CN114841774A - Accurate service recommendation method based on economic operation uncertainty factor analysis - Google Patents
Accurate service recommendation method based on economic operation uncertainty factor analysis Download PDFInfo
- Publication number
- CN114841774A CN114841774A CN202210507284.8A CN202210507284A CN114841774A CN 114841774 A CN114841774 A CN 114841774A CN 202210507284 A CN202210507284 A CN 202210507284A CN 114841774 A CN114841774 A CN 114841774A
- Authority
- CN
- China
- Prior art keywords
- model
- sample
- user
- data
- system model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000000556 factor analysis Methods 0.000 title claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000002159 abnormal effect Effects 0.000 claims abstract description 6
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 235000008733 Citrus aurantifolia Nutrition 0.000 claims abstract 2
- 235000011941 Tilia x europaea Nutrition 0.000 claims abstract 2
- 239000004571 lime Substances 0.000 claims abstract 2
- 239000013598 vector Substances 0.000 claims description 26
- 230000003993 interaction Effects 0.000 claims description 23
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 238000005070 sampling Methods 0.000 claims description 10
- 210000002569 neuron Anatomy 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 238000003066 decision tree Methods 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 abstract description 15
- 230000006870 function Effects 0.000 description 25
- 230000006399 behavior Effects 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 238000005065 mining Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000007477 logistic regression Methods 0.000 description 2
- 230000006386 memory function Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 235000013361 beverage Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 229910052731 fluorine Inorganic materials 0.000 description 1
- 125000001153 fluoro group Chemical group F* 0.000 description 1
- 235000013305 food Nutrition 0.000 description 1
- 235000021022 fresh fruits Nutrition 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 210000002364 input neuron Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000002759 z-score normalization Methods 0.000 description 1
Images
Classifications
-
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides an accurate service recommendation method based on uncertainty factor analysis of economic operation, which comprises the following steps of: step 1, processing original data. And preprocessing the data, including missing value processing, abnormal value processing and standardization, extracting effective features from the original data, and embedding the features. And 2, initializing the model. And initializing a recommendation system model according to the data characteristics. And 3, training the model. And (3) inputting the samples processed in the step (1) into the model in the step (2), and selecting a proper loss function training model. And 4, collecting a difficult negative sample. And obtaining the attribution values of the characteristics by using a LIME algorithm, and collecting the difficult negative samples according to the attribution values. And 5, training the model again and outputting the final model. And adding a difficult negative sample into the negative sample, and training the model again to improve the robustness of the model.
Description
Technical Field
The invention relates to the field of recall of recommendation systems, in particular to an accurate service recommendation method based on uncertainty factor analysis of economic operation.
Background
The 'family economy' and 'lazy people' jointly promote the development of online consumption. According to the report of the Ehry consultation and release, the online consumption frequency of the user is obviously increased before and after the epidemic situation is compared. For example, the fresh fruits and food and beverage products with the highest user increment increase by 27.6 percent and 17.3 percent respectively after the epidemic situation on-line consumption compared with before the epidemic situation; during 618, the whole volume of the Jingdong appliance monopoly store is increased by 240 percent.
On one hand, the online market rapidly expands, and each e-commerce platform needs to provide accurate service recommendation for users, and recalls a batch of goods interested by the users from a mass of candidate goods so as to facilitate online shopping of the users. On the other hand, e-commerce recommendation systems should be robust. The recommendation system captures the consumption preference of the user according to historical data such as the user portrait, the user consumption record and the like, so that the user is matched with the proper commodities. The existing recommendation system excessively emphasizes the accuracy of the historical consumption data, and the higher the accuracy of the historical consumption data is, the more the future consumption behavior of the user can be predicted. However, this may result in the recommendation system being overly reliant on individual features learned from historical data, lacking robustness to cope with the effects of instability factors that may appear in the future. In addition, in an unstable environment, the user interest changes at any time, and the model should sharply capture the change of the user interest.
A robust recommendation system should not rely on individual features in the data and should improve discrimination as much as possible, especially for positive and negative samples that are difficult to distinguish. Usually, the negative samples in the recall stage of the recommendation system are randomly sampled, most of the negative samples are simple negative samples, namely, the negative samples have a large difference with the positive samples, and the recommendation system can easily complete judgment according to a small number of characteristics. In order to improve the robustness of the model, difficult negative samples are added, the negative samples are closer to the positive samples, and the model can learn more information.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the technical problems of improving the existing E-commerce recommendation system and improving the robustness of the recommendation system so as to reduce the influence caused by unstable factors of economy, and particularly provides an accurate service recommendation method based on uncertainty factor analysis of economic operation, which comprises the following steps:
step 3, inputting the samples processed in the step 1 into the recommendation system model in the step 2, and training the recommendation system model;
step 4, collecting a difficult negative sample;
and 5, training the recommendation system model again, outputting a final recommendation system model, and obtaining a recommendation result by using the final recommendation system model.
In step 1, the raw data includes data describing the user, data describing the item, and interaction data between the user and the item.
In the step 1, missing value processing, abnormal value processing and standardization are respectively carried out on original data, effective features are finally selected by using priori knowledge and used for training a model, one-hot coding is used for carrying out feature embedding on discrete features, and numerical features are kept unchanged.
In step 2, the recommendation system model includes a linear model and a deep neural network, and the linear model is expressed as:
y＝w T x+b
wherein x is [ x ] 1 x 2 …x d ] T For input samples, x i The value of i is 1-d for the ith sample characteristic; weight w ═ w 1 w 2 …w d ] T And bias b is a parameter that the linear model needs to learn; y is the output of the linear model;
the deep neural network comprises a feature combination layer and a feature interaction layer, wherein the feature combination layer is expressed as follows:
wherein, V x Is a collection of all feature embedding vectors; x is the number of i Denotes the ith feature, v i The embedding vector of the ith feature has n features; an element product vector obtained by multiplying corresponding dimensions of two vectors is indicated, and the time complexity is reduced by the following formula:
in step 2, the feature interaction layer includes two or more hidden layers, and the first hidden layer is represented as follows:
h (1) ＝f(W (1) f(V x )+b (1) )+f(V x )
wherein h is (1) Is the output of the first hidden layer; w (1) And b (1) Respectively, the weight and the bias parameter of the first hidden layer; f (V) x ) Is the output of the feature composition layer; f (.) is the Relu activation function, expressed as follows:
f(x)＝max(0,x)
the remaining layers of the feature interaction layer are represented as follows:
h (l+1) ＝f(W (l) h (l) +b (l) )+h (l)
wherein h is (l) Is the output of the l layer of the deep neural network; w (l) And b (l) Respectively are the weight and the bias parameter of the l layer; f (.) is the Relu activation function.
In step 2, the output P (Y ═ 1| x) of the recommended system model is:
wherein Y represents a data tag, and in the click prediction task, Y-1 represents that the user clicked on the item,
for values on the last layer neurons of a deep neural network，
σ (.) is a sigmoid function, which is a weight parameter for the last layer of neurons, and is expressed as follows:
where e is a natural base number and x is an input value on the last layer of neurons.
The step 3 comprises the following steps: defining a probability p that the user matches the positive sample more than the user matches the negative sample:
p＝σ(P(u,i + )-P(u,i - ))
wherein sigma is a sigmoid function; u denotes a user, i + And i - Respectively representing items in a positive sample and items in a negative sample; σ (.) is sigmoid function; p (.) is the output of the recommended system model;
combining the probability p and the cross entropy loss function to obtain a Bayesian Personalized Ranking (BPR) loss function L BPR Comprises the following steps:
L BPR ＝-logp
one negative sample is randomly sampled for each positive sample, and the recommended system model is trained using the BPR loss function.
In step 4, a Local Interpretetable Model-Agnostic extension (LIME) algorithm is used for collecting a difficult negative sample:
setting G e G as an interpretable model, wherein G represents a set of interpretable models, omega () is defined as model complexity, the model complexity of a linear model is nonzero weight, and the model complexity of a decision tree model is the depth of a tree; define the neighborhood of sample x as pi x ，L(f,g,π x ) Is represented at pi x The error between the output value of an interpretable model g and a recommended system model f on a sample in a defined neighborhood, the LIME algorithm is targeted to be in the neighborhood of pi x Obtaining an interpretable model xi (x) table closest to the recommended system modelShown below:
ξ(x)＝arg min g∈G L(f,g,π x )+Ω(g)。
step 4 also includes the following steps:
step 4-1, defining neighborhood pi of sample x to be interpreted x At pi x Sampling samples similar to x;
4-2, predicting on a sample close to x by using a recommendation system model;
4-3, training an interpretable model by using a sample similar to the x to obtain attribution values of all characteristics of the sample x, and then averaging to obtain the characteristic x with the maximum attribution value max Let its ascribed value be phi max (ii) a For each positive sample i of user u + With a certain probability p sample Randomly sampling a sample having a characteristic x max As a difficult negative sample, probability p sample Comprises the following steps:
wherein n is the characteristic number of the article; phi is a unit of i Represents the average ascribed value of each feature over all positive samples for user u; alpha is a hyperparameter.
And 5, training the recommendation system model again by using the collected difficult negative sample to finally obtain a recommendation system model with robustness.
Compared with the prior art, the invention has the beneficial effects that:
(1) the recommendation system is high in accuracy. The recommendation model consists of a memory part and a generalization part, and the memory part can better memorize a high-frequency mode appearing in historical interactive data by first-order interaction of learning characteristics; and generalizing high-order interaction among the learning features displayed by the parts, and mining the interest of the user.
(2) The recommendation system is strong in robustness. The method in the game theory is used for collecting difficult negative samples to train the model again, the robustness of the model is improved, and disturbance caused by uncertain factors is reduced.
Drawings
The foregoing and/or other advantages of the invention will become further apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
FIG. 1 is a flow chart of the steps of the present invention.
FIG. 2 is a basic framework diagram of the recommendation system model of the present invention.
Fig. 3 is a schematic diagram of a data format of a data set published by a certain e-commerce platform.
FIG. 4 is a graph showing the normalized display of four characteristics of an article to a value attributable to the output of a model.
Detailed Description
The invention provides an accurate service recommendation method based on uncertainty factor analysis of economic operation, which comprises the following steps of:
And 2, initializing the model. And initializing a recommendation system model according to the data characteristics.
And 3, training the model. And (3) inputting the samples processed in the step (1) into the model in the step (2), and selecting a proper loss function training model.
And 4, collecting a difficult negative sample. And obtaining the attribution values of the characteristics by using a LIME algorithm, and collecting the difficult negative samples according to the attribution values.
And 5, training the model again and outputting the final model. And adding a difficult negative sample into the negative sample, and training the model again to improve the robustness of the model.
In the present invention, there are first data describing the user, data describing the item, and interaction data between the user and the item. The data describing the user comprises the gender, age, city and the like of the user, the data describing the article comprises the category, price, label and the like of the article, and the interactive data between the user and the article comprises the interactive behavior data of the user clicking, scoring or purchasing the article.
The raw data is then processed for missing values, outliers, and normalized, respectively. Missing values in the data may be due to objective storage failure, subjective user hiding, etc., and the processing principle is to fill in the missing values with the most likely values. The similar mean interpolation method can be used, the category to which the missing value belongs is predicted by a clustering method, and then the mean value of the category is used for filling; an approximate alternative method can also be used, wherein the euclidean distance is used to determine the samples closest to the sample containing the missing value, and the samples are weighted and averaged to fill in the missing value. The abnormal value in the data refers to data deviating from most of the data, and may be specifically set as data deviating from the average value in the data by more than three times the standard deviation. These data may be deleted directly or modified by a method in missing value processing. The data is normalized by converting numerical data, such as the age of a user, the price of an article, etc., into data with a mean value of 0 and a variance of 1 by using a z-score normalization method, and the formula is as follows:
wherein, a represents new data, b represents original data, c represents the average value of the original data, and d represents the standard deviation of the original data;
and finally, effective features can be selected from the prior knowledge and used for training a model, one-hot coding is used for embedding the discrete features, and the numerical features are kept unchanged.
In the invention, the used recommendation system model is divided into two parts, and the memory function and the generalization function are respectively realized. The memory part of the model is used for memorizing high-frequency modes common in historical data, and is a linear model, and the first-order action of the learning characteristic is expressed as:
y＝w T x+b
wherein x is [ x ] 1 x 2 …x d ] T For input samples, x i Is a sample feature; weight w ═ w 1 w 2 …w d ] T And bias b is the parameter that the linear model needs to learn; y is the output of the linear model.
The generalization part of the model is used for mining the interest of the user from historical data, and is a deep neural network, and the high-order interaction of learning features is divided into a feature combination layer and a feature interaction layer. The feature combination layer combines the features two by two, the feature interaction layer is a multilayer perceptron, the output of the feature combination layer is used as the input, and the layer automatically learns the high-order interaction between the features. The feature combination layer is represented as:
wherein, V x Is a collection of all feature embedding vectors; x is a radical of a fluorine atom i Denotes the ith feature, v i A total of n features are embedded vectors for the features; an element product vector obtained by multiplying corresponding dimensions of two vectors is indicated. It is apparent that the temporal complexity of the above formula is O (n) 2 ) The time complexity can be reduced to o (n) by converting to the following formula:
the feature interaction layer comprises a plurality of hidden layers, and the first hidden layer is represented as follows:
h (1) ＝f(W (1) f(V x )+b (1) )+f(V x )
wherein h is (1) Is the output of the first hidden layer; w (1) And b (1) Respectively, the weight and the bias parameter of the first hidden layer; f (V) x ) Is the output of the feature composition layer; f (.) is the Relu activation function, which is expressed as follows:
f(x)＝max(0,x)
the remaining layers of the feature interaction layer are represented as follows:
h (l+1) ＝f(W (k) h (k) +b (k) )+h (l)
wherein，h (l) Is the output of the l layer of the neural network; w (l) And b (l) Respectively are the weight and the bias parameter of the l layer; f (.) is the Relu activation function. Here, a residual network is used, i.e. every hidden layer learns h (l+1) -h (l) This solves the problem of degraded accuracy of the deepened network.
Finally, the output of the recommended system model is a logistic regression of the two parts:
the above equation is a logarithmic probability that the label of the model prediction sample x is Y ═ 1, and can be regarded as a prediction value of the model for the sample x. Wherein, w T And b is a parameter of the memory portion model,
to generalize the values on the last layer of neurons of the partial model,
σ (.) is a sigmoid function, which is a weight parameter of neurons in the layer and is expressed as follows:
the model is initialized, and the schematic structure of the model is shown in fig. 3.
The recommendation system is mainly divided into a recall stage and a sorting stage, wherein the recall stage selects about hundreds of articles which are interested by users from all articles, the sorting stage sorts the articles, and the top tens of articles are exposed to the users. The positive samples in the recommendation system are items clicked and purchased by the user, the ranking stage of the recommendation system usually uses the sample without exposure click as a negative sample, but in the recall stage of the recommendation system, thousands of candidate items cannot be completely exposed to the user, and then the sample without user click is selected. The negative samples are usually randomly sampled during the recall stage, so that it cannot be guaranteed that the negative samples obtained by random sampling are true negative samples. The recommender system typically performs a binary problem, with the model predicting items of interest to the user as 1 and items of non-interest to the user as 0. The cross entropy loss function commonly used in the binary problem calculates the absolute accuracy of the sample, and requires that the negative sample is a true negative sample. So here the cross entropy loss function is not used, the BPR loss function is used. The loss function models < user, positive sample, negative sample > triplets with the goal of ensuring that the user matches the positive sample more than the user matches the negative sample. First, a probability is defined that the user matches the positive sample more than the negative sample:
p＝σ(P(u,i + )-P(u,i - ))
where u denotes a user, i + And i - Respectively representing items in a positive sample and items in a negative sample; σ (.) is sigmoid function; p (.) is the model output. Then combining the probability and the cross entropy loss function to obtain a BPR loss function as follows:
L BPR ＝-logp
one negative sample is randomly sampled for each positive sample and the model is trained using the BPR loss function. However, the negative samples obtained by random sampling are usually simple negative samples, which are often too far apart from the positive samples of the user, and the model can easily distinguish them. Therefore, after the model is trained, the model needs to be trained again by using difficult negative samples in order to enhance the robustness of the model.
The difficult negative sample is a sample which is closer to the positive sample, and the purpose is to enable the model to learn more information as much as possible, and avoid the model from being classified only depending on individual characteristics, so that the robustness of the model to unknown data is improved, and the disturbance caused by uncertain factors is reduced. Here, the LIME algorithm based on the sharey value method in the game theory is used, and the LIME algorithm can obtain the contribution of each feature to the output value of the model, namely the attribution value of each feature. If the attribution value of a feature of the positive sample is large, the model is shown to be classified mainly according to the feature. However, under the influence of uncertainty factors, the preference of a user changes, and the simple negative sample training model is adopted, so that the model is classified according to the features in a lazy manner, that is, samples with the features have high probability to be judged as positive samples, and information in other features is ignored. So negative sampling is performed on features attributed to larger values, reducing the dependence of the model on that feature.
If a linear model y ═ w is used 1 x 1 +w 2 x 2 +…+w m x m + b prediction, for each feature x of the sample x i The value of the factor is the corresponding coefficient w i . The method for obtaining the characteristic cause value by the LIME algorithm is to use interpretable models such as a linear model and a decision tree model to explain complex models such as a neural network. The core idea is that if a cause value of each feature of a sample x is to be obtained, a neighborhood of the sample x may be focused, an interpretable model is trained using samples in the neighborhood, and the trained interpretable model is used to obtain the cause value of each feature of the sample x. And setting G e to G as an interpretable model, wherein G represents a set of interpretable models, omega (G) is defined as model complexity, omega (G) of a linear model is non-zero weight, and omega (G) of a decision tree model is defined as tree depth. Define neighborhood of sample x as pi x ， L(f,g,π x ) Is represented at pi x In the defined neighborhood, an interpretable model g and the recommended system model f have an error between their output values on the samples. The LIME algorithm targets pi in the neighborhood x The above obtains an interpretable model that is closest to the recommended system model, and is expressed as follows:
the LIME algorithm includes the steps of:
1. defining a neighborhood pi of a sample x to be interpreted x At pi x Sampling samples similar to x;
2. making predictions on these samples with a recommendation system model;
3. an interpretable model, such as a linear model, is trained with these samples to obtain the attributed values of the features of its sample x.
Solving all positive samples i of user u by using LIME algorithm + Then averaging the attribution values of the features of (1) to obtain the feature x with the maximum attribution value max Let its ascribed value be phi max . For each positive sample i of user u + Randomly sampling a feature x with a certain probability max The probability is:
wherein n is the characteristic number of the article; phi is a i Represents the average ascribed value of each feature over all positive samples for user u; alpha is a hyperparameter which can be set to
And (5) training the model again by using the collected difficult negative sample to obtain a robust recommendation system model.
Examples
The embodiment discloses an accurate service recommendation method based on uncertainty factor analysis of economic operation, and a flow chart of the method is shown in fig. 1, and the method comprises the following steps:
And 2, initializing the model. And initializing a recommendation system model according to the data characteristics.
And 3, training the model. And (3) inputting the samples processed in the step (1) into the model in the step (2), and selecting a proper loss function training model.
And 4, collecting a difficult negative sample. And obtaining the attribution values of the characteristics by using a LIME algorithm, and collecting the difficult negative samples according to the attribution values.
And 5, training the model again and outputting the final model. And adding a difficult negative sample into the negative sample, and training the model again to improve the robustness of the model.
In step 1, user data, item data, and interaction data between a user and an item are first obtained from a server. The user data comprises the id, age, gender, city and the like of the user, the item data comprises the id, category, brand, price and the like of the item, and the interaction data of the user and the item comprises whether the user clicks the item, the score of the user on the item and the like. The data is then pre-processed, including missing value processing, outlier processing, and normalization. Then, according to the recommended target, proper features are extracted from the original data.
Here, a data set published by a certain e-commerce platform is used, and the data format of the data set is shown in fig. 3;
the data set classifies the age of the user; the interaction time represents that the user clicked on the item at that time; the items browsed by the user every half hour are all set to be the same session id. Since all users are from the same country, the feature of "country" is not used here; to avoid model overfitting, the "session id" is used here to take into account the time factor when the user clicks on the item, and the feature "interaction time" is not used.
Finally, embedding the characteristics, wherein continuous data can be directly input into the model after standardization, and discrete data such as user id, city, article id and category are input into the model after one-hot coding.
Wherein, in step 2, the number of neurons of each layer of the model is set according to the dimension of feature embedding. The recommendation system model used by the invention is divided into two parts, and the memory function and the generalization function are respectively realized. The memory part of the model is used for memorizing high-frequency modes which are common in historical data and is a linear model; the generalization part of the model is used for mining the interest of the user from historical data, and is a deep neural network; the output of the recommendation system model is a logistic regression of the two outputs. The linear model and the depth model have the same number of input neurons, which is equal to the dimension of the embedded vector. After the model is constructed, the parameters are initialized randomly.
In step 3, if the current task is click prediction, the positive sample item is the item clicked by the user, and the negative sample item is randomly sampled from all candidate items. Firstly, splicing a user characteristic vector and a positive sample characteristic vector together, and inputting the user characteristic vector and the positive sample characteristic vector into a model to obtain an output value p; and splicing the user characteristic vector and the negative sample characteristic vector together, and inputting the spliced user characteristic vector and the negative sample characteristic vector into the model to obtain an output value n. The model is trained using a BPR loss function that models < user, positive sample, negative sample > triplet with the goal of ensuring that the user matches the positive sample more than the negative sample, i.e. the output p is greater than the output n.
In step 4, a LIME algorithm based on a sharey value method in the game theory is used, and the algorithm can obtain the contribution of each feature to the output value of the model, namely the attribution value of each feature. And if the user u clicks the prediction task, setting that the user u clicks m articles, respectively splicing the user characteristic vectors and the characteristic vectors of the m articles to form m positive sample characteristic vectors, inputting the positive sample characteristic vectors into a recommendation system model, and obtaining m outputs. Here, a linear model is selected to explain the recommended system model, and the linear model is trained by the m positive sample feature vectors and the m output values. Then the characteristic x with the largest cause value is obtained max Let its ascribed value be phi max . Then for each positive sample i of user u + Randomly sampling a feature x with a certain probability max As a difficult negative sample.
The items of the data set have four characteristics, respectively item id, brand, store id, category. The LIME algorithm is used for a user with the user id of 2399402, attribution values of the four characteristics of the article to the model output are obtained, and then the attribution values are normalized and displayed as shown in FIG. 4;
the cause value of the store id is the largest, and it can be seen that the store id can most influence the click behavior of the user. To avoid model overfitting and further improve model robustness, to
Randomly collecting difficult negative samples, wherein the commodity ids of the negative samples are the same as the commodity ids in the user positive sample.
In step 5, the model is trained again according to the method in step 3 by using the difficult negative samples collected in step 4. And finally, the model can predict a new sample, and the click prediction rate of the model is improved by adding a difficult negative sample for training.
In some embodiments, if the task is to predict the score of the user on the item, the interaction data of the user and the item needs to include the score of the user on the item, the item with the score exceeding a certain threshold value can be set as a positive sample item of the user, and the item with the score lower than the certain threshold value can be set as a negative sample item of the user.
It should be noted that in step 3 and step 5, the training model needs to set a proper learning rate, and the learning rate determines the training speed of the model and can be set to 0.01; an L2 regularization term can be added into the loss function, so that the generalization capability of the model is further improved; a random gradient descent method may be selected to update the model parameters.
It should be noted that, in step 4, if there are too few items clicked by the user and an interpretable model cannot be trained by using the LIME algorithm, the user may be classified, for example, by clustering according to the features of the user, and then the interpretable model is trained by using the interaction data of all users in the same category.
The invention provides an accurate service recommendation method based on uncertainty factor analysis of economic operation, and a plurality of methods and ways for implementing the technical scheme, and the above description is only a preferred embodiment of the invention, and it should be noted that, for those skilled in the art, a plurality of improvements and modifications can be made without departing from the principle of the invention, and these improvements and modifications should also be regarded as the protection scope of the invention. All the components not specified in the present embodiment can be realized by the prior art.
Claims (10)
1. An accurate service recommendation method based on economic operation uncertainty factor analysis is characterized by comprising the following steps:
step 1, preprocessing original data to obtain data characteristics;
step 2, initializing a recommendation system model according to the data characteristics;
step 3, inputting the samples processed in the step 1 into the recommendation system model in the step 2, and training the recommendation system model;
step 4, collecting a difficult negative sample;
and 5, training the recommendation system model again, outputting a final recommendation system model, and obtaining a recommendation result by using the final recommendation system model.
2. The method according to claim 1, wherein in step 1, the raw data comprises data describing the user, data describing the item, and interaction data between the user and the item.
3. The method according to claim 2, wherein in step 1, the raw data is processed with missing values, processed with abnormal values and normalized, and finally the effective features are selected by using the prior knowledge for training the model, and the discrete features are embedded by using one-hot coding, and the numerical features remain unchanged.
4. The method of claim 3, wherein in step 2, the recommendation system model comprises a linear model and a deep neural network, the linear model being represented as:
y＝w T x+b
wherein x is [ x ] 1 x 2 … x d ] T For input samples, x i The value of i is 1-d for the ith sample characteristic; weight w ═ w 1 w 2 … w d ] T And bias b is a parameter that the linear model needs to learn; y is the output of the linear model;
the deep neural network comprises a feature combination layer and a feature interaction layer, wherein the feature combination layer is expressed as:
wherein, V x Is a collection of all feature embedding vectors; x is the number of i Denotes the ith feature, v i The embedded vector of the ith feature has n features in total; an element product vector obtained by multiplying corresponding dimensions of two vectors is indicated, and the time complexity is reduced by the following formula:
5. the method of claim 4, wherein in step 2, the feature interaction layer comprises more than two hidden layers, and a first hidden layer is represented as follows:
h (1) ＝f(W (1) f(V x )+b (1) )+f(V x )
wherein h is (1) Is the output of the first hidden layer; w (1) And b (1) Respectively, the weight and the bias parameter of the first hidden layer; f (V) x ) Is the output of the feature composition layer; f (.) is the Relu activation function, expressed as follows:
f(x)＝max(0,x)
the remaining layers of the feature interaction layer are represented as follows:
h (l+1) ＝f(W (l) h (l) +b (l) )+h (l)
wherein h is (l) Is the output of the l layer of the deep neural network; w (l) And b (l) Respectively are the weight and the bias parameter of the l layer; f (.) is the Relu activation function.
6. The method of claim 5, wherein in step 2, the output P (Y-1 | x) of the recommended system model is:
wherein Y represents a data tag, and in the click prediction task, Y-1 represents that the user clicked on the item,
for values on the last layer neurons of the deep neural network,
σ (.) is a sigmoid function, which is a weight parameter for the last layer of neurons, and is expressed as follows:
where e is a natural base number and x is an input value on the last layer of neurons.
7. The method of claim 6, wherein step 3 comprises: defining a probability p that the user matches the positive sample more than the user matches the negative sample:
p＝σ(P(u,i + )-P(u,i - ))
wherein sigma is a sigmoid function; u denotes a user, i + And i - Respectively representing items in a positive sample and items in a negative sample; σ (.) is a sigmoid function; p (.) is the output of the recommended system model;
combining the probability p and the cross entropy loss function to obtain a Bayes personalized ranking BPR loss function L BPR Comprises the following steps:
L BPR ＝-log p
one negative sample is randomly sampled for each positive sample, and the recommended system model is trained using the BPR loss function.
8. The method according to claim 7, characterized in that in step 4, the difficult negative examples are collected using a model-independent locally interpretable algorithm:
setting G e G as an interpretable model, wherein G represents a set of interpretable models, omega () is defined as model complexity, the model complexity of a linear model is nonzero weight, and the model complexity of a decision tree model is the depth of a tree; define neighborhood of sample x as pi x ，L(f,g,π x ) Is represented at pi x The error between the output value of an interpretable model g and a recommended system model f on a sample in a defined neighborhood, the LIME algorithm is targeted to be in the neighborhood of pi x The above obtains an interpretable model xi (x) closest to the recommended system model, expressed as follows:
ξ(x)＝arg min g∈G L(f,g,π x )+Ω(g)。
9. the method of claim 8, wherein step 4 further comprises the steps of:
step 4-1, defining neighborhood pi of sample x to be interpreted x At pi x Sampling samples similar to x;
4-2, predicting on a sample close to x by using a recommendation system model;
4-3, training an interpretable model by using a sample similar to the x to obtain attribution values of all characteristics of the sample x, and then averaging to obtain the characteristic x with the maximum attribution value max Let its ascribed value be phi max (ii) a For each positive sample i of user u + With a certain probability p sample Randomly sampling a sample having a characteristic x max Is taken as a difficult negative sample, probability p sample Comprises the following steps:
wherein n is the characteristic number of the article; phi is a i Represents the average ascribed value of each feature over all positive samples for user u; alpha is a hyperparameter.
10. The method of claim 9, wherein in step 5, the recommendation system model is retrained again by using the collected difficult negative samples, and finally a robust recommendation system model is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210507284.8A CN114841774A (en) | 2022-05-10 | 2022-05-10 | Accurate service recommendation method based on economic operation uncertainty factor analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210507284.8A CN114841774A (en) | 2022-05-10 | 2022-05-10 | Accurate service recommendation method based on economic operation uncertainty factor analysis |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114841774A true CN114841774A (en) | 2022-08-02 |
Family
ID=82570857
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210507284.8A Pending CN114841774A (en) | 2022-05-10 | 2022-05-10 | Accurate service recommendation method based on economic operation uncertainty factor analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114841774A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116776160A (en) * | 2023-08-23 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Data processing method and related device |
-
2022
- 2022-05-10 CN CN202210507284.8A patent/CN114841774A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116776160A (en) * | 2023-08-23 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Data processing method and related device |
| CN116776160B (en) * | 2023-08-23 | 2023-11-10 | 腾讯科技（深圳）有限公司 | Data processing method and related device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110969516B (en) | Commodity recommendation method and device | |
| CN112598462B (en) | Personalized recommendation method and system based on collaborative filtering and deep learning | |
| CN111222332B (en) | Commodity recommendation method combining attention network and user emotion | |
| CN111932336A (en) | Commodity list recommendation method based on long-term and short-term interest preference | |
| CN110781409B (en) | Article recommendation method based on collaborative filtering | |
| CN113256367B (en) | Commodity recommendation method, system, equipment and medium for user behavior history data | |
| CN112231584B (en) | Data pushing method and device based on small sample transfer learning and computer equipment | |
| CN110955826B (en) | Recommendation system based on improved cyclic neural network unit | |
| CN111259140B (en) | False comment detection method based on LSTM multi-entity feature fusion | |
| CN108665311B (en) | Electric commercial user time-varying feature similarity calculation recommendation method based on deep neural network | |
| CN110796313B (en) | Session recommendation method based on weighted graph volume and item attraction model | |
| CN112364976B (en) | User preference prediction method based on session recommendation system | |
| CN109903103B (en) | Method and device for recommending articles | |
| CN112508613A (en) | Commodity recommendation method and device, electronic equipment and readable storage medium | |
| CN112487199B (en) | User characteristic prediction method based on user purchasing behavior | |
| CN112700274A (en) | Advertisement click rate estimation method based on user preference | |
| CN114693397A (en) | Multi-view multi-modal commodity recommendation method based on attention neural network | |
| CN113570398A (en) | Promotion data processing method, model training method, system and storage medium | |
| CN114841774A (en) | Accurate service recommendation method based on economic operation uncertainty factor analysis | |
| CN114529077A (en) | Click rate prediction method based on intra-session heterogeneous behaviors | |
| CN117408735A (en) | Client management method and system based on Internet of things | |
| CN110851694A (en) | Personalized recommendation system based on user memory network and tree structure depth model | |
| CN116320626A (en) | Method and system for calculating live broadcast heat of electronic commerce | |
| CN115375408A (en) | Article information recommendation method and device and electronic equipment | |
| CN113033090B (en) | Push model training method, data push device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |