WO2019210695A1 - 模型训练和业务推荐 - Google Patents

模型训练和业务推荐 Download PDF

Info

Publication number
WO2019210695A1
WO2019210695A1 PCT/CN2018/121950 CN2018121950W WO2019210695A1 WO 2019210695 A1 WO2019210695 A1 WO 2019210695A1 CN 2018121950 W CN2018121950 W CN 2018121950W WO 2019210695 A1 WO2019210695 A1 WO 2019210695A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample data
type
user
weight
data
Prior art date
Application number
PCT/CN2018/121950
Other languages
English (en)
French (fr)
Inventor
王子伟
Original Assignee
北京三快在线科技有限公司
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 北京三快在线科技有限公司 filed Critical 北京三快在线科技有限公司
Publication of WO2019210695A1 publication Critical patent/WO2019210695A1/zh
Priority to US17/077,416 priority Critical patent/US20210042664A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/12—Hotels or restaurants
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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
    • G06Q30/0203—Market surveys; Market polls
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations

Definitions

  • Embodiments of the present disclosure relate to the field of artificial intelligence technologies, and in particular, to a model training method and apparatus, a method and apparatus for service recommendation, and an electronic device.
  • Machine Learning is the core of artificial intelligence. Its application spans all fields of artificial intelligence and is the fundamental way to make computers intelligent.
  • the training set samples need to be physically consistent with the prediction set, such as the training set sample and the prediction set are for the same type of user, to ensure that the trained machine learning model is performed for the prediction set.
  • the prediction set such as the training set sample and the prediction set are for the same type of user
  • the number of training set samples that are physically consistent with the prediction set is small. Because the number of training samples is small, the parameter optimization of the machine learning model is insufficient, resulting in accurate subsequent prediction. The rate drops.
  • Embodiments of the present disclosure provide a new model training method, the method comprising: acquiring a sample data set, wherein the sample data set includes first type of sample data and second type of sample data; acquiring the first a first weight corresponding to the class sample data, and a second weight corresponding to the second type of sample data; according to the first weight, the second weight, a loss function corresponding to the first type of sample data, and a The loss function corresponding to the second type of sample data is subjected to a weighting operation to obtain an overall loss function; and based on the overall loss function, the sample data set is used to train the machine learning model.
  • acquiring the first weight corresponding to the first type of sample data, and the second weight corresponding to the second type of sample data comprising: determining a first ratio and a second ratio, where The first ratio is a probability that the behavior of the first type of sample data is a specified behavior, and the second ratio is a probability that the behavior of the second type of sample data is a specified behavior; A weight is used as the second weight.
  • acquiring the first weight corresponding to the first type of sample data, and the second weight corresponding to the second type of sample data comprising: determining the first type of sample data and the second Classification information of the class sample data; in the preset classification information and the weight mapping relationship, matching the classification information to obtain the first weight corresponding to the first type of sample data and the second type of sample data corresponding to The second weight.
  • the sample data set is a user data set for a liquor business
  • the first type of sample data includes user data of a first level user and a feature label of the user data
  • the second type of sample The data includes user data of a second level user and a feature tag of the user data, the user data including attribute data and behavior data, the level of the first level user being higher than the level of the second level user
  • the feature tag is used to indicate a correspondence between the user data and the purchase behavior.
  • the embodiment of the present disclosure further provides a method for service recommendation, the method comprising: training a target machine learning model by using the model training method according to any one of the preceding claims; acquiring a candidate user list, wherein the candidate user list includes multiple User data of each candidate user; respectively inputting the user data of each candidate user into the trained target machine learning model, obtaining a predicted value corresponding to the user data of each candidate user; and detecting the user When the predicted value corresponding to the data is greater than the first preset threshold, the candidate user corresponding to the user data is used as the target user, and the target service is recommended to the target user.
  • An embodiment of the present disclosure further provides an apparatus for model training, the apparatus comprising: a sample data set obtaining module, configured to acquire a sample data set, wherein the sample data set includes a first type of sample data and a second type of sample a data acquisition module, configured to acquire a first weight corresponding to the first type of sample data, and a second weight corresponding to the second type of sample data; an overall loss function determining module, configured to use the first weight And the second weight, the loss function corresponding to the first type of sample data, and the loss function corresponding to the second type of sample data are weighted to obtain an overall loss function; the model training module is configured to be based on the total A loss function that uses the sample data set to train the machine learning model.
  • the weight obtaining module includes: a ratio determining submodule, configured to determine a first ratio and a second ratio, wherein the first ratio is a probability that the behavior of the first type of sample data is a specified behavior, The second ratio is a probability that the behavior of the second type of sample data is a specified behavior; the weight is used as a submodule, and the first ratio is used as the first weight, and the second ratio is used as the Second weight.
  • a ratio determining submodule configured to determine a first ratio and a second ratio, wherein the first ratio is a probability that the behavior of the first type of sample data is a specified behavior, The second ratio is a probability that the behavior of the second type of sample data is a specified behavior; the weight is used as a submodule, and the first ratio is used as the first weight, and the second ratio is used as the Second weight.
  • the weight obtaining module includes: a classification information determining submodule, configured to determine classification information of the first type of sample data and the second type of sample data; and a weight matching submodule for presetting In the classification information and the weight mapping relationship, the classification information is matched, and the first weight corresponding to the first type of sample data and the second weight corresponding to the second type of sample data are obtained.
  • a classification information determining submodule configured to determine classification information of the first type of sample data and the second type of sample data
  • a weight matching submodule for presetting In the classification information and the weight mapping relationship, the classification information is matched, and the first weight corresponding to the first type of sample data and the second weight corresponding to the second type of sample data are obtained.
  • the sample data set is a user data set for a liquor business
  • the first type of sample data includes user data of a first level user and a feature label of the user data
  • the second type of sample The data includes user data of a second level user and a feature tag of the user data, the user data including attribute data and behavior data, the level of the first level user being higher than the level of the second level user
  • the feature tag is used to indicate a correspondence between the user data and the purchase behavior.
  • the embodiment of the present disclosure further provides an apparatus for service recommendation, the apparatus comprising: a model training module, configured to train a target machine learning model by using the model training method according to any one of the above; a candidate user list obtaining module, configured to: Obtaining a candidate user list, where the candidate user list includes user data of a plurality of candidate users; and a predicted value calculation module, configured to respectively input the user data of each candidate user into the trained target machine learning model, Obtaining a predicted value corresponding to the user data of each candidate user; the service recommendation module, configured to: when the predicted value corresponding to the user data is greater than a first preset threshold, The candidate user is referred to as the target user, and the target service is recommended to the target user.
  • a model training module configured to train a target machine learning model by using the model training method according to any one of the above
  • a candidate user list obtaining module configured to: Obtaining a candidate user list, where the candidate user list includes user data of a plurality of candidate users
  • Embodiments of the present disclosure also provide an electronic device including a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the program as described above method.
  • Embodiments of the present disclosure also provide a computer readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method as described above.
  • the embodiment of the present disclosure includes the following advantages: in the embodiment of the present disclosure, by acquiring a sample data set, the sample data set includes the first type of sample data, and the second type of sample data, and may obtain the first corresponding to the first type of sample data.
  • the weight, and the second weight corresponding to the second type of sample data may be weighted according to the first weight, the second weight, the loss function corresponding to the first type of sample data, and the loss function corresponding to the second type of sample data.
  • the overall loss function based on the overall loss function, uses the sample data set to train the machine learning model.
  • the model training is implemented by using various types of sample data, so that when the quantity of the first type of sample data is small, the second type of sample data is supplemented as the first type of sample data, which helps to improve the effect of the model training.
  • FIG. 1 is a flow chart showing the steps of a method of model training according to an embodiment of the present disclosure
  • FIG. 2 is a flow chart showing the steps of a method of model training according to another embodiment of the present disclosure
  • FIG. 3 is a function diagram of a cost function of an embodiment of the present disclosure.
  • FIG. 4 is a flow chart showing the steps of a method for service recommendation according to an embodiment of the present disclosure
  • FIG. 5 is a structural block diagram of an apparatus for model training according to an embodiment of the present disclosure.
  • FIG. 6 is a structural block diagram of an apparatus for service recommendation according to an embodiment of the present disclosure.
  • FIG. 1 a flow chart of steps of a method for model training according to an embodiment of the present disclosure is shown, which may specifically include the following steps:
  • Step 101 Acquire a sample data set, wherein the sample data set includes a first type of sample data, and a second type of sample data.
  • the sample data set may be a user data set for a liquor business
  • the first type of sample data may include user data of a first level user and feature tags of user data
  • the second type of sample data may include a second level user.
  • the user data and the feature tag of the user data the level of the first level user is higher than the level of the second level user.
  • the feature tag can be used to indicate a correspondence between the corresponding user data and the purchase behavior. For example, some feature tags corresponding to user data generate purchase behavior, and some feature tags corresponding to user data do not generate purchase behavior.
  • the user data may include attribute data and behavior data, such as the attribute data may include a user tag of the sample user, a consumption level, etc., and the behavior data may be purchase behavior data of the sample user, browsing behavior data, and the like.
  • attribute data may include a user tag of the sample user, a consumption level, etc.
  • behavior data may be purchase behavior data of the sample user, browsing behavior data, and the like.
  • the embodiment of the present disclosure not only acquires the first type of sample data having strong correlation with the prediction target, but also acquires the second weak correlation with the prediction target.
  • Class sample data The generalization ability of the model is improved by the introduction of more sample data that is related to the predicted target.
  • the prediction target may be user data for the user to be predicted
  • the first type of sample data may be sample data that is physically consistent with the prediction target, such as the first type of sample data and the prediction target are first-level users.
  • the second type of sample data is inconsistent with the prediction target in physical sense, but is related to a part of the data of the prediction target.
  • the predicted target is the user data of the first level user
  • the second type of sample data is the user data of the second level user
  • the second level user is the second level user with the specified behavior.
  • a designated behavior can refer to a purchase behavior.
  • the first-level users can be high-star users
  • the second-level users can be low-star users with purchase behaviors for high-star hotel services.
  • the sample data in the sample data set can also be divided into positive sample data and negative sample data, so that during the training of the model, the positive sample data and the negative sample data can be differentially trained, such as a positive sample.
  • the data set can be a user data set with purchase behavior for the high star hotel business
  • the negative sample data set can be a user data set with browsing behavior for the high star hotel business but not for purchase.
  • Step 102 Acquire a first weight corresponding to the first type of sample data, and a second weight corresponding to the second type of sample data.
  • the first weight corresponding to the first type of sample data and the second weight corresponding to the second type of sample data may be obtained.
  • the first weight may be directly set to the second weight, so that different sample data are treated differently during the model training process, and the first type of sample data having a larger weight is focused.
  • step 102 by analyzing the sample data, the probability that the user corresponding to the sample data performs the specified behavior may be determined, and then the weight corresponding to each sample data may be determined according to the probability. Then step 102 can include:
  • the first ratio and the second ratio are determined.
  • the first ratio may be a probability that the behavior of the first type of sample data is a specified behavior
  • the second ratio may be a probability that the behavior of the second type of sample data is a specified behavior.
  • the behavior of the first type of sample data can be determined, and then the probability that the specified behavior accounts for the behavior of the first type of sample data can be counted as the first ratio.
  • the behavior of the second type of sample data can be determined, and the probability of specifying the behavior of the second type of sample data can be counted as the second ratio.
  • the second sample user data including 10 purchase behavior data for the hotel business, the purchase behavior data for the high-star hotel business accounts for one, and the purchase behavior data for the low-star hotel accounts for nine, It is determined that the purchase behavior data for the high-star hotel business accounts for 10% of all purchase behavior data for the hotel business, that is, the second ratio is 10%.
  • the first ratio is taken as a first weight
  • the second ratio is taken as a second weight
  • the first ratio is used as the first weight corresponding to the first type of sample data
  • the second ratio is used as the second weight corresponding to the second type of sample data.
  • the weight corresponding to the first ratio may be obtained from the preset ratio and the weight mapping relationship, and the first weight corresponding to the first type of sample data is obtained, and the weight corresponding to the second proportion is obtained, which is corresponding to the second type of sample data. Two weights.
  • the sample data may be analyzed to determine the category to which the sample data belongs, and the weight corresponding to the sample data may be determined according to the category. Then step 102 can include:
  • Determining classification information of the first type of sample data and the second type of sample data Determining classification information of the first type of sample data and the second type of sample data.
  • the classification information may include a user type corresponding to the sample data, for example, the user type corresponding to the first type of sample data is a first level user, and the user type corresponding to the second type of sample data is a second level user.
  • the sample data may have classification information, and for each sample data, the classification information may be obtained from the specified field.
  • the classification information is matched, and the first weight corresponding to the first type of sample data and the second weight of the second type of sample data are obtained.
  • the classification information may be matched in the preset classification information and the weight mapping relationship, and then the weight corresponding to the classification information of the first type of sample data may be determined, as the first corresponding to the first type of sample data. Weights, and can determine the weight corresponding to the classification information of the second type of sample data, as the second weight corresponding to the second type of sample data.
  • Step 103 Perform a weighting operation according to the first weight, the second weight, a loss function corresponding to the first type of sample data, and a loss function corresponding to the second type of sample data, to obtain an overall loss function.
  • a loss function may be separately set for the first type of sample data and the second type of sample data to calculate the prediction loss of the first type of sample data and the second type of sample data, respectively.
  • the predicted loss is the difference between the predicted value and the pre-acquired real value when predicting a certain type of sample data.
  • the weight function corresponding to the first type of sample data may be weighted by using the first weight, and the weight function corresponding to the second type of sample data may be weighted by using the second weight, and then the weighted
  • the loss function is organized as an overall loss function to calculate the overall predicted loss corresponding to the sample data set.
  • the loss function can be as follows:
  • L is the predicted loss corresponding to the first type of sample data or the second type of sample data
  • y i is a real value corresponding to the pre-acquired i-th first-class sample data or the i-th second-class sample data
  • w i is the first weight corresponding to the i-th first type of sample data or the second weight corresponding to the i-th second type of sample data
  • abs is the absolute value operation.
  • the overall loss function can be as shown in the following formula (2):
  • J is the overall predicted loss
  • n is the number of data of the first type of sample data
  • m is the number of data of the second type of sample data
  • W 1 is the first weight
  • W 2 is the second weight
  • ⁇ is the summation Operation.
  • the overall loss function can also be as shown in equation (3) below:
  • J 1 is a prediction loss corresponding to the first type of sample data
  • W 1 is a first weight
  • J 2 is a prediction loss corresponding to the second type of sample data
  • W 2 is a second weight
  • J 1 and J 2 can be calculated by the following formula (4):
  • Step 104 using the sample data set to perform training of the machine learning model based on the overall loss function.
  • the overall loss function can calculate the overall predicted loss of the sample data. Then the model data is trained according to the overall predicted loss. For example, the Gradient Boosting Decision Tree (GBDT) algorithm can be used to iteratively train the model. The goal of the iteration may be to make the overall loss function of the sample data as small as possible. The result is finally to find the optimal parameters of the model.
  • GBDT Gradient Boosting Decision Tree
  • the sample data set includes the first type of sample data, and the second type of sample data, and the first weight corresponding to the first type of sample data, and the second The second weight corresponding to the class sample data, and then the weight function corresponding to the loss function corresponding to the first type of sample data and the loss function corresponding to the second type of sample data may be weighted by using the first weight and the second weight, respectively, to obtain an overall loss function, based on
  • the overall loss function uses model data sets for model training.
  • the model training is implemented by using various types of sample data, so that when the quantity of the first type of sample data is small, the second type of sample data is supplemented as the first type of sample data, thereby improving the effect of the model training.
  • the method may include the following steps:
  • Step 201 Acquire a sample data set, where the sample data set includes a first type of sample data, and a second type of sample data.
  • the embodiment of the present disclosure can not only obtain the first type of sample data that is strongly correlated with the prediction target, but also obtain the second weak correlation with the prediction target.
  • Class sample data Improve the generalization ability of the model by introducing more sample data that is related to the predicted target.
  • Step 202 Acquire a first weight corresponding to the first type of sample data, and a second weight corresponding to the second type of sample data.
  • the first weight corresponding to the first type of sample data and the second weight corresponding to the second type of sample data may be obtained.
  • Step 203 Obtain an overall loss function according to the first weight, the second weight, a first loss function corresponding to the first type of sample data, and a second loss function corresponding to the second type of sample data.
  • the overall loss function can be determined.
  • the weight function corresponding to the loss function corresponding to the first type of sample data may be weighted by using the first weight, and the weight function corresponding to the second type of sample data may be weighted by using the second weight, and then the weighted loss function may be organized into a total Loss function.
  • Step 204 initializing a machine learning model.
  • the machine learning model can have multiple model parameters, and the model parameters of the machine learning model can be initialized before the model training begins.
  • Step 205 In the machine learning model, calculate an overall prediction loss corresponding to the sample data set by using an overall loss function according to the first type of sample data and the second type of sample data.
  • the sample data set can be input into the machine learning model, and the machine learning model can predict the sample data set to obtain a predicted value for each sample data.
  • the total loss function can be used to calculate the predicted value and the true value of each sample data in the sample data set respectively, and the overall prediction loss corresponding to the sample data set is obtained.
  • the overall loss function may calculate the overall predicted loss corresponding to the sample data set as follows:
  • the real value corresponding to each first type of sample data may be collected in advance, and then the absolute value of the difference between the predicted value and the true value may be calculated to obtain a first prediction loss, and then the first weight may be used.
  • the first predicted loss is weighted to obtain a first weighted prediction loss.
  • the real value corresponding to each second type of sample data may be pre-acquired, and then the absolute value of the difference between the predicted value and the true value may be calculated to obtain a second predicted loss, and then the second weight may be used.
  • the second predicted loss is weighted to obtain a second weighted prediction loss.
  • the weighted prediction loss of each sample data may be averaged to obtain an overall prediction loss corresponding to the sample data set.
  • Step 206 Iteratively adjust parameters of the machine learning model, and recalculate the overall prediction loss corresponding to the sample data set.
  • the model parameters can be iterated, the iterative machine learning model is obtained, and then the iterative model is used to recalculate the overall prediction loss.
  • each sample data may have one or more sample features
  • the model parameters in the machine learning model may be sample weights set for each sample feature, and the sample weights are iterated to obtain different The overall forecast loss.
  • the overall loss function can calculate multiple overall prediction losses, as shown in Figure 3, where p is the p-th model parameter and J(p) is the overall prediction loss as a function of the overall loss function.
  • p is the p-th model parameter
  • J(p) is the overall prediction loss as a function of the overall loss function.
  • step 207 the machine learning model corresponding to the smallest overall prediction loss is determined as the target machine learning model.
  • the sample data set is first acquired, the sample data set includes the first type of sample data, and the second type of sample data, and the first weight corresponding to the first type of sample data is obtained, and the first The second weight corresponding to the second type of sample data. Then, the weighting operation may be performed according to the first weight, the second weight, the loss function corresponding to the first type of sample data, and the loss function corresponding to the second type of sample data, to obtain an overall loss function, and the sample data set is used to perform the model based on the overall loss function. training.
  • the model training is implemented by using various types of sample data, so that when the quantity of the first type of sample data is small, the second type of sample data is supplemented as the first type of sample data, thereby improving the effect of the model training.
  • the target machine learning model corresponding to the smallest overall prediction loss is determined, which ensures the accuracy of the model prediction and reduces the model.
  • the predicted loss
  • FIG. 4 a flow chart of method steps of a service recommendation according to an embodiment of the present disclosure is shown, which may specifically include the following steps:
  • Step 401 Acquire a candidate user list, where the candidate user list includes user data of multiple candidate users.
  • the candidate user may be a user whose user level is higher than a second preset threshold in the business of the wine business, for example, the candidate user is a first-level user, that is, a high-star user.
  • the user type corresponding to the target service may be determined, and then multiple candidate users that match the user type are filtered out from the background data, and user data of multiple candidate users are obtained, and a candidate user list is obtained.
  • the target service to be recommended is a wine-based business for a high-star user
  • all high-star users can be used as candidate users, and user data of the high-star user can be obtained.
  • Step 402 Enter the user data of each candidate user into the trained target machine learning model to obtain a predicted value corresponding to the user data of each candidate user.
  • the user data of each candidate user may be input into a target machine learning model, and the target machine model predicts the user data of each candidate user, and obtains a predicted value corresponding to the user data of each candidate user.
  • the target machine learning model is trained by using the model training method described in any of the above embodiments, and the main steps include: acquiring a sample data set; wherein the sample data set includes the first type of sample data, and the second type of sample data; a first weight corresponding to the first type of sample data, and a second weight corresponding to the second type of sample data; and a loss corresponding to the first weight, the second weight, and the first type of sample data
  • the function, the loss function corresponding to the second type of sample data is subjected to a weighting operation to obtain an overall loss function; and based on the total loss function, the sample data set is used to train the machine learning model to obtain a target machine learning model.
  • Step 403 When detecting that the predicted value corresponding to the user data is greater than the first preset threshold, the candidate user corresponding to the user data is used as the target user, and the associated target service is recommended to the target user.
  • the target service may be a liquor business associated with the candidate user, such as a liquor business for the first-level user (high-star user).
  • Target business such as issuing subsidized coupons to target users.
  • the candidate user list may include user data of the plurality of candidate users, and input user data of each candidate user into the trained target machine learning model, respectively, to obtain each candidate user's
  • the candidate user corresponding to the user data is used as the target user, and then the associated target service is recommended to the target user.
  • the target machine learning model is used for prediction, and the business recommendation is based on the prediction result, which improves the success rate of the business recommendation.
  • FIG. 5 a structural block diagram of an apparatus for model training according to an embodiment of the present disclosure is shown, which may specifically include the following modules:
  • the sample data set obtaining module 501 is configured to acquire a sample data set, wherein the sample data set includes a first type of sample data, and a second type of sample data.
  • the weight obtaining module 502 is configured to acquire a first weight corresponding to the first type of sample data, and a second weight corresponding to the second type of sample data.
  • the total loss function determining module 503 is configured to perform a weighting operation according to the first weight, the second weight, a loss function corresponding to the first type of sample data, and a loss function corresponding to the second type of sample data, Get the overall loss function.
  • the model training module 504 is configured to perform training of the machine learning model using the sample data set based on the overall loss function.
  • the weight obtaining module 502 includes: a ratio determining sub-module, configured to determine a first ratio and a second ratio; wherein the first ratio is behavior of the first type of sample data For the probability of specifying the behavior, the second ratio is a probability that the behavior of the second type of sample data is a specified behavior; the weight is used as a submodule, and the first ratio is used as the first weight, and the second is The ratio is taken as the second weight.
  • the weight obtaining module 502 includes: a classification information determining sub-module, configured to determine classification information of the first type of sample data and the second type of sample data; a weight matching sub-module, And matching the classification information in the preset classification information and the weight mapping relationship, and obtaining a first weight corresponding to the first type of sample data and a second weight corresponding to the second type of sample data.
  • the sample data set is a user data set for a liquor business
  • the first type of sample data includes user data of a first level user
  • the second type of sample data includes a first User data of the second-level user
  • the level of the first-level user is higher than the level of the second-level user
  • the feature tag is used to indicate a correspondence between the user data and the purchase behavior.
  • the user data includes attribute data and behavior data.
  • the sample data set includes the first type of sample data, and the second type of sample data, and the first weight corresponding to the first type of sample data, and the second The second weight corresponding to the class sample data may then be weighted according to the first weight, the second weight, the loss function corresponding to the first type of sample data, and the loss function corresponding to the second type of sample data, to obtain an overall loss function, based on The overall loss function uses the sample data set to train the machine learning model.
  • the model training is implemented by using various types of sample data, so that when the quantity of the first type of sample data is small, the second type of sample data is supplemented as the first type of sample data, thereby improving the effect of the model training.
  • FIG. 6 it is a structural block diagram of an apparatus for service recommendation according to an embodiment of the present disclosure, which may specifically include the following modules:
  • a model training module 601 configured to train a target machine learning model by using the model training method described in any of the above method embodiments;
  • a candidate user list obtaining module 602 configured to acquire a candidate user list, where the candidate user list includes user data of multiple candidate users;
  • the predicted value calculation module 603 is configured to respectively input the user data of each candidate user into the trained target machine learning model, and obtain a predicted value corresponding to the user data of each candidate user;
  • the service recommendation module 604 is configured to: when the predicted value corresponding to the user data is greater than the first preset threshold, use the candidate user corresponding to the user data as the target user, and recommend the target service to the target user.
  • the target service is a tour-type service associated with the candidate user
  • the candidate user is a user whose user level is higher than a second preset threshold in the tour service.
  • the user data of each candidate user is input into the trained target machine learning model by obtaining the candidate user list, and the predicted value corresponding to the user data of each candidate user is obtained.
  • the candidate user corresponding to the user data is used as the target user, and then the associated target service is recommended to the target user.
  • the target machine learning model is used for prediction, and the business recommendation is based on the prediction result, which improves the success rate of the business recommendation.
  • An embodiment of the present disclosure also discloses an electronic device including a memory, a processor, and a computer program stored on the memory and operable on the processor, the processor implementing the program as described above method.
  • Embodiments of the present disclosure also disclose a computer readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method as described above.
  • embodiments of the disclosed embodiments can be provided as a method, apparatus, or computer program product.
  • embodiments of the present disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware.
  • embodiments of the present disclosure may take the form of a computer program product embodied on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.
  • Embodiments of the present disclosure are described with reference to flowchart illustrations and/or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present disclosure. It will be understood that each flow and/or block of the flowchart illustrations and/or FIG.
  • These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing terminal device to produce a machine such that instructions are executed by a processor of a computer or other programmable data processing terminal device
  • Means are provided for implementing the functions specified in one or more of the flow or in one or more blocks of the flow chart.
  • the computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device.
  • the instruction device implements the functions specified in one or more blocks of the flowchart or in a flow or block of the flowchart.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Strategic Management (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Software Systems (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • General Business, Economics & Management (AREA)
  • Mathematical Physics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Primary Health Care (AREA)
  • Human Resources & Organizations (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)

Abstract

一种模型训练方法和装置、业务推荐的方法和装置、电子设备，该模型训练方法，包括：获取样本数据集；其中，所述样本数据集包括第一类样本数据，以及第二类样本数据(101)；获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重(102)；根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数，进行加权运算，得到总体损失函数(103)；基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练(104)。

Description

模型训练和业务推荐
相关申请的交叉引用
本专利申请要求于2018年05月02日提交的、申请号为201810411497.4、发明名称为“模型训练方法和装置、业务推荐的方法和装置、电子设备”的中国专利申请的优先权，该申请的全文以引用的方式并入本文中。
技术领域
本公开的实施例涉及人工智能技术领域，特别是涉及一种模型训练方法和装置、业务推荐的方法和装置、电子设备。
背景技术
机器学习(Machine Learning，ML)是人工智能的核心，其应用遍及人工智能的各个领域，是使计算机具有智能的根本途径。通常，在训练机器学习的模型时，训练集样本需要与预测集在物理意义上一致，如训练集样本与预测集针对的是同一类型用户，以保证训练的机器学习的模型在针对预测集进行预测时，有一个比较好的精度。而在某些业务场景中，与预测集在物理意义上一致的训练集样本的数量较少，由于训练样本数较少，会导致机器学习的模型的参数优化不充分，从而导致后续的预测准确率下降。
发明内容
本公开的实施例提供了一种新的模型训练方法，所述方法包括：获取样本数据集，其中，所述样本数据集包括第一类样本数据以及第二类样本数据；获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重；根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数；基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
可选地，获取所述第一类样本数据对应的所述第一权重，以及所述第二类样本数据对应的所述第二权重包括：确定第一比例和第二比例，其中，所述第一比例为所述第一类样本数据的行为为指定行为的概率，所述第二比例为所述第二类样本数据的行为为指 定行为的概率；将所述第一比例作为所述第一权重，将所述第二比例作为所述第二权重。
可选地，获取所述第一类样本数据对应的所述第一权重，以及所述第二类样本数据对应的所述第二权重包括：确定所述第一类样本数据和所述第二类样本数据的分类信息；在预置的分类信息与权重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的所述第一权重和所述第二类样本数据对应的所述第二权重。
可选地，所述样本数据集为针对酒旅类业务的用户数据集，所述第一类样本数据包括第一等级用户的用户数据以及所述用户数据的特征标签，所述第二类样本数据包括第二等级用户的用户数据以及所述用户数据的特征标签，所述用户数据包括属性数据和行为数据，所述第一等级用户的级别高于所述第二等级用户的级别，所述特征标签用于指示所述用户数据与购买行为的对应关系。
本公开实施例还提供了一种业务推荐的方法，所述方法包括：利用上述任一项所述的模型训练方法训练目标机器学习模型；获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据；分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值；在检测到所述用户数据对应的所述预测值大于第一预设阈值时，将所述用户数据对应的所述候选用户作为目标用户，并向所述目标用户推荐目标业务。
本公开实施例还提供了一种模型训练的装置，所述装置包括：样本数据集获取模块，用于获取样本数据集，其中，所述样本数据集包括第一类样本数据以及第二类样本数据；权重获取模块，用于获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重；总体损失函数确定模块，用于根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数；模型训练模块，用于基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
可选地，所述权重获取模块包括：比例确定子模块，用于确定第一比例和第二比例，其中，所述第一比例为所述第一类样本数据的行为为指定行为的概率，所述第二比例为所述第二类样本数据的行为为指定行为的概率；权重作为子模块，用于将所述第一比例作为所述第一权重，将所述第二比例作为所述第二权重。
可选地，所述权重获取模块包括：分类信息确定子模块，用于确定所述第一类样本数据和所述第二类样本数据的分类信息；权重匹配子模块，用于在预置的分类信息与权 重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的所述第一权重和所述第二类样本数据对应的所述第二权重。
可选地，所述样本数据集为针对酒旅类业务的用户数据集，所述第一类样本数据包括第一等级用户的用户数据以及所述用户数据的特征标签，所述第二类样本数据包括第二等级用户的用户数据以及所述用户数据的特征标签，所述用户数据包括属性数据和行为数据，所述第一等级用户的级别高于所述第二等级用户的级别，所述特征标签用于指示所述用户数据与购买行为的对应关系。
本公开实施例还提供了一种业务推荐的装置，所述装置包括：模型训练模块，用于利用上述任一项所述的模型训练方法训练目标机器学习模型；候选用户列表获取模块，用于获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据；预测值计算模块，用于分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值；业务推荐模块，用于在检测到所述用户数据对应的所述预测值大于第一预设阈值时，将所述用户数据对应的所述候选用户作为目标用户，并向所述目标用户推荐目标业务。
本公开实施例还提供了一种电子设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述程序时实现如上所述方法。
本公开实施例还提供了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如上所述方法的步骤。
本公开实施例包括以下优点：在本公开实施例中，通过获取样本数据集，样本数据集包括第一类样本数据，以及第二类样本数据，并可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重，然后可以根据第一权重、第二权重、第一类样本数据对应的损失函数、和第二类样本数据对应的损失函数进行加权运算，得到总体损失函数，基于总体损失函数，使用样本数据集进行机器学习模型的训练。实现了采用多种类型的样本数据进行模型训练，以在第一类样本数据的数量较少时，将第二类样本数据作为第一类样本数据的补充，有助于提升模型训练的效果。
附图说明
为了更清楚地说明本公开实施例的技术方案，下面将对本公开实施例的描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本公开的一些实施 例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本公开一个实施例的一种模型训练的方法的步骤流程图；
图2是本公开另一个实施例的一种模型训练的方法的步骤流程图；
图3是本公开实施例的一种成本函数的函数示意图；
图4是本公开实施例的一种业务推荐的方法的步骤流程图；
图5是本公开实施例的一种模型训练的装置的结构框图；
图6是本公开实施例的一种业务推荐的装置的结构框图。
具体实施方式
为使本公开的上述目的、特征和优点能够更加明显易懂，下面结合附图和具体实施方式对本公开作进一步详细的说明。显然，所描述的实施例是本公开一部分实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本公开保护的范围。
参照图1，示出了本公开一个实施例的一种模型训练的方法的步骤流程图，具体可以包括如下步骤：
步骤101，获取样本数据集，其中，所述样本数据集包括第一类样本数据，以及第二类样本数据。
例如，样本数据集可以为针对酒旅类业务的用户数据集，第一类样本数据可以包括第一等级用户的用户数据以及用户数据的特征标签，第二类样本数据可以包括第二等级用户的用户数据以及用户数据的特征标签，第一等级用户的级别高于第二等级用户的级别。所述特征标签可以用于指示对应的用户数据与购买行为的对应关系。比如一些用户数据对应的特征标签为产生了购买行为，一些用户数据对应的特征标签为没有产生购买行为。
作为一种示例，用户数据可以包括属性数据和行为数据，如属性数据可以包括样本用户的用户标签、消费水平等，行为数据可以为样本用户的购买行为数据、浏览行为数据等。
针对每个机器学习的模型，具有对应的预测目标，即预测集，本公开实施例不仅获 取与预测目标关联性较强的第一类样本数据，还获取与预测目标关联性较弱的第二类样本数据。由于引入了更多与预测目标具有关联性的样本数据，提升了模型的泛化能力。
实际上，预测目标可以为针对待预测用户的用户数据，则第一类样本数据可以为与预测目标在物理意义上一致的样本数据，如第一类样本数据与预测目标均为第一等级用户的用户数据。而第二类样本数据与预测目标在物理意义上不一致，但与预测目标的一部分数据具有关联性。如预测目标为第一等级用户的用户数据，而第二类样本数据为第二等级用户的用户数据，但该第二等级用户为具有指定行为的第二等级用户。比如在酒旅类业务的环境下，指定行为可以是指购买的行为。
例如，在酒店业务中，当某个用户针对高星级酒店业务的购买行为次数超过预设次数，则该用户为高星用户，否则为低星用户，则当预测目标中待预测的用户为高星用户时，则第一等级用户可以为高星用户，第二等级用户可以为具有针对高星级酒店业务的购买行为的低星用户。
在一种示例中，还可以将样本数据集中的样本数据区分为正样本数据和负样本数据，以使得在模型训练的过程中，能够对正样本数据、负样本数据进行区别训练，如正样本数据集可以为具有针对高星酒店业务的购买行为的用户数据集，负样本数据集可以为具有针对高星酒店业务的浏览行为但不进行购买的用户数据集。
步骤102，获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重。
在获取样本数据集后，可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重。
在一种实施方式中，可以直接设置第一权重大于第二权重，以使得在模型训练过程中区别对待不同的样本数据，并关注于权重较大的第一类样本数据。
在另一种实施方式中，可以通过分析样本数据，确定样本数据对应的用户进行指定行为的概率，进而可以根据概率确定各类样本数据对应的权重。则步骤102可以包括：
确定第一比例和第二比例。
其中，第一比例可以为第一类样本数据的行为为指定行为的概率，第二比例可以为第二类样本数据的行为为指定行为的概率。针对每个第一类样本数据，可以确定该第一类样本数据的行为，然后可以统计指定行为占第一类样本数据的行为的概率，作为第一比例。针对每个第二类样本数据，可以确定该第二类样本数据的行为，可以统计指定行 为占第二类样本数据的行为的概率，作为第二比例。
例如，在第二样本用户数据中，包含10条针对酒店业务的购买行为数据，针对高星级酒店业务的购买行为数据占1条，针对低星级酒店的购买行为数据占9条，则可以确定针对高星级酒店业务的购买行为数据占所有针对酒店业务的购买行为数据的比值为10％，即第二比例为10％。
将所述第一比例作为第一权重，将所述第二比例作为第二权重。
在确定比例后，将第一比例作为第一类样本数据对应的第一权重，将第二比例作为第二类样本数据对应的第二权重。
还可以从预置的比例和权重映射关系中，获取第一比例对应的权重，作为第一类样本数据对应的第一权重，获取第二比例对应的权重，作为第二类样本数据对应的第二权重。
在另一种实施方式中，还可以通过分析样本数据，确定样本数据所属的类别，进而可以根据该类别确定样本数据对应的权重。则步骤102可以包括：
确定所述第一类样本数据和所述第二类样本数据的分类信息。
其中，分类信息可以包括样本数据对应的用户类型，如第一类样本数据对应的用户类型为第一等级用户，第二类样本数据对应的用户类型为第二等级用户。
在具体实现中，样本数据可以具有分类信息，则针对每个样本数据，可以从指定字段获取分类信息。
在预置的分类信息与权重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的第一权重和所述第二类样本数据的第二权重。
在获得分类信息后，可以在预置的分类信息和权重映射关系中，对分类信息进行匹配，然后可以确定第一类样本数据对应分类信息对应的权重，作为第一类样本数据对应的第一权重，并可以确定第二类样本数据对应分类信息对应的权重，作为第二类样本数据对应的第二权重。
步骤103，根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数。
在本公开实施例中，可以分别为第一类样本数据、第二类样本数据设置损失函数，以分别计算第一类样本数据、第二类样本数据的预测损失。
其中，预测损失为针对某类样本数据进行预测时，预测值与预先采集的真实值之间的差值。
在获得权重后，可以使用第一权重对第一类样本数据对应的损失函数进行加权运算，并可以使用第二权重对第二类样本数据对应的损失函数进行加权运算，然后可以将加权后的损失函数组织为总体损失函数，以计算样本数据集对应的总体预测损失。
损失函数可以如下公式所示：
Figure PCTCN2018121950-appb-000001
其中，L为第一类样本数据或第二类样本数据对应的预测损失，
Figure PCTCN2018121950-appb-000002
为第i个第一类样本数据或第i个第二类样本数据对应的预测值，y i为预先采集的第i个第一类样本数据或第i个第二类样本数据对应的真实值，w i为第i个第一类样本数据对应的第一权重或第i个第二类样本数据对应的第二权重，对于一类样本数据来讲，每个数据的w i的值都一样，都等于该类数据对应的权重，abs为求绝对值运算。
总体损失函数可以如下公式(2)所示：
Figure PCTCN2018121950-appb-000003
其中，J为总体预测损失，n为第一类样本数据的数据个数，m为第二类样本数据的数据个数，W 1为第一权重，W 2为第二权重，Σ为求和运算。
在一种示例中，总体损失函数也可以如下公式(3)所示：
J＝W 1J 1+W 2J 2       (3),
其中，J 1为第一类样本数据对应的预测损失，W 1为第一权重，J 2为第二类样本数据对应的预测损失，W 2为第二权重。
其中，J 1和J 2可以采用如下公式(4)计算：
Figure PCTCN2018121950-appb-000004
步骤104，基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
在获得总体损失函数后，总体损失函数可以计算样本数据的总体预测损失。然后根据总体预测损失对样本数据进行模型训练，如可以采用梯度提升决策树(Gradient Boosting Decision Tree，GBDT)算法迭代进行模型训练。迭代的目标可以是使样本数据的总体损失函数尽可能变小。从而最终找到模型的最优参数。
在以上任一本公开实施例中，通过获取样本数据集，样本数据集包括第一类样本数据，以及第二类样本数据，并可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重，然后可以依次使用第一权重、第二权重对第一类样本数据对应的损失函数、第二类样本数据对应的损失函数进行加权运算，得到总体损失函数，基于总体损失函数，使用样本数据集进行模型训练。实现了采用多种类型的样本数据进行模型训练，以在第一类样本数据的数量较少时，将第二类样本数据作为第一类样本数据的补充，提升了模型训练的效果。
参照图2，示出了本公开实施例的另一种模型训练的方法的步骤流程图，具体可以包括如下步骤：
步骤201，获取样本数据集，其中，所述样本数据集包括第一类样本数据，以及第二类样本数据。
针对每个机器学习模型，具有对应的的预测目标，即预测集，本公开实施例不仅可以获取与预测目标关联性强的第一类样本数据，还可以获取与预测目标关联性弱的第二类样本数据。由于引入更多与预测目标具有关联性的样本数据，提升模型的泛化能力。
步骤202，获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重。
在获取样本数据集后，可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重。
步骤203，根据所述第一权重、所述第二权重、所述第一类样本数据对应的第一损失函数、所述第二类样本数据对应的第二损失函数，得到总体损失函数。
在获得权重后，可以确定总体损失函数。可以使用第一权重对第一类样本数据对应的损失函数进行加权运算，并可以使用第二权重对第二类样本数据对应的损失函数进行加权运算，然后可以将加权后的损失函数组织为总体损失函数。
步骤204，初始化机器学习模型。
在实际应用中，机器学习模型可以具有多个模型参数，在模型训练开始前，可以初始化机器学习模型的模型参数。
步骤205，在所述机器学习模型中，根据所述第一类样本数据、所述第二类样本数据，采用总体损失函数计算所述样本数据集对应的总体预测损失。
在初始化后，可以将样本数据集输入机器学习模型，机器学习模型可以对样本数据集进行预测，得到每个样本数据对应预测值。根据预先采集的每个样本数据的真实值，可以采用总体损失函数，分别对样本数据集中每个样本数据的预测值和真实值进行计算，得到样本数据集对应的总体预测损失。
在本公开一种实施例中，所述总体损失函数可以采用如下方式计算样本数据集对应的总体预测损失：
计算所述第一类样本数据对应的预测值的第一预测损失，并采用所述第一权重对所述第一预测损失进行加权，得到第一加权预测损失。
针对每个第一类样本数据，可以预先采集每个第一类样本数据对应的真实值，然后计算预测值和真实值的差值的绝对值，得到第一预测损失，然后可以采用第一权重对第一预测损失进行加权，得到第一加权预测损失。
计算所述第二类样本数据对应的预测值的第二预测损失，并采用所述第二权重对所述第二预测损失进行加权，得到第二加权预测损失。
针对每个第二类样本数据，可以预先采集每个第二类样本数据对应的真实值，然后计算预测值和真实值的差值的绝对值，得到第二预测损失，然后可以采用第二权重对第二预测损失进行加权，得到第二加权预测损失。
对所述第一加权预测损失和所述第二加权预测损失进行均值计算，得到所述样本数据对应的预测值的总体预测损失。
在获得加权预测损失后，可以将每个样本数据的加权预测损失进行均值计算，得到样本数据集对应的总体预测损失。
步骤206，迭代调整所述机器学习模型的参数，重新计算所述样本数据集对应的总体预测损失。
在获得总体预测损失后，可以对模型参数进行迭代，得到迭代后机器学习模型，然后采用迭代后的模型重新计算总体预测损失。
更进一步的，在实际应用中，每个样本数据可以具有一个或多个样本特征，机器学习模型中的模型参数可以为针对每个样本特征设置的样本权重，通过对样本权重进行迭代，得到不同的总体预测损失。
随着机器学习模型的迭代，总体损失函数可以计算得到多个总体预测损失，如图3 所示，p为第p组模型参数，J(p)为总体预测损失，沿着总体损失函数的函数值下降最快的方向，对总体损失函数进行收敛，可以得到使总体损失函数最小的模型参数，进而建立目标机器学习模型。
步骤207，将最小的总体预测损失对应的机器学习模型，确定为目标机器学习模型。
在以上任一本公开的实施例中，首先获取样本数据集，样本数据集包括第一类样本数据，以及第二类样本数据，并可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重。然后可以根据第一权重、第二权重、第一类样本数据对应的损失函数、第二类样本数据对应的损失函数进行加权运算，得到总体损失函数，基于总体损失函数，使用样本数据集进行模型训练。实现了采用多种类型的样本数据进行模型训练，以在第一类样本数据的数量较少时，将第二类样本数据作为第一类样本数据的补充，提升了模型训练的效果。
而且，通过在模型迭代的过程中，采用总体损失函数计算样本数据集对应的总体预测损失，进而确定最小的总体预测损失对应的目标机器学习模型，保证了模型预测的准确性，减小了模型的预测损失。
参照图4，示出了本公开实施例的一种业务推荐的方法步骤流程图，具体可以包括如下步骤：
步骤401，获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据。
其中，候选用户可以为在酒旅类业务中用户级别高于第二预设阈值的用户，如候选用户为第一等级用户，即高星用户。
当需要推荐目标业务时，可以确定该目标业务对应的用户类型，然后从后台数据中，筛选出符合该用户类型的多个候选用户，并获取多个候选用户的用户数据，得到候选用户列表。
例如，当需要推荐的目标业务为针对高星用户的酒旅类业务，可以将所有的高星用户作为候选用户，并获取高星用户的用户数据。
步骤402，分别将每个候选用户的用户数据输入训练好的目标机器学习模型，获得每个候选用户的用户数据对应的预测值。
在获得候选用户列表后，可以将每个候选用户的用户数据输入目标机器学习模型， 目标机器模型对每个候选用户的用户数据进行预测，获得每个候选用户的用户数据对应的预测值。
其中，利用以上任意实施例所述的模型训练方法训练目标机器学习模型，主要步骤包括：获取样本数据集；其中，所述样本数据集包括第一类样本数据，以及第二类样本数据；获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重；根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数；基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练，以得到目标机器学习模型。
对于建立目标机器学习模型的过程，可以参考上述各个实施例关于模型训练的方法的描述，这里不再赘述。
步骤403，在检测到所述用户数据对应的预测值大于第一预设阈值时，将所述用户数据对应的候选用户作为目标用户，并向所述目标用户推荐关联的目标业务。
其中，目标业务可以为与候选用户关联的酒旅类业务，如针对第一等级用户(高星用户)的酒旅类业务。
在获得预测值后，可以判断用户数据对应的预测值是否大于第一预设阈值，若是，则将该用户数据对应的候选用户作为目标用户，在确定目标用户后，可以向目标用户推荐关联的目标业务，如向目标用户发放补贴优惠券等。
在本公开实施例中，通过获取候选用户列表，候选用户列表可以包括多个候选用户的用户数据，分别将每个候选用户的用户数据输入训练好的目标机器学习模型，获得每个候选用户的用户数据对应的预测值，在检测到用户数据对应的预测值大于第一预设阈值时，将用户数据对应的候选用户作为目标用户，然后向目标用户推荐关联的目标业务。采用目标机器学习模型进行预测，并基于预测结果进行业务推荐，提升了业务推荐的成功率。
需要说明的是，对于方法实施例，为了简单描述，故将其都表述为一系列的动作组合，但是本领域技术人员应该知悉，本公开实施例并不受所描述的动作顺序的限制，因为依据本公开实施例，某些步骤可以采用其他顺序或者同时进行。其次，本领域技术人员也应该知悉，说明书中所描述的实施例均属于优选实施例，所涉及的动作并不一定是本公开实施例所必须的。
参照图5，示出了本公开实施例的一种模型训练的装置的结构框图，具体可以包括 如下模块：
样本数据集获取模块501，用于获取样本数据集，其中，所述样本数据集包括第一类样本数据，以及第二类样本数据。
权重获取模块502，用于获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重。
总体损失函数确定模块503，用于根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数。
模型训练模块504，用于基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
在本公开一种实施例中，所述权重获取模块502包括：比例确定子模块，用于确定第一比例和第二比例；其中，所述第一比例为所述第一类样本数据的行为为指定行为的概率，所述第二比例为所述第二类样本数据的行为为指定行为的概率；权重作为子模块，用于将所述第一比例作为第一权重，将所述第二比例作为第二权重。
在本公开一种实施例中，所述权重获取模块502包括：分类信息确定子模块，用于确定所述第一类样本数据和所述第二类样本数据的分类信息；权重匹配子模块，用于在预置的分类信息与权重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的第一权重和所述第二类样本数据对应的第二权重。
在本公开一种实施例中，所述样本数据集为针对酒旅类业务的用户数据集，所述第一类样本数据包括第一等级用户的用户数据，所述第二类样本数据包括第二等级用户的用户数据，所述第一等级用户的级别高于所述第二等级用户的级别，所述特征标签用于指示所述用户数据与购买行为的对应关系。
在本公开一种实施例中，所述用户数据包括属性数据和行为数据。
在以上任一本公开实施例中，通过获取样本数据集，样本数据集包括第一类样本数据，以及第二类样本数据，并可以获取第一类样本数据对应的第一权重，以及第二类样本数据对应的第二权重，然后可以根据第一权重、第二权重、第一类样本数据对应的损失函数、和第二类样本数据对应的损失函数进行加权运算，得到总体损失函数，基于总体损失函数，使用样本数据集进行机器学习模型的训练。实现了采用多种类型的样本数据进行模型训练，以在第一类样本数据的数量较少时，将第二类样本数据作为第一 类样本数据的补充，提升了模型训练的效果。
参照图6，示出了本公开实施例的一种业务推荐的装置的结构框图，具体可以包括如下模块：
模型训练模块601，用于利用以上任意的方法实施例所述的模型训练方法训练目标机器学习模型；
候选用户列表获取模块602，用于获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据；
预测值计算模块603，用于分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值；
业务推荐模块604，用于在检测到所述用户数据对应的预测值大于第一预设阈值时，将所述用户数据对应的候选用户作为目标用户，并向所述目标用户推荐目标业务。
在本公开的一个实施例中，所述目标业务为与所述候选用户关联的酒旅类业务，所述候选用户为在所述酒旅类业务中用户级别高于第二预设阈值的用户。
在本公开实施例中，通过获取候选用户列表，分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值，在检测到用户数据对应的预测值大于第一预设阈值时，将用户数据对应的候选用户作为目标用户，然后向目标用户推荐关联的目标业务。采用目标机器学习模型进行预测，并基于预测结果进行业务推荐，提升了业务推荐的成功率。
本公开各实施例的装置可用于对应的执行上述各个实施例提供的方法，相关术语和描述可以参考关于方法的描述，这里不再赘述。
本公开实施例还公开了一种电子设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述程序时实现如上所述方法。
本公开实施例还公开了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如上所述方法的步骤。
本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见即可。
本领域内的技术人员应明白，本公开实施例的实施例可提供为方法、装置、或 计算机程序产品。因此，本公开实施例可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本公开实施例可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本公开实施例是参照根据本公开实施例的方法、终端设备(***)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理终端设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理终端设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理终端设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理终端设备上，使得在计算机或其他可编程终端设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程终端设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本公开实施例的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例做出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本公开实施例范围的所有变更和修改。
最后，还需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者终端设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者终端设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者终端设备中还存在另外的相同要素。
以上对本公开所提供的一种模型训练及业务推荐的方法和装置，进行了详细介绍，本文中应用了具体个例对本公开的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本公开的方法及其核心思想；同时，对于本领域的一般技术人员，依据本公开的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本公开的限制。

Claims (14)

  1. 一种模型训练的方法，包括：
    获取样本数据集，其中，所述样本数据集包括第一类样本数据以及第二类样本数据；
    获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重；
    根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数；
    基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
  2. 根据权利要求1的方法，其特征在于，获取所述第一类样本数据对应的所述第一权重，以及所述第二类样本数据对应的所述第二权重包括：
    确定第一比例和第二比例，其中，所述第一比例为所述第一类样本数据的行为为指定行为的概率，所述第二比例为所述第二类样本数据的行为为指定行为的概率；
    将所述第一比例作为所述第一权重，将所述第二比例作为所述第二权重。
  3. 根据权利要求1所述的方法，其特征在于，获取所述第一类样本数据对应的所述第一权重，以及所述第二类样本数据对应的所述第二权重包括：
    确定所述第一类样本数据和所述第二类样本数据的分类信息；
    在预置的分类信息与权重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的所述第一权重和所述第二类样本数据对应的所述第二权重。
  4. 根据权利要求3所述的方法，其特征在于，
    所述样本数据集为针对酒旅类业务的用户数据集，
    所述第一类样本数据包括第一等级用户的用户数据以及所述用户数据的特征标签，
    所述第二类样本数据包括第二等级用户的用户数据以及所述用户数据的特征标签，所述第一等级用户的级别高于所述第二等级用户的级别，
    所述特征标签用于指示所述用户数据与购买行为的对应关系。
  5. 根据权利要求4所述的方法，其特征在于，所述用户数据包括属性数据和行为数据。
  6. 一种业务推荐的方法，包括：
    利用如权利要求1-5任一项所述的方法训练目标机器学习模型；
    获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据；
    分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值；
    在检测到所述用户数据对应的所述预测值大于第一预设阈值时，将所述用户数据对应的所述候选用户作为目标用户，并向所述目标用户推荐目标业务。
  7. 一种模型训练的装置，包括：
    样本数据集获取模块，用于获取样本数据集，其中，所述样本数据集包括第一类样本数据以及第二类样本数据；
    权重获取模块，用于获取所述第一类样本数据对应的第一权重，以及所述第二类样本数据对应的第二权重；
    总体损失函数确定模块，用于根据所述第一权重、所述第二权重、所述第一类样本数据对应的损失函数、和所述第二类样本数据对应的损失函数进行加权运算，得到总体损失函数；
    模型训练模块，用于基于所述总体损失函数，使用所述样本数据集进行机器学习模型的训练。
  8. 根据权利要求7的装置，其特征在于，所述权重获取模块包括：
    比例确定子模块，用于确定第一比例和第二比例，其中，所述第一比例为所述第一类样本数据的行为为指定行为的概率，所述第二比例为所述第二类样本数据的行为为指定行为的概率；
    权重作为子模块，用于将所述第一比例作为所述第一权重，将所述第二比例作为所述第二权重。
  9. 根据权利要求7所述的装置，其特征在于，所述权重获取模块包括：
    分类信息确定子模块，用于确定所述第一类样本数据和所述第二类样本数据的分类信息；
    权重匹配子模块，用于在预置的分类信息与权重映射关系中，匹配所述分类信息，得到所述第一类样本数据对应的所述第一权重和所述第二类样本数据对应的所述第二权重。
  10. 根据权利要求9所述的装置，其特征在于，
    所述样本数据集为针对酒旅类业务的用户数据集，
    所述第一类样本数据包括第一等级用户的用户数据以及所述用户数据的特征标签，
    所述第二类样本数据包括第二等级用户的用户数据以及所述用户数据的特征标签，
    所述第一等级用户的级别高于所述第二等级用户的级别，
    所述特征标签用于指示所述用户数据与购买行为的对应关系。
  11. 根据权利要求10所述的装置，其特征在于，所述用户数据包括属性数据和行 为数据。
  12. 一种业务推荐的装置，包括：
    模型训练模块，用于利用如权利要求1-5任一项所述的方法训练目标机器学习模型；
    候选用户列表获取模块，用于获取候选用户列表，其中，所述候选用户列表包括多个候选用户的用户数据；
    预测值计算模块，用于分别将每个候选用户的所述用户数据输入训练好的所述目标机器学习模型，获得每个候选用户的所述用户数据对应的预测值；
    业务推荐模块，用于在检测到所述用户数据对应的所述预测值大于第一预设阈值时，将所述用户数据对应的所述候选用户作为目标用户，并向所述目标用户推荐目标业务。
  13. 一种电子设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述程序时实现权利要求1至6任一项所述方法。
  14. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求1至6任一项所述方法的步骤。
PCT/CN2018/121950 2018-05-02 2018-12-19 模型训练和业务推荐 WO2019210695A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/077,416 US20210042664A1 (en) 2018-05-02 2020-10-22 Model training and service recommendation

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810411497.4A CN108875776B (zh) 2018-05-02 2018-05-02 模型训练方法和装置、业务推荐的方法和装置、电子设备
CN201810411497.4 2018-05-02

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/077,416 Continuation US20210042664A1 (en) 2018-05-02 2020-10-22 Model training and service recommendation

Publications (1)

Publication Number Publication Date
WO2019210695A1 true WO2019210695A1 (zh) 2019-11-07

Family

ID=64327115

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/121950 WO2019210695A1 (zh) 2018-05-02 2018-12-19 模型训练和业务推荐

Country Status (3)

Country Link
US (1) US20210042664A1 (zh)
CN (1) CN108875776B (zh)
WO (1) WO2019210695A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110968695A (zh) * 2019-11-18 2020-04-07 罗彤 基于弱监督技术主动学***台
WO2022022163A1 (zh) * 2020-07-30 2022-02-03 腾讯科技（深圳）有限公司 文本分类模型的训练方法、装置、设备及存储介质

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108875776B (zh) * 2018-05-02 2021-08-20 北京三快在线科技有限公司 模型训练方法和装置、业务推荐的方法和装置、电子设备
CN110008984B (zh) * 2019-01-22 2023-07-25 创新先进技术有限公司 一种基于多任务样本的目标欺诈交易模型训练方法和装置
CN110163245A (zh) * 2019-04-08 2019-08-23 阿里巴巴集团控股有限公司 业务类别预测方法及***
CN110481561B (zh) * 2019-08-06 2021-04-27 北京三快在线科技有限公司 无人驾驶车辆自动控制信号生成方法和装置
CN110533489B (zh) * 2019-09-05 2021-11-05 腾讯科技（深圳）有限公司 应用于模型训练的样本获取方法及装置、设备、存储介质
CN110732139B (zh) * 2019-10-25 2024-03-05 腾讯科技（深圳）有限公司 检测模型的训练方法和用户数据的检测方法、装置
US11442459B2 (en) * 2019-12-11 2022-09-13 Uatc, Llc Systems and methods for training predictive models for autonomous devices
CN111695036B (zh) * 2020-06-11 2024-03-08 北京百度网讯科技有限公司 内容推荐方法及装置
CN112597356B (zh) * 2020-12-02 2023-09-05 京东科技控股股份有限公司 模型训练方法、个性化数据推荐方法、装置及电子设备
CN112733729B (zh) * 2021-01-12 2024-01-09 北京爱笔科技有限公司 模型训练、回归分析的方法、装置、存储介质和设备
CN112925926B (zh) * 2021-01-28 2022-04-22 北京达佳互联信息技术有限公司 多媒体推荐模型的训练方法、装置、服务器以及存储介质
CN113191812B (zh) * 2021-05-12 2024-02-02 深圳索信达数据技术有限公司 业务推荐方法、计算机设备及计算机可读存储介质
CN113360777B (zh) * 2021-08-06 2021-12-07 北京达佳互联信息技术有限公司 内容推荐模型训练方法、内容推荐方法及相关设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105989374A (zh) * 2015-03-03 2016-10-05 阿里巴巴集团控股有限公司 一种在线训练模型的方法和设备
CN106778820A (zh) * 2016-11-25 2017-05-31 北京小米移动软件有限公司 识别模型确定方法及装置
CN107578294A (zh) * 2017-09-28 2018-01-12 北京小度信息科技有限公司 用户行为预测方法、装置及电子设备
CN107798390A (zh) * 2017-11-22 2018-03-13 阿里巴巴集团控股有限公司 一种机器学习模型的训练方法、装置以及电子设备
CN108875776A (zh) * 2018-05-02 2018-11-23 北京三快在线科技有限公司 模型训练方法和装置、业务推荐的方法和装置、电子设备

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2010092266A (ja) * 2008-10-08 2010-04-22 Nec Corp 学習装置、学習方法及びプログラム
CN103678329B (zh) * 2012-09-04 2018-05-04 中兴通讯股份有限公司 推荐方法及装置
CN104077306B (zh) * 2013-03-28 2018-05-11 阿里巴巴集团控股有限公司 一种搜索引擎的结果排序方法及***
US10956995B1 (en) * 2015-07-23 2021-03-23 Expedia, Inc. User-specific travel offers
CN107153630B (zh) * 2016-03-04 2020-11-06 阿里巴巴集团控股有限公司 一种机器学习***的训练方法和训练***
CN106296305A (zh) * 2016-08-23 2017-01-04 上海海事大学 大数据环境下的电商网站实时推荐***与方法
CN106776873A (zh) * 2016-11-29 2017-05-31 珠海市魅族科技有限公司 一种推荐结果生成方法以及装置
US10970553B2 (en) * 2017-11-15 2021-04-06 Uatc, Llc Semantic segmentation of three-dimensional data
CN107918922B (zh) * 2017-11-15 2020-10-27 中国联合网络通信集团有限公司 业务推荐方法及业务推荐装置
US11106896B2 (en) * 2018-03-26 2021-08-31 Intel Corporation Methods and apparatus for multi-task recognition using neural networks

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105989374A (zh) * 2015-03-03 2016-10-05 阿里巴巴集团控股有限公司 一种在线训练模型的方法和设备
CN106778820A (zh) * 2016-11-25 2017-05-31 北京小米移动软件有限公司 识别模型确定方法及装置
CN107578294A (zh) * 2017-09-28 2018-01-12 北京小度信息科技有限公司 用户行为预测方法、装置及电子设备
CN107798390A (zh) * 2017-11-22 2018-03-13 阿里巴巴集团控股有限公司 一种机器学习模型的训练方法、装置以及电子设备
CN108875776A (zh) * 2018-05-02 2018-11-23 北京三快在线科技有限公司 模型训练方法和装置、业务推荐的方法和装置、电子设备

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110968695A (zh) * 2019-11-18 2020-04-07 罗彤 基于弱监督技术主动学***台
WO2022022163A1 (zh) * 2020-07-30 2022-02-03 腾讯科技（深圳）有限公司 文本分类模型的训练方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN108875776B (zh) 2021-08-20
CN108875776A (zh) 2018-11-23
US20210042664A1 (en) 2021-02-11

Similar Documents

Publication Publication Date Title
WO2019210695A1 (zh) 模型训练和业务推荐
Sarstedt et al. Treating unobserved heterogeneity in PLS-SEM: A multi-method approach
Yuan et al. Marginalized average attentional network for weakly-supervised learning
CN110852447B (zh) 元学习方法和装置、初始化方法、计算设备和存储介质
WO2021204269A1 (zh) 分类模型的训练、对象分类
US11816183B2 (en) Methods and systems for mining minority-class data samples for training a neural network
Goodfellow On distinguishability criteria for estimating generative models
US10218716B2 (en) Technologies for analyzing uniform resource locators
CN108287864B (zh) 一种兴趣群组划分方法、装置、介质及计算设备
CN116261731A (zh) 基于多跳注意力图神经网络的关系学习方法与***
CN110659744A (zh) 训练事件预测模型、评估操作事件的方法及装置
CN108985190B (zh) 目标识别方法和装置、电子设备、存储介质
CN112733995B (zh) 训练神经网络的方法、行为检测方法及行为检测装置
CN110162609B (zh) 用于向用户推荐咨询问题的方法及装置
CN114298851A (zh) 基于图表征学习的网络用户社交行为分析方法、装置及存储介质
CN113779380B (zh) 跨域推荐、内容推荐方法、装置及设备
CN110008348A (zh) 结合节点和边进行网络图嵌入的方法和装置
CN111753151B (zh) 一种基于互联网用户行为的服务推荐方法
CN107070702B (zh) 基于合作博弈支持向量机的用户账号关联方法及其装置
CN112541010B (zh) 一种基于逻辑回归的用户性别预测方法
He et al. Semi-supervised dialogue abstractive summarization via high-quality pseudolabel selection
CN116342906A (zh) 一种跨域小样本图像识别方法及***
CN104572820B (zh) 模型的生成方法及装置、重要度获取方法及装置
US11601509B1 (en) Systems and methods for identifying entities between networks
CN107665244B (zh) 一种获取粮食产量与化肥使用量之间关系的方法及装置

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: 18917505

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18917505

Country of ref document: EP

Kind code of ref document: A1