Object prediction method, device, computer equipment and storage medium

Abstract

The present application relates to an object prediction method, device, computer equipment and storage medium. The method includes: obtaining user behavior data within a preset time period, wherein the user behavior data includes object information of the object to be analyzed; extracting data features from the user behavior data; inputting the data features into a time information model to obtain a first prediction result of the object to be analyzed; inputting the first prediction result and data features of the object to be analyzed into a first prediction model to obtain a first prediction probability of the object to be analyzed; inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed; inputting the first prediction probability of the object to be analyzed and the second prediction probability of the object to be analyzed into a logistic regression model to output the probability of the object to be analyzed. The data features are analyzed by multiple different models to identify the probability of the object to be analyzed, and the push object is determined according to the probability of each object to be analyzed, thereby improving the accuracy of data push.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an object prediction method, apparatus, computer device and storage medium. Background Art

[0002] Data push refers to each data content provider actively pushing data, such as pushing corresponding data content to each user, such as pushing news, pushing products, etc. Common data push is based on association rules, that is, finding similar users based on the user's rating of the data, and using the data of similar users and the user itself to push data to the user. Pushing data based only on the user's corresponding data cannot effectively predict the user's intention, thus affecting the accuracy of data push. Summary of the invention

[0003] In order to solve the above technical problems, the present application provides an object prediction method, apparatus, computer device and storage medium.

[0004] The present application provides an object prediction method, comprising:

[0005] Acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed;

[0006] Extracting data features from the user behavior data;

[0007] Inputting the data feature into the time information model to obtain a first prediction result of the object to be analyzed;

[0008] Inputting the first prediction result of the object to be analyzed and the data feature into a first prediction model to obtain a first prediction probability of the object to be analyzed;

[0009] Inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed;

[0010] The first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed are input into the logistic regression model, and the probability of the object to be analyzed is output.

[0011] The present application provides an object prediction device, comprising:

[0012] A data acquisition module, used to acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed;

[0013] A feature extraction module, used to extract data features from the user behavior data;

[0014] A time information model, used to receive the data feature and obtain a first prediction result of the object to be analyzed;

[0015] A first prediction model, used to receive a first prediction result of the object to be analyzed and the data feature, and output a first prediction probability of the object to be analyzed;

[0016] A second prediction model, used to receive the data features and output a second prediction probability of the object to be analyzed;

[0017] The logistic regression model is used to receive the first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed, and output the probability of the object to be analyzed.

[0018] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following steps are implemented:

[0019] Acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed;

[0020] Extracting data features from the user behavior data;

[0021] Inputting the data feature into the time information model to obtain a first prediction result of the object to be analyzed;

[0022] Inputting the first prediction result of the object to be analyzed and the data feature into a first prediction model to obtain a first prediction probability of the object to be analyzed;

[0023] Inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed;

[0024] The first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed are input into the logistic regression model, and the probability of the object to be analyzed is output.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0026] Acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed;

[0027] Extracting data features from the user behavior data;

[0028] Inputting the data feature into the time information model to obtain a first prediction result of the object to be analyzed;

[0029] Inputting the first prediction result of the object to be analyzed and the data feature into a first prediction model to obtain a first prediction probability of the object to be analyzed;

[0030] Inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed;

[0031] The first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed are input into the logistic regression model, and the probability of the object to be analyzed is output.

[0032] The object prediction method, device, computer equipment and storage medium mentioned above, the method includes: obtaining user behavior data within a preset time period, the user behavior data including object information of the object to be analyzed; extracting data features from the user behavior data; inputting the data features into the time information model to obtain the first prediction result of the object to be analyzed; inputting the first prediction result of the object to be analyzed and the data features into the first prediction model to obtain the first prediction probability of the object to be analyzed; inputting the data features into the second prediction model to obtain the second prediction probability of the object to be analyzed; inputting the first prediction probability of the object to be analyzed and the second prediction probability of the object to be analyzed into the logistic regression model to output the probability of the object to be analyzed. The data features of the user behavior data are analyzed by multiple different models, the probability of the object to be analyzed is identified, and the push object is determined according to the probability of each object to be analyzed, thereby improving the accuracy of the push. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0035] Figure 1 A diagram of an application environment of an object prediction method in one embodiment;

[0036] Figure 2 is a schematic flow chart of an object prediction method in one embodiment;

[0037] Figure 3 A schematic diagram for preparing user history data in one embodiment;

[0038] Figure 4 A schematic diagram of feature extraction of a conversion rate feature group in one embodiment;

[0039] Figure 5 A schematic diagram of feature extraction for ranking feature groups in one embodiment;

[0040] Figure 6 A schematic diagram of feature extraction of a proportional feature group in one embodiment;

[0041] Figure 7 A schematic diagram of feature extraction of an interactive feature group in one embodiment;

[0042] Figure 8 A schematic diagram of a method for dividing an input data set in one embodiment;

[0043] Fig. 9 A schematic diagram of a first method for inputting a data set in an embodiment;

[0044] Fig.10 A schematic diagram of a second method for inputting a data set in an embodiment;

[0045] Fig.11 A schematic diagram of a third method for inputting a data set in an embodiment;

[0046] Fig.12 A diagram of a multi-layer model fusion architecture in one embodiment;

[0047] Fig.13 is a structural block diagram of an object prediction device in one embodiment;

[0048] Fig.14 is a structural block diagram of a second prediction model in one embodiment;

[0049] Fig.15 is a structural block diagram of a global data sub-model in an embodiment;

[0050] Fig.16 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0052] Figure 1 FIG. 1 is an application environment diagram of an object prediction method in an embodiment. Figure 1, the object prediction method is applied to the object prediction system. The object prediction system includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected via a network. The server 120 obtains the user behavior data of each terminal 110 within a preset time period, the user behavior data includes the object information of the object to be analyzed, extracts the data features in the user behavior data, inputs the data features into the time information model, and obtains the first prediction result of the object to be analyzed; inputs the first prediction result and the data features of the object to be analyzed into the first prediction model, and obtains the first prediction probability of the object to be analyzed; inputs the data features into the second prediction model, and obtains the second prediction probability of the object to be analyzed; inputs the first prediction probability of the object to be analyzed and the second prediction probability of the object to be analyzed into the logistic regression model, and outputs the probability of the object to be analyzed.

[0053] The terminal 110 may be a desktop terminal or a mobile terminal, and the mobile terminal may be at least one of a mobile phone, a tablet computer, a notebook computer, etc. The server 120 may be implemented by an independent server or a server cluster composed of multiple servers.

[0054] like Figure 2 As shown, in one embodiment, an object prediction method is provided. This embodiment mainly applies the method to the above Figure 1 The server 120 in FIG. Figure 2 , the object prediction method specifically includes the following steps:

[0055] Step S201, obtaining user behavior data within a preset time period.

[0056] In this embodiment, the user behavior data includes object information of the object to be analyzed.

[0057] Specifically, the preset time period is a user-defined time period, such as a week, half a month, or a month, etc. User behavior data includes user shopping data, browsed product data, browsed store information, searched product information, etc. User behavior data includes the predicted day data on the predicted day and the historical behavior data before the predicted day, or only includes the historical behavior data.

[0058] Step S202: extracting data features from user behavior data.

[0059] In this embodiment, the user behavior data includes object information of the object to be analyzed.

[0060] Specifically, data features include user features, store features, product features and search term features. Data features can be divided into basic feature sets, statistical feature sets and cross feature sets. The basic feature set includes but is not limited to the number of store products, the number of product attributes, the number of product categories that are the same as the search term categories, the number of product attributes that are the same as the search term attributes, the number of search term categories, and the number of search term attributes. The statistical feature set includes but is not limited to the conversion rate feature subset, the ranking feature subset, the ratio feature subset, the trend feature subset, and the category feature subset. The cross feature set includes the interaction feature subset, the competition feature subset, and the business feature subset. According to the time information of the user behavior data, the data features can be divided into historical data features corresponding to the historical behavior data and predicted day data features corresponding to the predicted day data. The object to be analyzed refers to the object to be analyzed, and the object to be analyzed can be a product, a store, etc. The features of the store, product, and search term are extracted when extracting features because the store information and product information such as the store ranking, store type, product ranking and type will affect the predicted probability of the object to be analyzed, that is, the probability of the store being clicked, and the probability of the product being purchased and clicked.

[0061] In one embodiment, data features are extracted from user behavior data, including: when the user behavior data includes predicted day behavior data on the predicted day and historical behavior data before the predicted day, features of the predicted day behavior data are extracted to obtain predicted day data features, features of historical behavior data are extracted to obtain historical data features, and the predicted day data features and historical data features constitute data features; when the user behavior data includes historical behavior data before the predicted day, features in the historical behavior data are extracted to obtain historical data features, and the historical data features are used as data features.

[0062] Step S203: input data features into the time information model to obtain a first prediction result of the object to be analyzed.

[0063] Specifically, the time information model is a model for analyzing the input data features. The model is a machine learning model, which is obtained by training the data features of a large amount of data. The classification algorithm of the machine learning model can be a Boosting algorithm, a random forest (Randomforest), or an SVM (support vector machine), wherein the Boosting algorithm includes LightGBM (lightweight gradient boosting machine algorithm) or XGBoost (extreme gradient boosting), etc. LightGBM has the advantages of fast calculation speed and high accuracy. The algorithm idea is to continuously add trees and continuously perform feature splitting to grow a tree. Each time a tree is added, it actually learns a new function to fit the residual of the last prediction. XGBoost grows a tree by continuously performing feature splitting. Each time a tree is added, it learns a new function to fit the residual of the last prediction. When the training is completed and k trees are obtained, to predict the score of an object, it is based on the characteristics of the object, and each tree will fall into a corresponding leaf node. Each leaf node corresponds to a score, and finally the scores corresponding to each tree are added up to the predicted value of the object.

[0064] The time information model uses the data from day 1 to day N as training samples, extracts features from the data from day 1 to day N, and predicts all samples on the prediction day. The time information model analyzes the data features to obtain the first prediction result of the object to be analyzed, that is, the probability value of the object to be analyzed. For example, the probability of placing an order or clicking a product is predicted.

[0065] Step S204: input the first prediction result and data features of the object to be analyzed into the first prediction model to obtain a first prediction probability of the object to be analyzed.

[0066] Specifically, the first prediction model is also a machine learning model. The algorithm used by the machine learning model may be the same as or different from the algorithm used by the time information model. The input data of the first prediction model is the first prediction result and data features of the object to be analyzed. The prediction is performed through the first prediction model to obtain the first prediction probability of the object to be analyzed. The data features used by the first prediction model include historical data features and prediction day data features. If there are no data features on the prediction day, the historical data features are used. The prediction day data features are used because recent data have a higher relevance for predicting future behavior.

[0067] In one embodiment, the first prediction model uses the data of the morning of the prediction day as training samples, and uses the data from the 1st day to the Nth day to extract features, where 1-N is the time period corresponding to the prediction time period.

[0068] Step S205: input the data features into the second prediction model to obtain a second prediction probability of the object to be analyzed.

[0069] Step S206, inputting the first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed into the logistic regression model, and outputting the probability of the object to be analyzed.

[0070] Specifically, the second prediction model can be an independent model or an integrated model of multiple sub-models. If it is an independent model, the classification algorithm of the model can be the same as or different from the classification algorithm of the first prediction model. When the model includes multiple sub-models, the classification algorithm of each sub-model can be the same or different, and each sub-model is a different model.

[0071] The second prediction model directly performs feature analysis on the data features to obtain the second prediction probability of the object to be analyzed. The first prediction probability is the probability obtained by analyzing the data after combining the output results of other models and the data features, and the second prediction probability is the probability obtained by directly analyzing the data features. The logistic regression model is a generalized linear regression model. The probability of the object to be analyzed is calculated in two different ways, and the probability is input into the logistic regression model to regress the probability of the object to be analyzed. The probability of each object to be analyzed being clicked or purchased is determined by the probability, and data push is performed according to the ranking of the probability of each object to be analyzed.

[0072] The object prediction method includes: obtaining user behavior data within a preset time period, wherein the user behavior data includes object information of the object to be analyzed; extracting data features from the user behavior data; inputting the data features into a time information model to obtain a first prediction result of the object to be analyzed; inputting the first prediction result and data features of the object to be analyzed into a first prediction model to obtain a first prediction probability of the object to be analyzed; inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed; inputting the first prediction probability of the object to be analyzed and the second prediction probability of the object to be analyzed into a logistic regression model to output the probability of the object to be analyzed. Two different models are used to analyze the data features corresponding to the user behavior data to obtain the probability of the object to be analyzed, and the probability of each object to be analyzed being clicked and purchased is predicted based on the obtained probability, and the object to be pushed to each user is determined based on the probability of being clicked and purchased.

[0073] In one embodiment, if the second prediction model includes multiple sub-models of different types, step S205 includes: inputting the data features into each sub-model respectively to obtain the reference probability of the object to be analyzed corresponding to each sub-model; and performing weighted summation on the reference probability of the object to be analyzed corresponding to each sub-model to obtain the second prediction probability of the object to be analyzed.

[0074] Specifically, when the second prediction model includes multiple sub-models, the data features are input into each sub-model respectively, and the data features are analyzed by each sub-model to obtain the reference probability of the object to be analyzed of each sub-model. Each sub-model can adopt different classification algorithms, and different sub-models can also adopt the same classification algorithm. And the sub-model can be a sub-model composed of an independent prediction unit, or a sub-model including multiple independent prediction units. When multiple independent prediction units are included, the independent prediction unit is used to predict the data features separately to obtain the prediction probability corresponding to each independent unit. The prediction probabilities corresponding to each independent unit are summed to obtain the reference probability of the sub-model. The reference probability of each sub-model is weighted to obtain the second prediction probability of the object to be analyzed. The sub-models included in the second prediction model can be models of the same type using different classification algorithms, or models of multiple different types but using the same algorithm, or at least three sub-models, of which at least two sub-models are models of the same type, and the classification algorithms corresponding to the models of the same type are different.

[0075] In one embodiment, the second prediction model includes a prediction day model and a global data sub-model, the reference probability of the object to be analyzed includes a first reference sub-probability and a second reference sub-probability, and the data features are input into each sub-model respectively to obtain the reference probability of the object to be analyzed corresponding to each sub-model, including: inputting the data features into the prediction day model to obtain the first reference sub-probability; inputting the data features into the global data sub-model to obtain the second reference sub-probability.

[0076] Specifically, the prediction day model refers to a model that uses the data of the morning of the prediction day as training samples and uses the data from the 1st to the Nth day (historical user data) to extract features. The global data sub-model refers to a model that uses all labeled samples as training samples and uses all data to extract features. The prediction day model is used to analyze the historical data features and the prediction day data features to obtain the first reference sub-probability. The global data sub-model analyzes the historical data features to obtain the second reference sub-probability. Different data features are used for analysis to obtain the corresponding reference sub-probabilities, and the reference sub-probabilities obtained by each sub-model are weighted and summed to obtain the second prediction probability of the object to be analyzed. Different data features and different types of models are used for data analysis, that is, data is analyzed from multiple angles and levels, and the data is mined more comprehensively, thereby obtaining a more accurate probability.

[0077] In one embodiment, the global data sub-model includes a first global data unit and a second global data unit, and inputting data features into the global data sub-model to obtain a second reference sub-probability includes: inputting the data features into the first global data unit to obtain a sub-probability corresponding to the first global data unit; inputting the data features into the second global data unit to obtain a sub-probability corresponding to the second global data unit; and weighted summing the sub-probability corresponding to the first global data unit and the sub-probability of the second global data unit to obtain the second reference sub-probability.

[0078] Specifically, the first global data unit and the second global data unit are data processing units of the same type, and the classification algorithms of the two different global data units are different. Through the same type of data model, different classification algorithms perform data analysis, and the same data can be classified in different ways to obtain different classification results, that is, the sub-probabilities of the first global data unit and the second global data unit are obtained respectively, and the sub-probabilities of the first global data unit and the second global data unit are weighted and summed to obtain the second reference sub-probability of the global data sub-model. Different classification algorithms perform data analysis, and the analysis is more comprehensive.

[0079] In one embodiment, when the user behavior data includes the predicted day behavior data on the predicted day and the historical behavior data before the predicted day, the data features input by the time information model, the first global data unit, and the second global data unit are historical data features, and the data features input by the predicted day model and the first prediction model include historical data features and predicted day data features. Different models analyze different data features, and the analysis of different dimensional feature data is more in-depth, thereby obtaining more accurate analysis results.

[0080] In one embodiment, the data features include a basic feature set, a statistical feature set, and a cross feature set. The basic feature set and the cross feature set are input into a third global data model to obtain a third predicted probability of the object to be analyzed; the weighted sum of the third predicted probability of the object to be analyzed and the probability of the object to be analyzed is calculated to obtain the target probability of the object to be analyzed.

[0081] Specifically, the third global data model refers to a model that uses all labeled samples as training samples and all data to extract features. The probability of the object to be analyzed obtained by the third global data model and the probability of the object to be analyzed calculated in the above embodiment are weighted to obtain the target probability of the object to be analyzed. The features input by the time information model, the first prediction model and the second prediction model include a basic feature set, a statistical feature set and a cross feature set, and the features input by the third global data model are a basic feature set and a cross feature set. That is, a combination analysis of different types of feature sets is performed to analyze the association between data more accurately. Subsequent object push is performed according to the target probability, and the push is more accurate.

[0082] In one embodiment, a basic feature set and a statistical feature set are input into a fourth global data model to obtain a fourth predicted probability of the object to be analyzed; a weighted sum of the fourth predicted probability of the object to be analyzed and the target probability of the object to be analyzed is calculated to obtain a weighted probability of the object to be analyzed.

[0083] Specifically, after adding the third global data model, after adding the fourth global data model, the fourth global data model is used to analyze the basic feature set and the statistical feature set to obtain the fourth predicted probability of the object to be analyzed. The fourth predicted probability of the object to be analyzed output by the fourth global data model is weighted and summed with the target probability to obtain the weighted probability of the object to be analyzed, and subsequent object push is performed according to the weighted probability. Consistent with the reason for adding the third global data model, the fourth global data model is added to further dig into the association between the data, thereby improving the accuracy of object push.

[0084] In one embodiment, the data features include a basic feature set, a statistical feature set, and a cross feature set. The basic feature set and the statistical feature set are input into the fifth global data model to obtain the fifth predicted probability of the object to be analyzed; the weighted sum of the fifth predicted probability of the object to be analyzed and the probability of the object to be analyzed is calculated to obtain the target probability of the object to be analyzed.

[0085] Specifically, the fifth global data model and the fourth global data model are the same data model. The fifth predicted probability of the object to be analyzed output by the fifth global data model and the probability of the object to be analyzed are used to jointly determine the target probability of the object to be analyzed. Subsequent object push is performed according to the target probability, and the push is more accurate.

[0086] In one embodiment, when the user behavior data includes predicted day behavior data on the predicted day and historical behavior data before the predicted day, the basic feature sets input into the third global data model and the fourth global data model are the basic feature sets contained in the historical data features, the cross feature set input into the third global data model is the cross feature set contained in the historical data features, and the statistical feature set input into the fourth global data model is the statistical feature set contained in the historical data features.

[0087] In one embodiment, the method includes a first global data unit, a second global data unit and a third global data model, the classification algorithms of the time information model, the first prediction model, the prediction day model, the first global data unit, the third global data model and the fourth global data model are lightweight gradient boosting machine algorithms (LightGBM), and the classification algorithm of the second global data unit is extreme gradient boosting (XGBoost).

[0088] In a specific embodiment, the object prediction method includes:

[0089] Obtain user behavior data, which includes user data, product data, store data and search keywords.

[0090] Extract the initial data features of user behavior data, including UserFeatures (user features), ItemFeatures (product features), ShopFeatures (shop features), and QueryFeatures (search features). Preprocess the extracted features, use the majority value to fill in missing values ​​(identification features), and use the mean to fill in numerical features. The majority value filling refers to counting the number of occurrences of label features, and filling the missing features with the features that appear the most times. Mean filling refers to calculating the average value of numerical features, and using the average value to fill in missing feature values. Reconstruct ID-type (identification-type) features, and take the ID before the preset position to generate new ID-type features for the search result sorting information.

[0091] The initial data features are a basic feature set (Feature AGroup), which includes but is not limited to the number of store products, the number of product attributes, the number of product categories that are the same as the search term categories, the number of product attributes that are the same as the search term attributes, the number of search term categories, and the number of search term attributes.

[0092] The statistical feature set (FeatureBGroup) is obtained by performing statistics on all or part of the features in the basic feature set, wherein the statistical feature set includes but is not limited to a conversion rate feature subset, a ranking feature subset, a ratio feature subset, a trend feature subset, and a category feature subset.

[0093] Conversion rate feature subset:

[0094] User ID features, number of clicks, number of purchases, conversion rate, and the proportion of the day before the prediction day to all previous days.

[0095] Reference Figure 4 , Figure 4 It is a schematic diagram of feature extraction of conversion rate feature group, such as click-through conversion rate from 1 to N-1 days, click-through conversion rate of day N, and click-through conversion of Forenoon-Predict on the predicted morning.

[0096] Ranked feature subsets:

[0097] Feature extraction is performed from two parts of data: (1) all the data on the day before the prediction; (2) the data one day before the prediction, i.e. the Nth day.

[0098] Reference Figure 5 , Figure 5 The figure is a schematic diagram of feature extraction for ranking feature groups. Extract ID-category features (ID CateCVR Features) from the two parts of data, such as user conversion rate features (User CVR), product conversion rate features (ItemCVR), product category conversion rate features (Brand CVR), search term conversion rate features (Qurey CVR), etc.; extract ID-category features (ID-Category) from the two parts of data, such as store ID features (shop), product ID features (Item), category ID features (Brand), etc.; combine and sort the conversion rate features and ID features to obtain the Rank ranking feature set.

[0099] Scale feature subset:

[0100] Feature extraction is performed from two parts of data: (1) all the data on the day before the prediction; (2) the data one day before the prediction, i.e. the Nth day.

[0101] Reference Figure 6 , Figure 6Schematic diagram of feature extraction of proportional feature group. Extract ID class combination feature click counts (ID Click Group) from the two parts of data, such as (user, product) click counts (click-count(User, Item)), (user, shop) click counts (click-count(User, Brand)), (user, category) click counts (click-count(User, Shop)), (user, search term) click counts (click-count(User, Query)), etc.; extract ID class feature click counts from the two parts of data, such as shop click counts, product click counts, category click counts, etc.; calculate the ratio of ID class combination feature click counts and ID class feature click counts to obtain a proportional feature set.

[0102] Trend feature subset:

[0103] Count the ID features from day 1 to day N, the number of clicks and purchases each day, and divide the data of two consecutive days to get the ratio as the trend feature, such as the number on day N divided by the number on day N-1.

[0104] Extract cross-feature sets (Feature CGroup), and interactive feature sets include interactive feature sets, competitive feature sets, and business feature sets.

[0105] Interaction feature subsets:

[0106] Reference Figure 7 , Figure 7 Schematic diagram of feature extraction for interactive feature groups. In the figure, the search terms (Query) include: search brand 1 (QueryBrand1), search brand 2 (QueryBrand2) and search brand 3 (QueryBrand3), etc. Multiple search brand 2s correspond to products (Item). The extracted search term interaction features include whether the current search term has been searched before, how many search terms are used to search for this product / category / store, the maximum and minimum click intervals, the average click interval, the ratio of adjacent click intervals, the time interval from the first click, how many ID features have been clicked before, and the proportion of each ID feature.

[0107] Competing feature subsets:

[0108] Before the current time point, how many products with lower prices did the user click on, how many products with higher sales did the user click on, how many stores with higher reviews did the user click on, how many stores with higher positive reviews did the user click on, how many stores with higher star ratings did the user click on, how many stores with better service attitudes did the user click on, how many stores with better logistics did the user click on, how many stores with higher description ratings did the user click on, etc.

[0109] Business feature subset:

[0110] The feature extraction time period is divided into two cases: one is from the 1st day to the Nth day, that is, [1, N]; the other is the Nth day.

[0111] The extracted features include how many times the user clicked continuously without purchasing, how many products the user clicked continuously without purchasing, how many stores the user clicked continuously without purchasing, the percentage of the product appearing on each display page, etc.

[0112] The features extracted above are used to obtain the features of the business feature set using the features of the [1, N] time period / the features of the Nth day.

[0113] Model design consists of two parts: one is the data division method, and the other is the model fusion method.

[0114] Reference Figure 8 , Figure 8 The following is a schematic diagram of the input data set partitioning method. Given the historical data from day 1 to day N (historical user behavior data), the user's intention on day N+1, i.e. the prediction day DayPredict, is predicted.

[0115] There are three ways to divide the data set:

[0116] Reference Fig. 9 , Fig. 9 The diagram is a diagram of the input data set partitioning method 1. The global data model Model_MData uses all labeled samples as training samples and uses all data to extract features.

[0117] Reference Fig.10 , Fig.10 The following is a schematic diagram of the second input data set division method. The forecast day model Model_PDay uses the data of the forecast day morning as training samples and uses the data from the 1st day to the Nth day to extract features.

[0118] Reference Fig.11 , Fig.11 The diagram is a diagram of the third way to divide the input data set. The time information model Model_TData uses the data from day 1 to day N as training samples, extracts features from the data from day 1 to day N, and predicts all samples on the prediction day. The prediction results carrying time information are used as new features to make up for the lack of time description in the model.

[0119] Model fusion architecture such as Fig.12 As shown:

[0120] The interactive feature set, competitive feature set, and business feature set in the basic feature set (Feature AGroup) and the cross feature set (Feature CGroup) are used as the feature input of model Model_Z (the third global model). The data set partitioning form of the global data model Model_MData is adopted, and the LightGBM algorithm (lightweight gradient boosting machine algorithm) is selected as the classification algorithm to predict user intentions.

[0121] The conversion rate feature set, ranking feature set, proportion feature set, trend feature set, category feature set and interactive feature set, competition feature set and business feature set of the cross feature set (CrossFeatGroup) of the basic feature set and statistical feature set (Feature BGroup) are used as the feature input of the model. Among them, the model Model_PDay (prediction day model) adopts the LightGBM algorithm as the classification algorithm in addition to its data set division method; Molel_Mdata-1 (first global data unit) adopts the LightGBM algorithm as the classification algorithm in addition to its data set division method; Molel_Mdata-2 (second global data unit) adopts the XGBoost algorithm (extreme gradient boosting) as the classification algorithm in addition to its data set division method; Model_TData (time information model) adopts the LightGBM algorithm as the classification algorithm in addition to its data set division method. At the same time, the output of the model is used as a new feature, and the feature group basic feature set, statistical feature set and cross feature set are used as the feature input of Model_PDay (first prediction model) to obtain a new prediction result.

[0122] Finally, the results of the three models, Model_PDay (prediction day model), Molel_Mdata-1 (first global data unit), and Molel_Mdata-2 (second global data unit), are linearly weighted and the results of Model_TData (time information model) + Model_PDay (first prediction model) are used as the input of model Model_X (logistic regression model). The logistic regression model uses the LogisticsRegression algorithm to sort the search results.

[0123] The basic feature set, the conversion rate feature set of the statistical feature set, the ranking feature set, the proportion feature set, the trend feature set, and the category feature set are used as the input of the model's feature Model_Y (the fourth global data model / fifth global data model). The data set partitioning form of the global data model Model_MData is adopted, and LightGBM is selected as the classification algorithm to predict user intentions.

[0124] Finally, the results of Model_Z (the third global data model), Model_X (logistic regression model) and Model_Y (the fourth global data model) are linearly integrated to obtain the click conversion status of the search results.

[0125] The above-mentioned model based on search terms and subsequent behaviors aims to improve click-through conversion and solve the problem of sorting massive data after search. The massive product sorting model based on search terms and subsequent behaviors, when active search is triggered, displays relevant product information results based on search terms and performs personalized sorting on the displayed results, thereby improving user satisfaction and conversion rate.

[0126] Figure 2 FIG. 1 is a flow chart of an object prediction method in one embodiment. It should be understood that although Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0127] In one embodiment, Fig.13 As shown, an object prediction device 200 is provided, comprising:

[0128] The data acquisition module 201 is used to acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed;

[0129] A feature extraction module 202 is used to extract data features from user behavior data;

[0130] The time information model 203 is used to receive data features and obtain a first prediction result of the object to be analyzed;

[0131] The first prediction model 204 is used to receive the first prediction result and data features of the object to be analyzed, and output a first prediction probability of the object to be analyzed;

[0132] The second prediction model 205 is used to receive data features and output a second prediction probability of the object to be analyzed;

[0133] The logistic regression model 206 is used to receive the first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed, and output the probability of the object to be analyzed.

[0134] In one embodiment, if the second prediction model 205 includes multiple sub-models of different types, the second prediction model 205 is specifically used to input data features into each sub-model respectively, obtain the reference probability of the object to be analyzed corresponding to each sub-model, and perform weighted summation on the reference probability of the object to be analyzed corresponding to each sub-model to obtain the second prediction probability of the object to be analyzed.

[0135] In one embodiment, Fig.14 As shown, the second prediction model 205 includes a prediction day model 2051 and a global data sub-model 2052, and the reference probability of the object to be analyzed includes a first reference sub-probability and a second reference sub-probability.

[0136] The prediction day model 2051 is used to receive data features and obtain a first reference sub-probability.

[0137] The global data sub-model 2052 is used to receive data features and obtain a second reference sub-probability.

[0138] In one embodiment, Fig.14 As shown, the global data sub-model 2052 includes a first global data unit 20521 and a second global data unit 20522.

[0139] The first global data unit 20521 is used to receive the data feature and obtain the sub-probability corresponding to the first global data unit;

[0140] The second global data unit 20522 is used to receive the data feature and obtain the sub-probability corresponding to the second global data unit;

[0141] The global data sub-model 2052 specifically performs a weighted summation of the sub-probability corresponding to the first global data unit and the sub-probability of the second global data unit to obtain a second reference sub-probability.

[0142] In one embodiment, the feature extraction module 202 is specifically used to extract the features of the predicted day behavior data to obtain the predicted day data features, extract the features of the historical behavior data to obtain the historical data features, and the predicted day data features and the historical data features constitute the data features when the user behavior data includes the predicted day behavior data on the predicted day and the historical behavior data before the predicted day.

[0143] In one embodiment, the feature extraction module 202 is further used to extract features from the historical behavior data when the user behavior data includes historical behavior data before the prediction date, obtain historical data features, and use the historical data features as data features.

[0144] In one embodiment, when the user behavior data includes predicted day behavior data on the predicted day and historical behavior data before the predicted day, the data features input by the time information model 203, the first global data unit 20521, and the second global data unit 20522 are historical data features, and the data features input by the predicted day model 2051 and the first prediction model 204 include historical data features and predicted day data features.

[0145] In one embodiment, the data features include a basic feature set, a statistical feature set, and a cross feature set. The object prediction device 200 further includes:

[0146] A third global data model is used to receive the basic feature set and the cross feature set to obtain a third prediction probability of the object to be analyzed;

[0147] The first weighting module is used to calculate the weighted sum of the third predicted probability of the object to be analyzed and the probability of the object to be analyzed to obtain the target probability of the object to be analyzed.

[0148] In one embodiment, the object prediction device 200 further includes:

[0149] A fourth global data model, used to receive the basic feature set and the statistical feature set, and obtain a fourth predicted probability of the object to be analyzed;

[0150] The second weighting module is used to calculate the weighted sum of the fourth predicted probability of the object to be analyzed and the target probability of the object to be analyzed to obtain the weighted probability of the object to be analyzed.

[0151] In one embodiment, the data features include a basic feature set, a statistical feature set, and a cross feature set. The object prediction device 200 further includes:

[0152] a fifth global data model, used to receive the basic feature set and the statistical feature set, and obtain a fifth predicted probability of the object to be analyzed;

[0153] The third weighting module is used to calculate the weighted sum of the fifth predicted probability of the object to be analyzed and the probability of the object to be analyzed to obtain the target probability of the object to be analyzed.

[0154] In one embodiment, the fifth global data model and the fourth global data model are the same data model.

[0155] In one embodiment, the above-mentioned object prediction device 200 includes: a third global data model and a fourth global data model. When the user behavior data includes predicted day behavior data on the prediction day and historical behavior data before the prediction day, the basic feature sets input into the third global data model and the fourth global data model are the basic feature sets contained in the historical data features, the cross feature set input into the third global data model is the cross feature set contained in the historical data features, and the statistical feature set input into the fourth global data model is the statistical feature set contained in the historical data features.

[0156] In one embodiment, the object prediction device 200 includes: the time information model 203, the first prediction model 204, the prediction day model 2051, the first global data unit 20521, the third global data model and the fourth global data model. The classification algorithm is a lightweight gradient boosting algorithm, and the classification algorithm of the second global data unit 20522 is extreme gradient boosting.

[0157] Fig.16 The internal structure diagram of a computer device in one embodiment is shown. The computer device may specifically be Figure 1 The terminal 110 (or server 120) in FIG. Fig.16 As shown, the computer device includes a processor, a memory, a network interface, an input device and a display screen connected via a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program, which, when executed by the processor, enables the processor to implement the object prediction method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to execute the object prediction method. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0158] Those skilled in the art will understand that Fig.16 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0159] In one embodiment, the object prediction device provided by the present application can be implemented in the form of a computer program. The computer program can be Fig.16The computer device is run on the computer device shown. The memory of the computer device can store various program modules constituting the object prediction device, such as: Fig.13 The data acquisition module 201, feature extraction module 202, time information model 203, first prediction model 204, second prediction model 205 and logistic regression model 206 are shown. The computer program composed of various program modules enables the processor to execute the steps of the object prediction method of each embodiment of the present application described in this specification.

[0160] For example, Fig.16 The computer device shown can be Fig.13 The data acquisition module 201 in the object prediction device shown is used to obtain user behavior data within a preset time period, and the user behavior data includes object information of the object to be analyzed. The computer device can extract data features from the user behavior data through the feature extraction module 202. The computer device can receive data features through the time information model 203 to obtain a first prediction result of the object to be analyzed. The computer device can receive the first prediction result and data features of the object to be analyzed through the first prediction model 204, and output the first prediction probability of the object to be analyzed. The computer device can receive data features through the second prediction model 205, and output the second prediction probability of the object to be analyzed. The computer device can receive the first prediction probability of the object to be analyzed and the second prediction probability of the object to be analyzed through the logistic regression model 206, and output the probability of the object to be analyzed.

[0161] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in any one of the above-mentioned object prediction methods when executing the computer program.

[0162] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in any embodiment of the above-mentioned object prediction method are implemented.

[0163] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0164] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0165] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. An object prediction method, characterized in that: The method comprises: Acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed; Extracting data features from the user behavior data; Inputting the data feature into the time information model to obtain a first prediction result of the object to be analyzed; Inputting the first prediction result of the object to be analyzed and the data feature into a first prediction model to obtain a first prediction probability of the object to be analyzed; Inputting the data features into a second prediction model to obtain a second prediction probability of the object to be analyzed; The first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed are input into the logistic regression model, and the probability of the object to be analyzed is output.

2. The method according to claim 1, characterized in that If the second prediction model includes a plurality of sub-models of different types, inputting the data feature into the second prediction model to obtain a second prediction probability of the object to be analyzed includes: Inputting the data features into each of the sub-models respectively to obtain a reference probability of the object to be analyzed corresponding to each of the sub-models; A weighted sum is performed on the reference probability of the object to be analyzed corresponding to each of the sub-models to obtain a second predicted probability of the object to be analyzed.

3. The method according to claim 2, characterized in that The second prediction model includes a prediction day model and a global data sub-model, the reference probability of the object to be analyzed includes a first reference sub-probability and a second reference sub-probability, and the data features are input into each of the sub-models respectively to obtain the reference probability of the object to be analyzed corresponding to each of the sub-models, including: Inputting the data feature into the prediction day model to obtain the first reference sub-probability; The data feature is input into the global data sub-model to obtain the second reference sub-probability.

4. The method according to claim 3, characterized in that The global data sub-model includes a first global data unit and a second global data unit, and the step of inputting the data feature into the global data sub-model to obtain the second reference sub-probability includes: Inputting the data feature into the first global data unit to obtain a sub-probability corresponding to the first global data unit; Inputting the data feature into the second global data unit to obtain a sub-probability corresponding to the second global data unit; The sub-probability corresponding to the first global data unit and the sub-probability of the second global data unit are weighted and summed to obtain the second reference sub-probability.

5. The method according to claim 4, characterized in that The extracting data features from the user behavior data includes: When the user behavior data includes the predicted day behavior data on the predicted day and the historical behavior data before the predicted day, extracting the features of the predicted day behavior data to obtain the predicted day data features, extracting the features of the historical behavior data to obtain the historical data features, the predicted day data features and the historical data features constitute the data features; When the user behavior data includes historical behavior data before the prediction date, features in the historical behavior data are extracted to obtain historical data features, and the historical data features are used as the data features.

6. The method according to claim 5, characterized in that When the user behavior data includes predicted day behavior data on the predicted day and historical behavior data before the predicted day, the data features input by the time information model, the first global data unit, and the second global data unit are the historical data features, and the data features input by the predicted day model and the first prediction model include the historical data features and the predicted day data features.

7. The method according to any one of claims 1 to 6, characterized in that The data features include a basic feature set, a statistical feature set and a cross feature set, and the method further includes: Inputting the basic feature set and the cross feature set into a third global data model to obtain a third predicted probability of the object to be analyzed; A weighted sum of the third predicted probability of the object to be analyzed and the probability of the object to be analyzed is calculated to obtain a target probability of the object to be analyzed.

8. The method according to claim 7, characterized in that The method further comprises: Inputting the basic feature set and the statistical feature set into a fourth global data model to obtain a fourth predicted probability of the object to be analyzed; A weighted sum of the fourth predicted probability of the object to be analyzed and the target probability of the object to be analyzed is calculated to obtain a weighted probability of the object to be analyzed.

9. The method according to any one of claims 1 to 6, characterized in that: The data features include a basic feature set, a statistical feature set and a cross feature set, and the method further includes: Inputting the basic feature set and the statistical feature set into a fifth global data model to obtain a fifth predicted probability of the object to be analyzed; A weighted sum of the fifth predicted probability of the object to be analyzed and the probability of the object to be analyzed is calculated to obtain a target probability of the object to be analyzed.

10. The method according to claim 8, characterized in that The method includes a third global data model and a fourth global data model, When the user behavior data includes the predicted day behavior data on the predicted day and the historical behavior data before the predicted day, The basic feature sets input into the third global data model and the fourth global data model are the basic feature sets contained in the historical data features, the cross feature set input into the third global data model is the cross feature set contained in the historical data features, and the statistical feature set input into the fourth global data model is the statistical feature set contained in the historical data features.

11. The method according to claim 8, characterized in that The method includes a first global data unit, a second global data unit and a third global data model, The classification algorithms of the time information model, the first prediction model, the prediction day model, the first global data unit, the third global data model and the fourth global data model are lightweight gradient boosting algorithms, and the classification algorithm of the second global data unit is extreme gradient boosting.

12. An object prediction device, characterized in that: The device comprises: A data acquisition module, used to acquire user behavior data within a preset time period, wherein the user behavior data includes object information of an object to be analyzed; A feature extraction module, used to extract data features from the user behavior data; A time information model, used to receive the data feature and obtain a first prediction result of the object to be analyzed; A first prediction model, used to receive a first prediction result of the object to be analyzed and the data feature, and output a first prediction probability of the object to be analyzed; A second prediction model, used to receive the data features and output a second prediction probability of the object to be analyzed; The logistic regression model is used to receive the first predicted probability of the object to be analyzed and the second predicted probability of the object to be analyzed, and output the probability of the object to be analyzed.

13. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images