WO2024060587A1 - 自监督学习模型的生成方法和转化率预估模型的生成方法 - Google Patents

自监督学习模型的生成方法和转化率预估模型的生成方法 Download PDF

Info

Publication number
WO2024060587A1
WO2024060587A1 PCT/CN2023/087290 CN2023087290W WO2024060587A1 WO 2024060587 A1 WO2024060587 A1 WO 2024060587A1 CN 2023087290 W CN2023087290 W CN 2023087290W WO 2024060587 A1 WO2024060587 A1 WO 2024060587A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample
user
target
model
browsed
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/087290
Other languages
English (en)
French (fr)
Inventor
程维
范聪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Jingdong Century Trading Co Ltd
Beijing Wodong Tianjun Information Technology Co Ltd
Original Assignee
Beijing Jingdong Century Trading Co Ltd
Beijing Wodong Tianjun Information Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Jingdong Century Trading Co Ltd, Beijing Wodong Tianjun Information Technology Co Ltd filed Critical Beijing Jingdong Century Trading Co Ltd
Publication of WO2024060587A1 publication Critical patent/WO2024060587A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0242—Determining effectiveness of advertisements
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange

Definitions

  • Embodiments of the present disclosure relate to the field of artificial intelligence technology, and specifically relate to a method of generating a self-supervised learning model and a method of generating a conversion rate estimation model.
  • Conversion Rate was originally an indicator to measure the effectiveness of advertising, that is, calculating the proportion of users who clicked on an ad and paid for it among all users who viewed the ad. With the continuous development of related technologies, CVR is also often used to evaluate the effects of advertising, product recommendation, and search. Conversion rate estimation technology generally uses artificial intelligence algorithms to model each user's purchase conversion of online products, advertisements, etc., and make predictions to determine whether the user will purchase or give a probability value of purchase. In e-commerce systems, users are usually shown ads, products, etc. that have a higher CVR for the user. It can be seen that CVR prediction is often an important technology.
  • some embodiments of the present disclosure provide a method for generating a self-supervised learning model, including: inputting sample input data in the first sample data into a first initial model, and predicting sample item information browsed at the target time , where the sample input data includes the attribute information of the sample user, the attribute information of each sample item browsed by the sample user, and the sample item information sequence of the target period browsed by the sample user before the target time; determine the predicted sample item information and the target sample item The loss function value between the information, where the target sample item information is the sample item information browsed by the sample user at the target moment in the first sample data; adjust the first initial model according to the comparison result of the loss function value and the preset threshold The relevant parameters in are trained based on the adjusted parameters until the training is completed; the first initial model that has been trained is determined as the self-supervised learning model.
  • some embodiments of the present disclosure provide a device for generating a self-supervised learning model, including: an input unit configured to input the sample input data in the first sample data into the first initial model to predict the target Sample item information browsed at all times, where the sample input data includes the attribute information of the sample user, the attribute information of each sample item browsed by the sample user, and the sample item information sequence of the target period browsed by the sample user before the target time; the loss value determination unit , configured to determine the loss function value between the predicted sample item information and the target sample item information, where the target sample item information is the sample item information browsed by the sample user at the target time in the first sample data; parameter adjustment unit , is configured to adjust the relevant parameters in the first initial model based on the comparison result of the loss function value and the preset threshold, and performs training based on the adjusted parameters until the training is completed; the generation unit is configured to generate the first training completed
  • the initial model is determined to be a self-supervised learning model.
  • some embodiments of the present disclosure provide a method for generating a conversion rate prediction model, including: inputting the sample input data in the second sample data into the second initial model to obtain whether the sample user obtains the sample items on the target page.
  • the prediction results where the sample input data includes the attribute information of the sample user, the attribute information of the sample items on each target page browsed by the sample user in the target display page, and the sequence of sample item information browsed by the sample user within the set time period,
  • the second initial model is generated based on the generation method described in any implementation in the first aspect; the prediction result is compared with the target page in the second sample data.
  • the second initial model is determined as the conversion rate estimation model corresponding to the target display page.
  • some embodiments of the present disclosure provide a device for generating a conversion rate prediction model, including: an input unit configured to input the sample input data in the second sample data into the second initial model to obtain the sample user Whether to obtain the prediction results of sample items on the target page, where the sample input data includes the attribute information of the sample user, the attribute information of the sample items on each target page that the sample user has browsed in the target display page, and the sample items browsed by the sample user within the set time period.
  • the second initial model is generated based on the generation method described in any implementation manner in the first aspect;
  • the analysis unit is configured to compare the prediction result with the target page sample item in the second sample data
  • the tag information is analyzed, where the tag information is used to characterize whether the item has been acquired by the user;
  • the parameter adjustment unit is configured to adjust relevant parameters in the second initial model according to the analysis results, and train based on the adjusted parameters until the training is completed.
  • the generation unit is configured to determine the trained second initial model as a conversion rate estimation model corresponding to the target display page.
  • some embodiments of the present disclosure provide an electronic device, including: at least one processor; a storage device having at least one program stored thereon, and when the at least one program is executed by the at least one processor, the at least one process
  • the device implements the method described in any implementation manner of the first aspect and the third aspect.
  • some embodiments of the present disclosure provide a computer-readable medium on which a computer program is stored, wherein when the program is executed by a processor, the implementation described in any of the first and third aspects is implemented. Methods.
  • some embodiments of the present disclosure provide a computer program product, including a computer program. When executed by a processor, the computer program implements the method described in any implementation manner of the first aspect or the third aspect.
  • Figure 1 is a flow chart of some embodiments of a method for generating a self-supervised learning model according to the present disclosure
  • FIG2 is a schematic diagram of the structure of some embodiments of the first initial model of the present disclosure.
  • Figure 3 is a schematic structural diagram of some embodiments of a self-supervised learning model generation device according to the present disclosure
  • Figure 4 is a flow chart of some embodiments of a method for generating a conversion rate prediction model according to the present disclosure
  • Figure 5 is a structural schematic diagram of some embodiments of the second initial model of the present disclosure.
  • FIG6 is a schematic diagram of the structure of some embodiments of a device for generating a conversion rate estimation model according to the present disclosure
  • FIG. 7 is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure.
  • some embodiments of the present disclosure propose a method and device for generating a self-supervised learning model.
  • the self-supervised learning algorithm can be applied to conversion rate prediction modeling, thereby improving the accuracy of model prediction. .
  • FIG. 1 shows a process 100 of some embodiments of a method for generating a self-supervised learning model according to the present disclosure.
  • the method includes the following steps:
  • Step 101 Input the sample input data in the first sample data into the first initial model to predict the sample item information browsed at the target time.
  • the execution subject (such as a server) of the method for generating a self-supervised learning model can input the sample input data in the first sample data into the first initial model through a wired connection or a wireless connection.
  • the sample input data may include attribute information of the sample user, attribute information of each sample item browsed by the sample user, and a target period sample item information sequence browsed by the sample user before the target time.
  • the time can be any time when the sample user has browsed the sample item information before.
  • the attribute information of the sample item may include at least one of the following: item identification (such as code), price, specification model, category, etc.
  • item identification such as code
  • the sample item information sequence is usually used to represent the sample item information browsed by the user in chronological order. Among them, the time order can be positive order (that is, from far to near from the current time), or it can be reverse order (that is, from recent to far).
  • the sample item information sequence in the target period is usually the sample item information sequence browsed by the sample user before the target time.
  • the execution subject may predict the sample item information browsed by the sample user at the target time through the first initial model.
  • the sample item information here is usually the information of the sample item input into the model.
  • the first initial model can output the probability that each input sample item is browsed by the sample user at the target time.
  • the execution subject can determine the information of the sample item (such as the item identification) with the largest probability value as the predicted sample item information. If there are sample items with the same or similar probability values, the time points at which these sample items were browsed can be compared based on the sample item information sequence. Thus, the information of sample items with a relatively late time point (that is, recently browsed) is determined as predicted sample item information.
  • the first initial model can also directly output the information of the sample items browsed by the sample user at the target time.
  • the above-mentioned first sample data can be obtained by collecting historical behavior data of users.
  • it can be obtained through the following collection methods: First, the user's historical behavior data on different display pages can be obtained.
  • the display page here is usually the item information display page in different scenarios, such as the display page of the application homepage, or the display page of a store in the application, etc.
  • the historical behavior data here may include behavior data characterizing browsing item information.
  • the user can be determined as a sample user and the attribute information of the sample user can be obtained.
  • the items indicated by each item information browsed by the user can be determined as sample items, and attribute information of the sample items can be obtained.
  • sample item information in the sample item information sequence may include the browsed time point and the identification of the sample item.
  • the above set time period is usually a time period including the target time.
  • the target time is usually a certain time within a set time period.
  • the above browsed time point may be the time point when the user performs the instruction information display operation, such as the time when a certain item information displayed on the page is clicked, or it may also be the starting time of displaying the item information page, etc. .
  • the collection process of sample data can be completed by the execution subject, and then stored locally or in a database, etc.; it can also be completed by other electronic devices, and then the execution subject obtains it from it.
  • Step 102 Determine the loss function value between the predicted sample item information and the target sample item information.
  • the execution subject may determine a loss function value between the prediction result and the target sample item information.
  • the target sample item information may be the sample item information browsed by the sample user at the target time in the first sample data.
  • the loss function here can also be set according to needs. As an example, the cross-entropy loss function can be adopted.
  • self-supervised learning mainly uses auxiliary tasks to mine its own supervisory information from large-scale unsupervised data, and trains deep neural networks through this constructed supervisory information, so that it can learn valuable information for downstream tasks. characterization. That is, the supervision information of self-supervised learning is often not manually labeled.
  • the algorithm automatically constructs the supervision information in large-scale unsupervised data for supervised learning or training.
  • the "annotation" of self-supervised learning usually comes from the data itself, and its conventional operation is to improve the quality of the learned representation through respective auxiliary tasks, thereby improving the quality of downstream tasks.
  • the model can learn vector representations of various features of different user information and item information without any bias.
  • Step 103 Adjust the relevant parameters in the first initial model according to the comparison result between the loss function value and the preset threshold, and perform training based on the adjusted parameters until the training is completed.
  • the execution subject may compare and analyze the loss function value with a preset threshold. Thus, based on the comparison results between the two, it is determined whether the first initial model has been trained. become. For example, it can be determined whether the loss function value is less than a preset threshold. If it is less than, it means that the training is completed; if not, it means that the training is not completed.
  • the relevant parameters in the first initial model can be adjusted. Thus training is performed based on the adjusted parameters. That is, repeat the above steps and continue training the adjusted first initial model until the training is completed.
  • Step 104 Determine the first initial model that has been trained as a self-supervised learning model.
  • the execution subject may determine the first initial model that has been trained as a self-supervised learning model. At this point, a self-supervised learning algorithm can be obtained that can be used for pre-training of the conversion rate estimation model.
  • the first initial model can be stored locally in the execution subject or in other electronic devices.
  • the first initial model can adopt a commonly used machine learning or deep learning model structure.
  • the first initial model may include an input layer, an embedding layer, a transformation layer (Transformer Layer), a splicing layer, a fully connected layer and an output layer.
  • the execution subject can input the above sample input data, that is, the attribute information of the sample user, the attribute information of the sample items, and the sequence of sample item information in the target period, as input features into the input layer.
  • the input layer can send the received data to the embedding layer.
  • the main purpose of the embedding layer is to convert discrete features into vectors.
  • the discrete data in the sample input data such as the gender in the attribute information of the sample user, and the item identification in the sample item information sequence in the target period
  • the discrete data in the sample input data can be converted into vectors, so that the attribute information vector and sample of the sample user can be obtained.
  • the embedding layer can send the obtained embedding vector, that is, the attribute information vector of the sample user and the attribute information vector of the sample item, to the splicing layer.
  • the target period sample item information sequence vector can be sent to the conversion layer.
  • the conversion layer can perform feature extraction on the sample item information sequence vector of the target period, that is, extract the temporal features, item identification features, etc., so that the feature vector of the sample item information sequence of the target period can be obtained. Afterwards, the feature vectors can be sent to the concatenation layer. At this time, the splicing layer can splice the output of the conversion layer and the output of the embedding layer to obtain spliced features. Then, the spliced features can be sent to the fully connected layer for analysis and processing. The prediction results are thus output through the output layer.
  • the above conversion layer can use a recurrent neural network or a self-attention network.
  • recurrent neural networks often need to process data one by one or layer by layer.
  • the self-attention network can process data in parallel, that is, it can process multiple sample item information in the sequence at the same time. This helps improve the data processing efficiency and training efficiency of the model.
  • the method for generating a self-supervised learning model proposes a self-supervised learning algorithm based on a user behavior sequence (ie, a browsed sample item information sequence). Therefore, the self-supervised learning algorithm can be applied to the construction and pre-training of the conversion rate estimation model. In this way, the deviation problem caused by labels in traditional model training can be better avoided when using multiple scene data to train the model. In addition, combining data from all scenarios to train the model can improve the accuracy of the model.
  • the present disclosure provides some embodiments of a device for generating a self-supervised learning model. These device embodiments correspond to those method embodiments shown in Figure 1 , the device can be applied in various electronic devices.
  • the self-supervised learning model generation device 300 of some embodiments may include: an input unit 301 configured to input the sample input data in the first sample data into the first initial model to predict browsing at the target time.
  • sample item information where the sample input data includes the attribute information of the sample user, the attribute information of each sample item browsed by the sample user, and the target period sample item information sequence browsed by the sample user before the target time;
  • the loss value determination unit 302 Configured to determine the loss function value between the predicted sample item information and the target sample item information, where the target sample item information is the sample item information browsed by the sample user at the target time in the first sample data;
  • the parameter adjustment unit 303 is configured to adjust the relevant parameters in the first initial model based on the comparison result of the loss function value and the preset threshold, and perform training based on the adjusted parameters until the training is completed;
  • the forming unit 304 is configured to determine the first initial model that has been trained as a self-supervised learning model.
  • the first initial model may include: an input layer, used to receive sample input data and send it to the embedding layer; an embedding layer, used to convert discrete data in the sample input data into vectors to obtain the sample user's Attribute information vector, attribute information vector of sample items and target period sample item information sequence vector; conversion layer, used to perform feature extraction on the target period sample item information sequence vector to obtain the feature vector of the target period sample item information sequence; and splicing layer , used to splice the feature vector of the sample item information sequence in the target period, the attribute information vector of the sample user, and the attribute information vector of the sample item to obtain splicing features; the fully connected layer is used to analyze and process the splicing features to obtain prediction results. ; Output layer, used to output prediction results.
  • the transformation layer may employ a self-attention network.
  • the device 300 may also include a data collection unit (not shown in Figure 3), configured to obtain the user's historical behavior data on different display pages; determine the user as a sample user, and obtain the attributes of the sample user information; determine the items indicated by each item information that the user has browsed as sample items, and obtain the attribute information of the sample items; based on the order of browsing time, generate a sequence of sample item information that the sample user browsed within the set time period, where , The sample item information in the sample item information sequence includes the browsed time point and the identification of the sample item.
  • a data collection unit not shown in Figure 3
  • the units recorded in the device 300 correspond to various steps in the method described with reference to FIG. 1 . Therefore, the operations, features and beneficial effects described above for the method are also applicable to the device 300 and the units included therein, and will not be described again here.
  • FIG. 4 illustrates a process 400 of some embodiments of a method for generating a conversion rate prediction model according to the present disclosure.
  • the method includes the following steps:
  • Step 401 Input the sample input data in the second sample data into the second initial model to obtain a prediction result of whether the sample user obtains the sample item.
  • the execution subject of the generation method can also input the sample input data in the second sample data into the second initial model through a wired connection or a wireless connection.
  • the second initial model it is obtained whether the sample user obtains Prediction results for sample items on the target page.
  • the acquisition here can include at least one of the following: placing an order to purchase, adding to the shopping cart, collecting, etc.
  • the second initial model can output the probability of the sample user acquiring each sample item.
  • the execution subject can determine whether the user will obtain each sample item based on the probability value. If the probability value is greater than the set value, the item will be obtained.
  • the second initial model can also directly output predicted label information (such as yes or no) indicating whether the sample user obtains each sample item.
  • the sample input data may include the attribute information of the sample user, the attribute information of the sample items of each target page, and the sample item information sequence browsed by the sample user within a set time period.
  • the target page sample items are usually items indicated by the item information browsed by the sample user in the target display page.
  • the second initial model may be generated based on the generation method described in any implementation in the embodiment of FIG. 1. Please refer to the relevant description in step 101 of the embodiment of FIG. 1, which will not be repeated here.
  • the target display page here can be one of the different display pages mentioned above, or it can be an item information display page in other scenarios.
  • the execution subject in this embodiment may be the same as the execution subject in the embodiment of FIG. 1 , or may be different.
  • the second initial model can also be stored locally in the execution subject or in other electronic devices.
  • the model parameters of the self-supervised learning model (multi-scene data on the left in Figure 5 combined with self-supervised pre-training) can be determined as the model (single-scene data fine-tuning supervision on the right in Figure 5 training) initialization parameters.
  • the self-supervised learning model can be generated using the generation method described in any implementation manner in the embodiment of Figure 1 . That is, each model parameter value in the self-supervised learning model obtained above, such as weight values, hyperparameter values, etc. in different network layers, is used as the initial value of each model parameter of the second initial model. That is to say, after self-supervised pre-training using all scene data is completed, the parameters obtained by pre-training of all models except the output layer can be saved as initialization of the model for supervised training using labels in each different scene. parameter.
  • the output layer of the model can be adjusted to output the probability of the user obtaining the item, and the adjusted model can be used as the second initial model.
  • the main model layer remains unchanged and the output layer is changed.
  • the architecture of the second initial model is obtained. This can improve the efficiency of model construction.
  • each model Models can use data from their respective scenarios to conduct supervised training using labels of whether users have purchased. This enables training of different conversion rate prediction models for different scenarios.
  • the second sample data here can also be obtained by collecting the user's historical behavior data.
  • the user's historical behavior data on the target display page can be obtained first.
  • the user is determined as a sample user, and the attribute information of the sample user is obtained.
  • the item acquisition data in the historical behavior data the items indicated by the various item information browsed by the user are selected as the target page sample items, the attribute information of the target page sample items is obtained, and the label information is added to the target page sample items.
  • the label information can be used to characterize whether the sample item is acquired by the sample user.
  • a sample item information sequence browsed by the sample user within a set time period can be generated in the order of browsing time.
  • the sample item information in the sample item information sequence can include the browsed time point and the identification of the sample item.
  • the above-mentioned sample item information sequence can also be a target page sample item information sequence.
  • the sample items on the target page can include positive sample items and negative sample items.
  • positive sample items represent items obtained by users.
  • Negative sample items represent items that have not been acquired by the user.
  • the ratio of positive and negative sample items can be set according to the actual situation.
  • the second sample data here can be further divided into training sample data and test sample data. The above sample input data and the corresponding label information of the sample items on the target page can belong to the training sample data.
  • the second sample data in this embodiment and the first sample data in the embodiment of FIG. 1 may be data of the same sample user. And can be collected at the same time. This can improve data collection efficiency.
  • the set time periods in the two types of sample data can usually be the same time period or different time periods.
  • Step 402 Analyze the prediction results with the label information of the sample items on the target page in the second sample data.
  • the execution subject may analyze the prediction result of step 401 with the tag information of the sample items on the target page in the second sample data.
  • Tag information here It can be used to represent whether the item is obtained by the user. It can be a probability value or text information of "yes” or "no". As an example, the execution agent may determine whether the two are the same or similar.
  • Step 403 Adjust relevant parameters in the second initial model according to the analysis results, and perform training based on the adjusted parameters until the training is completed.
  • the execution subject can determine whether the second initial model has been trained based on the analysis results of step 402. For example, if the two are the same or similar, it means that the training is completed. If the two are different or the difference is large, such as the difference between the two probability values is greater than a certain value, it means that the training is not completed. At this time, the execution subject can adjust the relevant parameters in the second initial model. Thereby training is performed based on the adjusted parameters. That is, the above steps are repeated to continue training the adjusted second initial model until the training is completed.
  • Step 404 Determine the trained second initial model as the conversion rate estimation model corresponding to the target display page.
  • the execution entity may determine the trained second initial model as the conversion rate estimation model corresponding to the target display page. In other words, a conversion rate estimation model in the scenario of the target display page is obtained. At this time, the conversion rate estimation model can be used to determine the item information and/or display order to be displayed in the target display page.
  • the attribute information of the target user, the attribute information of multiple candidate items, and the historical item information sequence browsed by the target user can be input into the conversion rate prediction model, so that the prediction result of whether the target user obtains each candidate item can be obtained.
  • the information display order and display of each candidate item can be determined.
  • the execution subject can display the information of each candidate item on the page in order of probability value from high to low.
  • the execution subject can send the information display sequence to the front end, thereby controlling it to display the information of each candidate item in order.
  • Some embodiments of the present disclosure provide a method for generating a conversion rate prediction model, which proposes a conversion rate prediction model of a self-supervised learning algorithm.
  • This model can solve the problem of insufficient data volume in small traffic scenarios in multi-scenario CVR modeling without causing bias during the training process of the CVR prediction model, thus improving the overall CVR prediction accuracy.
  • this method uses data from multiple scenarios together for pre-training. During the pre-training process, the bias problem caused by using purchase tags when using multi-scene data intensively is avoided. For scenarios where the amount of data is insufficient, pre-training can be used to ensure that it is fully trained even if the sample size is small. For scenarios with sufficient data volume, pre-training can be used to further improve the model's feature extraction capabilities, thereby further improving accuracy.
  • the present disclosure provides some embodiments of a device for generating a conversion rate prediction model. These device embodiments correspond to the method embodiments shown in FIG. 4 , and the device can be specifically applied to various electronic devices.
  • the generation device 600 of some embodiments may include: an input unit 601 configured to input the sample input data in the second sample data into the second initial model to obtain a prediction of whether the sample user obtains the sample item on the target page.
  • the sample input data includes the attribute information of the sample user, the attribute information of sample items on each target page browsed by the sample user in the target display page, and the sequence of sample item information browsed by the sample user within the set time period.
  • the second The initial model is generated based on the generation method described in any implementation in the embodiment of Figure 1;
  • the analysis unit 602 is configured to analyze the prediction results with the tag information of the target page sample items in the second sample data, where , the label information is used to represent whether the item is acquired by the user;
  • the parameter adjustment unit 603 is configured to adjust the relevant parameters in the second initial model according to the analysis results, and perform training based on the adjusted parameters until the training is completed;
  • the generation unit 604, It is configured to determine the second initial model that has been trained as a conversion rate estimation model corresponding to the target display page.
  • the device 600 may also include an initial model generation unit (not shown in the figure) configured to determine the model parameters of the self-supervised learning model as initialization parameters of the model, where the self-supervised learning model adopts The generation method described in any implementation in the embodiment of Figure 1 is generated; the output layer of the model is adjusted to output the concept of the user acquiring the item. rate, using the adjusted model as the second initial model.
  • an initial model generation unit (not shown in the figure) configured to determine the model parameters of the self-supervised learning model as initialization parameters of the model, where the self-supervised learning model adopts The generation method described in any implementation in the embodiment of Figure 1 is generated; the output layer of the model is adjusted to output the concept of the user acquiring the item. rate, using the adjusted model as the second initial model.
  • the device 600 also includes a data collection unit (not shown in the figure) configured to obtain the user's historical behavior data on the target display page; determine the user as a sample user and obtain the attribute information of the sample user; According to the item acquisition data in the historical behavior data, select the items as the target page sample items from the items indicated by the item information browsed by the user, obtain the attribute information of the target page sample items, and add label information to the target page sample items.
  • a data collection unit (not shown in the figure) configured to obtain the user's historical behavior data on the target display page; determine the user as a sample user and obtain the attribute information of the sample user; According to the item acquisition data in the historical behavior data, select the items as the target page sample items from the items indicated by the item information browsed by the user, obtain the attribute information of the target page sample items, and add label information to the target page sample items.
  • the target page sample items include positive sample items and negative sample items; based on the sample user's historical browsing behavior data on different display pages, and in the order of browsing time, the sample item information browsed by the sample user within the set time period is generated Sequence, wherein the sample item information in the sample item information sequence includes the browsed time point and the identification of the sample item.
  • the device 600 further includes a display unit (not shown in the figure) configured to convert the attribute information of the target user, the attribute information of multiple candidate items, and the sequence of historical item information browsed by the target user.
  • the rate prediction model is used to obtain the prediction result of whether the target user obtains each candidate item; based on the prediction result, the information display order and display of each candidate item are determined.
  • the units recorded in the device 600 correspond to various steps in the method described with reference to FIG. 4 . Therefore, the operations, features and beneficial effects described above for the method are also applicable to the device 600 and the units included therein, and will not be described again here.
  • FIG. 7 a schematic structural diagram of an electronic device (eg, server) 700 suitable for implementing some embodiments of the present disclosure is shown.
  • the electronic device shown in FIG. 7 is only an example and should not bring any limitations to the functions and scope of use of the embodiments of the present disclosure.
  • the electronic device 700 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703.
  • a processing device e.g., a central processing unit, a graphics processing unit, etc.
  • RAM random access memory
  • various programs and data required for the operation of the electronic device 700 are also stored.
  • the processing device 701, the ROM 702, and the RAM 703 are connected to each other via a bus 704.
  • An input/output (I/O) interface 705 is also connected to the bus 704.
  • the following devices can be connected to the I/O interface 705: input devices 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 707 including, for example, a speaker, a vibrator, etc.; storage devices 708 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 709.
  • the communication device 709 can allow the electronic device 700 to communicate with other devices wirelessly or wired to exchange data.
  • FIG. 7 shows an electronic device 700 with various devices, it should be understood that it is not required to implement or have all the devices shown. More or fewer devices may be implemented or provided alternatively. Each box shown in FIG. 7 may represent one device, or may represent multiple devices as needed.
  • the processes described above with reference to the flowcharts may be implemented as a computer software program.
  • some embodiments of the present disclosure include a computer program product including a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart.
  • the computer program may be downloaded and installed from the network via communication device 709, or from storage device 708, or from ROM 702.
  • the processing device 701 the above-described functions defined in the methods of some embodiments of the present disclosure are performed.
  • the computer-readable medium recorded in some embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of computer readable storage media may include, but are not limited to: an electrical connection having at least one conductor, a portable computer disk, a hard drive, random access memory (RAM), read only memory (ROM), erasable programmable memory Read memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program therein. code. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wire, optical cable, RF (radio frequency), etc., or any suitable combination of the above.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device; it may also exist independently without being assembled into the electronic device.
  • the computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: inputs the sample input data in the first sample data into the first initial model, predicts Sample item information browsed at the target time, where the sample input data includes the attribute information of the sample user, the attribute information of each sample item browsed by the sample user, and the sample item information sequence of the target period browsed by the sample user before the target time; determine the prediction The loss function value between the sample item information and the target sample item information, where the target sample item information is the sample item information browsed by the sample user at the target time in the first sample data; according to the loss function value and the preset threshold Compare the results, adjust relevant parameters in the first initial model, and perform training based on the adjusted parameters until the training is completed; determine the first initial model that has been trained as a self-supervised learning model.
  • sample input data in the second sample data into the second initial model to obtain the prediction result of whether the sample user obtains the sample item on the target page
  • the sample input data includes the attribute information of the sample user, the sample user's information on the target display page
  • the second initial model is generated based on the generation method described in any implementation in the embodiment of Figure 1 ; Analyze the prediction results with the label information of the sample items on the target page in the second sample data, where the label information is used to characterize whether the items are acquired by the user; according to the analysis results, adjust the relevant parameters in the second initial model, based on the adjustment The subsequent parameters are trained until the training is completed; the second initial model that has been trained is determined as the conversion rate estimation model corresponding to the target display page.
  • Programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or the like. Design language.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as an Internet service provider). connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider such as an Internet service provider
  • each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains at least one operable function for implementing the specified logical function.
  • Execute instructions may also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown one after another may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved.
  • each block of the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration can be implemented by special purpose hardware-based systems that perform the specified functions or operations. , or can be implemented using a combination of specialized hardware and computer instructions.
  • the units described in some embodiments of the present disclosure may be implemented in software or hardware.
  • the described unit may also be provided in a processor.
  • a processor includes an input unit, a loss value determination unit (or analysis unit), a parameter adjustment unit and a generation unit.
  • the names of these units do not constitute a limitation on the unit itself under certain circumstances.
  • the input unit can also be described as "a unit that inputs the sample input data in the sample data into the initial model.”
  • exemplary types of hardware logic components include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on Chips (SOCs), Complex Programmable Logical device (CPLD), etc. wait.
  • FPGAs Field Programmable Gate Arrays
  • ASICs Application Specific Integrated Circuits
  • ASSPs Application Specific Standard Products
  • SOCs Systems on Chips
  • CPLD Complex Programmable Logical device
  • Some embodiments of the present disclosure also provide a computer program product, including a computer program.
  • the computer program When executed by a processor, the computer program implements any of the above methods for generating a self-supervised learning model or a method for generating a conversion rate prediction model. .

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Data Mining & Analysis (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Game Theory and Decision Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Technology Law (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本公开的实施例公开了自监督学习模型的生成方法和转化率预估模型的生成方法。该生成方法的一具体实施方式包括：将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息；确定预测的样本物品信息与目标样本物品信息之间的损失函数值；根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；将训练完成的第一初始模型确定为自监督学习模型。

Description

自监督学习模型的生成方法和转化率预估模型的生成方法
相关申请的交叉引用
本申请要求于申请日为2022年09月19日提交的，申请号为202211138878.2、发明名称为“自监督学习模型的生成方法和转化率预估模型的生成方法”的中国专利申请的优先权，其全部内容作为整体并入本申请中。
技术领域
本公开的实施例涉及人工智能技术领域，具体涉及自监督学习模型的生成方法和转化率预估模型的生成方法。
背景技术
转化率(Conversion Rate，CVR)最初是一种衡量广告投放效果的指标，即计算通过点击广告并对其付费的用户，在所有浏览过广告的用户中所占的比例。随着相关技术不断地发展，CVR也通常用于广告投放、商品推荐和搜索等效果的评估。转化率预估技术一般是利用人工智能算法，对用户每次对在线商品、广告等的购买转化的情况进行建模，并作出预测，判断用户是否购买或者给出购买的概率值。在电子商务系统中，通常会给用户展示对该用户来说CVR较高的广告、商品等。由此可见，CVR预估往往是一项重要的技术。
发明内容
本公开的内容部分用于以简要的形式介绍构思，这些构思将在后面的具体实施方式部分被详细描述。本公开的内容部分并不旨在标识要求保护的技术方案的关键特征或必要特征，也不旨在用于限制所要求的保护的技术方案的范围。本公开的一些实施例提出了自监督学习模型的生成方法、转化率预估模型的生成方法、装置、电子设备和计 算机可读介质，来解决以上背景技术部分提到的技术问题中的一项或多项。
第一方面，本公开的一些实施例提供了一种自监督学习模型的生成方法，包括：将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，样本输入数据包括样本用户的属性信息、样本用户浏览过的各样本物品的属性信息、样本用户在目标时刻之前浏览的目标时段样本物品信息序列；确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，目标样本物品信息为第一样本数据中，样本用户在目标时刻浏览的样本物品信息；根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；将训练完成的第一初始模型确定为自监督学习模型。
第二方面，本公开的一些实施例提供了一种自监督学习模型的生成装置，包括：输入单元，被配置成将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，样本输入数据包括样本用户的属性信息、样本用户浏览过的各样本物品的属性信息、样本用户在目标时间之前浏览的目标时段样本物品信息序列；损失值确定单元，被配置成确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，目标样本物品信息为第一样本数据中，样本用户在目标时间浏览的样本物品信息；参数调整单元，被配置成根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；生成单元，被配置成将训练完成的第一初始模型确定为自监督学习模型。
第三方面，本公开的一些实施例提供了一种转化率预估模型的生成方法，包括：将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，样本输入数据包括样本用户的属性信息、样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、样本用户在设定时间段内浏览的样本物品信息序列，第二初始模型是基于第一方面中任一实现方式所描述的生成方法而生成的；将预测结果与第二样本数据中，目标页面 样本物品的标签信息进行分析，其中，标签信息用于表征物品是否被用户获取；根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。
第四方面，本公开的一些实施例提供了一种转化率预估模型的生成装置，包括：输入单元，被配置成将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，样本输入数据包括样本用户的属性信息、样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、样本用户在设定时间段内浏览的样本物品信息序列，第二初始模型是基于第一方面中任一实现方式所描述的生成方法而生成的；分析单元，被配置成将预测结果与第二样本数据中，目标页面样本物品的标签信息进行分析，其中，标签信息用于表征物品是否被用户获取；参数调整单元，被配置成根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；生成单元，被配置成将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。
第五方面，本公开的一些实施例提供了一种电子设备，包括：至少一个处理器；存储装置，其上存储有至少一个程序，当至少一个程序被至少一个处理器执行，使得至少一个处理器实现上述第一方面、第三方面中任一实现方式所描述的方法。
第六方面，本公开的一些实施例提供了一种计算机可读介质，其上存储有计算机程序，其中，程序被处理器执行时实现上述第一方面、第三方面中任一实现方式所描述的方法。
第七方面，本公开的一些实施例提供了一种计算机程序产品，包括计算机程序，计算机程序在被处理器执行时实现上述第一方面、第三方面中任一实现方式所描述的方法。
附图说明
结合附图并参考以下具体实施方式，本公开各实施例的上述和其他特征、优点及方面将变得更加明显。贯穿附图中，相同或相似的附 图标记表示相同或相似的元素。应当理解附图是示意性的，元件和元素不一定按照比例绘制。
图1是根据本公开的自监督学习模型的生成方法的一些实施例的流程图；
图2是本公开的第一初始模型的一些实施例的结构示意图；
图3是根据本公开的自监督学习模型的生成装置的一些实施例的结构示意图；
图4是根据本公开的转化率预估模型的生成方法的一些实施例的流程图；
图5是本公开的第二初始模型的一些实施例的结构示意图；
图6是根据本公开的转化率预估模型的生成装置的一些实施例的结构示意图；
图7是适于用来实现本公开的一些实施例的电子设备的结构示意图。
具体实施方式
下面将参照附图更详细地描述本公开的实施例。虽然附图中显示了本公开的某些实施例，然而应当理解的是，本公开可以通过各种形式来实现，而且不应该被解释为限于这里阐述的实施例。相反，提供这些实施例是为了更加透彻和完整地理解本公开。应当理解的是，本公开的附图及实施例仅用于示例性作用，并非用于限制本公开的保护范围。
另外还需要说明的是，为了便于描述，附图中仅示出了与有关发明相关的部分。在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互组合。
需要注意，本公开中提及的“第一”、“第二”等概念仅用于对不同的装置、模块或单元进行区分，并非用于限定这些装置、模块或单元所执行的功能的顺序或者相互依存关系。
需要注意，本公开中提及的“一个”、“多个”的修饰是示意性而非限制性的，本领域技术人员应当理解，除非在上下文另有明确指出， 否则应该理解为“至少一个”。
本公开实施方式中的多个装置之间所交互的消息或者信息的名称仅用于说明性的目的，而并不是用于对这些消息或信息的范围进行限制。
本公开中所涉及的用户个人信息的收集、存储、使用等操作，例如用户的属性信息、浏览访问操作行为、下单数据等，在执行相应操作之前，相关组织或个人尽到包括开展个人信息安全影响评估、向个人信息主体履行告知义务、事先征得个人信息主体的授权同意等义务。
相关的自监督学习模型的生成方法，例如，在CVR预估建模过程中，首先，一般是获取到用户对商品的购买转化行为。从而将用户是否购买作为标签，将商品和用户的属性等作为特征，构建数据集。然后，利用这些数据集，对CVR预估模型进行训练，从而使模型能判断每一个用户，针对不同的商品是否会进行购买等经常会存在如下技术问题：不同场景的数据分布通常不同。例如在应用主页的搜索场景中，某用户没有购买某商品。但是在店铺页面的搜索场景中，该用户可能购买了该商品。因此，若将所有场景的数据合并，来联合训练模型通常会带来较大的偏差，影响模型预估的准确度。
为了解决以上所阐述的问题，本公开的一些实施例提出了自监督学习模型的生成方法及装置，可以将自监督学习算法应用在转化率预估建模上，从而提升模型预估的准确度。
下面将参考附图并结合实施例来详细说明本公开。
图1示出了根据本公开的自监督学习模型的生成方法的一些实施例的流程100。该方法包括以下步骤：
步骤101，将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息。
在一些实施例中，自监督学习模型的生成方法的执行主体(例如服务器)可以通过有线连接方式或无线连接方式，将第一样本数据中的样本输入数据输入第一初始模型。其中，样本输入数据可以包括样本用户的属性信息、样本用户浏览过的各样本物品的属性信息、样本用户在目标时刻之前浏览的目标时段样本物品信息序列。这里的目标 时刻可以是样本用户在此前浏览过样本物品信息的任意时刻。
在这里，样本用户的属性信息和样本物品的属性信息的内容不限制。样本物品的属性信息可以包括以下至少一项：物品标识(如编码)、价格、规格型号、所属类别等。样本物品信息序列通常用于表示按照时间顺序用户浏览过的样本物品信息。其中，时间顺序可以是正序(即距离当前时间由远至近)，也可以是倒序(即由近到远)。而目标时段样本物品信息序列通常为样本用户，在目标时刻之前所浏览的样本物品信息序列。
在一些实施例中，执行主体可以通过第一初始模型，来预测样本用户在目标时刻浏览的样本物品信息。这里的样本物品信息通常是输入模型中的样本物品的信息。
可以理解的是，第一初始模型可以输出所输入的各样本物品，在目标时刻被样本用户浏览的概率。此时，执行主体可以将概率值最大的样本物品的信息(如物品标识)，确定为预测的样本物品信息。若存在概率值相同或相近的样本物品，则可以根据样本物品信息序列，来比较这些样本物品被浏览的时间点。从而将时间点相对晚(即最近被浏览)的样本物品的信息，确定为预测的样本物品信息。可选地，第一初始模型也可以直接输出，在目标时刻被样本用户浏览的样本物品的信息。
在一些实施例中，上述第一样本数据可以通过收集用户的历史行为数据而得到的。作为示例，可以通过以下采集方法得到：首先，可以获取用户在不同展示页面的历史行为数据。这里的展示页面通常为不同场景下的物品信息展示页面，如应用主页的展示页面，或者应用中某个店铺展示页面等。这里的历史行为数据可以包括表征浏览物品信息的行为数据。此时，可以将用户确定为样本用户，得到样本用户的属性信息。接着，可以将用户浏览过的各物品信息所指示的物品确定为样本物品，得到样本物品的属性信息。之后，根据浏览时间的先后顺序，生成样本用户在设定时间段(如24小时)内浏览的样本物品信息序列。其中，样本物品信息序列中的样本物品信息可以包括被浏览的时间点和样本物品的标识。
可以理解的是，上述设定时间段通常为包含目标时刻的时间段。也就是说，目标时刻通常为设定时间段内的某时刻。另外，上述被浏览的时间点可以是用户执行指示信息展示操作的时间点，如点击页面中所展示的某个物品信息所在位置的时间，或者，也可以是显示物品信息页面的起始时间等。样本数据的采集过程可以是由执行主体完成的，进而存储在本地或数据库中等；也可以是由其他电子设备完成的，进而执行主体从中获取到。
步骤102，确定预测的样本物品信息与目标样本物品信息之间的损失函数值。
在一些实施例中，基于步骤101预测的样本物品信息，执行主体可以确定预测结果与目标样本物品信息之间的损失函数值。其中，目标样本物品信息可以为第一样本数据中，样本用户在目标时刻浏览的样本物品信息。这里的损失函数同样可以根据需求进行设置。作为示例，可以采用交叉熵损失函数。
可以理解的是，自监督学习主要是利用辅助任务从大规模的无监督数据中挖掘自身的监督信息，通过这种构造的监督信息对深度神经网络进行训练，从而可以学习到对下游任务有价值的表征。即自监督学习的监督信息往往不是人工标注的，通常是算法在大规模无监督数据中自动构造监督信息，来进行监督学习或训练。自监督学习的“标注”通常来自于数据本身，其常规操作是通过各自的辅助任务来提高学习表征的质量，从而提高下游任务的质量。
由此可见，在本公开的实施例中，由于没有直接使用表征用户是否购买的标签，而是利用序列数据自身的时序性质。即利用t-1及之前时刻的信息序列，来预测t时刻用户的行为。因此，在使用多场景数据进行模型训练时，可以使模型在没有任何偏差的情况下，学习到不同用户信息、物品信息的各类特征的向量表征。
步骤103，根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成。
在一些实施例中，执行主体可以将损失函数值与预设阈值进行比较分析。从而根据两者的比较结果，来确定第一初始模型是否训练完 成。例如，可以确定损失函数值是否小于预设阈值。若小于，说明训练完成；若不小于，则可以说明未训练完成。在这里，如果确定未训练完成，可以调整第一初始模型中的相关参数。从而基于调整后的参数进行训练。即重复上述步骤，继续训练调整后的第一初始模型，直至训练完成。
步骤104，将训练完成的第一初始模型确定为自监督学习模型。
在一些实施例中，若确定第一初始模型训练完成，则执行主体可以将训练完成的第一初始模型，确定为自监督学习模型。此时，可以得到一种可用于转化率预估模型预训练的自监督学习算法。
需要说明的是，第一初始模型可以存储在执行主体本地，也可以存储在其他电子设备中。第一初始模型可以采用常用的机器学习或深度学习模型结构。
在一些实施例中，如图2所示，第一初始模型可以包括输入层、嵌入层、转换层(Transformer Layer)、拼接层、全连接层和输出层。从图2中可以看出，执行主体可以将上述的样本输入数据，即样本用户的属性信息、样本物品的属性信息和目标时段样本物品信息序列，作为输入特征，输入到输入层。输入层可以将接收的数据发送给嵌入层。嵌入层的主要目的是把离散特征转换为向量。即可以将样本输入数据中的离散数据，如样本用户的属性信息中的性别，以及目标时段样本物品信息序列中的物品标识等特征，转换为向量，从而可以得到样本用户的属性信息向量、样本物品的属性信息向量和目标时段样本物品信息序列向量。之后，嵌入层可以将得到的嵌入向量，即样本用户的属性信息向量和样本物品的属性信息向量，发送给拼接层。同时，可以将目标时段样本物品信息序列向量发送给转换层。
在这里，转换层可以对目标时段样本物品信息序列向量进行特征提取，即提取其中的时序特征、物品标识特征等，从而可以得到目标时段样本物品信息序列的特征向量。之后，可以将特征向量发送给拼接层。此时，拼接层可以将转换层的输出与嵌入层的输出进行拼接，从而得到拼接特征。接着，可以将拼接特征送入全连接层，进行分析处理。从而将预测结果通过输出层进行输出。
需要说明的是，上述转换层可以采用循环神经网络，也可以采用自注意力网络。其中，循环神经网络往往需要对数据进行逐个或逐层处理。而自注意力网络可以进行数据的并行处理，即可以同时处理序列中的多个样本物品信息。这样有助于提高模型的数据处理效率和训练效率。
可以理解的是，现有的自监督学习方法，大多是运用于自然语言处理和计算机视觉领域。因为该领域的数据都是非结构化数据，其数据本身存在丰富的上下文信息或时序信息。而在推荐搜索等领域，数据大多为结构化表格数据，可利用的数据本身的上下文和时序信息较少，难以直接将自监督学习方法应用过来。
而本公开的一些实施例提供的自监督学习模型的生成方法，提出了一种基于用户行为序列(即浏览的样本物品信息序列)的自监督学习算法。从而可以将自监督学习算法，应用于转化率预估模型的构建和预训练中。这样，可以较好地避免传统模型训练中，使用多种场景数据训练模型，由标签带来的偏差问题。另外，联合所有场景中的数据来训练模型，可以提升模型的准确度。
继续参考图3，作为对上述图1所示方法的实现，本公开提供了一种自监督学习模型的生成装置的一些实施例，这些装置实施例与图1所示的那些方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图3所示，一些实施例的自监督学习模型的生成装置300可以包括：输入单元301，被配置成将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，样本输入数据包括样本用户的属性信息、样本用户浏览过的各样本物品的属性信息、样本用户在目标时间之前浏览的目标时段样本物品信息序列；损失值确定单元302，被配置成确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，目标样本物品信息为第一样本数据中，样本用户在目标时间浏览的样本物品信息；参数调整单元303，被配置成根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；生 成单元304，被配置成将训练完成的第一初始模型确定为自监督学习模型。
在一些实施例中，第一初始模型可以包括：输入层，用于接收样本输入数据，以及发送给嵌入层；嵌入层，用于将样本输入数据中的离散数据转换为向量，得到样本用户的属性信息向量、样本物品的属性信息向量和目标时段样本物品信息序列向量；转换层，用于对目标时段样本物品信息序列向量进行特征提取，得到目标时段样本物品信息序列的特征向量；以及拼接层，用于对目标时段样本物品信息序列的特征向量、样本用户的属性信息向量、样本物品的属性信息向量进行拼接，得到拼接特征；全连接层，用于对拼接特征进行分析处理，得到预测结果；输出层，用于输出预测结果。
在一些实施例中，转换层可以采用自注意力网络。
在一些实施例中，该装置300还可以包括数据采集单元(图3中未示出)，被配置成获取用户在不同展示页面的历史行为数据；将用户确定为样本用户，得到样本用户的属性信息；将用户浏览过的各物品信息所指示的物品确定为样本物品，得到样本物品的属性信息；根据浏览时间的先后顺序，生成样本用户在设定时间段内浏览的样本物品信息序列，其中，样本物品信息序列中的样本物品信息包括被浏览的时间点和样本物品的标识。
可以理解的是，该装置300中记载的诸单元与参考图1描述的方法中的各个步骤相对应。由此，上文针对方法描述的操作、特征以及产生的有益效果同样适用于装置300及其中包含的单元，在此不再赘述。
请参考图4，示出了根据本公开的转化率预估模型的生成方法的一些实施例的流程400。该方法包括以下步骤：
步骤401，将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取样本物品的预测结果。
在一些实施例中，生成方法的执行主体(例如服务器)同样可以通过有线连接或无线连接的方式，将第二样本数据中的样本输入数据输入第二初始模型。从而通过第二初始模型，得到样本用户是否获取 目标页面样本物品的预测结果。这里的获取可以包括以下至少一种：下单购买、加购物车、收藏等。
同样地，第二初始模型可以输出样本用户获取各样本物品的概率。此时，执行主体可以根据概率值，确定用户是否会获取各样本物品。如概率值大于设定值，说明会获取该物品。可选地，第二初始模型也可以直接输出表征样本用户是否获取各样本物品的预测标签信息(如是或否)。
其中，样本输入数据可以包括样本用户的属性信息、各目标页面样本物品的属性信息、样本用户在设定时间段内浏览的样本物品信息序列。其中，目标页面样本物品通常为样本用户，在目标展示页面中浏览过的物品信息所指示的物品。而第二初始模型可以是基于图1实施例中，任一实现方式所描述的生成方法而生成的。可以参见图1实施例步骤101中的相关描述，此处不再赘述。
这里的目标展示页面可以是上述不同展示页面中的一种，也可以是其他场景下的物品信息展示页面。而本实施例中的执行主体与图1实施例中的执行主体可以相同，也可以不同。
在一些实施例中，第二初始模型同样可以存储在执行主体本地，或者其他电子设备中。在这里，如图5所示，可以将自监督学习模型(图5中左侧的多场景数据联合自监督预训练)的模型参数，确定为模型(图5中右侧的单场景数据微调监督训练)的初始化参数。其中，自监督学习模型可以采用图1实施例中任一实现方式所描述的生成方法生成。即利用上述得到的自监督学习模型中的各模型参数值，如不同网络层中的权重值、超参数值等，作为第二初始模型的各模型参数的初始值。也就是说，在使用所有场景数据自监督预训练完成之后，可以将除输出层之外的所有模型预训练得到的参数保存，从而作为在每个不同场景下，使用标签进行监督训练的模型初始化参数。
之后，可以将模型的输出层调整为输出用户获取物品的概率，将调整后的模型作为第二初始模型。也就是说，主要的模型层不变，改变输出层。通过微调自监督学习模型的架构，来得到第二初始模型的架构。这样可以提高模型的构建效率。之后，针对不同场景，每个模 型可以使用各自场景的数据，利用用户是否购买的标签进行有监督训练。从而实现为不同场景训练不同的转化率预估模型。
在一些应用场景中，这里的第二样本数据同样可以通过收集用户的历史行为数据得到。作为示例，首先可以获取用户在目标展示页面的历史行为数据。此时，将用户确定为样本用户，得到样本用户的属性信息。接着，根据历史行为数据中的物品获取数据，从用户浏览过的各物品信息所指示的物品中，选取物品作为目标页面样本物品，得到目标页面样本物品的属性信息，以及给目标页面样本物品添加标签信息。其中，标签信息可以用于表征样本物品是否被样本用户获取。之后，根据样本用户在不同展示页面的历史浏览行为数据，可以按照浏览时间的先后顺序，生成样本用户在设定时间段内浏览的样本物品信息序列。其中，样本物品信息序列中的样本物品信息，可以包括被浏览的时间点和样本物品的标识。可选地，在目标展示页面的历史行为数据数据量充足的情况下，上述样本物品信息序列也可以为目标页面样本物品信息序列。
需要说明的是，为了保证数据的多样性，提升模型的准确度，这里的目标页面样本物品可以包括正样本物品和负样本物品。其中，正样本物品表征被用户获取的物品。而负样本物品表征没有被用户获取的物品。正负样本物品的比例可以根据实际情况进行设置。为了进一步提升模型的准确度，这里的第二样本数据可以进一步分为训练样本数据和测试样本数据。上述样本输入数据和对应的目标页面样本物品的标签信息，可以属于训练样本数据。
另外，本实施例中的第二样本数据与图1实施例中的第一样本数据，可以为同一样本用户的数据。并且可以同时采集得到。这样可以提高数据的采集效率。此外，两种样本数据中的设定时间段通常可以为同一时间段，也可以为不同的时间段。
步骤402，将预测结果与第二样本数据中，目标页面样本物品的标签信息进行分析。
在一些实施例中，执行主体可以将步骤401的预测结果，与第二样本数据中目标页面样本物品的标签信息进行分析。这里的标签信息 可以用于表征物品是否被用户获取，其可以为概率值，也可以为“是”或“否”的文字信息。作为示例，执行主体可以确定两者是否相同或相近。
步骤403，根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成。
在一些实施例中，执行主体可以根据步骤402的分析结果，来确定第二初始模型是否训练完成。例如，若两者相同或相近，说明训练完成。若两者不同或差距较大，如两个概率值的差值大于某值，说明未训练完成。此时，执行主体可以调整第二初始模型中的相关参数。从而基于调整后的参数进行训练。即重复上述步骤，对调整后的第二初始模型继续训练，直至训练完成。
步骤404，将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。
在一些实施例中，若确定训练完成，执行主体可以将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。也就是说，得到了在目标展示页面这种场景下的转化率预估模型。此时，可以利用该转化率预估模型，来确定在目标展示页面中，所要显示的物品信息和/或显示顺序等等。
作为示例，可以将目标用户的属性信息、多个候选物品的属性信息、目标用户浏览过的历史物品信息序列输入到转化率预估模型，从而可以得到目标用户是否获取各候选物品的预测结果。此时，根据预测结果，可以确定各候选物品的信息显示顺序及显示。例如，执行主体可以按照概率值由高到低的顺序，在页面中显示各候选物品的信息。又例如，执行主体可以将信息显示顺序发送给前端，从而控制其按序显示各候选物品的信息。
可以理解的是，为了解决背景技术中所描述的偏差，现有方法中通常是将不同场景的数据划分开。对于某个场景，只利用该场景下的数据单独训练一个模型。这种方式在一定程度上能缓解偏差问题。但不同场景的流量不同，数据量也不尽相同。例如，电商应用主页中的搜索数据量往往是店铺页面中的搜索数据量的几十至几百倍，这就导 致了在单独为店铺页面搜索场景训练CVR模型时，存在数据量较小，模型训练不充分的情况。进而影响模型预估的准确度。
而本公开的一些实施例提供的转化率预估模型的生成方法，提出了一种自监督学习算法的转化率预估模型。该模型能够在CVR预估模型的训练过程中，在不产生偏差的情况下，解决多场景CVR建模中，小流量场景数据量不足的问题，从而提升整体的CVR预估准确度。也就是说，该方法将多场景的数据集中使用，进行预训练。预训练的过程中，避免了在集中使用多场景数据时，使用购买标签而带来的偏差问题。对于数据量不足的场景，可以通过预训练，保证其在样本量较小的情况下，也得到充分的训练。对于数据量充足的场景，可以通过预训练进一步提升模型特征提取的能力，从而进一步提升准确率。
参考图6，作为对上述图4所示方法的实现，本公开提供了一种转化率预估模型的生成装置的一些实施例，这些装置实施例与图4所示的那些方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图6所示，一些实施例的生成装置600可以包括：输入单元601，被配置成将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，样本输入数据包括样本用户的属性信息、样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、样本用户在设定时间段内浏览的样本物品信息序列，第二初始模型是基于图1实施例中任一实现方式所描述的生成方法而生成的；分析单元602，被配置成将预测结果与第二样本数据中，目标页面样本物品的标签信息进行分析，其中，标签信息用于表征物品是否被用户获取；参数调整单元603，被配置成根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；生成单元604，被配置成将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。
在一些实施例中，该装置600还可以包括初始模型生成单元(图中未示出)，被配置成将自监督学习模型的模型参数，确定为模型的初始化参数，其中，自监督学习模型采用图1实施例中任一实现方式所描述的生成方法生成；将模型的输出层调整为输出用户获取物品的概 率，将调整后的模型作为第二初始模型。
在一些实施例中，该装置600还包括数据采集单元(图中未示出)，被配置成获取用户在目标展示页面的历史行为数据；将用户确定为样本用户，得到样本用户的属性信息；根据历史行为数据中的物品获取数据，从用户浏览过的各物品信息所指示的物品中，选取物品作为目标页面样本物品，得到目标页面样本物品的属性信息，以及给目标页面样本物品添加标签信息，其中，目标页面样本物品包括正样本物品和负样本物品；根据样本用户在不同展示页面的历史浏览行为数据，按照浏览时间的先后顺序，生成样本用户在设定时间段内浏览的样本物品信息序列，其中，样本物品信息序列中的样本物品信息包括被浏览的时间点和样本物品的标识。
在一些实施例中，该装置600还包括显示单元(图中未示出)，被配置成将目标用户的属性信息、多个候选物品的属性信息、目标用户浏览过的历史物品信息序列输入转化率预估模型，得到目标用户是否获取各候选物品的预测结果；根据预测结果，确定各候选物品的信息显示顺序及显示。
可以理解的是，该装置600中记载的诸单元与参考图4描述的方法中的各个步骤相对应。由此，上文针对方法描述的操作、特征以及产生的有益效果同样适用于装置600及其中包含的单元，在此不再赘述。
下面参考图7，其示出了适于用来实现本公开的一些实施例的电子设备(例如服务器)700的结构示意图。图7示出的电子设备仅仅是一个示例，不应对本公开的实施例的功能和使用范围带来任何限制。
如图7所示，电子设备700可以包括处理装置(例如中央处理器、图形处理器等)701，其可以根据存储在只读存储器(ROM)702中的程序或者从存储装置708加载到随机访问存储器(RAM)703中的程序而执行各种适当的动作和处理。在RAM 703中，还存储有电子设备700操作所需的各种程序和数据。处理装置701、ROM 702以及RAM 703通过总线704彼此相连。输入/输出(I/O)接口705也连接至总线704。
通常，以下装置可以连接至I/O接口705：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置706；包括例如扬声器、振动器等的输出装置707；包括例如磁带、硬盘等的存储装置708；以及通信装置709。通信装置709可以允许电子设备700与其他设备进行无线或有线通信以交换数据。虽然图7示出了具有各种装置的电子设备700，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。图7中示出的每个方框可以代表一个装置，也可以根据需要代表多个装置。
特别地，根据本公开的一些实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的一些实施例包括一种计算机程序产品，其包括承载在计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的一些实施例中，该计算机程序可以通过通信装置709从网络上被下载和安装，或者从存储装置708被安装，或者从ROM 702被安装。在该计算机程序被处理装置701执行时，执行本公开的一些实施例的方法中限定的上述功能。
需要说明的是，本公开的一些实施例中记载的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有至少一个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开的一些实施例中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开的一些实施例中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序 代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，样本输入数据包括样本用户的属性信息、样本用户浏览过的各样本物品的属性信息、样本用户在目标时刻之前浏览的目标时段样本物品信息序列；确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，目标样本物品信息为第一样本数据中，样本用户在目标时刻浏览的样本物品信息；根据损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；将训练完成的第一初始模型确定为自监督学习模型。
或者，将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，样本输入数据包括样本用户的属性信息、样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、样本用户在设定时间段内浏览的样本物品信息序列，第二初始模型是基于图1实施例中任一实现方式所描述的生成方法而生成的；将预测结果与第二样本数据中，目标页面样本物品的标签信息进行分析，其中，标签信息用于表征物品是否被用户获取；根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；将训练完成的第二初始模型，确定为与目标展示页面对应的转化率预估模型。
此外，可以以一种或多种程序设计语言或其组合来编写用于执行 本公开的一些实施例的操作的计算机程序代码，程序设计语言包括面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言—诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)——连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含至少一个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开的一些实施例中的单元可以通过软件的方式实现，也可以通过硬件的方式来实现。所描述的单元也可以设置在处理器中，例如，可以描述为：一种处理器包括输入单元、损失值确定单元(或分析单元)、参数调整单元和生成单元。其中，这些单元的名称在某种情况下并不构成对该单元本身的限定，例如，输入单元还可以被描述为“将样本数据中的样本输入数据输入初始模型的单元”。
本文中以上描述的功能可以至少部分地由至少一个硬件逻辑部件来执行。例如，非限制性地，可以使用的示范类型的硬件逻辑部件包括：现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、片上系统(SOC)、复杂可编程逻辑设备(CPLD)等 等。
本公开的一些实施例还提供一种计算机程序产品，包括计算机程序，计算机程序在被处理器执行时，实现上述的任一种自监督学习模型的生成方法，或转化率预估模型的生成方法。
以上描述仅为本公开的一些较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开的实施例中所涉及的发明范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述发明构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开的实施例中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。

Claims (13)

  1. 一种自监督学习模型的生成方法，包括：
    将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，所述样本输入数据包括样本用户的属性信息、所述样本用户浏览过的各样本物品的属性信息、所述样本用户在所述目标时刻之前浏览的目标时段样本物品信息序列；
    确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，所述目标样本物品信息为所述第一样本数据中，所述样本用户在所述目标时刻浏览的样本物品信息；
    根据所述损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；
    将训练完成的第一初始模型确定为自监督学习模型。
  2. 根据权利要求1所述的方法，其中，所述第一初始模型包括：
    输入层，用于接收所述样本输入数据，以及发送给嵌入层；
    嵌入层，用于将所述样本输入数据中的离散数据转换为向量，得到样本用户的属性信息向量、样本物品的属性信息向量和目标时段样本物品信息序列向量；
    转换层，用于对所述目标时段样本物品信息序列向量进行特征提取，得到所述目标时段样本物品信息序列的特征向量；以及
    拼接层，用于对所述目标时段样本物品信息序列的特征向量、所述样本用户的属性信息向量、所述样本物品的属性信息向量进行拼接，得到拼接特征；
    全连接层，用于对所述拼接特征进行分析处理，得到预测结果；
    输出层，用于输出所述预测结果。
  3. 根据权利要求2所述的方法，其中，所述转换层采用自注意力网络。
  4. 根据权利要求1-3之一所述的方法，其中，所述第一样本数据通过以下采集方法得到：
    获取用户在不同展示页面的历史行为数据；
    将所述用户确定为样本用户，得到样本用户的属性信息；
    将所述用户浏览过的各物品信息所指示的物品确定为样本物品，得到样本物品的属性信息；
    根据浏览时间的先后顺序，生成所述样本用户在设定时间段内浏览的样本物品信息序列，其中，所述样本物品信息序列中的样本物品信息包括被浏览的时间点和样本物品的标识。
  5. 一种自监督学习模型的生成装置，包括：
    输入单元，被配置成将第一样本数据中的样本输入数据输入第一初始模型，预测在目标时刻浏览的样本物品信息，其中，所述样本输入数据包括样本用户的属性信息、所述样本用户浏览过的各样本物品的属性信息、所述样本用户在所述目标时间之前浏览的目标时段样本物品信息序列；
    损失值确定单元，被配置成确定预测的样本物品信息与目标样本物品信息之间的损失函数值，其中，所述目标样本物品信息为所述第一样本数据中，所述样本用户在所述目标时刻浏览的样本物品信息；
    参数调整单元，被配置成根据所述损失函数值与预设阈值的比较结果，调整第一初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；
    生成单元，被配置成将训练完成的第一初始模型确定为自监督学习模型。
  6. 一种转化率预估模型的生成方法，包括：
    将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，所述样本输入数据包括所述样本用户的属性信息、所述样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、所述样本用户在设定时间段 内浏览的样本物品信息序列，所述第二初始模型是基于如权利要求1-4之一所述的生成方法而生成的；
    将所述预测结果与所述第二样本数据中，所述目标页面样本物品的标签信息进行分析，其中，所述标签信息用于表征物品是否被用户获取；
    根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；
    将训练完成的第二初始模型，确定为与所述目标展示页面对应的转化率预估模型。
  7. 根据权利要求6所述的方法，其中，所述第二初始模型是基于如权利要求1-4之一所述的生成方法而生成的，包括：
    将自监督学习模型的模型参数，确定为模型的初始化参数，其中，所述自监督学习模型采用如权利要求1-4之一所述的生成方法生成；
    将模型的输出层调整为输出用户获取物品的概率，将调整后的模型作为第二初始模型。
  8. 根据权利要求6或7所述的方法，其中，所述第二样本数据通过以下采集方法得到：
    获取用户在目标展示页面的历史行为数据；
    将所述用户确定为样本用户，得到样本用户的属性信息；
    根据所述历史行为数据中的物品获取数据，从所述用户浏览过的各物品信息所指示的物品中，选取物品作为目标页面样本物品，得到目标页面样本物品的属性信息，以及给目标页面样本物品添加标签信息，其中，所述目标页面样本物品包括正样本物品和负样本物品；
    根据所述样本用户在不同展示页面的历史浏览行为数据，按照浏览时间的先后顺序，生成所述样本用户在设定时间段内浏览的样本物品信息序列，其中，所述样本物品信息序列中的样本物品信息包括被浏览的时间点和样本物品的标识。
  9. 根据权利要求6-8之一所述的方法，所述方法还包括：
    将目标用户的属性信息、多个候选物品的属性信息、所述目标用户浏览过的历史物品信息序列输入所述转化率预估模型，得到所述目标用户是否获取各候选物品的预测结果；
    根据预测结果，确定各候选物品的信息显示顺序及显示。
  10. 一种转化率预估模型的生成装置，包括：
    输入单元，被配置成将第二样本数据中的样本输入数据输入第二初始模型，得到样本用户是否获取目标页面样本物品的预测结果，其中，所述样本输入数据包括所述样本用户的属性信息、所述样本用户在目标展示页面中浏览过的各目标页面样本物品的属性信息、所述样本用户在设定时间段内浏览的样本物品信息序列，所述第二初始模型是基于如权利要求1-4之一所述的生成方法而生成的；
    分析单元，被配置成将所述预测结果与所述第二样本数据中，所述目标页面样本物品的标签信息进行分析，其中，所述标签信息用于表征物品是否被用户获取；
    参数调整单元，被配置成根据分析结果，调整第二初始模型中的相关参数，基于调整后的参数进行训练，直至训练完成；
    生成单元，被配置成将训练完成的第二初始模型，确定为与所述目标展示页面对应的转化率预估模型。
  11. 一种电子设备，包括：
    至少一个处理器；
    存储装置，其上存储有至少一个程序，
    当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现如权利要求1-4、6-9中任一所述的方法。
  12. 一种计算机可读介质，其上存储有计算机程序，其中，所述程序被处理器执行时实现如权利要求1-4、6-9中任一所述的方法。
  13. 一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现根据权利要求1-4、6-9中任一项所述的方法。
PCT/CN2023/087290 2022-09-19 2023-04-10 自监督学习模型的生成方法和转化率预估模型的生成方法 Ceased WO2024060587A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211138878.2 2022-09-19
CN202211138878.2A CN117764633A (zh) 2022-09-19 2022-09-19 自监督学习模型的生成方法和转化率预估模型的生成方法

Publications (1)

Publication Number Publication Date
WO2024060587A1 true WO2024060587A1 (zh) 2024-03-28

Family

ID=90316659

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/087290 Ceased WO2024060587A1 (zh) 2022-09-19 2023-04-10 自监督学习模型的生成方法和转化率预估模型的生成方法

Country Status (2)

Country Link
CN (1) CN117764633A (zh)
WO (1) WO2024060587A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118172146A (zh) * 2024-05-14 2024-06-11 腾讯科技（深圳）有限公司 物品数据处理方法、装置、计算机设备和存储介质
CN119006060A (zh) * 2024-10-15 2024-11-22 中国科学技术大学 一种转化率模型更新方法、装置、电子设备及存储介质
CN119510709A (zh) * 2024-11-01 2025-02-25 中国科学院自动化研究所 一种基于编码机制的土壤多参数传感器时序数据校准方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210027178A1 (en) * 2019-07-26 2021-01-28 Ricoh Company, Ltd. Recommendation method and recommendation apparatus based on deep reinforcement learning, and non-transitory computer-readable recording medium
CN112288146A (zh) * 2020-10-15 2021-01-29 北京沃东天骏信息技术有限公司 页面显示方法、装置、系统、计算机设备以及存储介质
CN113393299A (zh) * 2021-06-17 2021-09-14 中国工商银行股份有限公司 推荐模型训练方法、装置、电子设备和存储介质
CN114896515A (zh) * 2022-04-02 2022-08-12 哈尔滨工程大学 基于时间间隔的自监督学习协同序列推荐方法、设备和介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210027178A1 (en) * 2019-07-26 2021-01-28 Ricoh Company, Ltd. Recommendation method and recommendation apparatus based on deep reinforcement learning, and non-transitory computer-readable recording medium
CN112288146A (zh) * 2020-10-15 2021-01-29 北京沃东天骏信息技术有限公司 页面显示方法、装置、系统、计算机设备以及存储介质
CN113393299A (zh) * 2021-06-17 2021-09-14 中国工商银行股份有限公司 推荐模型训练方法、装置、电子设备和存储介质
CN114896515A (zh) * 2022-04-02 2022-08-12 哈尔滨工程大学 基于时间间隔的自监督学习协同序列推荐方法、设备和介质

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118172146A (zh) * 2024-05-14 2024-06-11 腾讯科技（深圳）有限公司 物品数据处理方法、装置、计算机设备和存储介质
CN119006060A (zh) * 2024-10-15 2024-11-22 中国科学技术大学 一种转化率模型更新方法、装置、电子设备及存储介质
CN119006060B (zh) * 2024-10-15 2024-12-24 中国科学技术大学 一种转化率模型更新方法、装置、电子设备及存储介质
CN119510709A (zh) * 2024-11-01 2025-02-25 中国科学院自动化研究所 一种基于编码机制的土壤多参数传感器时序数据校准方法

Also Published As

Publication number Publication date
CN117764633A (zh) 2024-03-26

Similar Documents

Publication Publication Date Title
CN109460514B (zh) 用于推送信息的方法和装置
JP2024503774A (ja) 融合パラメータの特定方法及び装置、情報推奨方法及び装置、パラメータ測定モデルのトレーニング方法及び装置、電子機器、記憶媒体、並びにコンピュータプログラム
US11521221B2 (en) Predictive modeling with entity representations computed from neural network models simultaneously trained on multiple tasks
EP4242955A1 (en) User profile-based object recommendation method and device
WO2024060587A1 (zh) 自监督学习模型的生成方法和转化率预估模型的生成方法
CN112541122A (zh) 推荐模型的训练方法、装置、电子设备及存储介质
CN110263938A (zh) 用于生成信息的方法和装置
CN112115257A (zh) 用于生成信息评估模型的方法和装置
CN113609397B (zh) 用于推送信息的方法和装置
CN108520470A (zh) 用于生成用户属性信息的方法和装置
CN114418093B (zh) 训练路径表征模型、输出信息的方法和装置
CN113392920B (zh) 生成作弊预测模型的方法、装置、设备、介质及程序产品
CN111783810A (zh) 用于确定用户的属性信息的方法和装置
CN110457476A (zh) 用于生成分类模型的方法和装置
CN116401372A (zh) 知识图谱表示学习方法、装置、电子设备及可读存储介质
CN115374862A (zh) 信息处理方法及装置、电子设备和存储介质
CN117670366A (zh) 风险预测方法、装置、设备、介质和程序产品
CN117591734A (zh) 点击率预估模型的训练方法及装置、物品推荐方法及装置
CN117573973A (zh) 资源推荐方法、装置、电子设备以及存储介质
CN109948773A (zh) 生成信息的方法和装置
EP4293534A1 (en) Blockchain address classification method and apparatus
CN115392230A (zh) 对话文本标签识别模型训练方法、对话文本标签识别方法
CN113792952B (zh) 用于生成模型的方法和装置
CN112200602B (zh) 用于广告推荐的神经网络模型训练方法及装置
CN118071385B (zh) 一种面向全物联网架构的营销业务融合方法和系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23866884

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 26.06.2025)

122 Ep: pct application non-entry in european phase

Ref document number: 23866884

Country of ref document: EP

Kind code of ref document: A1