavior features, and time features are used to generate a first feature sequence. After that, the attribute features, behavior features, and time features are subjected to a second feature extraction, to obtain feature data after the second extraction, and are combined together as a second feature sequence. Finally, the first feature sequence and the second feature sequence are fused together to obtain the behavior pattern features of the first object at the first time. Among them, each feature in the first feature sequence and the second feature sequence can be added bit by bit to obtain the behavior pattern features; or, the features in the first feature sequence and the second feature sequence can be directly spliced ​​at the first position to obtain the behavior pattern features.

[0055] In some embodiments, the attribute information, the first behavior information, and the time information can be first mapped to the feature space through an embedding layer to obtain attribute features, behavior features, and time features, and then the attribute features, behavior features, and time features are formed into a feature sequence in a random order, or in the order in which the features are generated, thereby obtaining a first feature sequence; then the attribute features, behavior features, and time features are calculated through an MLP layer to obtain new features corresponding to each feature, and these calculated new features are combined in the form of a sequence to generate a second feature sequence; then the first feature sequence and the second feature sequence are fused through a concatenation layer, and the output result is used as a behavior pattern feature. Among them, the concatenation layer belongs to the operation layer, which is used to implement multi-dimensional vector concatenation operations, merging data from different sources or features of different dimensions into a single output.

[0056] In the embodiment of the present application, a first feature sequence and a second feature sequence are first extracted from the attribute information, the first behavior information, and the time information. Then, based on the first feature sequence and the second feature sequence, the behavioral pattern characteristics of the first subject at the first time are determined. In this way, by combining the features of the attribute information, the first behavior information, and the time information from different sources into a single behavioral pattern characteristic, the relationships and characteristics between the various data can be more accurately captured, thereby improving the accuracy of the analysis of the first behavior information of the first subject.

[0057] In step 1012, second behavior information of the first object at the first time is determined based on the behavior pattern feature.

[0058] Here, the second behavior information can be expressed as "will receive information" or "will not receive information"; or it can be expressed as the probability of receiving information, for example, the probability of receiving information is 17%, the probability of receiving information is 80%, etc. Among them, receiving information corresponds to a probability of receiving information of 100%, and not receiving information corresponds to a probability of receiving information of 0%.

[0059] In some embodiments, the correlation between the first object and the first behavior may be determined based on the behavior pattern characteristics, and then the behavior information of the first object at the first time may be determined based on the correlation.

[0060] Here, the first behavior refers to the behavior of receiving information. The higher the correlation between the first object and the first behavior, the more likely it is that the first object will receive the information at the first time; the lower the correlation between the first object and the first behavior, the more likely it is that the first object will not receive the information at the first time.

[0061] The relevance threshold can be pre-set based on marketing needs. If the efficiency of information recommendation is high, the relevance threshold can be increased; if the efficiency of information recommendation is low, the relevance threshold can be appropriately lowered. When the relevance reaches the relevance threshold, the second behavior information is determined to be information reception; when the relevance does not reach the relevance threshold, the second behavior information is determined to be information non-reception. Alternatively, the relevance is used as a probability value to determine the second behavior information. For example, if the relevance is 30%, the probability of determining the second behavior information as information reception is 30%.

[0062] In this way, the correlation between the first subject and the first behavior is determined based on the behavioral pattern characteristics, and the behavioral information of the first subject at the first time is determined based on the correlation. In this way, by analyzing the correlation between the behavioral pattern characteristics of the first subject and its first behavior (such as answering a call, checking a short message, etc.), the first subject's future behavior can be more accurately predicted, thereby improving the accuracy of the prediction of the first subject's information reception.

[0063] In the embodiment of the present application, based on the attribute information of the first subject, the first behavior information, and the time information of the first time, the behavioral pattern characteristics of the first subject at the first time are determined, and then based on the behavioral pattern characteristics, the second behavior information of the first subject at the first time is determined. In this way, the behavioral habits of the first subject are more comprehensively analyzed through multiple aspects of user-related information, thereby improving the accuracy of the prediction of the first subject's behavior.

[0064] Continue to see Figure 2A In step 102, based on the attribute information of the first object, the state information of the first object at the first time is determined.

[0065] Here, the state information is a prediction result of the idle state of the first object at the first time. The idle state can be expressed as "idle" or "not idle"; or can be expressed as the probability of idleness, for example, the probability of idleness is 80%, the probability of idleness is 30%, etc. The probability of idleness corresponding to idleness is 100%, and the probability of not idleness corresponding to idleness is 0%.

[0066] In some embodiments, see Figure 2C , the above step 102 can be implemented through steps 1021 to 1023, that is, based on the attribute information of the first object, the state information of the first object at the first time is determined. Figure 2C To explain:

[0067] In step 1021 , feature extraction is performed on the attribute information to obtain attribute features of the first object.

[0068] Here, the original attribute information is mapped into the feature space through an embedding operation to obtain attribute features. The attribute features can capture the similarities and relationships between the attributes. Here, the attribute features can be features in the form of vectors or matrices, which is not limited in this embodiment of the present application.

[0069] In step 1022, feature extraction is performed on the time information of the first time to obtain a time feature.

[0070] Here, the original time information is converted into a continuous vector representation through embedding operation to obtain time features, which can capture the periodicity, seasonality and relationship between time points.

[0071] In step 1023 , based on the attribute feature and the time feature, the state information of the first object at the first time is determined.

[0072] In some embodiments, the state information of the first object at the first time is determined by the feature similarity between the attribute feature and the time feature.

[0073] Here, the attribute features and time features can be processed through the MLP layer to obtain attribute representations and time representations. Similarity is then calculated between the attribute representations and time representations to obtain feature similarity. A higher feature similarity indicates a higher probability that the first object was idle during the first time period; a lower feature similarity indicates a higher probability that the first object was not idle during the first time period.

[0074] A similarity threshold can be pre-set based on marketing needs. If the efficiency of information push is high, the similarity threshold can be increased; if the efficiency is low, the similarity threshold can be appropriately lowered. When the feature similarity reaches the similarity threshold, the status information is determined to be idle; when the correlation does not reach the similarity threshold, the second behavior information is determined to be not idle. Alternatively, feature similarity can be used as a probability value to determine the status information. For example, if the correlation is 30%, the probability of determining the status information as idle is 30%.

[0075] In this way, by determining the state information of the first object at the first time through the feature similarity between the attribute feature and the time feature, the behavior and needs of the first object can be better understood, thereby making the predicted state information more accurate.

[0076] In the embodiment of the present application, feature extraction is performed on the attribute information to obtain the attribute features of the first object, and feature extraction is performed on the time information of the first time to obtain the time features. Based on the attribute features and the time features, the state information of the first object at the first time is determined. In this way, the attribute information and time information can better understand and predict the state of the first object, thereby improving the accuracy of the prediction of the user's idle state.

[0077] Continue to see Figure 2C In step 103, information is pushed to the first object based on the second behavior information and the state information.

[0078] Here, the second behavior information and the state information are used to simultaneously determine whether to push information to the first object at the first time. For example, when the second behavior information indicates that the device will receive information and the state information indicates that the device is idle, information is pushed to the first object at the first time; or when the second behavior information indicates that the probability of receiving information is greater than a first threshold (e.g., 60%) and the state information indicates that the device is idle at the first time, information is pushed to the first object at the first time.

[0079] In some embodiments, see Figure 2D , the above step 103 can be implemented through steps 1031 to 1032, that is, information is pushed to the first object based on the second behavior information and state information. Figure 2D To explain:

[0080] In step 1031 , based on the second behavior information and the state information, the information interest level of the first object at the first time is determined.

[0081] Here, information interest indicates whether the first subject is interested in the information and can be understood as the final prediction of the probability that the first subject will receive the information in the first place. When determining information interest, the importance of the second behavior information and the status information in determining information interest can be determined based on marketing needs. For example, if the weight parameter of the second behavior information is preset to 2 and the weight parameter of the status information is preset to 1, then when the probability of the second behavior information receiving the information is 30% and the probability of the status information being idle is 90%, the corresponding value of the information interest is (2*30%+1*90%) / 3=60%. During calculation, if the second behavior information indicates that the information will be received, it can be represented as a probability of receiving the information of 100%; if the second behavior information indicates that the information will not be received, it can be represented as a probability of receiving the information of 0%; if the idle information indicates that the idle information is idle, it can be represented as a probability of idle of 100%; if the idle information indicates that the idle information is not idle, it can be represented as a probability of idle of 0%, thereby facilitating calculation and determining information interest.

[0082] In step 1032, when the information interest level satisfies the information push condition, the information is pushed to the first object at the first time.

[0083] Here, the information push condition can be pre-set according to marketing needs. For example, if the information push condition is that the information interest reaches 80%, then when the value corresponding to the information interest reaches 80%, the information will be pushed to the first target at the first time.

[0084] In some embodiments, when the information reception tendency does not meet the information push condition, a second time can be determined for the first object, and information can be pushed to the first object at the second time.

[0085] Here, the second time is different from the first time, and the information interest level of the first object at the second time meets the information push condition, or the information interest level of the first object at the second time does not meet the information push condition but is greater than the information interest level at the first time. For example, the first time is 9:00 AM, and the corresponding information interest level is 50%, but the information push condition requires that the information interest level reach 70%. Therefore, information push cannot be performed on the first object at 9:00 AM. In this case, the information interest level of the first object at 10:00 AM is determined. If the information interest level of the first object at 10:00 AM reaches 70%, then 10:00 AM is determined as the second time. If the information interest level of the first object at 10:00 AM still does not reach 70%, then the information interest level of the first object at 1:00 PM can be determined, and so on, until the information interest level of the first object at a certain time reaches 70%, and this time is determined as the second time. Alternatively, if the information interest level of the first object does not reach 70% at any time, but the information interest level of the first object at 11:00 AM is the highest, then 11:00 AM can be determined as the second time.

[0086] Thus, in this embodiment of the present application, based on the behavior information and state information, the information interest of the first subject at the first time is determined. If the information reception tendency meets the information push conditions, the information is pushed to the first subject at the first time. If the information reception tendency does not meet the information push conditions, a second time is determined for the first subject, and the information is pushed to the first subject at the second time. In this way, information is pushed to the user at the time when the user is most likely to receive information, allowing the user to smoothly receive the pushed information, thereby improving the accuracy of information push.

[0087] In the embodiment of the present application, first, the first behavior information that can reflect the first object, that is, the behavior taken by the first object in response to the pushed information in the past period of time, and the attribute information are combined to predict the behavior taken by the first object in response to the information in the first time; secondly, because in the information push scenario, the first object's tendency to receive information is closely related to the state of the first object, the first object is usually more likely to accept information push when it is idle. Therefore, in the embodiment of the present application, the state of the first object in the first time is also predicted based on the attribute information of the first object. Finally, the two types of information, the behavior taken by the first object in response to the information and the state of the first object, are combined to jointly push information to the first object. In this way, it is possible to simultaneously push information by combining two different factors, the behavior and state of the first object, so that the information push is more in line with the personalized preferences of the first object, and ultimately the accuracy of the information push is improved.

[0088] In some embodiments, the electronic device can determine the second behavior information and state information using a pre-trained first behavior model and a first state model. The first behavior model is used to predict a user's reception of information, and the first state model is used to predict a user's idle state. The first behavior model is used to determine the second behavior information of the first object at a first time based on the first object's attribute information and the first behavior information; and the first state model is used to determine the first object's state information at a first time based on the first object's attribute information.

[0089] In this way, determining the second behavior information and state information of the first object at the first time through the model can improve data processing efficiency and improve the accuracy of the predicted second behavior information and state information.

[0090] In some embodiments, see Figure 2E , the model can be trained in advance through steps 201 to 204 to obtain the first behavior model and the first state model. Figure 2E To explain:

[0091] In step 201, the training second behavior information of the second object at the training time is determined based on the training attribute information and the training first behavior information of the second object through the second behavior model.

[0092] Here, the second behavior model is a model that has not been trained (for example, a deep learning model obtained after parameter initialization), or a model that has not been trained (for example, a deep learning model obtained by pre-training using unlabeled samples). The second behavior model that has completed training is the first behavior model. The second object is the sample object during training, and is a user who has received information push in the past. The training time includes multiple time points within a day, such as 8 o'clock, 19 o'clock, etc. The training attribute information is used to describe the basic characteristics of the second object, including user-related information such as the age, gender, and occupation of the second object. The occurrence time corresponding to the training first behavior information (that is, the first behavior information during training) is earlier than the training time, and is used to describe the behavior information related to the information to be pushed that occurred in the past period of time for the second object. The training second behavior information (that is, the second behavior information during training) is the predicted result of the first object's reception of information during the training time, including the probability of the second object receiving the pushed information during the training time. The training attribute information and the training first behavior information are used as inputs to the second behavior model, and the training second behavior information of the second object during the training time is output after prediction by the second behavior model.

[0093] In step 202, the training state information of the second object at the training time is determined based on the training attribute information through the second state model.

[0094] Here, the training attribute information is used as the input of the second state model, and the second state model predicts and outputs the training state information of the second object during the training time. The training state information is the prediction result of the idleness of the second object, including the probability of the second object being idle during the training time.

[0095] In step 203 , a loss value is calculated based on the difference between the training state information and the state label at the training time, and the difference between the training second behavior information and the behavior label at the training time.

[0096] Here, the state label includes the real state information corresponding to the second object at each training time, and the behavior label includes the real behavior information corresponding to the second object at each training time. The difference between the training state information and the state label at the training time can be used to determine the prediction error of the second state model for the idle state of the second object. The difference between the training second behavior information and the behavior label at the training time can be used to determine the prediction error of the second behavior model for the case where the second object receives information. The electronic device can obtain the weight parameters of the trained second behavior model and the second state model, and perform weighted calculation on the difference between the training state information and the state label and the difference between the training second behavior information and the state label through the weight parameters to obtain a loss value, and then use the loss value to measure the difference between the prediction result of the model and the actual situation.

[0097] In some possible implementations, the second subject's hourly connection rate for marketing calls and online time on the software corresponding to the marketing business are collected for 30 days. Based on all the connection rate and online time information, the average connection rate and average online time of the second subject are determined. Hours with a connection rate greater than the average connection rate are labeled 1, and hours with a connection rate less than or equal to the average connection rate are labeled 0, thereby obtaining a behavior label. Hours with an online time greater than the average online time are labeled 1, and hours with an online time less than or equal to the average online time are labeled 0, thereby obtaining a state label.

[0098] In step 204 , parameters of the second behavior model and the second state model are adjusted based on the loss value, and when the training end condition is met, the first behavior model and the first state model are obtained.

[0099] Here, the second behavior model and the second state model are trained by directional propagation using the loss value, thereby adjusting the model parameters of the second behavior model and the second state model, and obtaining the first behavior model and the first state model when the training end condition is met. The training end condition refers to the difference between the training state information output by the second behavior model and the state label at the training time being less than a preset difference, and the difference between the training second behavior information and the behavior label at the training time being less than a preset difference. At this time, the accuracy of the second behavior model and the second state model meets the requirements and can be used as the first behavior model and the first state model.

[0100] In an embodiment of the present application, the training second behavior information of the second object at the training time is determined through the second behavior model based on the training attribute information and the training first behavior information of the second object; the training state information of the second object at the training time is determined through the second state model based on the training attribute information; then, based on the difference between the training state information and the state label at the training time, and the difference between the training second behavior information and the behavior label at the training time, the loss value is calculated, and the parameters of the second behavior model and the second state model are adjusted based on the loss value. When the training end condition is met, the first behavior model and the first state model are obtained, the prediction accuracy of the first behavior model and the first state model is improved, thereby improving the efficiency of information recommendation.

[0101] The information recommendation method provided in the embodiments of this application can be applied to marketing scenarios, product recommendation scenarios, personalized information scenarios, search scenarios, and the like. The information recommendation method provided in the embodiments of this application can combine the user's personal reception preferences and idle state to push information to the user when the user is most likely to receive information, thereby improving the efficiency of information push.

[0102] Below, we will explain the exemplary application of the embodiment of the present application in an actual application scenario, taking the telephone sales scenario as an example, and explain it in detail below.

[0103] The modeling of user behavior prediction models in related technologies is to find features that can characterize the user's performance in a single scenario and perform single-target modeling. In this way, the user's characteristics can only single-handedly characterize the user's historical performance in a certain aspect, the model prediction accuracy is low, and it will also make it difficult to accurately estimate the user's preferences when there is a lack of user behavior data. The auxiliary module introduced in the embodiment of the present application can simultaneously predict the user's connection behavior and idle state, and better characterize the user's behavioral characteristics. In addition, by utilizing the advantages of the dual-tower model in learning and mining homogeneous users, the mining of representation vectors between users can be enhanced, thereby ensuring that users who lack historical data can also better use group characteristics to characterize their idle time.

[0104] 1. Data preprocessing

[0105] 1. Construct state labels and behavior labels:

[0106] Collect the second subject's hourly connection rate for marketing calls and the duration of online use of the marketing software for each hour over a 30-day period. Based on all this connection rate and duration information, determine the second subject's average connection rate and average duration. Label hours with a connection rate greater than the average connection rate as 1, and label hours with a connection rate less than or equal to the average connection rate as 0, thereby obtaining the behavior label 1. Label hours with an online duration greater than the average as 1, and label hours with an online duration less than or equal to the average as 0, thereby obtaining the state label 2.

[0107] 2. Collect training attribute information and training first behavior information:

[0108] The age, gender, occupation, etc. of the second object are counted as training attribute information X u ; Statistics the user's historical marketing call connection rate, connection times, number of calls, call duration and other information for each hour in a month as training history telemarketing information, recorded as X dx ; Count the number of times the user logs into the software corresponding to the marketing business at each hour in the past month, the login duration and other information as the training history login information, recorded as X ap ; Statistical time information, such as time, day of the week, month, etc., as training time, recorded as X t Among them, the training history telemarketing information X dx and training history login information X ap Combined together, it is the information for the first behavior of training.

[0109] 3. Embedding operation:

[0110] For training attribute information X u , training history telemarketing information X dx , training history login information X ap and training time X t Perform embedding mapping separately, and record the embedded features as X′ u , X′ dx , X′ ap , X′ t .

[0111] Here, the i-th feature x i For example, through embedding operation, we can get where g i is the length of the encoding query list of the i-th feature. Finally, we get X={x1,x2…,xf}, where i∈f represents the embedded representation of f features.

[0112] 2. Building the model network structure

[0113] 1. Build the second behavior model (using MLP+wide side structure to learn timing preference):

[0114] Among them, MLP+Wide is another expression of the Wide&Deep model, which combines deep learning with traditional linear models to achieve the memory and generalization capabilities of the recommendation system.

[0115] First, the interactive hidden information in the features is learned through the MLP layer, and feature cross-learning is performed, as shown in formula (1):

[0116]

[0117] Here, l represents the lth hidden layer in the MLP, W l 、b l is the learnable weight parameter in the MLP structure; h(l) represents the vector expression obtained by the lth MLP hidden layer.

[0118] Then, the feature expressiveness is enhanced by the wide side, and the generalization ability of the model is increased by residual learning, as shown in formula (2):

[0119] Y(1)=W1·[h(l),X]+b1 (2)

[0120] Here, W1 and b1 are the learnable weight parameters of the last linear layer in the timing target learning, and Y(1) is the output result of target learning 1, that is, the training second behavior information.

[0121] 2. Construct the second state model: Learning during idle time based on the dual-tower model:

[0122] Here, since the state label label2 learns the representation of homogeneous users, only the training attribute information X is used u As the vector expression of homogeneous users. In the dual-tower model of idle period learning, the training attribute information X u Learn to get the attribute representation u(l), using the training time X t Learning the temporal representation t(l):

[0123]

[0124] here, and are the weight and bias learnable parameter matrices in the attention mechanism, respectively.

[0125] After obtaining the training attribute representation u(l) and the training time representation t(l), we can calculate the similarity between these two representations and then obtain the probability of being idle at the current time, that is, the training state information. This process can be seen in formula (4):

[0126] Y(2)=u(l)*t(l)(4)

[0127] 3. Multi-objective fusion training:

[0128] Here, the loss calculation and parameter transmission are performed on the second behavior information of training and the training status information respectively, as shown in formula (5):

[0129]

[0130] Where L() is the loss function, using the cross-entropy loss function. w1 and w2 are the weights assigned to the second training behavior information and the training state information, and are hyperparameters. Loss1 is calculated by applying the loss function to the second training behavior information and label1, and loss2 is calculated by applying the loss function to the training state information and label2. Weighted operations are then applied to each to obtain the loss value, loss. Backpropagation learning is performed on the model using this loss value to obtain the first behavior model and the first state model.

[0131] 3. Model training process

[0132] The second object's connection performance in each time period is used as the behavior label label1, and the second object's login performance in each time period is used as the state label label2 for training. The multi-layer perceptron (MLP) involved uses a three-layer network structure with 256 neurons and a learning rate of 0.003. The Adam algorithm learner is used for training. The cross-entropy loss function is used to calculate the index between the training information interestingness output by the model and the real information interestingness, and then the loss value is determined to perform positive feedback training on the model. The specific mathematical expression is shown in Equation (6):

[0133]

[0134] Among them, y i and They represent the real information interest of the second object and the training information interest output by the model respectively.

[0135] See also Figure 3 , Figure 3 This is another flow chart of the model training method provided in the embodiment of the present application. Figure 3 To explain:

[0136] Training attribute information 411, training history telemarketing information 412, training history login information 413, and training first time 414 are input into shared embedding layer 42. Shared embedding layer 42 outputs training attribute features 431, training history telemarketing features 432, training history login features 433, and training time features 434. Training attribute features 431, training history telemarketing features 432, training history login features 433, and training time features 434 are then used as input to the second behavior model 44 on the left. MLP1 in second behavior model 44 performs calculations to obtain representation vectors for all features. The output of MLP1 is then input into concat layer 441 along with training attribute features 431, training history telemarketing features 432, training history login features 433, and training time features 434 for fusion and concatenation. The fusion result output by concat layer 441 is then input into linear layer 442, which outputs the result, obtaining training second behavior information 443. Afterwards, loss1 is determined based on the training second behavior information 443 and the behavior label label1. The training attribute information 411 and the training first time 414 are used as inputs to the second state model 45. The MLP2 in the second state model 45 is used for calculation to obtain the training attribute representation 451 and the training time representation 452, respectively. The similarity between the training attribute representation 451 and the training time representation 452 is then calculated to obtain the training state information 453. Afterwards, loss2 is determined based on the training state information 453 and the state label label2. Finally, loss1 and loss2 are weighted by the assigned weight w1 of the second behavior model 44 and the assigned weight w2 of the second state model 45 to obtain the loss value loss. The model is back-propagated using the loss value loss to obtain the first behavior model and the first state model.

[0137] In the embodiment of the present application, considering that in real marketing scenarios, the user's connection preference is often closely related to whether the user is currently idle, by simultaneously introducing the connection time preference and the user's idle time, the information learned about the user's idle time reversely affects the connection time preference learning, so that the model can perceive time information, provide users with time-aware personalized services, and significantly improve the user experience. Among them, a prediction model that integrates multi-objective modeling of idle state and connected state is designed. The user's interest is predicted through a deep learning framework, and the user's idle period is learned through a twin-tower model, so that the model learns idle information. Using the two types of behavioral information of the connected state and the idle state, the user's multi-dimensional time expression is performed. Through different features, the representation of different information is learned, and then the user's overall willingness prediction is obtained. Among them, the second state model based on the twin-tower model makes the homogeneous performance of different user groups in idle state explicit. Through user-side information and time-side information, the density relationship between user characteristics and idle time is found, and the user's idle probability is obtained through similarity calculation, thereby improving the prediction accuracy. The embodiment of the present application adds idle rate learning on the basis of individual interest modeling, which can make use of the representation of homogeneous users, and the group information has a certain stability, thus effectively avoiding individual disturbances caused by random factors and improving the robustness of the model.

[0138] Next, the structure of the electronic device according to the embodiment of the present application is described.

[0139] See also Figure 4 , Figure 4 is a structural diagram of an electronic device provided in an embodiment of the present application, Figure 4 The electronic device 200 shown includes: at least one processor 210, a memory 230 and at least one network interface 220. The various components in the server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 240 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 4 Various buses are labeled as bus system 240 .

[0140] The processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0141] The memory 230 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, a hard drive, an optical drive, etc. The memory 230 may optionally include one or more storage devices that are physically remote from the processor 210.

[0142] The memory 230 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 230 described in the embodiments of the present application is intended to include any suitable type of memory.

[0143] In some embodiments, memory 230 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0144] Operating system 231, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0145] The network communication module 232 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 220 . Exemplary network interfaces 220 include Bluetooth, Wireless Authentication (WiFi), and Universal Serial Bus (USB).

[0146] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 4 The information push device 233 stored in the memory 230 is shown. This device can be software in the form of a program or plug-in, and includes the following software modules: a determination module 2331 and a push module 2332. These modules are logical and can be arbitrarily combined or further separated according to the functions they implement. The functions of each module will be described below.

[0147] The following continues to describe the exemplary structure of the information push device 233 provided in the embodiment of the present application as a software module. In some embodiments, such as Figure 4 As shown, the software modules stored in the information push device 233 of the memory 230 may include:

[0148] Determination module 2331 is used to determine the second behavior information of the first object at the first time based on the attribute information and first behavior information of the first object; the occurrence time corresponding to the first behavior information is earlier than the first time; based on the attribute information of the first object, determine the status information of the first object at the first time.

[0149] The push module 2332 is configured to push information to the first object based on the second behavior information and the state information.

[0150] In some embodiments, the determination module 2331 is further used to perform feature extraction on the attribute information to obtain the attribute features of the first object; perform feature extraction on the time information of the first time to obtain the time features; and determine the status information of the first object at the first time based on the attribute features and the time features.

[0151] In some embodiments, the determination module 2331 is further configured to determine the state information of the first object at the first time based on feature similarity between the attribute feature and the time feature.

[0152] In some embodiments, the determination module 2331 is also used to determine the behavior pattern characteristics of the first object at the first time based on the attribute information of the first object, the first behavior information and the time information of the first time; and determine the behavior information of the first object at the first time based on the behavior pattern characteristics.

[0153] In some embodiments, the determination module 2331 is also used to extract a first feature sequence and a second feature sequence from the attribute information, the first behavior information and the time information; and determine the behavior pattern characteristics of the first object at the first time based on the first feature sequence and the second feature sequence.

[0154] In some embodiments, the determination module 2331 is further used to determine the correlation between the first object and the first behavior based on the behavior pattern characteristics; and determine the behavior information of the first object at the first time based on the correlation.

[0155] In some embodiments, the push module 2332 is also used to determine the information interest of the first object at the first time based on the behavior information and the status information; and push information to the first object at the first time when the information reception tendency meets the information push conditions.

[0156] In some embodiments, the push module 2332 is further configured to determine a second time for the first object and push information to the first object at the second time if the information reception tendency does not meet the information push condition.

[0157] In some embodiments, the determination module 2331 is also used to determine the second behavior information of the first object at the first time based on the attribute information and the first behavior information of the first object through a first behavior model; and to determine the state information of the first object at the first time based on the attribute information of the first object through a first state model.

[0158] In some embodiments, the determination module 2331 is also used to determine the training second behavior information of the second object at the training time through a second behavior model based on the training attribute information and the training first behavior information of the second object; determine the training state information of the second object at the training time through a second state model based on the training attribute information; calculate the loss value based on the difference between the training state information and the state label at the training time, and the difference between the training second behavior information and the behavior label at the training time; adjust the parameters of the second behavior model and the second state model based on the loss value, and obtain the first behavior model and the first state model when the training end condition is met.

[0159] An embodiment of the present application provides a computer program product, which includes a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the information push method described in the embodiment of the present application.

[0160] The embodiment of the present application provides a computer-readable storage medium in which computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the information push method provided by the embodiment of the present application, for example, Figure 4 The information push method shown.

[0161] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0162] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0163] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored in part of a file that stores other programs or data, e.g., in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0164] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0165] In summary, through the embodiments of the present application, first, based on the first object's attribute information and first behavior information, the second behavior information of the first object at a first time is determined; then, based on the first object's attribute information, the first object's state information at the first time is determined; and finally, based on the second behavior information and state information, information is pushed to the first object. The first behavior information corresponds to an occurrence time earlier than the first time. The first behavior information can be used to analyze the first object's information reception preference over the past period of time, thereby predicting the first object's information reception within the first time period. In marketing scenarios, a user's information reception tendency is closely related to whether the user is currently idle; users are generally more receptive to information recommendations when they are idle. Therefore, in the embodiments of the present application, whether the first object is idle at a first time is predicted based on attribute information. Then, information is pushed to the first object based on both the first object's information reception and whether it is idle. In this way, combining the second behavior information and state information for information push ensures that information is pushed to the user when the user is most receptive to information, allowing the user to smoothly receive the pushed information, thereby improving the efficiency of information push.

[0166] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. An information push method, characterized in that: The method comprises: Determining, based on attribute information and first behavior information of a first object, second behavior information of the first object at a first time; the occurrence time corresponding to the first behavior information is earlier than the first time; determining, based on the attribute information of the first object, state information of the first object at the first time; Based on the second behavior information and the state information, information is pushed to the first object.

2. The method according to claim 1, characterized in that The determining, based on the attribute information of the first object, the state information of the first object at the first time includes: Performing feature extraction on the attribute information to obtain attribute features of the first object; Extracting features from the time information of the first time to obtain time features; The state information of the first object at the first time is determined based on the attribute feature and the time feature.

3. The method according to claim 2, characterized in that The determining, based on the attribute feature and the time feature, the state information of the first object at the first time includes: The state information of the first object at the first time is determined based on the feature similarity between the attribute feature and the time feature.

4. The method according to any one of claims 1 to 3, characterized in that The determining, based on the attribute information and the first behavior information of the first object, the second behavior information of the first object at the first time includes: determining a behavior pattern feature of the first object at the first time based on the attribute information of the first object, the first behavior information, and the time information of the first time; Based on the behavior pattern characteristics, the second behavior information of the first object at the first time is determined.

5. The method according to claim 4, characterized in that The determining, based on the attribute information of the first object, the first behavior information, and the time information of the first time, a behavior pattern feature of the first object at the first time includes: extracting a first feature sequence and a second feature sequence from the attribute information, the first behavior information, and the time information; The behavior pattern feature of the first object at the first time is determined based on the first feature sequence and the second feature sequence.

6. The method according to claim 4, characterized in that The determining, based on the behavior pattern feature, the second behavior information of the first object at the first time includes: determining, based on the behavior pattern characteristics, a degree of association between the first object and the first behavior; Based on the association degree, the second behavior information of the first object at the first time is determined.

7. The method according to any one of claims 1 to 3, characterized in that The pushing information to the first object based on the second behavior information and the state information includes: determining, based on the second behavior information and the state information, information interest of the first object at the first time; When the information interest level satisfies an information push condition, information is pushed to the first object at the first time.

8. The method according to claim 7, characterized in that After determining the information interest of the first object at the first time based on the second behavior information and the state information, the method further includes: When the information interest level does not satisfy the information push condition, a second time is determined for the first object, and information is pushed to the first object at the second time.

9. The method according to any one of claims 1 to 3, characterized in that The determining, based on the attribute information and the first behavior information of the first object, the second behavior information of the first object at the first time includes: determining, by a first behavior model, the second behavior information of the first object at the first time based on the attribute information and the first behavior information of the first object; The determining, based on the attribute information of the first object, the state information of the first object at the first time includes: The state information of the first object at the first time is determined based on the attribute information of the first object through a first state model.

10. The method according to claim 9, characterized in that The method further comprises determining, using the first behavior model and based on the attribute information and the first behavior information of the first object, that the first object occurs before the second behavior information at the first time: Determining, by the second behavior model, the trained second behavior information of the second object at the training time based on the trained attribute information and the trained first behavior information of the second object; determining, by a second state model, training state information of the second object at a training time based on the training attribute information; Calculating a loss value based on a difference between the training state information and the state label at the training time, and a difference between the training second behavior information and the behavior label at the training time; Parameters of the second behavior model and the second state model are adjusted based on the loss value, and when a training end condition is met, the first behavior model and the first state model are obtained.

11. An information push device, characterized in that: The device comprises: a determination module, configured to determine, based on attribute information and first behavior information of a first object, second behavior information of the first object at a first time, wherein the occurrence time corresponding to the first behavior information is earlier than the first time; and determine, based on the attribute information of the first object, state information of the first object at the first time; A push module is used to push information to the first object based on the second behavior information and the status information.

12. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions or computer programs; A processor, configured to implement the method according to any one of claims 1 to 10 when executing computer-executable instructions or computer programs stored in the memory.

13. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer executable instructions or computer program are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

14. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer executable instructions or computer program are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

Images