Recommendation model training method, recommendation method, device and computer-readable storage medium

Abstract

The present invention provides a recommendation model training method, a recommendation method, an apparatus, and a computer-readable storage medium. The recommendation model includes: a first BERT network, a second BERT network, and a prediction network. The method includes: obtaining training samples, the training samples including: historical user response behaviors to historical marketing activities, text data of candidate marketing activities, and user response labels for the candidate marketing activities; constructing time series data based on the historical response behaviors, inputting the time series data into the first BERT network to extract time series features with time series dynamic weights; constructing marketing activity features based on the text data of the candidate marketing activities, inputting the marketing activity features into the second BERT network to extract semantic features; inputting the time series features and semantic features into the prediction network to obtain the user's response probability for the candidate marketing activities, and using the response probability and response label to train the recommendation model. Using the above method, accurate recommendations can be achieved.

Description

Technical Field

[0001] The present invention belongs to the field of recommendation, and specifically relates to a recommendation model training method, a recommendation method, a device and a computer-readable storage medium. Background Art

[0002] This section is intended to provide a background or context to embodiments of the invention that are recited in the claims. No statement herein is admitted to be prior art by inclusion in this section.

[0003] With the development of recommendation algorithms, how to accurately recommend marketing activities of interest to users has become an important issue. However, user preferences may change at any time due to various factors, and current recommendation algorithms have problems of lag and low accuracy.

[0004] Therefore, how to achieve accurate marketing activity recommendations is an urgent problem to be solved. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned prior art, a recommendation model training method, a recommendation method, a device and a computer-readable storage medium are proposed. By using this method, the above-mentioned problems can be solved.

[0006] The present invention provides the following solutions.

[0007] In a first aspect, a recommendation model training method is provided, wherein the recommendation model includes: a first BERT network, a second BERT network and a prediction network, including: obtaining training samples, wherein the training samples include: historical response behaviors of users to historical marketing activities, text data of candidate marketing activities and response labels of users to the candidate marketing activities; constructing time series data according to the historical response behaviors, and inputting the time series data into the first BERT network to extract time series features with time series dynamic weights; constructing marketing activity features according to the text data of the candidate marketing activities, and inputting the marketing activity features into the second BERT network to extract semantic features; inputting the time series features and the semantic features into the prediction network to obtain the response probability of the user to the candidate marketing activities, and using the response probability and the response label to train the recommendation model.

[0008] In one embodiment, constructing time series data based on the historical response behavior also includes: obtaining the user's N historical response behaviors for the most recent N historical marketing activities to generate first time series data; performing an MLM pre-training task on the behavior sequence to output second time series data; performing word embedding processing on the second behavior sequence to output third time series data as the time series data.

[0009] In one embodiment, the training sample further includes: static attribute information of the user, and the method further includes: performing word embedding processing on the static attribute information of the user and adding the time series data.

[0010] In one embodiment, generating the first time series data also includes: obtaining the user's N historical response behaviors to the most recent N historical marketing activities and performing vectorized processing; determining the correlation between each of the historical marketing activities and the candidate marketing activities based on the time relationship; determining the weight corresponding to each of the historical marketing activities based on the correlation; and generating the first time series data based on the historical response behaviors and the weights.

[0011] In one embodiment, determining the correlation between each of the historical marketing activities and the candidate marketing activities based on the time relationship also includes: defining a time decay condition based on the time difference between each of the historical marketing activities and the current one; and determining the correlation based on the time decay condition.

[0012] In one embodiment, determining the correlation between each of the historical marketing activities and the candidate marketing activities based on the time relationship also includes: using the position coding mechanism of the Transformer encoder to time-series encode the behavior sequence to obtain a time-series coding vector, and determining the correlation based on the time-series coding vector.

[0013] In one embodiment, the position encoding mechanism of the Transformer encoder is used to temporally encode the behavior sequence, further comprising: wherein the temporal encoding rule is defined using the following formula:

[0014]

[0015]

[0016] Among them, TE(t,2i) is the 2i-th dimension of the temporal encoding vector of time series t, TE(t,2i+1) is the 2i+1-th dimension of the temporal encoding vector of time series t, and d model is the dimension of the temporal encoding vector.

[0017] In one embodiment, the text data of the candidate marketing campaign includes: obtaining one or more of text information, rights and interests information, and attribute information of the candidate marketing campaign.

[0018] In one embodiment, after the temporal features and the semantic features are input into the prediction network, the method further includes: learning the temporal features V using the following formula: u With the specified candidate marketing campaign V s The correlation relationship between: Among them, the V i Refers to the i-th consumption response behavior of the user u, and its value is a positive integer between 1-N. s For a specific marketing campaign, i =g(V i ,V s ) refers to the V to be learned i and V s The correlation relationship between them.

[0019] In one embodiment, the recommendation model is trained using the response probability and the response label, further comprising: defining a dynamic weight W i =f(V u ,V s ), the W i is the model parameter to be trained, which refers to the behavior sequence V u With the candidate marketing campaign V s The correlation weight between the dynamic weight W i An improved loss function LOSS is added to train the recommendation model.

[0020] In one embodiment, obtaining training samples also includes: building a marketing network based on the user's historical consumption information on different consumption platforms, and forming a marketing group division scheme through a graph algorithm; obtaining multi-platform historical consumption information of multiple users in the marketing group, and obtaining the training samples to generate a sample set.

[0021] In a second aspect, a recommendation method is provided, comprising: obtaining a user's historical marketing activity response sequence over a previous period of time, inputting the recommendation model trained as in the first aspect, and obtaining a recommended target marketing activity.

[0022] In a third aspect, a training device for a recommendation model is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute: the method of the first aspect.

[0023] In a fourth aspect, a recommendation device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute: the method of the second aspect.

[0024] In a fifth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a program, and when the program is executed by a multi-core processor, the multi-core processor executes the method of the first aspect or the method of the second aspect.

[0025] One of the advantages of the above implementation is that it can achieve more accurate marketing activity recommendations.

[0026] Other advantages of the present invention will be explained in more detail with reference to the following description and accompanying drawings.

[0027] It should be understood that the above description is only an overview of the technical solution of the present invention, so that the technical means of the present invention can be more clearly understood and implemented in accordance with the contents of the description. In order to make the above and other objects, features and advantages of the present invention more clearly understood, the following examples are given to illustrate specific embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The advantages and benefits described herein, as well as other advantages and benefits, will be apparent to those skilled in the art upon reading the detailed description of the exemplary embodiments below. The accompanying drawings are provided for illustration purposes only and are not to be construed as limiting the present invention. The same reference numerals are used throughout the accompanying drawings to denote the same components. In the accompanying drawings:

[0029] Figure 1 Schematic diagram of the structure of a recommendation model training device according to one embodiment of the present invention;

[0030] Figure 2 A flowchart of a method for performing recommendation model training according to an embodiment of the present disclosure is shown.

[0031] Figure 3a -b is a schematic diagram before and after the introduction of timing weights according to an embodiment of the present invention;

[0032] Figure 4 A flowchart of a method for performing recommendation model training according to an embodiment of the present disclosure is shown.

[0033] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. DETAILED DESCRIPTION

[0034] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0035] In the description of the embodiments of the present application, it should be understood that terms such as "including" or "having" are intended to indicate the presence of features, numbers, steps, actions, components, parts or combinations thereof disclosed in this specification, and are not intended to exclude the possibility of the presence of one or more other features, numbers, steps, actions, components, parts or combinations thereof.

[0036] Unless otherwise specified, “ / ” means or. For example, A / B can mean A or B. “and / or” in this article is merely a way to describe the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0037] The terms "first," "second," etc., are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Thus, a feature defined as "first," "second," etc. may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0038] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0039] See first Figure 1 , which schematically illustrates a diagram of an environment 100 in which exemplary implementations according to the present disclosure may be used.

[0040] Figure 1 FIG1 shows a schematic diagram of an example of a computing device 100 according to an embodiment of the present disclosure. It should be noted that, Figure 1 This is a schematic diagram of the hardware operating environment of the training method for the recommendation model. The embodiment of the present invention is based on the XX device, which can be a terminal device such as a PC or a portable computer.

[0041] like Figure 1As shown, the training device of the recommendation model may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may optionally also be a storage device independent of the aforementioned processor 1001.

[0042] Those skilled in the art will understand that Figure 1 The training device structure of the recommended model shown in does not constitute a limitation on the XX device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0043] like Figure 1 As shown, memory 1005, a computer storage medium, may include an operating system, a network communication module, a user interface module, and a recommendation model training program. The operating system is a program that manages and controls the hardware and software resources of the XX device and supports the operation of the recommendation model training program and other software or programs.

[0044] exist Figure 1 In the training device for the recommendation model shown, the user interface 1003 is mainly used to receive requests, data, etc. sent by the first terminal, the second terminal, and the supervisory terminal; the network interface 1004 is mainly used to connect to the backend server for data communication; and the processor 1001 can be used to call the training program for the recommendation model stored in the memory 1005 and perform the following operations:

[0045] Acquire training samples, wherein the training samples include: historical response behaviors of users to historical marketing activities, text data of candidate marketing activities, and response labels of users to the candidate marketing activities; construct time series data based on the historical response behaviors, and input the time series data into a first BERT network to extract time series features with time series dynamic weights; construct marketing activity features based on the text data of the candidate marketing activities, and input the marketing activity features into a second BERT network to extract semantic features; input the time series features and the semantic features into a prediction network to obtain the response probability of the user to the candidate marketing activities, and use the response probability and the response label to train the recommendation model.

[0046] Therefore, this solution captures users' temporal consumption habits by adding time series information for modeling, dynamically constructs weights for different time series, and simultaneously labels real-time transaction data through a comprehensive consumption platform to conduct precise marketing for users, respecting and interpreting users' consumption habits and preferences at different life cycles, thereby maximizing the use of marketing resources.

[0047] Figure 2 FIG1 shows a flow chart of a method for performing a recommendation model training according to an embodiment of the present disclosure. The recommendation model includes: a first BERT network, a second BERT network and a prediction network. The method can be performed as follows: Figure 1 It should be understood that the method 200 may include additional blocks not shown and / or may omit blocks shown, and the scope of the present disclosure is not limited in this respect.

[0048] Step 210: Obtain training samples.

[0049] The training samples include: historical response behaviors of users to historical marketing activities, text data of candidate marketing activities, and response labels of users to the candidate marketing activities;

[0050] Optionally, the training samples may also include user static attribute information, such as user ID, gender, age, device model, occupation, etc. Such user static attribute information may also be related to the user's consumption habits.

[0051] The training sample set includes several black and white samples, where black samples refer to training samples with a response label of "no" and white samples refer to training samples with a response label of "yes", and each training sample is obtained based on the user side and / or the activity side.

[0052] For example, training samples can be: user static features (user A, gender, age, occupation), user historical response behavior to historical marketing activities (act1, t1, act2, t2, ... t N-1 ,act N ), where act n is the historical marketing activity identifier, the above t1, t2, t N-1 The corresponding interval between the user's response to two adjacent historical marketing activities can be seen from this, which shows the user's preference for marketing actions. S , text data, such as the logo, merchant information and copy information of candidate marketing activities. It can be understood that different copy information of candidate marketing activities will also lead to different response behaviors of users. Therefore, using such information as training samples is beneficial to the recommendation accuracy of the recommendation model.

[0053] Step 220: construct time series data based on the historical response behavior, and input the time series data into a first BERT network to extract time series features with time series dynamic weights;

[0054] Traditional recommendation solutions fail to fully capture the temporal information of user consumption transactions, and the weight of marketing information received by users at different stages of their lifecycles is not factored into marketing decisions. In this implementation, each historical marketing activity is assumed to be a single word, and the user's temporal data is treated as a sentence. This transforms the problem into a natural language processing (NLP) text classification problem. The first BERT network is used as the encoder for encoding, and this solution employs a six-layer transformer. Each transformer layer takes the latent variables of the previous layer as input, passes through multi-head self-attention and layer norm, and finally outputs latent variables with higher interactivity.

[0055] BERT's outstanding performance in the NLP field stems primarily from the extensive implicit semantic information captured through pre-training. Therefore, this solution is well-suited to capturing the semantic meaning of historical marketing text. However, unlike NLP, historical marketing campaigns and consumer scenarios are sparse, resulting in a vocabulary of only seven or eight million words. This presents two challenges. The first is that such a large vocabulary is difficult to fit on a GPU. The second is that even on a CPU, training is difficult due to the large number of parameters. To address these issues, this solution uses a two-stage pre-training approach to train BERT.

[0056] In the first stage, word2vec is used to pre-train embeddings for marketing campaigns. The goal is to cluster similar marketing campaigns, input the user's consumer behavior sequence for participating in marketing campaigns into the embedding layer, and then predict the campaign ID. However, this prediction does not utilize the attributes of the marketing campaign itself, such as category information. For example, if the campaign is a new type of marketing campaign, it is easier to predict the current campaign. The model is also encouraged to combine the attributes of the recently clicked marketing campaign and the current marketing campaign to make predictions, thereby aligning the embeddings of the campaign ID and its attributes.

[0057] The second phase improves Masked Language Modeling (MLM). The original MLM directly masks the entire marketing campaign, including its attributes, but this solution only masks a portion of the marketing campaign (campaign ID or attributes). Specifically, 20% of the tokens can be randomly masked, and the masked positions can be replaced with a 0 vector or an attribute ID can be sampled for replacement. When using the improved MLM to predict these activities, not only can the most recently clicked activities be used, but also their attributes, so that the embedding expressions (embedding) of different attributes can be aligned in the semantic space. At the same time, this solution reduces the vocabulary to 100,000 (extracting the top 10w and the rest as unk), which allows BERT to run more efficiently.

[0058] In one embodiment, in step 220, constructing time series data based on the historical response behavior may further include:

[0059] First, the user's N historical responses to the most recent N marketing campaigns are obtained to generate first time series data. Then, an MLM pre-training task is performed on this sequence of behaviors, outputting second time series data. The MLM pre-training task is an unsupervised task that randomly selects a certain percentage of tokens in a sentence and replaces them with "[MASK]." A classification model is then used to predict the actual word "[MASK]." Afterwards, word embedding is performed on this second sequence of behaviors. This allows for sufficient pre-processing of the input data for the first BERT network, allowing the first BERT network to extract more implicit information from this third time series data.

[0060] In one embodiment, in step 220, constructing time series data based on the historical response behavior may further include:

[0061] Step 221: Obtain the user's N historical responses to the most recent N marketing campaigns and perform vectorization processing. For example, (1) fine-tune the results to directly embed the text; (2) perform knowledge distillation as data enhancement to add semantic information. It is also possible to set the newly added campaigns that fail to capture relevant information as unk.

[0062] Step 222: determining the relevance between each of the historical marketing activities and the candidate marketing activities based on the time relationship;

[0063] In one embodiment, the above step 222 may include: defining a time decay condition according to the time difference between each of the historical marketing activities and the current one; and determining the correlation according to the time decay condition.

[0064] In another embodiment, the above step 222 may further include: performing temporal encoding on the behavior sequence using the positional encoding mechanism of the Transformer encoder to obtain a temporal encoding vector, and determining the correlation according to the temporal encoding vector.

[0065] For example, the user's response behavior to historical marketing activities: (act1, t1, act2, t2, act3…t N-1 ,act N ), at this time, if the user clicks to respond to a marketing action that was responded to a short time ago, the correlation between the two is high. For example, the marketing action that was responded to a short time before the target activity is highly correlated, while the marketing action that was responded to a long time before the target activity is less correlated. For example, the following formula is used to set each historical marketing action: action n Relevance to current candidate campaign:

[0066] At the same time, considering the user's usage habits, in addition to the absolute time relationship, relative time is also very important. Therefore, the position encoding mechanism of the Transformer encoder can be referred to to perform temporal encoding on the usage timing information.

[0067] For example, the following timing encoding rules can be used:

[0068]

[0069]

[0070] Among them, TE(t,2i) is the 2i-th dimension of the temporal encoding vector of time series t, TE(t,2i+1) is the 2i+1-th dimension of the temporal encoding vector of time series t, and d model is the dimension of the temporal encoding vector.

[0071] According to the above formula, the time series vector at time t+t1 can be obtained by linearly changing time t, which makes it easier for the model to capture changes between relative time series.

[0072] Step 223: determining a weight corresponding to each of the historical marketing activities according to the correlation;

[0073] Step 224: Generate the time series data according to the historical response behavior and the weight.

[0074] Because of the addition of the attention mechanism, feature concatenation has changed from direct summation to a weighted sum of the benefits of different scenarios and activities, and its weight is determined by the relationship between user features and activity features. For example, given a candidate marketing activity V s , follow each historical marketing campaign Vi With this candidate marketing campaign V s The correlation g(V i ,V s ), to realize user time series data V u expression.

[0075]

[0076] Among them, the V i Refers to the i-th consumption response behavior of the user u, and its value is a positive integer between 1-N. s For a specific marketing campaign, i =g(V i ,V s ) refers to the V to be learned i and V s The correlation relationship between them. For example, the time series data V u As shown, this embodiment avoids using the same vector to express the different consumption habits of all users, but instead considers the correlation of historical behaviors w i Adaptively calculate the representation vector of user interests (for different given scenarios and candidate marketing activities V s ), the user's time series data represents the vector V u Varies with different scenarios and marketing interests.

[0077] It is understandable that the fixed-length user representation vector in the traditional model remains unchanged for different candidate marketing plans. However, in business, behaviors related to consumption scenarios and merchants greatly affect users' responses and click actions to marketing benefits. The following are examples of how this can be done:

[0078] (1) Non-sequential method: see Figure 3a The non-sequential approach is simple and requires less computing resources, and pushes content based on the user's historical static clicks. However, if a user has a history of frequent high-calorie fast food consumption, high-calorie fast food coupons may be prioritized. However, if the user has recently started using the gym or eating a diet meal, the push of these high-calorie fast food coupons may be significantly delayed.

[0079] (2) Add timing mode: see Figure 3b It can be seen that at this time, push notifications are not only based on the static click content of the user's history, but also based on the dynamic time series content of the user's historical consumption. The weight of the most recent consumption items is increased, and the weight of the earlier consumption items is reduced. Finally, the gym marketing project is pushed to the user, and the user click-through rate is increased.

[0080] The time series method has high requirements for real-time computing. If a simple time series response is selected to push notifications based on the user's recent consumption habits, for example, if the user's long-term consumption preferences for offline transactions have not changed, but a coupon is pushed based solely on online Pinduoduo consumption behavior based on time series results, the probability of click-through consumption is low;

[0081] The problem with the above approach is that users may no longer be interested in the promotional benefits or their consumption habits may not align with their current lifecycle, leading them to not click or consume. Therefore, a dynamic user consumption profile needs to be constructed. Using adaptive dynamic weights, we can learn the weights assigned to different consumption patterns, thereby capturing short-term shifts in consumption preferences. This increases the likelihood that users will click to consume sports and exercise coupons in the future.

[0082] This embodiment avoids using the same vector to express the different consumption habits of all users, but instead considers the correlation of historical behaviors w i Adaptively calculate the representation vector of user interests (for different given scenarios and candidate marketing activities V s ), the user's time series data represents the vector V u This varies depending on different scenarios and marketing benefits. For example, if there is a higher correlation between fat-loss meals and exercise coupons, then if users' short-term consumption preferences change, this correlation between consumption habits and coupons will be further strengthened.

[0083] Step 230: construct marketing campaign features based on the text data of the candidate marketing campaign, and input the marketing campaign features into a second BERT network to extract semantic features;

[0084] It is understandable that the current semantic features of product recommendations are relatively simple, and users themselves have different sensitivities to different copywriting. For example, young people are more likely to click on lively copywriting and symbolic emoticons, such as "Send you a cup of coffee and milk tea to start a good mood~". Through pre-training and encoding, the model can learn these features to supplement the user's response to the activity.

[0085] The text data of the candidate marketing activities include copy information, attribute information, activity information, etc.

[0086] Step 240: Input the temporal features and the semantic features into a prediction network to obtain the user's response probability to the candidate marketing activity, and train the recommendation model using the response probability and the response label.

[0087] In the training objective, by defining the dynamic weight W i =f(V u ,V s), add an improved loss function, for high-value users or key scenarios that are important in business goals, multiply the cumulative factor in the cross entropy loss function by the response weight intensity:

[0088]

[0089] For example, for a specific customer group V u , can improve its related weight = f(V u ,V s ), through this adjustment, more accurate predictions can be provided for high-value users or key scenarios, achieving more user recommendations while bringing more clicks and conversions, and improving the efficiency of precision marketing.

[0090] For example, the prediction network may include: pooling layer, batch normalization layer, softmax layer, etc. Continuous features are added through the pooling layer, entering the batch normalization layer, and finally output through a multi-layer perceptron combined with softmax to obtain the user's response probability prediction for a single activity. Separately measuring multiple activities simultaneously can obtain the logits output of the response to multiple activities. Setting the corresponding threshold can perform a rough screening and recall of the results, and the return result of the ranked activity priority can be obtained through probabilistic sorting.

[0091] In one embodiment, it is also possible to: construct a marketing network based on the historical consumption information of users on different consumption platforms, form a marketing group division scheme through graph algorithms; obtain the multi-platform historical consumption information of multiple users in the marketing group, obtain the training samples to generate a sample set. In this way, targeted training can be carried out for different groups of customers, for example, the V of a specific customer group can be improved. u The training weights can improve the efficiency of precision marketing, and multi-platform data can better express the changes in user consumption preferences.

[0092] Through the above embodiments, the present application can achieve the following technical effects: First, the user's consumption time series information on each platform is used in a linked manner, and attention is paid to the user's consumption habits at different time series to form a dynamic weight expression. The multi-dimensional data of cross-platform transactions is integrated to form a dynamic user portrait, quantify the consumption preferences for different consumption scenarios and merchants, and provide a quantitative basis for precision marketing push; second, while paying attention to the behavior information of the platform, cross-platform information is integrated to obtain the user's latest consumption habits in the first place, and the platform labeling feature calculation is constructed through real-time features. By incorporating feature model training based on the attention method, the accuracy of the weight expression of the user's consumption habits at different time series in the life cycle is improved.

[0093] It should be noted that the steps not described in detail in this embodiment can be referred to Figure 1 The description of the relevant steps in the illustrated embodiment will not be repeated here.

[0094] In the description of this specification, the description with reference to the terms "some possible embodiments", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0095] About the method flow chart of the present application embodiment, some operations are described as the different steps performed in a certain order.Such flow chart belongs to illustrative and non-restrictive.Some steps described in this article can be grouped together and performed in a single operation, some steps can be divided into multiple sub-steps and can be performed in a different order than shown in this article.The various steps shown in the flow chart can be realized in any way by any circuit structure and / or tangible mechanism (for example, by the software, hardware (for example, the logical function realized by processor or chip) etc. running on computer equipment and / or its any combination).

[0096] Based on the same technical concept, an embodiment of the present invention also provides a recommendation method, including: obtaining a user's historical marketing activity response sequence over a previous period of time, inputting the recommendation model trained using the above embodiment, and obtaining a recommended target marketing activity.

[0097] Based on the same technical concept, an embodiment of the present invention also provides a training device for a recommendation model, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute: the method described in the above embodiment.

[0098] Based on the same technical concept, an embodiment of the present invention also provides a recommendation device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute: the method described in the above-mentioned recommendation method embodiment.

[0099] It should be noted that the device in the embodiment of the present application can implement each process of the embodiment of the aforementioned method and achieve the same effects and functions, which will not be repeated here.

[0100] According to some embodiments of the present application, a non-volatile computer storage medium for a training method of a recommendation model is provided, on which computer executable instructions are stored. The computer executable instructions are configured to execute the method described in the above embodiment when executed by a processor.

[0101] Each embodiment of the present application is described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the descriptions of the apparatus, device, and computer-readable storage medium embodiments are simplified because they are generally similar to the method embodiments. For relevant portions, reference can be made to the descriptions of the method embodiments.

[0102] The apparatus, device and computer-readable storage medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the apparatus, device and computer-readable storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, device and computer-readable storage medium will not be repeated here.

[0103] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, apparatus (device or system), or computer-readable storage media. Thus, the present invention may take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer-readable storage medium implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (devices or systems), and computer-readable storage media according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0107] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0108] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0109] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information using any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. In addition, although the operations of the present method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order or that all illustrated operations must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into a single step, and / or a single step may be broken down into multiple steps.

[0110] Although the spirit and principles of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features of these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A recommendation model training method, characterized in that: The recommendation model includes: a first BERT network, a second BERT network and a prediction network, including: Acquire training samples, where the training samples include: historical response behaviors of users to historical marketing activities, text data of candidate marketing activities, and response labels of users to the candidate marketing activities; Constructing time series data based on the historical response behavior, and inputting the time series data into a first BERT network to extract time series features with time series dynamic weights; constructing a marketing campaign feature based on the text data of the candidate marketing campaign, and inputting the marketing campaign feature into a second BERT network to extract semantic features; The temporal features and the semantic features are input into a prediction network to obtain the user's response probability to the candidate marketing activity, and the recommendation model is trained using the response probability and the response label.

2. The method according to claim 1, characterized in that Constructing time series data based on the historical response behavior also includes: Obtaining N historical response behaviors of the user to the most recent N historical marketing activities to generate first time series data; Perform MLM pre-training tasks on the user's behavior sequence of participating in marketing activities and output the second time series data; Perform word embedding processing on the second time series data, and output third time series data as the time series data.

3. The method according to claim 1, characterized in that The training sample further includes: static attribute information of the user, and the method further includes: The user's static attribute information is word-embedded and added with the time series data.

4. The method according to claim 2, characterized in that Generating the first time series data also includes: Obtaining N historical response behaviors of the user for the most recent N historical marketing activities and performing vectorized processing; Determining the relevance between each of the historical marketing activities and the candidate marketing activities based on a time relationship; determining a weight corresponding to each of the historical marketing activities according to the correlation; The time series data is generated according to the historical response behavior and the weight.

5. The method according to claim 4, characterized in that Determining the relevance between each of the historical marketing activities and the candidate marketing activities based on the time relationship further includes: defining a time decay condition based on the time difference between each of the historical marketing activities and the current one; The correlation is determined according to the timing attenuation condition.

6. The method according to claim 4, characterized in that Determining the relevance between each of the historical marketing activities and the candidate marketing activities based on the time relationship further includes: The behavior sequence is temporally encoded using a position encoding mechanism of a Transformer encoder to obtain a temporal encoding vector, and the correlation is determined based on the temporal encoding vector.

7. The method according to claim 6, characterized in that The position encoding mechanism of the Transformer encoder is used to temporally encode the behavior sequence. include: The following formula is used to define the timing coding rule: Among them, TE(t,2i) is the 2i-th dimension of the temporal encoding vector of time series t, TE(t,2i+1) is the 2i+1-th dimension of the temporal encoding vector of time series t, and d model is the dimension of the temporal encoding vector.

8. The method according to claim 1, characterized in that The text data of the candidate marketing activities includes: obtaining one or more of the copywriting information, rights and interests information and attribute information of the candidate marketing activities.

9. The method according to claim 1, characterized in that After inputting the temporal features and the semantic features into the prediction network, the method further includes: The following formula is used to learn the time series feature V u With the specified candidate marketing campaign V s The correlation relationship between: Among them, the V i Refers to the i-th consumption response behavior of the user u, and its value is a positive integer between 1 and N, where N represents the number of consumption response behaviors. s is the candidate marketing activity, w i =g(V i ,V s ) refers to the V to be learned i and V s The correlation relationship between them.

10. The method according to claim 1, characterized in that Training the recommendation model using the response probability and the response label also includes: Define dynamic weight W i =f(V u ,V s ), the W i is the model parameter to be trained, which refers to the time series feature V u With the candidate marketing campaign V s The correlation weight between them; The dynamic weight W i An improved loss function LOSS is added to train the recommendation model.

11. The method according to claim 1, wherein Obtaining training samples also includes: Build a marketing network based on users' historical consumption information on different consumption platforms, and form a marketing group division plan through graph algorithms; The multi-platform historical consumption information of multiple users in the marketing group is obtained, and the training samples are obtained to generate a sample set.

12. A recommendation method, characterized in that: include: Obtain the user's historical marketing activity response sequence over a previous period of time, input the recommendation model trained using the recommendation model training method according to any one of claims 1 to 11, and obtain the recommended target marketing activity.

13. A training device for a recommendation model, characterized in that: include: at least one processor; And, a memory communicatively connected to at least one processor; wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable the at least one processor to execute: the method according to any one of claims 1-11.

14. A recommendation device, characterized in that: include: at least one processor; And, a memory communicatively connected to at least one processor; wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable the at least one processor to perform: the method as claimed in claim 12. 15 . A computer-readable storage medium storing a program, wherein when the program is executed by a multi-core processor, the multi-core processor is caused to execute the method according to claim 1 or the method according to claim 12 .

Images