WO2020143409A1 - 预测业务指标的方法及装置 - Google Patents

预测业务指标的方法及装置 Download PDF

Info

Publication number
WO2020143409A1
WO2020143409A1 PCT/CN2019/125856 CN2019125856W WO2020143409A1 WO 2020143409 A1 WO2020143409 A1 WO 2020143409A1 CN 2019125856 W CN2019125856 W CN 2019125856W WO 2020143409 A1 WO2020143409 A1 WO 2020143409A1
Authority
WO
WIPO (PCT)
Prior art keywords
tested
business
training
indicators
entity
Prior art date
Application number
PCT/CN2019/125856
Other languages
English (en)
French (fr)
Inventor
张志强
周俊
李小龙
Original Assignee
阿里巴巴集团控股有限公司
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 阿里巴巴集团控股有限公司 filed Critical 阿里巴巴集团控股有限公司
Publication of WO2020143409A1 publication Critical patent/WO2020143409A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q30/00Commerce
    • G06Q30/02Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201Market modelling; Market analysis; Collecting market data
    • G06Q30/0202Market predictions or forecasting for commercial activities

Definitions

  • One or more embodiments of this specification relate to the field of machine learning, and in particular, to a method and apparatus for predicting business indicators of business entities using machine learning.
  • the e-commerce scenario it is often necessary to predict various indicator sequences of the merchant, such as the forecast of the daily turnover sequence of the merchant.
  • Some scenarios will involve the joint prediction of multiple indicator sequences, such as predicting the daily transaction unit price and daily transaction number of the merchant at the same time. In this way, you can warn in advance or notice some abnormal phenomena as soon as possible, such as the phenomenon of order slips.
  • the Internet lending scenario it is often necessary to predict various business indicators of users to predict risks.
  • One or more embodiments of this specification describe a method and device for predicting business indicators, in which a recurrent neural network is used for joint modeling of multi-business entities and multi-business indicators, and the resulting joint model can be applied to the prediction of multi-entity and multi-index. Improve the accuracy and efficiency of multi-index forecasting.
  • a computer-implemented method for predicting business indicators including:
  • the data set sequence includes m data sets corresponding to m consecutive time periods and arranged in chronological order, the i data set of the m data sets includes, in the i time period, The entity characteristics of the business entity to be tested and the respective index values of the plurality of business indicators to be tested of the business entity to be tested;
  • the business entity to be tested includes a merchant; correspondingly, the entity characteristics include at least one of the following: registration duration, merchant size, reputation level, number of commodities, commodity category, shelf time;
  • the business indicators to be tested include multiple of the following indicators: the number of transactions, the unit price of transactions, the total value of transactions, and the percentage of transactions.
  • the business entity to be tested includes a user; correspondingly, the plurality of business indicators to be tested include a plurality of indicators related to consumption behavior and/or a plurality of indicators related to borrowing behavior.
  • the entity characteristics of the business entity to be tested include a first characteristic and a second characteristic, the characteristic value of the first characteristic remains unchanged during the m time periods, and the second The feature value of the feature is not completely the same in the m time periods.
  • the m time periods include the current time period where the current moment is located, and m-1 time periods retrospectively forward from the current time period.
  • the recurrent neural network includes a long- and short-term memory neural network LSTM.
  • the recurrent neural network is pre-trained through the following process:
  • the first training sample includes a first training data sequence and a first label
  • the first training data sequence includes corresponding to m consecutive time periods, and according to time M training data sets arranged in sequence
  • the jth training data set in the m training data sets includes, during the jth time period, the entity characteristics of the first business entity, and the plurality of pending data of the first business entity Measuring the respective index values of the service indicators
  • the first label includes the respective index values of the collected multiple service indicators to be tested of the first service entity in the next time period after the m-th time period;
  • the recurrent neural network includes an input layer, a recurrent unit and an output layer; inputting the first training data sequence into the recurrent neural network and obtaining the first output result includes:
  • the m training data sets are sequentially input to the loop unit through the input layer, so that the loop unit sequentially processes the m training data sets to obtain a processing vector after m times of processing;
  • the loss function is determined by:
  • the sum of the total error and the regularization term is determined as the loss function.
  • a computer-executed device for predicting business indicators including:
  • a data sequence acquisition unit configured to acquire a data set sequence, the data set sequence includes m data sets corresponding to m consecutive time periods and arranged in chronological order, and the i data set of the m data sets includes In the i-th time period, the entity characteristics of the business entity to be tested and the respective index values of the plurality of business indicators to be tested of the business entity to be tested;
  • the output result obtaining unit is configured to input the data set sequence into a pre-trained recurrent neural network to obtain an output result
  • the indicator value determining unit is configured to determine the indicator values of the plurality of service indicators to be tested of the service entity to be tested in the next time period of the mth time period according to the output result.
  • the recurrent neural network is trained by a training unit, and the training unit includes:
  • the sample acquisition module is configured to acquire multiple training samples, including a first training sample, the first training sample includes a first training data sequence and a first label, and the first training data sequence includes m corresponding to consecutive M training data sets arranged in time sequence and in chronological order, the jth training data set in the m training data sets includes, during the jth time period, the entity characteristics of the first business entity and the first business entity
  • the first label includes the collected next time period of the plurality of service indicators to be tested of the first service entity after the mth time period
  • a result acquisition module configured to input the first training data sequence into the recurrent neural network to obtain a first output result
  • An error determination module configured to determine a first prediction error corresponding to the first training sample based on at least the first output result and the first label;
  • the loss determination module is configured to determine the loss function based on the prediction error corresponding to each of the plurality of training samples
  • the parameter adjustment module is configured to adjust the network parameters of the training neural network according to the loss function, so that the loss function is reduced compared to before the adjustment.
  • a computer-readable storage medium on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to perform the method of the first aspect.
  • a computing device including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method of the first aspect is implemented .
  • a joint model can be obtained.
  • the joint model can be applied to the prediction of multiple test business indicators of multiple test business entities. Because the training data based on the training process includes both the features of the entity and the index data, the joint model obtained by the training better considers and mines the association relationship between entities/indicators, and because of the rich training data, it also makes The prediction accuracy is higher and the effect is better.
  • FIG. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification
  • FIG. 2 shows a flowchart of a method for training a joint model based on a recurrent neural network according to an embodiment
  • FIG. 3 shows a schematic diagram of processing a training data sequence through an LSTM neural network according to an embodiment
  • FIG. 5 shows a schematic block diagram of an apparatus for predicting business indicators according to an embodiment.
  • joint prediction is proposed for multiple index sequences.
  • the efficiency of joint modeling is often too low due to too many entities (such as hundreds of millions of merchants involved in Alipay transactions), and ultimately only multiple indicators for a single entity For modeling.
  • a feature-based joint modeling method is proposed. According to this method, the use of a unified model for modeling multiple business entities and multiple business indicators not only avoids the problems of insufficient training data and too many models for modeling a single entity and a single index, but also fully considers the inter-entity/ The correlation between indicators makes modeling faster and more accurate predictions.
  • FIG. 1 is a schematic diagram of an implementation scenario disclosed in this specification.
  • historical data about multiple business indicators from multiple business entities is collected as training samples to train a joint model based on recurrent neural networks.
  • a training sample may correspond to a business entity, such as a merchant, and the feature data used for training by the sample includes the entity characteristics of the business entity in m consecutive time periods, and the to-be-measured in the m time periods The respective indicator values of multiple business indicators.
  • These feature data are organized into a data set in the form of m training vectors according to the above time period.
  • the label data of the training sample may be the index values of the multiple business indicators in the m+1th time period of the business entity.
  • the aforementioned joint model based on the recurrent neural network can be trained to determine model parameters.
  • the prediction of multi-service entities and multi-service indicators can be achieved.
  • the feature data of the business entity p is obtained.
  • the feature data may include a sequence of m feature vectors, and each feature vector includes the corresponding time period of the business entity p in m time periods.
  • FIG. 2 shows a flowchart of a method for training a joint model based on a recurrent neural network according to an embodiment. It can be understood that the method can be performed by any device, device, platform, or device cluster with computing and processing capabilities. As shown in Figure 2, the training process includes at least the following steps.
  • step 21 multiple training samples are obtained, the multiple training samples involving historical data of multiple business entities on multiple business indicators, more specifically, where each training sample includes training data related to one business entity Sequences and labels.
  • each training sample includes training data related to one business entity Sequences and labels.
  • any one of the samples is called the first training sample. It should be understood that the terms “first” and “second” in this document are only for describing and distinguishing similar objects, and are not limited by order or other aspects.
  • any first training sample in the plurality of training samples includes a first training data sequence and a first label related to the first business entity, where the first training data sequence includes m consecutive time periods corresponding to, and M training data sets arranged in chronological order, the jth training data set includes, in the jth time period, the entity characteristics of the first business entity, and the respective index values of the plurality of business indicators to be tested of the first business entity
  • the first label includes the collected respective index values of the plurality of service indicators to be tested of the first service entity in the next time period after the m-th time period.
  • the above first business entity is denoted as E1
  • the corresponding first training data sequence may be denoted as (D1, D2, ..., Dm), which includes m training data sets D1, D2 to Dm.
  • the m training data sets respectively correspond to m consecutive time periods
  • the jth training data set Dj corresponds to the jth time period Tj among the m time periods.
  • Each training data set contains two parts: entity features and index values to be tested.
  • the jth training data set Dj includes, in the jth time period Tj, the entity characteristic X of the first business entity E1, and the indicators of the k business indicators to be predicted by the business entity E1 in the Tj time period Value Y.
  • the entity characteristic X and the index value Y can respectively constitute a vector x (j) and a vector y (j) , where y (j) is a k-dimensional vector.
  • the jth training data set Dj can be expressed as the concatenation of vectors x (j) and y (j) [x (j) ; y (j) ].
  • the first training data sequence can be expressed as:
  • the first training sample also includes a first label corresponding to the first training data sequence.
  • the first label includes the respective index values of the collected k service indicators to be tested of the first service entity E1 in the m+1th time period. Therefore, the first label can be written as a vector y (m+1) .
  • the following is a description with specific examples.
  • the joint model to be trained is applicable to various business entities. Therefore, the first business entity as training data may be any specific business entity, such as a specific merchant or a specific user.
  • the entity characteristic X of the business entity may include the overall attribute characteristic of the merchant, such as the registration time, the scale of the merchant, the reputation level, etc., and may also include the related products sold by the merchant Features, such as the number of commodities, category of commodities, length of time on shelves, etc
  • some of the features may be fixed features, such as the attribute features of the merchant, and these features are the same in the m time periods collected; there may also be some features that change with time, such as commodities number. Such features are not exactly the same in the m time periods collected.
  • the k business indicators to be tested may include multiple of the following indicators: number of transactions, unit price of transactions, total transaction amount, transaction ratio, etc.
  • the length of the above time period can be set as required.
  • the time period can be one day, one hour, one week, one month, etc.
  • the number of m can also be set according to the needs of training and prediction, for example, m is 7, 10, 30, etc.
  • the first business entity in the first training sample is a merchant with a merchant ID equal to ID1
  • the entity feature X of the business entity selects 20 features such as registration duration, merchant size, reputation level, number of products, and product category.
  • the first training data sequence for the first business entity is 7 training data sets corresponding to 7 consecutive days
  • the training data set on the jth day includes the 20 dimensions composed of the entity's entity characteristics on the jth day Vector, and a 3-dimensional vector composed of the index values of the three business indicators to be tested on the jth day of the merchant.
  • the label data of the first training sample is the index value of the above-mentioned three business indicators to be tested on the 8th day.
  • Such a training sample is equivalent to using the training data sequence of 7 days a week to predict the business indicators to be tested on the 8th day.
  • the first business entity is a merchant.
  • the first business entity may also be a specific user.
  • the extracted entity characteristics may include the attribute characteristics of the user, such as registration time, gender, age, occupation category, income range, etc.; and may also include some behavior characteristics of the user , For example, the number of interactions with friends, the length of time spent browsing shopping sites, and so on.
  • the user's test business indicators are mainly determined according to the business needs and business scenarios.
  • the test business indicators can be multiple indicators related to the consumption behavior, for example, including the number of consumption, single The amount of consumption, the total amount of consumption, etc.; in the scenario of analyzing and predicting the user's borrowing behavior, the business indicators to be tested can be multiple indicators related to the borrowing behavior, such as the number of loans, total borrowing, total repayment, etc. Wait.
  • the first training sample described above is any one of a plurality of training samples.
  • Each training sample has a training data sequence and corresponding label with the same data structure as the first training sample.
  • step 22 the first training data sequence in the first training samples is input into the recurrent neural network to obtain a first output result.
  • Recurrent Neural Networks is a time-recurrent neural network that can be used to process sequence data.
  • the recurrent neural network includes an input layer, a recurrent unit, and an output layer; the input layer is used to obtain sequence data, and the recurrent unit corresponds to one or more hidden layers of the neural network, which is used to sequentially process the sequence
  • the output layer can further transform the processing vector obtained by the cyclic unit through cyclic processing to obtain the output result of the entire neural network.
  • the time recursiveness of the recurrent neural network is mainly reflected in the processing of the recurrent unit, that is, the hidden layer.
  • each neuron node of the hidden layer will remember the previous information and apply it to the calculation of the current output.
  • the input of the hidden layer neuron includes not only the input from the input layer, but also the state output of the hidden layer neuron at the previous moment.
  • the output of the hidden layer neuron is not only the input of the input layer of the current moment.
  • the vector of is related to the previous state of the node. In other words, the state of the hidden layer neuron at time t can be expressed as:
  • X t is the state of the input layer at time t
  • S t-1 is the state of the hidden layer neuron at t-1
  • f is the calculation function
  • W and U are the weights.
  • the loop unit loops the previous state back to the current input, taking into account the influence of the historical input, so it is suitable for data sequences with time series.
  • step 22 an RNN containing multiple hidden layers is used as the recurrent neural network.
  • LSTM Long Short-Term Memory
  • St f (U*X t +W1*S t-1 +W2*S t-2 +...+Wn*S tn ).
  • the LSTM model is proposed to solve the problem of long-term dependence.
  • LSTM some information that is no longer needed is discarded by setting a "forgetting gate” that allows information to pass selectively, so that unnecessary interference information input is judged and shielded, so that long-term data is better analyzed and processed .
  • the LSTM neural network includes an input layer, an LSTM unit, and an output layer, where the LSTM unit serves as a cyclic unit, and may contain several hidden layers, and sequentially processes each data set in the first training data sequence.
  • FIG. 3 shows a schematic diagram of processing a training data sequence through an LSTM neural network according to an embodiment.
  • the first training data sequence is sequentially input to the LSTM unit through the input layer, so that the LSTM unit sequentially processes each training data set in the first training data sequence.
  • the first training data sequence includes m training data sets arranged in the order of corresponding time periods, where the j-th training data set Dj can be expressed as vectors x (j) and y ( j) Splicing [x (j) ; y (j) ], where x (j) is the entity feature and y (j) is the index value of the business index to be tested.
  • the LSTM unit reads and processes the m training data sets in sequence according to the order of the data sets.
  • the LSTM unit processes D1 and h 0 to obtain a processing vector h 1 .
  • the LSTM unit uses the mth training data set and the processing result h m-1 of the m-1 training data set at the previous moment as the current input to obtain the processing vector h m .
  • the output layer linearly transforms the processing vector h m obtained by the LSTM unit after m times of processing, to obtain the first output result:
  • the first output result is based on the first training data sequence corresponding to the m consecutive time periods in the first training sample, and predicts the index value of the service indicator to be measured in the m+1th time period.
  • the first training sample acquired in step 21 also includes a first label corresponding to the first training data sequence. Therefore, next, at step 23, a first prediction error corresponding to the first training sample is determined based on at least the first output result and the first label.
  • the above-mentioned first prediction error can be determined by various specific algorithms.
  • both the first output result and the first label can be represented as k-dimensional vectors (as mentioned above, k is the number of business indicators to be tested), therefore, in an example, the first output result corresponding vector and the first
  • the distance between the corresponding vectors of the label takes this distance as the first prediction error.
  • the square of the above distance can also be used as the first prediction error.
  • the above prediction error may also be calculated in other ways.
  • the first prediction error corresponding to the first training sample is determined. It can be understood that the above first training sample is any one of a plurality of training samples. Therefore, each training sample can be processed in the manner of steps 22 and 23 to obtain the prediction error corresponding to each training sample among multiple training samples.
  • the loss function is determined based at least on the prediction error corresponding to each sample.
  • the loss function is a comprehensive measure of the error between the model prediction value of the batch of training samples and the standard label value. Therefore, in one embodiment, the total error is determined based on the prediction error corresponding to each training sample, and the loss function is determined based on the total error.
  • the loss function can be expressed as:
  • the prediction error of the i-th training sample is expressed as The loss function is obtained by summing the prediction errors of n training samples.
  • the regularization term of the network parameters of the recurrent neural network is also determined, and the sum of the total error and the regularization term , Determined as the loss function.
  • the loss function can be expressed as:
  • Reg( ⁇ ) is a regularization term, for example, it may be an L2 regularization term.
  • L2 regularization can be expressed as It can be obtained by summing the squares of all network parameters ⁇ (wi) 2 . Therefore, the regularization term can characterize the overall value range of the network parameters.
  • step 25 the network parameters of the recurrent neural network are adjusted according to the above-mentioned loss function, so that the loss function is reduced compared to before the adjustment.
  • the training goal of the neural network is to make the loss function as small as possible, so that the predicted value is as close as possible to the true label value.
  • the various network parameters are adjusted by means of error back propagation to minimize the loss function.
  • the regularization term describes the overall value range of the network parameters.
  • the overall value range of the network parameters will also be optimized to a smaller range, thereby avoiding overfitting where the parameter value difference is too large phenomenon.
  • FIG. 4 illustrates a method for predicting business indicators according to an embodiment.
  • the method can be performed by any device, device, platform, or device cluster with computing and processing capabilities.
  • the prediction method includes the following steps.
  • a data set sequence for a business entity to be tested is obtained, the data set sequence includes m data sets corresponding to m consecutive time periods and arranged in chronological order, the i data set in the m data sets
  • the data set includes, in the i-th time period, the entity characteristics of the business entity to be tested, and the respective index values of the plurality of business indicators to be tested of the business entity to be tested.
  • the aforementioned data set sequence for prediction has a data structure corresponding to the training data sequence in the training process. Specifically, the length of the time period in the above data set sequence, the number of m needs to be the same as the training data sequence, the entity characteristics of the business entity to be tested and the entity characteristics of the business entity in the training sample are in the same feature space, and The business indicators to be tested are also the same as the training samples.
  • the business entities in multiple training samples are various businesses.
  • the entity features include 20 features related to the business
  • the business indicators to be tested include 3 indicators
  • the time period is days
  • m is set to 7.
  • the business entity to be tested may be a specific business vendor to be tested Ep
  • the data set sequence of step 41 may include 7 data sets corresponding to 7 consecutive days
  • the data set of the i-th day includes the business vendor Ep
  • the 20-dimensional vector composed of the physical characteristics of the entity includes, for example, the registration time, the size of the merchant, the reputation level, the number of commodities, and the category of the commodity, as well as the 3-dimensional vector composed of the index values of the business indicators of the merchant Ep to be tested on the i-th day.
  • the business entity in the training sample is each user
  • the above-mentioned business entity to be tested corresponds to a specific user to be tested.
  • the entity features of the business entity to be tested may include features that remain unchanged in m time periods, or may include features that are not completely the same in m time periods.
  • the above m time periods may also correspond to the elapsed time period, which can be applied to the real value of the business indicator to be measured in the m+1th time period, which is obtained through the joint model The predicted value, and then through the difference between the predicted value and the real value, determine whether the real value is abnormal.
  • step 42 the data set sequence is input to the recurrent neural network trained by the method of FIG. 2 to obtain an output result.
  • the aforementioned recurrent neural network includes a long-term short-term memory neural network LSTM.
  • the recurrent neural network includes an input layer, a recurrent unit, and an output layer.
  • the m data sets in the data set sequence are sequentially input to the cyclic unit through the input layer, so that the cyclic unit processes the m data sets in sequence to obtain the processing vector after m times of processing.
  • the circulation unit i-th data set, and the processing result of the previous time i-1 th data sets H i-1, as co-current input the processing to obtain the vector h i.
  • the loop unit uses the mth data set and the processing result h m-1 of the m-1 data set at the previous moment as the current input to obtain the processing vector h m after the mth processing .
  • the output vector can be obtained by linearly transforming the processing vector h m through the output layer.
  • the network parameters of the above processing including the linear transformation matrix W, are determined and optimized through the training process.
  • step 43 based on the above output result, it can be determined that, in the next time period after the mth time period, that is, in the m+1th time period, the multiple business indicators to be tested of the business entity to be tested The respective index value.
  • the cyclic neural network obtains the output result by linearly transforming the processing vector of the cyclic unit, and the output result after the linear transformation is usually a k-dimensional vector. According to each element value of the k-dimensional vector, the corresponding index values of the k service indicators to be tested in the m+1th time period can be determined.
  • a joint model can be obtained.
  • the joint model can be applied to the prediction of multiple test business indicators of multiple test business entities. Because the training data based on the training process includes both the features of the entity and the index data, the joint model obtained by the training better considers and mines the association relationship between entities/indicators, and because of the rich training data, it also makes The prediction accuracy is higher and the effect is better.
  • FIG. 5 shows a schematic block diagram of an apparatus for predicting business indicators according to an embodiment.
  • the prediction device 500 includes:
  • the data sequence acquisition unit 51 is configured to acquire a data set sequence including m data sets corresponding to m consecutive time periods and arranged in chronological order, and the i data set in the m data sets Including, in the i-th time period, the entity characteristics of the business entity to be tested, and the respective index values of the plurality of business indicators to be tested of the business entity to be tested;
  • the output result obtaining unit 52 is configured to input the data set sequence into a pre-trained recurrent neural network to obtain an output result
  • the indicator value determination unit 53 is configured to determine the indicator values of the plurality of service indicators to be tested of the service entity to be tested in the next time period of the m-th time period according to the output result.
  • the business entity to be tested includes a merchant; in such a case, the entity characteristics include at least one of the following: registration duration, merchant size, reputation level, number of products, product category, shelf time;
  • the business indicators to be tested include multiple of the following indicators: the number of transactions, the unit price of transactions, the total value of transactions, and the proportion of transactions.
  • the above-mentioned business entity to be tested includes a user; the plurality of business indicators to be tested include a plurality of indicators related to consumption behavior, and/or a plurality of indicators related to borrowing behavior.
  • the entity characteristics of the business entity to be tested include a first characteristic and a second characteristic, the characteristic value of the first characteristic remains unchanged during the m time periods, and the second The feature value of the feature is not completely the same in the m time periods.
  • the m time periods include the current time period at the current moment, and m-1 time periods retrospectively forward from the current time period.
  • the aforementioned recurrent neural network includes a long-short-term memory neural network LSTM.
  • the recurrent neural network is trained by a training unit 54, and the training unit may be included in the prediction device 500 or may be located outside the prediction device 500.
  • the output result obtaining unit 52 can be connected to the training unit 54.
  • the training unit 54 includes:
  • the sample acquisition module 541 is configured to acquire multiple training samples, including a first training sample, the first training sample includes a first training data sequence and a first label, and the first training data sequence includes a corresponding m M training data sets arranged in chronological order over a period of time, the jth training data set in the m training data sets includes, during the jth time period, the entity characteristics of the first business entity and the first service
  • the first tag includes the next time after the mth time period of the collected plurality of business indicators to be tested of the first business entity The respective index value in the segment;
  • the result obtaining module 542 is configured to input the first training data sequence into the recurrent neural network to obtain a first output result
  • the error determination module 543 is configured to determine a first prediction error corresponding to the first training sample based on at least the first output result and the first label;
  • the loss determination module 544 is configured to determine a loss function based on the prediction error corresponding to each of the plurality of training samples
  • the parameter adjustment module 545 is configured to adjust the network parameters of the training neural network according to the loss function, so that the loss function is reduced compared to before adjustment.
  • the recurrent neural network includes an input layer, a recurrent unit and an output layer,
  • the above result obtaining module 542 is configured as:
  • the m training data sets are sequentially input to the loop unit through the input layer, so that the loop unit sequentially processes the m training data sets to obtain a processing vector after m times of processing;
  • the loss determination module 544 is configured as,
  • the sum of the total error and the regularization term is determined as the loss function.
  • the training unit 54 trains the recurrent neural network as a joint model, and the prediction device 500 realizes the prediction of the multi-service entity and multi-service indicators by using the recurrent neural network trained by the training unit 54.
  • a computer-readable storage medium on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to perform the method described in conjunction with FIGS. 2 and 4.
  • a computing device including a memory and a processor, where executable code is stored in the memory, and when the processor executes the executable code, the implementation is combined with FIG. 2 and FIG. 4 The method.

Abstract

一种计算机执行的预测业务指标的方法,首先获取数据集序列,该数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值。然后将该数据集序列输入预先训练的循环神经网络,获得输出结果;如此,根据输出结果,即可确定第m时间段的下一时间段中,待测业务实体的多个待测业务指标各自的指标值。

Description

预测业务指标的方法及装置 技术领域
本说明书一个或多个实施例涉及机器学习领域,尤其涉及利用机器学习预测业务实体的业务指标的方法和装置。
背景技术
随着计算机技术的发展,机器学习已经应用到各种各样的技术领域,用于分析、预测各种业务数据。在许多应用场景中,需要对业务实体的各种业务指标进行预测,便于在预测值出现较大波动,或预测值与真实值差异较大的情况下,提前预警,或发现异常。
例如,在电商场景中,常常需要对商家的各种指标序列进行预测,如商家的每日营业额序列预测。有些场景会涉及到多个指标序列的联合预测,比如同时预测商家的每日交易笔单价,每日交易笔数等,如此,可以提前预警或尽快注意到一些异常现象,例如刷单现象。又例如,在互联网借贷场景中,常常需要对用户的各种业务指标进行预测,从而预知风险。
然而,常规技术中,面对多个业务实体,多种业务指标的情况,常常存在建模效率低,预测准确率低等问题。因此,希望能有改进的方案,更加有效地实现多业务指标的准确预测。
发明内容
本说明书一个或多个实施例描述了预测业务指标的方法和装置,其中采用循环神经网络针对多业务实体多业务指标进行联合建模,得到的联合模型可以适用于多实体多指标的预测,全面提高多指标预测的准确性和效率。
根据第一方面,提供了一种计算机执行的预测业务指标的方法,包括:
获取数据集序列,所述数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值;
将所述数据集序列输入预先训练的循环神经网络,获得输出结果;
根据所述输出结果,确定第m时间段的下一时间段中,所述待测业务实体的所述多个待测业务指标各自的指标值。
在一个实施例中,上述待测业务实体包括商家;相应的,所述实体特征包括以下中的至少一项:注册时长、商家规模、信誉等级、商品数目、商品类别、上架时长;所述多个待测业务指标包括以下指标中的多个:交易笔数、交易笔单价、交易总额、成交比例。
在另一实施例中,所述待测业务实体包括用户;相应的,所述多个待测业务指标包括,与消费行为相关的多个指标和/或与借贷行为的相关的多个指标。
在一种实施方式中,所述待测业务实体的实体特征包括,第一特征和第二特征,所述第一特征的特征值在所述m个时间段中保持不变,所述第二特征的特征值在所述m个时间段中不完全相同。
根据一个实施例,所述m个时间段包括,当前时刻所处的当前时间段,和从所述当前时间段向前追溯的m-1个时间段。
在一个实施例中,所述循环神经网络包括,长短期记忆神经网络LSTM。
根据一种实施方式,所述循环神经网络通过以下过程预先训练:
获取多个训练样本,其中包括第一训练样本,所述第一训练样本包括第一训练数据序列和第一标签,所述第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,所述m个训练数据集中第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的所述多个待测业务指标各自的指标值;所述第一标签包括,采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值;
将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果;
至少基于所述第一输出结果和所述第一标签,确定与第一训练样本对应的第一预测误差;
基于与所述多个训练样本中各个训练样本对应的预测误差,确定损失函数;
根据所述损失函数调整所述训练神经网络的网络参数,使得所述损失函数相比调整前减小。
根据一个实施例,所述循环神经网络包括输入层、循环单元和输出层;将所述第 一训练数据序列输入所述循环神经网络,获得第一输出结果包括:
通过所述输入层将所述m个训练数据集依次输入到所述循环单元,使得所述循环单元依次处理所述m个训练数据集,得到m次处理后的处理向量;
通过所述输出层对所述处理向量进行线性变换,得到所述第一输出结果。
在一个实施例中,通过以下方式确定损失函数:
基于所述多个样本中各个样本对应的预测误差确定总误差;
确定所述网络参数的正则化项;
将所述总误差和所述正则化项的和,确定为所述损失函数。
根据第二方面,提供一种计算机执行的预测业务指标的装置,包括:
数据序列获取单元,配置为获取数据集序列,所述数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值;
输出结果获取单元,配置为将所述数据集序列输入预先训练的循环神经网络,获得输出结果;
指标值确定单元,配置为根据所述输出结果,确定第m时间段的下一时间段中,所述待测业务实体的所述多个待测业务指标各自的指标值。
在一个实施例中,所述循环神经网络通过训练单元训练,所述训练单元包括:
样本获取模块,配置为获取多个训练样本,其中包括第一训练样本,所述第一训练样本包括第一训练数据序列和第一标签,所述第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,所述m个训练数据集中第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的所述多个待测业务指标各自的指标值;所述第一标签包括,采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值;
结果获取模块,配置为将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果;
误差确定模块,配置为至少基于所述第一输出结果和所述第一标签,确定与第一 训练样本对应的第一预测误差;
损失确定模块,配置为基于与所述多个训练样本中各个训练样本对应的预测误差,确定损失函数;
参数调整模块,配置为根据所述损失函数调整所述训练神经网络的网络参数,使得所述损失函数相比调整前减小。
根据第三方面,提供了一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行第一方面的方法。
根据第四方面,提供了一种计算设备,包括存储器和处理器,其特征在于,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现第一方面的方法。
根据本说明书实施例提供的方法和装置,通过利用多个业务实体的实体特征和待测业务指标的已知值的数据序列作为训练样本,训练循环神经网络,可以得到一个联合模型。该联合模型可以适用于多个待测业务实体的多个待测业务指标的预测。由于训练过程中所基于的训练数据,即包含实体的特征又包含指标数据,因此训练得到的联合模型更好地考虑和挖掘了实体间/指标间的关联关系,并且由于训练数据丰富,也使得预测的准确度更高,效果更好。
附图说明
为了更清楚地说明本申请实施例的技术方案,下面将对实施例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申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1为本说明书披露的一个实施例的实施场景示意图;
图2示出根据一个实施例的训练基于循环神经网络的联合模型的方法流程图;
图3示出根据一个实施例的通过LSTM神经网络处理训练数据序列的示意图;
图4示出根据一个实施例的预测业务指标的方法;
图5示出根据一个实施例的预测业务指标的装置的示意性框图。
具体实施方式
下面结合附图,对本说明书提供的方案进行描述。
如前所述,在多种应用场景中,都需要对业务实体的多个业务指标进行预测。在一种方案中,考虑到对于不同的实体(例如不同的商家)的同一指标预测模型应该具有个性化的差异,针对单个实体单个指标进行单独建模。然而,只对单一的指标序列进行建模预测,难以充分挖掘序列间的关联关系,并且,单个实体单个指标单独建模会面临这单个模型训练数据不足和模型太多的问题。
在另一种方案中,提出针对多个指标序列进行联合预测。然而,面对多个实体多个指标的预测问题,往往会因为实体数量太多(如支付宝交易涉及的商家上亿)而导致联合建模效率太低,最终只能针对单个实体的多个指标进行建模。
针对多实体多指标序列联合预测的问题,进一步的,在本说明书中的实施例中,提出一种基于特征的联合建模方法。根据该方法,对多个业务实体多个业务指标使用统一模型进行建模,既避免了对单个实体单个指标建模遇到的训练数据不足和模型太多的问题,也充分考虑到了实体间/指标间的关联,使得建模更快,预测更准。
图1为本说明书披露的一个实施例的实施场景示意图。如图1所示,采集来自多个业务实体、关于多个业务指标的历史数据作为训练样本,训练基于循环神经网络的联合模型。更具体的,一个训练样本可以对应一个业务实体,例如一个商家,该样本用于训练的特征数据包括,该业务实体在m个连续时间段的实体特征,以及在该m个时间段中待测的多个业务指标各自的指标值。这些特征数据按照上述时间段整理成m个训练向量形式的数据集。该训练样本的标签数据可以是,上述业务实体在第m+1个时间段中,上述多个业务指标的指标值。
基于这样的多个训练样本,可以训练上述基于循环神经网络的联合模型,确定模型参数。
在模型训练完成后,就可以实现多业务实体、多业务指标的预测。具体的,对于待测的业务实体p,获取该业务实体p的特征数据,该特征数据可以包括m个特征向量构成的序列,各个特征向量包括,业务实体p在m个时间段的对应时间段中的实体特征,以及待测的多个业务指标各自的指标值。将这样的特征数据输入上述训练好的联合模型,即可得到,该业务实体p在下一时间段中,待测的多个业务指标的指标值,从而实现多个指标共同预测。
下面首先描述上述联合模型的训练过程。
图2示出根据一个实施例的训练基于循环神经网络的联合模型的方法流程图。可以理解,该方法可以通过任何具有计算、处理能力的装置、设备、平台、设备集群来执行。如图2所示,训练过程至少包括以下步骤。
首先,在步骤21,获取多个训练样本,所述多个训练样本涉及多个业务实体关于多个业务指标的历史数据,更具体的,其中每个训练样本包括与一个业务实体相关的训练数据序列和标签。为了描述的清楚,将其中任意的一个样本称为第一训练样本。需要理解,本文中的“第一”、“第二”的用语仅仅是为了描述和区分类似的对象,而不具有顺序或其他方面的限定。
具体而言,多个训练样本中任意的第一训练样本包括与第一业务实体相关的第一训练数据序列和第一标签,其中第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的多个待测业务指标各自的指标值;所述第一标签包括采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值。
更清楚的,将上述第一业务实体记为E1,对应的第一训练数据序列可以记为(D1,D2,…,Dm),其中包括了m个训练数据集D1,D2到Dm。这m个训练数据集分别对应于连续的m个时间段,其中第j个训练数据集Dj对应于m个时间段中的第j时间段Tj。每个训练数据集都包含实体特征和待测指标值两部分。具体地,第j个训练数据集Dj包括,在第j时间段Tj中,上述第一业务实体E1的实体特征X,以及该业务实体E1在Tj时间段中待预测的k个业务指标的指标值Y。实体特征X和指标值Y分别可以构成向量x (j),和向量y (j),其中y (j)为k维向量。第j训练数据集Dj可以表示为,向量x (j)和y (j)的拼接[x (j);y (j)]。如此,第一训练数据序列可以表示为:
{[x (1);y (1)],[x (2);y (2)],...,[x (j);y (j)],...,[x (m);y (m)]}
第一训练样本还包括与该第一训练数据序列相对应的第一标签。与前述第一训练数据序列的表示方式对应的,第一标签包括,采集到的第一业务实体E1的所述k个待测业务指标在第m+1时间段中各自的指标值。因此,第一标签可以记为向量y (m+1)。下面结合具体例子进行描述。
如前所述,要训练的联合模型适用于各种业务实体,因此,上述作为训练数据的 第一业务实体可以是任意的具体的业务实体,例如是一个具体商家,或者是一个具体用户。
在上述第一业务实体为某个商家的情况下,业务实体的实体特征X可以包括该商家的总体属性特征,例如注册时长、商家规模、信誉等级等,也可以包括该商家所售卖的商品相关特征,例如商品数目、商品类别、上架时长,等等。
上述这些特征中,一部分特征可以是固定不变的特征,例如商家的属性特征,这些特征在所采集的m个时间段中是相同的;还可以有一部分特征是随时间变化的特征,例如商品数目。这样的特征在所采集的m个时间段中不完全相同。
在上述第一业务实体为某个商家的情况下,k个待测业务指标可以包括以下指标中的多个:交易笔数、交易笔单价、交易总额、成交比例等等。
上述时间段的长短可以根据需要设定。例如,时间段可以是一天、1小时、1周、一个月等等,m的数目也可以根据训练和预测的需要而设定,例如m为7、10、30等等。
在一个具体例子中,第一训练样本中的第一业务实体为商户标识等于ID1的商家,业务实体的实体特征X选取注册时长、商家规模、信誉等级、商品数目、商品类别等20项特征,待测业务指标Y包括交易笔数、交易笔单价、交易总额3项指标,即k=3。并且,假定时间段设定为天,m设定为7。如此,针对该第一业务实体的第一训练数据序列即为,连续7天对应的7个训练数据集,第j天的训练数据集包括,第j天中该商家的实体特征构成的20维向量,和该商家在第j天中3个待测业务指标的指标值构成的3维向量。
对应的,第一训练样本的标签数据,即为上述商家在第8天中上述3个待测业务指标的指标值。这样的训练样本,相当于利用一周7天的训练数据序列,预测第8天的待测业务指标。
以上结合了第一业务实体为商家的例子进行了描述。在另一实施例中,第一业务实体也可以是具体的用户。在第一业务实体为某个用户的情况下,提取的实体特征可以包括该用户的属性特征,例如注册时长、性别、年龄、职业类别、收入范围等等;也可以包括该用户的一些行为特征,例如,与好友的交互次数,浏览购物网站的时长,等等。用户的待测业务指标主要根据业务需求和业务场景而定,在分析和预测用户的消费行为的场景下,待测业务指标可以是与消费行为相关的多个指标,例如包括消费笔数、单笔 消费金额、消费总金额等等;在分析和预测用户的借贷行为的场景下,待测业务指标可以是与借贷行为的相关的多个指标,例如包括借贷次数、借贷总额、还款总额等等。
尽管以上给出了业务实体、实体特征和业务指标的具体例子,但是,本说明书实施例的方案并不局限于这些具体例子,而是可以扩展到更多业务场景。
并且,可以理解,以上所描述的第一训练样本是多个训练样本中的任意一个。每个训练样本都具有上述第一训练样本同样数据结构的训练数据序列和对应标签。
在如上所述获取到多个训练样本的基础上,在步骤22,将上述第一训练样本中的第一训练数据序列输入循环神经网络,获得第一输出结果。
可以理解,循环神经网络(RNN,Recurrent Neural Networks)是一种时间递归神经网络,可用于处理序列数据。从网络结构的角度来说,循环神经网络包括输入层、循环单元和输出层;输入层用于获取序列数据,循环单元对应于神经网络的一个或多个隐含层,用于依次循环处理序列数据,输出层可以对循环单元经循环处理得到的处理向量进行进一步变换处理,得到整个神经网络的输出结果。
循环神经网络的时间递归性主要体现在循环单元,即隐含层的处理过程。在RNN中,隐含层的各个神经元节点会对之前的信息进行记忆并应用于当前输出的计算中。具体的,隐含层神经元的输入不仅包括来自输入层的输入,还包括该隐含层神经元上一时刻的状态输出,相应的,隐含层神经元的输出不仅与当前时刻输入层输入的向量相关,还与该节点之前的状态相关联。也就是说,t时刻隐含层神经元状态可以表示为:
St=f(U*X t+W*S t-1),
其中,X t为t时刻输入层的状态,S t-1为t-1时刻隐含层神经元的状态,f为计算函数,W,U为权重。如此,循环单元将之前的状态循环回当前输入,考虑了历史输入的影响,因而适合于具有时序的数据序列。
在一个实施例中,在步骤22,采用包含多个隐含层的RNN作为上述循环神经网络。
更进一步地,在一个实施例中,在传统RNN架构下,采用长短期记忆(LSTM,Long Short Term Memory)神经网络作为上述循环神经网络。
如前所述,在RNN中当前隐含层状态依赖于之前的状态输出,因此在处理长期记忆的情况下,需要将当前的隐含态的计算与前n次的计算关联,即St=f(U*X t+W1*S t-1+W2*S t-2+…+Wn*S t-n)。随着n的增大,计算量呈指数式增长,导致模型训练的时间 大幅增加。为此,提出LSTM模型来解决长期依赖的问题。
在LSTM中,通过设置让信息选择性通过的“遗忘门”来丢弃某些不再需要的信息,如此对输入的不必要的干扰信息进行判断和屏蔽,从而更好地对长期数据进行分析处理。
下面结合LSTM神经网络描述对第一训练数据序列进行处理的具体过程。
与前述常规循环神经网络类似的,LSTM神经网络包括输入层,LSTM单元和输出层,其中LSTM单元作为循环单元,可以包含若干隐含层,依次循环处理第一训练数据序列中的各个数据集。
图3示出根据一个实施例的通过LSTM神经网络处理训练数据序列的示意图。如图3所示,第一训练数据序列通过输入层被依次输入到LSTM单元,使得LSTM单元依次处理第一训练数据序列中的各个训练数据集。
具体而言,如前所述,第一训练数据序列包括按照对应的时间段顺序排列的m个训练数据集,其中,第j个训练数据集Dj可以表示为,向量x (j)和y (j)的拼接[x (j);y (j)],其中x (j)为实体特征,y (j)为待测业务指标的指标值。LSTM单元按照数据集排列顺序,依次读取和处理这m个训练数据集。
在第1次处理中,读取第1个训练数据集D1=[x (1);y (1)],将该训练数据集D1和初始的处理向量h 0共同输入到LSTM单元。LSTM单元对D1和h 0进行处理,得到处理向量h 1
在下一时刻进行第2次处理,读取第2个训练数据集D2=[x (2);y (2)],将该训练数据集D2和前一时刻对第1个训练数据集的处理结果h 1,共同作为当前输入,对D2和h 1进行处理,从而得到处理向量h 2。如此继续。
在第i次处理中,LSTM单元将第i个训练数据集,以及前一时刻对第i-1个训练数据集的处理结果h i-1,共同作为当前输入,得到处理向量h i。直到第m次处理,LSTM单元将第m个训练数据集,以及前一时刻对第m-1个训练数据集的处理结果h m-1,共同作为当前输入,得到处理向量h m
然后,输出层将LSTM单元经过m次处理后得到的处理向量h m进行线性变换,得到第一输出结果:
Figure PCTCN2019125856-appb-000001
其中,
Figure PCTCN2019125856-appb-000002
是一个待优化的参数矩阵。
可以理解,该第一输出结果是基于第一训练样本中与连续m个时间段对应的第一训练数据序列,对第m+1个时间段的待测业务指标的指标值的预测。
另一方面,步骤21获取的第一训练样本还包括与第一训练数据序列对应的第一标签。因此,接下来,在步骤23,至少基于上述第一输出结果和所述第一标签,确定与第一训练样本对应的第一预测误差。
可以通过多种具体算法确定上述第一预测误差。例如,第一输出结果和第一标签均可表示为k维向量(如前所述,k为待测业务指标数目),因此,在一个例子中,可以计算第一输出结果对应向量和第一标签对应向量之间的距离,例如余弦距离,欧式距离等,将该距离作为第一预测误差。还可以将上述距离的平方作为第一预测误差。在其他实施例中,还可以通过其他方式计算上述预测误差。
如此,确定出与第一训练样本对应的第一预测误差。可以理解,上述第一训练样本是多个训练样本中任意的一个样本。因此,可以按照步骤22和23的方式,处理每个训练样本,从而得到多个训练样本中各个训练样本对应的预测误差。
于是在步骤24,至少基于与各个样本对应的预测误差,确定损失函数。
可以理解,对于一批训练样本,损失函数是对这批训练样本的模型预测值与标准标签值之间误差的综合衡量。因此,在一个实施例中,基于各个训练样本对应的预测误差确定总误差,基于该总误差确定损失函数。例如,在一个例子中,损失函数可以表示为:
Figure PCTCN2019125856-appb-000003
在上式中,第i个训练样本的预测误差表示为
Figure PCTCN2019125856-appb-000004
损失函数是将n个训练样本的预测误差求和而得到。
进一步的,在一个实施例中,在基于各个训练样本对应的预测误差确定总误差之外,还确定循环神经网络的网络参数的正则化项,将所述总误差和所述正则化项的和,确定为损失函数。例如,损失函数可以表示为:
Figure PCTCN2019125856-appb-000005
其中,Reg(·)为正则化项,例如可以是L2正则化项。L2正则化又可以表示为
Figure PCTCN2019125856-appb-000006
可以通过对所有网络参数的平方求和∑(wi) 2而得到。因此,正则化项可以刻画网络参数整 体的数值范围。
可以理解,还可以采用其他具体公式或算法来确定损失函数。
接着,在步骤25,根据上述损失函数调整所述循环神经网络的网络参数,使得所述损失函数相比调整前减小。
可以理解,神经网络的训练目标是使得损失函数尽可能小,从而使得预测值与真实的标签值尽可能接近。在一个实施例中,通过误差反向传播的方式,调整各个网络参数,以最小化损失函数。
如前所述,正则化项刻画了网络参数整体的数值范围。在损失函数中添加正则化项的情况下,为了达到损失函数最小化的目标,网络参数的整体数值范围也会被优化到较小的范围,从而避免出现参数取值差异过大的过拟合现象。
综合以上,通过采集多个业务实体在连续的各个时间段中的实体特征数据和待测业务指标的指标数据作为特征数据序列,采集下一时间段中待测业务指标的指标数据作为标签数据,由此形成了多个训练样本,用这样的多个训练样本训练循环神经网络,就可以得到用于多实体多业务指标的联合模型。换而言之,在以上的循环神经网络训练完成后,就可以作为所需的联合模型,针对待测业务实体,预测多个待测业务指标的指标值。
图4示出根据一个实施例的预测业务指标的方法。该方法可以通过任何具有计算、处理能力的装置、设备、平台、设备集群来执行。如图4所示,该预测方法包括以下步骤。
在步骤41,获取用于待测业务实体的数据集序列,该数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值。
需要理解,上述预测用的数据集序列,与训练过程中的训练数据序列具有对应一致的数据结构。具体而言,以上数据集序列中时间段的长短,m的数目需要与训练数据序列中的相同,待测业务实体的实体特征与训练样本中的业务实体的实体特征处于相同的特征空间,且待测业务指标也与训练样本的相同。
例如,在一个例子中,多个训练样本中的业务实体为各个商家,实体特征包括与商家有关的20项特征,待测业务指标包括3项指标,时间段为天,m设定为7。
在这样的情况下,待测业务实体可以是某个具体的待测商家Ep,步骤41的数据集序列可以包括对应于连续7天的7个数据集,第i天的数据集包括,商家Ep的实体特征构成的20维向量,例如包括注册时长、商家规模、信誉等级、商品数目、商品类别等,以及商家Ep的待测业务指标在第i天的指标值构成的3维向量,例如包括交易笔数、交易笔单价、交易总额。
在训练样本中的业务实体为各个用户的情况下,上述待测业务实体对应也为某个具体的待测用户,其实体特征和待测业务指标可以参照对训练样本的描述,不再赘述。
与训练样本序列类似的,待测业务实体的实体特征可以包括,在m个时间段中保持不变的特征,也可以包括在m个时间段中不完全相同的特征。
在一个实施例中,为了对下一时间段的多个业务指标进行预测,一般地,采集从当前时间段向前追溯m-1个时间段,共m个时间段的m个数据集,用于预测下一时间段中多个待测业务指标的指标值。例如,在时间段为天,m=7的情况下,可以采集从今天向前追溯6天,加上今天一共是连续7天各自的数据集,用于预测明天的业务指标值。
然而,可以理解,这并不是必须的。在一个实施例中,上述m个时间段也可以均对应于已经过去的时间段,这可以适用于已知第m+1个时间段中待测业务指标的真实值,想要通过联合模型得到预测值,然后通过预测值与真实值的差异,判定真实值是否出现异常的情况。
在获取到上述的数据集序列后,在步骤42,将该数据集序列输入通过图2的方法训练的循环神经网络,获得输出结果。
如前所述,在一个实施例中,上述循环神经网络包括,长短期记忆神经网络LSTM。
更具体的,循环神经网络包括输入层、循环单元和输出层。相应的,在步骤42,通过输入层将数据集序列中的m个数据集依次输入到循环单元,使得循环单元依次处理所述m个数据集,得到m次处理后的处理向量。
上述处理过程与结合图4所述的训练过程的处理相同。其中,在第i次处理中,循环单元将第i个数据集,以及前一时刻对第i-1个数据集的处理结果h i-1,共同作为当前输入,得到处理向量h i。直到第m次处理,循环单元将第m个数据集,以及前一时刻对第m-1个数据集的处理结果h m-1,共同作为当前输入,得到m次处理后的处理向量h m
接着,通过输出层对处理向量h m进行线性变换,就可以得到输出结果。
以上处理过程的网络参数,包括线性变换矩阵W,均通过训练过程而确定和优化。
于是,在步骤43,根据上述输出结果,即可确定出,第m时间段后的下一时间段中,即第m+1时间段中,待测业务实体的所述多个待测业务指标各自的指标值。
具体的,如式(4)所示,循环神经网络通过将循环单元的处理向量进行线性变换得到输出结果,经过线性变换后的输出结果通常为k维向量。通过该k维向量的各个元素值,可以确定出k个待测业务指标在第m+1个时间段中分别对应的指标值。
Figure PCTCN2019125856-appb-000007
通过以上过程可以看到,通过利用多个业务实体的实体特征和待测业务指标的已知值的数据序列作为训练样本,训练循环神经网络,可以得到一个联合模型。该联合模型可以适用于多个待测业务实体的多个待测业务指标的预测。由于训练过程中所基于的训练数据,即包含实体的特征又包含指标数据,因此训练得到的联合模型更好地考虑和挖掘了实体间/指标间的关联关系,并且由于训练数据丰富,也使得预测的准确度更高,效果更好。
根据另一方面的实施例,提供了一种预测业务指标的装置,该装置可以部署在任何具有计算、处理能力的设备、平台或设备集群中。图5示出根据一个实施例的预测业务指标的装置的示意性框图。如图5所示,该预测装置500包括:
数据序列获取单元51,配置为获取数据集序列,所述数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值;
输出结果获取单元52,配置为将所述数据集序列输入预先训练的循环神经网络,获得输出结果;
指标值确定单元53,配置为根据所述输出结果,确定第m时间段的下一时间段中,所述待测业务实体的所述多个待测业务指标各自的指标值。
根据一个实施例,上述待测业务实体包括商家;在这样的情况下,实体特征包括以下中的至少一项:注册时长,商家规模,信誉等级,商品数目,商品类别,上架时长;所述多个待测业务指标包括以下指标中的多个:交易笔数,交易笔单价,交易总额,成 交比例。
根据另一实施例,上述待测业务实体包括用户;所述多个待测业务指标包括,与消费行为相关的多个指标,和/或,与借贷行为的相关的多个指标。
根据一种可能的实施方案,上述待测业务实体的实体特征包括,第一特征和第二特征,所述第一特征的特征值在所述m个时间段中保持不变,所述第二特征的特征值在所述m个时间段中不完全相同。
在一个实施例中,上述m个时间段包括,当前时刻所处的当前时间段,和从所述当前时间段向前追溯的m-1个时间段。
根据一个实施例,上述循环神经网络包括,长短期记忆神经网络LSTM。
在一种实施方式中,所述循环神经网络通过训练单元54训练,该训练单元可以包含在预测装置500之中,也可以位于预测装置500之外。相应的,输出结果获取单元52可连接到该训练单元54。在一个实施例中,训练单元54包括:
样本获取模块541,配置为获取多个训练样本,其中包括第一训练样本,所述第一训练样本包括第一训练数据序列和第一标签,所述第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,所述m个训练数据集中第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的所述多个待测业务指标各自的指标值;所述第一标签包括,采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值;
结果获取模块542,配置为将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果;
误差确定模块543,配置为至少基于所述第一输出结果和所述第一标签,确定与第一训练样本对应的第一预测误差;
损失确定模块544,配置为基于与所述多个训练样本中各个训练样本对应的预测误差,确定损失函数;
参数调整模块545,配置为根据所述损失函数调整所述训练神经网络的网络参数,使得所述损失函数相比调整前减小。
在一个实施例中,所述循环神经网络包括,输入层,循环单元和输出层,
上述结果获取模块542配置为:
通过所述输入层将所述m个训练数据集依次输入到所述循环单元,使得所述循环单元依次处理所述m个训练数据集,得到m次处理后的处理向量;
通过所述输出层对所述处理向量进行线性变换,得到所述第一输出结果。
根据一种实施方式,所述损失确定模块544配置为,
基于所述多个样本中各个样本对应的预测误差确定总误差;
确定所述网络参数的正则化项;
将所述总误差和所述正则化项的和,确定为所述损失函数。
如此,通过训练单元54训练循环神经网络作为联合模型,预测装置500通过利用训练单元54训练好的循环神经网络,实现多业务实体多业务指标的预测。
根据另一方面的实施例,还提供一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行结合图2和图4所描述的方法。
根据再一方面的实施例,还提供一种计算设备,包括存储器和处理器,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现结合图2和图4所述的方法。
本领域技术人员应该可以意识到,在上述一个或多个示例中,本发明所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时,可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式,对本发明的目的、技术方案和有益效果进行了进一步详细说明,所应理解的是,以上所述仅为本发明的具体实施方式而已,并不用于限定本发明的保护范围,凡在本发明的技术方案的基础之上,所做的任何修改、等同替换、改进等,均应包括在本发明的保护范围之内。

Claims (20)

  1. 一种计算机执行的预测业务指标的方法,包括:
    获取数据集序列,所述数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值;
    将所述数据集序列输入预先训练的循环神经网络,获得输出结果;
    根据所述输出结果,确定第m时间段的下一时间段中,所述待测业务实体的所述多个待测业务指标各自的指标值。
  2. 根据权利要求1所述的方法,其中,所述待测业务实体包括商家,
    所述实体特征包括以下中的至少一项:
    注册时长、商家规模、信誉等级、商品数目、商品类别、上架时长;
    所述多个待测业务指标包括以下指标中的多个:
    交易笔数、交易笔单价、交易总额、成交比例。
  3. 根据权利要求1所述的方法,其中,所述待测业务实体包括用户,
    所述多个待测业务指标包括,与消费行为相关的多个指标,和/或,与借贷行为的相关的多个指标。
  4. 根据权利要求1所述的方法,其中所述待测业务实体的实体特征包括,第一特征和第二特征,所述第一特征的特征值在所述m个时间段中保持不变,所述第二特征的特征值在所述m个时间段中不完全相同。
  5. 根据权利要求1所述的方法,其中所述m个时间段包括,当前时刻所处的当前时间段,和从所述当前时间段向前追溯的m-1个时间段。
  6. 根据权利要求1所述的方法,其中所述循环神经网络包括,长短期记忆神经网络LSTM。
  7. 根据权利要求1所述的方法,其中所述循环神经网络通过以下过程预先训练:
    获取多个训练样本,其中包括第一训练样本,所述第一训练样本包括第一训练数据序列和第一标签,所述第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,所述m个训练数据集中第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的所述多个待测业务指标各自的指标值;所述第一标签包括,采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值;
    将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果;
    至少基于所述第一输出结果和所述第一标签,确定与第一训练样本对应的第一预测误差;
    基于与所述多个训练样本中各个训练样本对应的预测误差,确定损失函数;
    根据所述损失函数调整所述训练神经网络的网络参数,使得所述损失函数相比调整前减小。
  8. 根据权利要求7所述的方法,其中,所述循环神经网络包括输入层、循环单元和输出层,
    将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果包括:
    通过所述输入层将所述m个训练数据集依次输入到所述循环单元,使得所述循环单元依次处理所述m个训练数据集,得到m次处理后的处理向量;
    通过所述输出层对所述处理向量进行线性变换,得到所述第一输出结果。
  9. 根据权利要求7所述的方法,其中,所述确定损失函数包括,
    基于所述多个样本中各个样本对应的预测误差确定总误差;
    确定所述网络参数的正则化项;
    将所述总误差和所述正则化项的和,确定为所述损失函数。
  10. 一种计算机执行的预测业务指标的装置,包括:
    数据序列获取单元,配置为获取数据集序列,所述数据集序列包括对应于连续的m个时间段、且按照时间顺序排列的m个数据集,所述m个数据集中第i个数据集包括,第i时间段中,待测业务实体的实体特征,以及该待测业务实体的多个待测业务指标各自的指标值;
    输出结果获取单元,配置为将所述数据集序列输入预先训练的循环神经网络,获得输出结果;
    指标值确定单元,配置为根据所述输出结果,确定第m时间段的下一时间段中,所述待测业务实体的所述多个待测业务指标各自的指标值。
  11. 根据权利要求10所述的装置,其中,所述待测业务实体包括商家,
    所述实体特征包括以下中的至少一项:
    注册时长、商家规模、信誉等级、商品数目、商品类别、上架时长;
    所述多个待测业务指标包括以下指标中的多个:
    交易笔数、交易笔单价、交易总额、成交比例。
  12. 根据权利要求10所述的装置,其中,所述待测业务实体包括用户,
    所述多个待测业务指标包括与消费行为相关的多个指标,和/或
    与借贷行为的相关的多个指标。
  13. 根据权利要求10所述的装置,其中所述待测业务实体的实体特征包括,第一特征和第二特征,所述第一特征的特征值在所述m个时间段中保持不变,所述第二特征的特征值在所述m个时间段中不完全相同。
  14. 根据权利要求10所述的装置,其中所述m个时间段包括,当前时刻所处的当前时间段,和从所述当前时间段向前追溯的m-1个时间段。
  15. 根据权利要求10所述的装置,其中所述循环神经网络包括,长短期记忆神经网络LSTM。
  16. 根据权利要求10所述的装置,其中所述循环神经网络通过训练单元训练,所述训练单元包括:
    样本获取模块,配置为获取多个训练样本,其中包括第一训练样本,所述第一训练样本包括第一训练数据序列和第一标签,所述第一训练数据序列包括对应于连续的m个时间段、且按照时间顺序排列的m个训练数据集,所述m个训练数据集中第j个训练数据集包括,第j时间段中,第一业务实体的实体特征,以及该第一业务实体的所述多个待测业务指标各自的指标值;所述第一标签包括,采集到的所述第一业务实体的所述多个待测业务指标在第m时间段后的下一时间段中各自的指标值;
    结果获取模块,配置为将所述第一训练数据序列输入所述循环神经网络,获得第一输出结果;
    误差确定模块,配置为至少基于所述第一输出结果和所述第一标签,确定与第一训练样本对应的第一预测误差;
    损失确定模块,配置为基于与所述多个训练样本中各个训练样本对应的预测误差,确定损失函数;
    参数调整模块,配置为根据所述损失函数调整所述训练神经网络的网络参数,使得所述损失函数相比调整前减小。
  17. 根据权利要求16所述的装置,其中,所述循环神经网络包括,输入层,循环单元和输出层,
    所述结果获取模块配置为:
    通过所述输入层将所述m个训练数据集依次输入到所述循环单元,使得所述循环单元依次处理所述m个训练数据集,得到m次处理后的处理向量;
    通过所述输出层对所述处理向量进行线性变换,得到所述第一输出结果。
  18. 根据权利要求16所述的装置,其中,所述损失确定模块配置为,
    基于所述多个样本中各个样本对应的预测误差确定总误差;
    确定所述网络参数的正则化项;
    将所述总误差和所述正则化项的和,确定为所述损失函数。
  19. 一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行权利要求1-9中任一项的所述的方法。
  20. 一种计算设备,包括存储器和处理器,其特征在于,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现权利要求1-9中任一项所述的方法。
PCT/CN2019/125856 2019-01-07 2019-12-17 预测业务指标的方法及装置 WO2020143409A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910012815.4A CN110009384A (zh) 2019-01-07 2019-01-07 预测业务指标的方法及装置
CN201910012815.4 2019-01-07

Publications (1)

Publication Number Publication Date
WO2020143409A1 true WO2020143409A1 (zh) 2020-07-16

Family

ID=67165351

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/125856 WO2020143409A1 (zh) 2019-01-07 2019-12-17 预测业务指标的方法及装置

Country Status (2)

Country Link
CN (1) CN110009384A (zh)
WO (1) WO2020143409A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112241351A (zh) * 2020-09-30 2021-01-19 中国银联股份有限公司 数据处理方法、装置、设备和介质
CN112308170A (zh) * 2020-11-10 2021-02-02 维沃移动通信有限公司 建模方法、装置及电子设备
CN112734195A (zh) * 2020-12-31 2021-04-30 平安科技(深圳)有限公司 数据处理方法、装置、电子设备及存储介质
CN112801465A (zh) * 2021-01-08 2021-05-14 上海画龙信息科技有限公司 一种交互式建模预测产品指标的方法、装置和电子设备
CN113378383A (zh) * 2021-06-10 2021-09-10 北京工商大学 一种食品供应链危害物预测方法及装置
CN113743643A (zh) * 2021-02-05 2021-12-03 北京京东振世信息技术有限公司 一种确定商品数据预测准确率的方法、装置、设备和介质
CN113780703A (zh) * 2020-09-27 2021-12-10 北京京东振世信息技术有限公司 一种指标调整方法和装置
CN114519610A (zh) * 2022-02-16 2022-05-20 支付宝(杭州)信息技术有限公司 信息预测方法以及装置
CN114707041A (zh) * 2022-04-11 2022-07-05 中国电信股份有限公司 消息推荐方法、装置、计算机可读介质及电子设备
CN115049060A (zh) * 2022-07-14 2022-09-13 南京理工大学 基于深度学习的业务过程任务执行时知识推荐方法
CN117455666A (zh) * 2023-10-16 2024-01-26 厦门国际银行股份有限公司 一种基于神经网络的交易技术指标预测方法、装置及设备
CN117557289A (zh) * 2023-11-20 2024-02-13 大头(广州)软件技术有限公司 一种基于智能营销场景的餐饮数据建模方法及系统
CN115049060B (zh) * 2022-07-14 2024-05-10 南京理工大学 基于深度学习的业务过程任务执行时知识推荐方法

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110009384A (zh) * 2019-01-07 2019-07-12 阿里巴巴集团控股有限公司 预测业务指标的方法及装置
CN110689070B (zh) * 2019-09-25 2023-04-07 第四范式(北京)技术有限公司 一种业务预测模型的训练方法及装置
CN110888857B (zh) * 2019-10-14 2023-11-07 平安科技(深圳)有限公司 基于神经网络的数据标签生成方法、装置、终端及介质
CN110796240A (zh) * 2019-10-31 2020-02-14 支付宝(杭州)信息技术有限公司 一种训练方法、特征提取方法、装置及电子设备
CN111339156B (zh) * 2020-02-07 2023-09-26 京东城市(北京)数字科技有限公司 业务数据的长期确定方法、装置和计算机可读存储介质
CN111461862B (zh) * 2020-03-27 2023-06-30 支付宝(杭州)信息技术有限公司 为业务数据确定目标特征的方法及装置
CN113822455A (zh) * 2020-06-18 2021-12-21 拉扎斯网络科技(上海)有限公司 一种时间预测方法、装置、服务器及存储介质
CN112528110A (zh) 2020-07-24 2021-03-19 支付宝(杭州)信息技术有限公司 确定实体业务属性的方法及装置
CN112215509A (zh) * 2020-10-21 2021-01-12 拉扎斯网络科技(上海)有限公司 资源参数的确定方法、装置及设备
CN113822371A (zh) * 2021-09-30 2021-12-21 支付宝(杭州)信息技术有限公司 训练分组模型,以及对时序数据进行分组的方法和装置
CN114970359A (zh) * 2022-06-06 2022-08-30 支付宝(杭州)信息技术有限公司 建立时间序列预测模型的方法、时间序列预测方法和装置
CN114971057A (zh) * 2022-06-09 2022-08-30 支付宝(杭州)信息技术有限公司 模型选择的方法及装置
CN115828075A (zh) * 2022-07-27 2023-03-21 京东城市(北京)数字科技有限公司 一种计算指标数据的方法和装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017093953A1 (en) * 2015-12-02 2017-06-08 Tata Consultancy Services Limited Method and system for purchase behavior prediction of customers
CN108038580A (zh) * 2017-12-30 2018-05-15 国网江苏省电力公司无锡供电公司 基于同步挤压小波变换的光伏功率多模型综合预测方法
CN108764539A (zh) * 2018-05-15 2018-11-06 中国长江电力股份有限公司 一种梯级电站的上下游水位预测方法
CN108932567A (zh) * 2018-08-10 2018-12-04 燕山大学 一种基于卷积神经网络的水泥烧成过程多能耗指标预测方法
CN109033450A (zh) * 2018-08-22 2018-12-18 太原理工大学 基于深度学习的电梯设备故障预测方法
CN110009384A (zh) * 2019-01-07 2019-07-12 阿里巴巴集团控股有限公司 预测业务指标的方法及装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107346464B (zh) * 2016-05-06 2021-04-16 腾讯科技(深圳)有限公司 业务指标预测方法及装置
US20180114171A1 (en) * 2016-10-24 2018-04-26 AINGEL Corp. Apparatus and method for predicting expected success rate for a business entity using a machine learning module
CN106682781A (zh) * 2016-12-30 2017-05-17 山东鲁能软件技术有限公司 一种电力设备多指标预测方法
CN108182634A (zh) * 2018-01-31 2018-06-19 国信优易数据有限公司 一种借贷预测模型的训练方法、借贷预测方法和装置
CN108446978A (zh) * 2018-02-12 2018-08-24 阿里巴巴集团控股有限公司 处理交易数据的方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017093953A1 (en) * 2015-12-02 2017-06-08 Tata Consultancy Services Limited Method and system for purchase behavior prediction of customers
CN108038580A (zh) * 2017-12-30 2018-05-15 国网江苏省电力公司无锡供电公司 基于同步挤压小波变换的光伏功率多模型综合预测方法
CN108764539A (zh) * 2018-05-15 2018-11-06 中国长江电力股份有限公司 一种梯级电站的上下游水位预测方法
CN108932567A (zh) * 2018-08-10 2018-12-04 燕山大学 一种基于卷积神经网络的水泥烧成过程多能耗指标预测方法
CN109033450A (zh) * 2018-08-22 2018-12-18 太原理工大学 基于深度学习的电梯设备故障预测方法
CN110009384A (zh) * 2019-01-07 2019-07-12 阿里巴巴集团控股有限公司 预测业务指标的方法及装置

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780703A (zh) * 2020-09-27 2021-12-10 北京京东振世信息技术有限公司 一种指标调整方法和装置
CN112241351A (zh) * 2020-09-30 2021-01-19 中国银联股份有限公司 数据处理方法、装置、设备和介质
CN112241351B (zh) * 2020-09-30 2023-12-22 中国银联股份有限公司 数据处理方法、装置、设备和介质
CN112308170A (zh) * 2020-11-10 2021-02-02 维沃移动通信有限公司 建模方法、装置及电子设备
CN112734195B (zh) * 2020-12-31 2023-07-07 平安科技(深圳)有限公司 数据处理方法、装置、电子设备及存储介质
CN112734195A (zh) * 2020-12-31 2021-04-30 平安科技(深圳)有限公司 数据处理方法、装置、电子设备及存储介质
CN112801465A (zh) * 2021-01-08 2021-05-14 上海画龙信息科技有限公司 一种交互式建模预测产品指标的方法、装置和电子设备
CN112801465B (zh) * 2021-01-08 2024-03-01 上海画龙信息科技有限公司 一种交互式建模预测产品指标的方法、装置和电子设备
CN113743643A (zh) * 2021-02-05 2021-12-03 北京京东振世信息技术有限公司 一种确定商品数据预测准确率的方法、装置、设备和介质
CN113743643B (zh) * 2021-02-05 2023-11-03 北京京东振世信息技术有限公司 一种确定商品数据预测准确率的方法、装置、设备和介质
CN113378383B (zh) * 2021-06-10 2024-02-27 北京工商大学 一种食品供应链危害物预测方法及装置
CN113378383A (zh) * 2021-06-10 2021-09-10 北京工商大学 一种食品供应链危害物预测方法及装置
CN114519610A (zh) * 2022-02-16 2022-05-20 支付宝(杭州)信息技术有限公司 信息预测方法以及装置
CN114707041A (zh) * 2022-04-11 2022-07-05 中国电信股份有限公司 消息推荐方法、装置、计算机可读介质及电子设备
CN114707041B (zh) * 2022-04-11 2023-12-01 中国电信股份有限公司 消息推荐方法、装置、计算机可读介质及电子设备
CN115049060A (zh) * 2022-07-14 2022-09-13 南京理工大学 基于深度学习的业务过程任务执行时知识推荐方法
CN115049060B (zh) * 2022-07-14 2024-05-10 南京理工大学 基于深度学习的业务过程任务执行时知识推荐方法
CN117455666A (zh) * 2023-10-16 2024-01-26 厦门国际银行股份有限公司 一种基于神经网络的交易技术指标预测方法、装置及设备
CN117557289A (zh) * 2023-11-20 2024-02-13 大头(广州)软件技术有限公司 一种基于智能营销场景的餐饮数据建模方法及系统
CN117557289B (zh) * 2023-11-20 2024-04-16 大头(广州)软件技术有限公司 一种基于智能营销场景的餐饮数据建模方法及系统

Also Published As

Publication number Publication date
CN110009384A (zh) 2019-07-12

Similar Documents

Publication Publication Date Title
WO2020143409A1 (zh) 预测业务指标的方法及装置
Sohaib et al. A PLS-SEM neural network approach for understanding cryptocurrency adoption
TWI759620B (zh) 確定用戶金融違約風險的方法及裝置與電腦可讀儲存媒體及計算設備
Carrión et al. Mediation analyses in partial least squares structural equation modeling: Guidelines and empirical examples
US6631362B1 (en) General decision-making support method and system
Kim et al. Bayesian synthetic control methods
US11861464B2 (en) Graph data structure for using inter-feature dependencies in machine-learning
US20090276368A1 (en) Systems and methods for providing personalized recommendations of products and services based on explicit and implicit user data and feedback
US20140289167A1 (en) System and method for determining retail-business-rule coefficients from current prices
Kargah-Ostadi et al. Developing performance prediction models for pavement management systems in local governments in absence of age data
CN110827143A (zh) 训练信用评分模型的方法、装置及设备
Park et al. A machine learning-based early warning system for the housing and stock markets
CN107133862A (zh) 动态产生增强信用评估的详细交易支付经历的方法和系统
Mintz et al. Information processing pattern and propensity to buy: An investigation of online point-of-purchase behavior
JP2019160089A (ja) 情報処理装置、情報処理方法及び情報処理プログラム
US20210117828A1 (en) Information processing apparatus, information processing method, and program
CN114756720B (zh) 时序数据的预测方法及装置
CN115471260A (zh) 基于神经网络的销售预测方法、装置、设备及介质
US11004156B2 (en) Method and system for predicting and indexing probability of financial stress
Chashmi et al. Predicting customer turnover using recursive neural networks
Angelini et al. Progressive visualization of epidemiological models for COVID-19 visual analysis
TW201506827A (zh) 用以從經整理及分析之一段時間內資料信號導出實質變化屬性以預測傳統預測因子之未來變化的系統及方法
Gaikwad et al. Unveiling Market Dynamics through Machine Learning: Strategic Insights and Analysis
Webb et al. Beyond accuracy: The advantages of the k-nearest neighbor algorithm for hotel revenue management forecasting
US20240113936A1 (en) Method and system for artificial intelligence-based acceleration of zero-touch processing

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: 19908555

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: 19908555

Country of ref document: EP

Kind code of ref document: A1