Recommended method, device, electronic equipment and readable storage medium

Abstract

The present disclosure provides a recommendation method and device, electronic equipment and readable storage medium, relating to the technical field of artificial intelligence such as cloud service and deep learning. The recommendation method comprises: obtaining target feature data and feature data of at least one candidate promotion information; generating at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information; obtaining prediction results corresponding to different preset targets respectively for the at least one candidate feature data pair; determining a recommendation score of the at least one candidate promotion information according to the prediction results corresponding to the different preset targets respectively, and determining a target promotion information from the at least one candidate promotion information according to the recommendation score. The present disclosure can improve the recommendation accuracy and efficiency of the promotion information by obtaining the prediction results corresponding to multiple preset targets to recommend the promotion information.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computers, in particular to the technical field of artificial intelligence such as cloud services and deep learning, and specifically provides a recommendation method and device, an electronic device, and a readable storage medium. BACKGROUND

[0002] In the prior art, when recommending promotion information, only a single target is considered, for example, only one of the click rate, conversion rate, and other targets is considered, so that the recommendation accuracy of the promotion information is low. SUMMARY

[0003] According to a first aspect of the present disclosure, a recommendation method is provided, including: obtaining target feature data and feature data of at least one candidate promotion information; generating at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information; obtaining prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively; determining a recommendation score of the at least one candidate promotion information according to the prediction results corresponding to the different preset targets respectively, and determining a target promotion information from the at least one candidate promotion information according to the recommendation score.

[0004] According to a second aspect of the present disclosure, a recommendation device is provided, including: an obtaining unit configured to obtain target feature data and feature data of at least one candidate promotion information; a generating unit configured to generate at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information; a prediction unit configured to obtain prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively; and a recommendation unit configured to determine a recommendation score of the at least one candidate promotion information according to the prediction results corresponding to the different preset targets respectively, and determine a target promotion information from the at least one candidate promotion information according to the recommendation score.

[0005] According to a third aspect of the present disclosure, an electronic device is provided, including: at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0006] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to perform the method described above.

[0007] According to a fifth aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the method as described above.

[0008] From the above technical solutions, it can be seen that the present disclosure obtains the prediction results of the candidate promotion information corresponding to different preset targets by constructing the feature data pair of the feature data comprising the target feature data and the candidate promotion information, and then determines the target promotion information according to the recommendation scores of the candidate promotion information determined according to the prediction results corresponding to different preset targets respectively, so as to recommend the promotion information by obtaining the prediction results corresponding to multiple preset targets, which can improve the recommendation accuracy and efficiency of the promotion information.

[0009] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0010] The accompanying drawings are used to better understand the present scheme and do not constitute a limitation on the present disclosure. Among them:

[0011] Figure 1 is a schematic diagram according to the first embodiment of the present disclosure;

[0012] Figure 2 is a schematic diagram according to the second embodiment of the present disclosure;

[0013] Figure 3 is a schematic diagram according to the third embodiment of the present disclosure;

[0014] Figure 4 is a schematic diagram according to the fourth embodiment of the present disclosure;

[0015] Figure 5 is a schematic diagram according to the fifth embodiment of the present disclosure;

[0016] Figure 6 is a schematic diagram according to the sixth embodiment of the present disclosure;

[0017] Figure 7 is a block diagram of an electronic device for implementing the recommendation method of the embodiments of the present disclosure. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure are described herein with reference to the accompanying drawings, which are cited as illustrative examples. Therefore, it will be recognized by those of ordinary skill in the art that various changes and modifications can be made to the embodiments described and subsets thereof without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and components are omitted for the sake of brevity and clarity.

[0019] Figure 1 is a schematic diagram according to a first embodiment of the present disclosure. As shown in Figure 1 , the recommendation method of the present embodiment specifically includes the following steps:

[0020] S101, obtaining target feature data and feature data of at least one candidate promotion information;

[0021] S102, generating at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information;

[0022] S103, obtaining prediction results corresponding to different preset targets respectively for the at least one candidate feature data pair;

[0023] S104, determining a recommendation score of the at least one candidate promotion information according to the prediction results corresponding to different preset targets respectively, and determining a target promotion information from the at least one candidate promotion information according to the recommendation score.

[0024] That is, the present embodiment obtains the prediction results corresponding to different preset targets for the candidate promotion information by constructing the feature data pair containing the target feature data and the feature data of the candidate promotion information, and then determines the target promotion information according to the recommendation score of the candidate promotion information determined by the prediction results corresponding to different preset targets. The recommendation of the promotion information is performed by obtaining the prediction results corresponding to multiple preset targets, which can improve the recommendation accuracy and efficiency of the promotion information.

[0025] The recommendation method of the present embodiment obtains the prediction results corresponding to different preset targets respectively for each candidate feature data pair after generating at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information, and then generates the recommendation score of the at least one candidate promotion information according to the prediction results, so as to achieve the purpose of determining the target promotion information according to the recommendation score, which can improve the recommendation accuracy and efficiency of the promotion information.

[0026] The target feature data obtained by the present embodiment in S101 is the feature data corresponding to the recommendation object of the candidate promotion information.

[0027] The candidate feature data pair generated by the embodiment of S102 corresponds to different candidate promotion information respectively.

[0028] The prediction result obtained by the embodiment of S103 corresponds to different candidate feature data pairs, that is, different prediction results correspond to different candidate promotion information. The different preset targets in the embodiment can include at least two of the click rate target, the conversion rate target, the complete playback rate target, the delay estimation target, and the like.

[0029] In the process of obtaining the prediction result of the at least one candidate feature data pair corresponding to the different preset targets by the embodiment of S103, the similarity between the target feature data in each candidate feature data pair and the feature data of the candidate promotion information can be calculated, and the similarity calculation result can be taken as the prediction result.

[0030] In order to further improve the accuracy of the obtained prediction result, in the process of obtaining the prediction result of the at least one candidate feature data pair corresponding to the different preset targets by the embodiment of S103, the optional implementation manner can be that the at least one candidate feature data pair is input into the recommendation model to obtain the prediction result output by the recommendation model for each candidate feature data pair, which corresponds to the different preset targets respectively.

[0031] The training process of the recommendation model used by the embodiment of S103 is described in detail below.

[0032] In the process of determining the recommendation score of the at least one candidate promotion information according to the prediction result corresponding to the different preset targets by the embodiment of S104, the prediction results corresponding to the different preset targets of one candidate feature data pair can be added or averaged, and then the addition result or the average value can be taken as the recommendation score of the corresponding candidate promotion information.

[0033] In the process of determining the target promotion information from the at least one candidate promotion information according to the recommendation score by the embodiment of S104, the candidate promotion information can be arranged in the order from high to low according to the recommendation score, and the candidate promotion information arranged in the top N positions can be taken as the target promotion information, where N is a positive integer greater than or equal to 1.

[0034] After the target promotion information is determined by the embodiment of S104, the determined target promotion information can be sent to the input end corresponding to the target feature data, so that the received target promotion information is displayed in the input end.

[0035] Figure 2 is a schematic diagram according to the second embodiment of the present disclosure. As shown in Figure 2 The recommendation model of the embodiment is trained in the following manner:

[0036] S201, obtain training data, the training data containing a plurality of feature data pairs and a plurality of feature data pairs respectively corresponding to different preset targets and a label result, each feature data pair containing first feature data and second feature data;

[0037] S202, construct a neural network model containing an embedding layer, two groups of deep neural network layers and an output layer, the embedding layer being used for respectively outputting embedding vectors of the first feature data corresponding to different preset targets and embedding vectors of the second feature data corresponding to different preset targets according to input feature data pairs, the two groups of deep neural network layers respectively corresponding to the first feature data and the second feature data, each group of deep neural network layers being used for respectively outputting feature vectors of the feature data corresponding to different preset targets according to embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, and the output layer being used for outputting prediction results of the feature data pairs respectively corresponding to different preset targets according to feature vectors corresponding to different preset targets output by the two groups of deep neural network layers;

[0038] S203, train the neural network model using the plurality of feature data pairs and the label result of the plurality of feature data pairs respectively corresponding to different preset targets, to obtain a recommendation model.

[0039] The training method of the recommendation model of the embodiment, by constructing a neural network model containing two groups of deep neural network layers, enables the neural network model to use different deep neural network layers to respectively obtain feature vectors of the first feature data and the second feature data corresponding to different preset targets, realizes the purpose of independently representing embedding vectors of different preset targets corresponding to different feature data, avoids the technical problem that training tasks corresponding to different preset targets will affect each other, thereby enhancing the training effect of the neural network model and improving the accuracy of the recommendation model obtained by training in obtaining prediction results.

[0040] The training data obtained by the embodiment S201 contains a plurality of feature data pairs, each feature data pair containing first feature data and second feature data; if the first feature data in the embodiment is input end feature data, the second feature data is promotion information feature data; if the first feature data in the embodiment is promotion information feature data, the second feature data is input end feature data.

[0041] Among them, the input end feature data in the embodiment can be model, name and other data of the input end, or gender, age, occupation and other data of the user corresponding to the input end; the promotion information feature data in the embodiment can be title, industry and other data of the promotion information.

[0042] The preset target in the embodiment can be set according to actual needs. The preset target can include at least two of a click rate target, a conversion rate target, a complete playback rate target, a deep conversion rate target, and a delay estimation target.

[0043] In the training data obtained by the embodiment in S201, the feature data pairs correspond to different preset target labels respectively, and the labels include at least two of the above targets, for example, the labels corresponding to the click rate target and the labels corresponding to the conversion rate target.

[0044] In addition, when obtaining the training data in S201, the embodiment can obtain feature data pairs corresponding to different application scenarios as training data, for example, multiple feature data pairs composed of input feature data and promotion information feature data in the recommendation scenario and the search scenario as training data, so as to realize the purpose that the trained recommendation model can predict results in different application scenarios, and improve the utilization rate of feature data in different application scenarios.

[0045] After obtaining the training data including multiple feature data pairs and multiple feature data pairs corresponding to different preset target labels in S201, the embodiment performs S202 to construct a neural network model including an embedding layer, two groups of deep neural network layers, and an output layer.

[0046] Specifically, the embedding layer in the neural network model constructed by the embodiment in S202 is used to output embedding vectors corresponding to different preset targets for the first feature data pair and embedding vectors corresponding to different preset targets for the second feature data pair according to the input feature data pair.

[0047] For example, if the preset target includes a click rate target and a conversion rate target, the embedding layer in the embodiment outputs embedding vectors corresponding to the click rate target and embedding vectors corresponding to the conversion rate target for the first feature data according to the first feature data; and outputs embedding vectors corresponding to the click rate target and embedding vectors corresponding to the conversion rate target for the second feature data according to the second feature data.

[0048] Specifically, the two groups of deep neural network layers included in the neural network model constructed by the embodiment in S202 correspond to the first feature data and the second feature data respectively; each group of deep neural network layers includes multiple deep neural network structures, and each deep neural network structure corresponds to a different preset target, so the number of deep neural network structures included in each group of deep neural network layers is the same as the number of preset targets.

[0049] For example, if the annotation result of the feature data pair obtained by the embodiment performing S201 corresponds to two preset targets, then in the neural network model constructed by the embodiment performing S202, each group of deep neural network layers includes two deep neural network structures; if the annotation result of the feature data pair obtained by the embodiment performing S201 corresponds to three preset targets, then in the neural network model constructed by the embodiment performing S202, each group of deep neural network layers includes three deep neural network structures.

[0050] Each group of deep neural network layers in the neural network model constructed by the embodiment performing S202 is configured to respectively output feature vectors corresponding to different preset targets of the same feature data (the first feature data or the second feature data) according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data.

[0051] Specifically, when each group of deep neural network layers in the embodiment respectively outputs feature vectors corresponding to different preset targets of the same feature data according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, an optional implementation manner that can be adopted is: for each group of deep neural network layers, determining the preset target corresponding to each deep neural network structure in the group of deep neural network layers; inputting the embedding vectors corresponding to different preset targets into the corresponding deep neural network structures respectively to obtain the feature vectors corresponding to different preset targets output by each deep neural network structure.

[0052] That is, each group of deep neural network layers in the embodiment respectively obtains feature vectors corresponding to different preset targets according to embedding vectors corresponding to different preset targets by using different deep neural network structures, thereby avoiding mutual influence between embedding vectors corresponding to different preset targets when obtaining feature vectors corresponding to different preset targets, and improving the accuracy of the obtained feature vectors corresponding to different preset targets.

[0053] For example, if the preset targets include a click rate target and a conversion rate target, then each group of deep neural network layers constructed by the embodiment performing S102 includes two deep neural network structures, one of which is configured to output a feature vector corresponding to the click rate target of the feature data according to the embedding vector corresponding to the click rate target of the feature data, and the other of which is configured to output a feature vector corresponding to the conversion rate target of the feature data according to the embedding vector corresponding to the conversion rate target of the feature data.

[0054] In addition, in the embodiment, each group of deep neural network layers can adopt an optional implementation manner that, for each group of deep neural network layers, the embedding vectors corresponding to different preset targets are spliced, and the spliced results are respectively input into each deep neural network structure in the group of deep neural network layers to obtain the feature vectors corresponding to different preset targets output by each deep neural network structure.

[0055] That is, in the embodiment, each group of deep neural network layers can also utilize different deep neural network structures to obtain the feature vectors corresponding to different preset targets from the embedding vectors corresponding to all preset targets of the feature data.

[0056] Specifically, the output layer in the neural network model constructed in S202 is used to output the prediction results of the feature data pair corresponding to different preset targets according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers.

[0057] In the embodiment, the output layer can adopt an optional implementation manner that, for each preset target, the feature vector corresponding to the preset target is obtained, and the prediction result of the feature data pair corresponding to the preset target is obtained according to the obtained feature vector, for example, the inner product of the obtained feature vector is calculated to obtain the prediction result of one preset target.

[0058] In addition, if the preset targets in the embodiment include the click rate target and the conversion rate target, after obtaining the prediction result corresponding to the click rate target and the prediction result corresponding to the conversion rate target, the output layer can also calculate the display conversion rate from the two prediction results as the prediction result of the feature data pair corresponding to the click rate target and the conversion rate target.

[0059] After constructing the neural network model including the embedding layer, the two groups of deep neural network layers and the output layer in S202, the embodiment trains the neural network model using the obtained multiple feature data pairs and the labeled results of the multiple feature data pairs corresponding to different preset targets in S203 to obtain a recommendation model.

[0060] Specifically, when performing S203, the neural network model is trained using the plurality of feature data pairs and the labeled results corresponding to different preset targets of the plurality of feature data pairs respectively, to obtain the recommendation model, the optional implementation manner that can be adopted is as follows: the plurality of feature data pairs are respectively input into the neural network model, to obtain the prediction results of the neural network model output for each feature data pair, the prediction results of the feature data pair corresponding to different preset targets respectively; the loss function value is calculated according to the prediction results and the labeled results corresponding to different preset targets of each feature data respectively; the parameters in the neural network model are adjusted according to the calculated loss function value, until the neural network model converges, to obtain the recommendation model.

[0061] It can be understood that, if the preset targets of the embodiment include the click rate target and the conversion rate target, when the loss function value is calculated according to the prediction results and the labeled results corresponding to different preset targets, for the click rate target, the prediction results and the labeled results corresponding to the click rate target are used to calculate the loss function value corresponding to the click rate target; for the conversion rate target, the prediction results and the labeled results corresponding to the conversion rate target, and the prediction results and the labeled results corresponding to the display conversion rate are used to calculate the loss function value together, which can improve the coverage of data.

[0062] The recommendation model trained by the embodiment can obtain the prediction results corresponding to different preset targets according to the input feature data and the promotion information feature data, and then determine the recommendation score of the promotion information according to the prediction results of each preset target, to achieve the purpose of recommending the promotion information to the input end according to the recommendation score of the promotion information.

[0063] Figure 3 is a schematic diagram according to the third embodiment of the disclosure. Figure 3 The architecture diagram of the neural network model constructed by the embodiment is shown; in the neural network model constructed by the embodiment, the preset targets are the click rate (CTR) target and the conversion rate (CVR) target, each group of deep neural network layers includes two deep neural network structures respectively, and each deep neural network structure outputs a feature vector (for example, a feature vector of a click rate target) corresponding to different preset targets according to an embedding vector (for example, an embedding vector of a click rate target) corresponding to different preset targets.

[0064] Figure 4 is a schematic diagram according to the fourth embodiment of the disclosure. As Figure 4 shown, when performing S203, the neural network model is trained using the plurality of feature data pairs and the labeled results corresponding to different preset targets of the plurality of feature data pairs respectively, to obtain the recommendation model, the embodiment specifically includes the following steps:

[0065] S401, training the neural network model using the plurality of feature data pairs and the annotation results corresponding to different preset targets respectively, to obtain a first neural network model, each deep neural network structure of each group of deep neural network layers in the neural network model being configured to output a feature vector corresponding to different preset targets according to embedding vectors corresponding to different preset targets;

[0066] S402, training the first neural network model using the plurality of feature data pairs and the annotation results corresponding to different preset targets respectively, to obtain the recommendation model, each deep neural network structure of each group of deep neural network layers in the first neural network model being configured to output a feature vector corresponding to different preset targets according to embedding vectors corresponding to all preset targets.

[0067] That is, the embodiment trains the neural network model to obtain the recommendation model through two training stages, the input of each deep neural network structure in the first training stage being the embedding vector of the preset target corresponding to the deep neural network structure, and the input of each deep neural network structure in the second training stage being the embedding vector of all preset targets. Through the manner of separately training each preset target and then jointly training the preset targets, the training effect of the neural network model is further improved, and the recommendation model obtained by training has higher recommendation accuracy.

[0068] Among them, the training processes of the two training stages of the embodiment are similar, both of which are to adjust the parameters of the network model through the loss function value calculated from the prediction result and the annotation result output by the neural network model for the feature data pair, until the network model converges, to obtain the first neural network model and the recommendation model.

[0069] Figure 5 is a schematic diagram according to the fifth embodiment of the disclosure. Figure 5 The flowchart of training the neural network model of the embodiment is shown; phase one is used to train the neural network model to obtain the first neural network model, and phase two is used to train the first neural network model to obtain the recommendation model.

[0070] Figure 6 is a schematic diagram according to the sixth embodiment of the disclosure. As Figure 6 shown, the recommendation device 600 of the embodiment includes:

[0071] The acquisition unit 601 is configured to acquire target feature data and feature data of at least one candidate promotion information;

[0072] The generation unit 602 is configured to generate at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information.

[0073] The prediction unit 603 obtains prediction results respectively corresponding to different preset targets for each candidate feature data pair.

[0074] The recommendation unit 604 determines a recommendation score of at least one candidate promotion information according to the prediction results respectively corresponding to different preset targets, and determines the target promotion information from the at least one candidate promotion information according to the recommendation score.

[0075] The target feature data obtained by the acquisition unit 601 is the feature data corresponding to the recommendation object of the candidate promotion information.

[0076] Each candidate feature data pair generated by the generation unit 602 respectively corresponds to different candidate promotion information.

[0077] The prediction results obtained by the prediction unit 603 correspond to different candidate feature data pairs, that is, different prediction results correspond to different candidate promotion information; the different preset targets can include at least two of the click rate target, the conversion rate target, the complete playback rate target, and the delay estimation target.

[0078] When obtaining the prediction results respectively corresponding to different preset targets for each candidate feature data pair, the prediction unit 603 can calculate the similarity between the target feature data in each candidate feature data pair and the feature data of the candidate promotion information, and take the similarity calculation result as the prediction result.

[0079] In order to further improve the accuracy of the obtained prediction results, when obtaining the prediction results respectively corresponding to different preset targets for each candidate feature data pair, the prediction unit 603 can adopt an optional implementation manner: inputting the at least one candidate feature data pair into a recommendation model to obtain the prediction results respectively corresponding to different preset targets output by the recommendation model for each candidate feature data pair.

[0080] When determining the recommendation score of at least one candidate promotion information according to the prediction results respectively corresponding to different preset targets, the recommendation unit 604 can add or average the prediction results respectively corresponding to different preset targets corresponding to one candidate feature data pair, and then take the addition result or the average value as the recommendation score of the corresponding candidate promotion information.

[0081] When determining the target promotion information from the at least one candidate promotion information according to the recommendation score, the recommendation unit 604 can arrange the candidate promotion information in the order from high to low according to the recommendation score, take the candidate promotion information arranged in the top N positions as the target promotion information, and N is a positive integer greater than or equal to 1.

[0082] The recommendation unit 604 can send the determined target promotion information to the input end corresponding to the target feature data, so as to display the received target promotion information in the input end.

[0083] The recommendation device 600 of the embodiment can further include a training unit 605 configured to train the recommendation model in the following manner: obtaining training data, the training data including a plurality of feature data pairs and a plurality of label results corresponding to the plurality of feature data pairs, each feature data pair including a first feature data and a second feature data; constructing a neural network model including an embedding layer, two groups of deep neural network layers, and an output layer, the embedding layer being configured to output embedding vectors corresponding to different preset targets for the first feature data and the second feature data according to an input feature data pair, each group of deep neural network layers corresponding to the first feature data and the second feature data, each group of deep neural network layers being configured to output feature vectors corresponding to different preset targets for the first feature data and the second feature data according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, and the output layer being configured to output prediction results corresponding to different preset targets for the feature data pair according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers; and training the neural network model using the plurality of feature data pairs and the plurality of label results corresponding to the plurality of feature data pairs to obtain the recommendation model.

[0084] The training data obtained by the training unit 605 includes a plurality of feature data pairs, each feature data pair including a first feature data and a second feature data; if the first feature data in the embodiment is input end feature data, the second feature data is promotion information feature data; if the first feature data in the embodiment is promotion information feature data, the second feature data is input end feature data.

[0085] The label results corresponding to different preset targets of the feature data pairs in the training data obtained by the training unit 605 include label results corresponding to at least two of the above targets, for example, label results corresponding to a click rate target and label results corresponding to a conversion rate target.

[0086] In addition, when obtaining the training data, the training unit 605 can obtain feature data pairs corresponding to different application scenarios as the training data, so as to achieve the purpose of the trained recommendation model being able to predict results in different application scenarios, and improve the utilization rate of feature data in different application scenarios.

[0087] The training unit 605, after obtaining the training data containing a plurality of feature data pairs and the labeling results that the plurality of feature data pairs correspond to different preset targets respectively, constructs a neural network model containing an embedding layer, two groups of deep neural network layers and an output layer.

[0088] Specifically, the embedding layer in the neural network model constructed by the training unit 605 is configured to output, according to the input feature data pair, the embedding vectors of the first feature data corresponding to different preset targets and the embedding vectors of the second feature data corresponding to different preset targets respectively.

[0089] Specifically, the two groups of deep neural network layers contained in the neural network model constructed by the training unit 605 correspond to the first feature data and the second feature data respectively; each group of deep neural network layers contains a plurality of deep neural network structures, and each deep neural network structure corresponds to a different preset target, so that the number of deep neural network structures contained in each group of deep neural network layers is the same as the number of preset targets.

[0090] Each group of deep neural network layers in the neural network model constructed by the training unit 605 is configured to output, according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data (the first feature data or the second feature data), the feature vectors of the feature data corresponding to different preset targets respectively.

[0091] Specifically, when each group of deep neural network layers constructed by the training unit 605 outputs, according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, the feature vectors of the feature data corresponding to different preset targets respectively, an optional implementation manner that can be adopted is that, for each group of deep neural network layers, the preset target corresponding to each deep neural network structure in the group of deep neural network layers is determined; the embedding vectors corresponding to different preset targets are input into the corresponding deep neural network structures respectively, and the feature vectors corresponding to different preset targets output by each deep neural network structure are obtained.

[0092] That is, each group of deep neural network layers constructed by the training unit 605 uses different deep neural network structures to obtain the feature vectors corresponding to different preset targets respectively according to the embedding vectors corresponding to different preset targets, thereby avoiding the mutual influence between the embedding vectors corresponding to different preset targets when the feature vectors corresponding to different preset targets are obtained, and improving the accuracy of the obtained feature vectors corresponding to different preset targets.

[0093] In addition, when each group of deep neural network layers constructed by the training unit 605 outputs the feature vectors corresponding to different preset targets of the feature data according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, an optional implementation manner that can be adopted is that, for each group of deep neural network layers, the embedding vectors corresponding to different preset targets are spliced; the spliced results are respectively input into each deep neural network structure in the group of deep neural network layers, so as to obtain the feature vectors corresponding to different preset targets output by each deep neural network structure.

[0094] That is, each group of deep neural network layers constructed by the training unit 605 can also obtain the feature vectors corresponding to different preset targets of the feature data according to the embedding vectors corresponding to all preset targets of the feature data by using different deep neural network structures.

[0095] Specifically, the output layer in the neural network model constructed by the training unit 605 is configured to output the prediction results of the feature data pairs corresponding to different preset targets respectively according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers.

[0096] When the output layer constructed by the training unit 605 outputs the prediction results of the feature data pairs corresponding to different preset targets respectively according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers, an optional implementation manner that can be adopted is that, for each preset target, the feature vector corresponding to the preset target is obtained; and the prediction result of the feature data pair corresponding to the preset target is obtained according to the obtained feature vector.

[0097] In addition, if the preset targets in the embodiment include the click rate target and the conversion rate target, after obtaining the prediction result corresponding to the click rate target and the prediction result corresponding to the conversion rate target, the output layer can also take the display conversion rate calculated from the two prediction results as the prediction result of the feature data pair corresponding to the click rate target and the conversion rate target.

[0098] After constructing the neural network model including the embedding layer, the two groups of deep neural network layers and the output layer, the training unit 605 trains the neural network model by using the obtained multiple feature data pairs and the label results of the multiple feature data pairs corresponding to different preset targets, so as to obtain the recommendation model.

[0099] Specifically, when training the neural network model using the plurality of feature data pairs and the label results corresponding to different preset targets respectively, the training unit 605 can adopt an optional implementation manner as follows: inputting the plurality of feature data pairs into the neural network model respectively to obtain prediction results of the neural network model output for each feature data pair, the prediction results corresponding to different preset targets respectively; calculating a loss function value according to the prediction results and the label results corresponding to different preset targets respectively; adjusting parameters in the neural network model according to the calculated loss function value until the neural network model converges, and obtaining the recommendation model.

[0100] It can be understood that, if the preset targets in the embodiment include the click rate target and the conversion rate target, when calculating the loss function value according to the prediction results and the label results corresponding to different preset targets, for the click rate target, the prediction results and the label results corresponding to the click rate target are used to calculate the loss function value corresponding to the click rate target; for the conversion rate target, the prediction results and the label results corresponding to the conversion rate target, and the prediction results and the label results corresponding to the display conversion rate are used to calculate the loss function value together, which can improve the coverage of data.

[0101] When training the neural network model using the plurality of feature data pairs and the label results corresponding to different preset targets respectively, the training unit 605 can also adopt the following manner: training the neural network model using the plurality of feature data pairs and the label results corresponding to different preset targets respectively to obtain a first neural network model, each deep neural network structure of each group of deep neural network layers in the neural network model is used to output a feature vector corresponding to different preset targets according to an embedding vector corresponding to different preset targets; training the first neural network model using the plurality of feature data pairs and the label results corresponding to different preset targets respectively to obtain a recommendation model, each deep neural network structure of each group of deep neural network layers in the first neural network model is used to output a feature vector corresponding to different preset targets according to an embedding vector corresponding to all preset targets.

[0102] That is, the training unit 605 also trains the neural network model to obtain the recommendation model through two training stages, in the first training stage, the input of each deep neural network structure is an embedding vector of a preset target corresponding to the deep neural network structure, and in the second training stage, the input of each deep neural network structure is an embedding vector of all preset targets, through the manner of separately training each preset target and then jointly training all preset targets, the training effect of the neural network model is further improved, and the recommendation model obtained by training has higher recommendation accuracy.

[0103] The training unit 605 adjusts the parameters of the neural network model according to the loss function value calculated based on the predicted result of the feature data output by the neural network model and the labeled result, in two training stages, until the neural network model converges, to obtain the first neural network model and the recommendation model.

[0104] In the technical solution of the present disclosure, the acquisition, storage and application of user personal information comply with relevant laws and regulations and do not violate public order and good customs.

[0105] According to embodiments of the present disclosure, the present disclosure further provides an electronic device, a readable storage medium and a computer program product.

[0106] As Figure 7 shown, is a block diagram of an electronic device for a recommendation method according to embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown in the figures, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit implementations of the present disclosure described and / or claimed in this document.

[0107] As Figure 7 shown, the device 700 includes a computing unit 701 that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 702 or a computer program loaded into a random access memory (RAM) 703 from a storage unit 708. In the RAM 703, various programs and data required for the operation of the device 700 can also be stored. The computing unit 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0108] Various components in the device 700 are connected to the I / O interface 705, including: an input unit 706, such as a keyboard, a mouse, etc.; an output unit 707, such as various types of displays, speakers, etc.; a storage unit 708, such as a magnetic disk, an optical disk, etc.; and a communication unit 709, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 709 allows the device 700 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunications networks.

[0109] The computing unit 701 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 701 performs various methods and processes described above, such as the recommendation method. For example, in some embodiments, the recommendation method can be implemented as a computer software program that is tangibly embodied in a machine-readable medium, such as the storage unit 708.

[0110] In some embodiments, portions or all of the computer program can be loaded onto the device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded onto the RAM 703 and executed by the computing unit 701, one or more steps of the recommendation method described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the recommendation method by any other appropriate means, such as by means of firmware.

[0111] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0112] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces the functions / operations specified in the flowcharts and / or the block diagrams. The program code can be embodied in whole or in part within a machine, executed partially on the machine, partially on one or more remote machines, and / or entirely on one or more remote machines or servers.

[0113] In the context of this disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0114] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0115] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0116] The computer system can include clients and servers. This relationship can be. The servers are typically remote from the clients with the interactions typically occurring over the network. The relationship between client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The servers can be cloud servers, also known as cloud computing servers or cloud hosts, which are a host product in the cloud computing service system to solve the defects of large management difficulty and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or simply "VPS"). The servers can also be servers of a distributed system, or servers combined with a blockchain.

[0117] It should be understood that the various forms of flow shown above can be reordered, steps added or removed. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, which are not limited herein.

[0118] The above detailed description does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A recommendation method, comprising: obtaining target feature data and feature data of at least one candidate promotion information; generating at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information, different candidate feature data pairs corresponding to different candidate promotion information; obtaining prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively; determining a recommendation score of at least one candidate promotion information according to the prediction results corresponding to different preset targets respectively, and determining a target promotion information from the at least one candidate promotion information according to the recommendation score; wherein the determining of the recommendation score of at least one candidate promotion information according to the prediction results corresponding to different preset targets respectively comprises: for each candidate feature data pair, adding or averaging the prediction results corresponding to different preset targets respectively output by the candidate feature data pair; taking the addition result or the average value result as the recommendation score of the candidate promotion information corresponding to the candidate feature data pair; wherein the obtaining of the prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively comprises: inputting the at least one candidate feature data pair into a recommendation model respectively to obtain the prediction results corresponding to different preset targets respectively output by the recommendation model for each candidate feature data pair; the recommendation model is obtained by training in the following manner: obtaining training data, the training data containing a plurality of feature data pairs and a plurality of label results of the plurality of feature data pairs corresponding to different preset targets respectively, each feature data pair containing first feature data and second feature data, the feature data containing input end feature data and promotion information feature data, the input end feature data being a model, a name of an input end, a gender, an age, and a profession of a user corresponding to the input end, the promotion information feature data being a title and an industry of the promotion information, the preset target containing at least two of a click rate target, a conversion rate target, a complete play rate target, a deep conversion rate target, and a delay estimation target; constructing a neural network model containing an embedding layer, two groups of deep neural network layers, and an output layer, the embedding layer being used to output embedding vectors of the first feature data corresponding to different preset targets and embedding vectors of the second feature data corresponding to different preset targets respectively according to the input feature data pair, the two groups of deep neural network layers respectively corresponding to the first feature data and the second feature data, each group of deep neural network layers being used to output feature vectors of the feature data corresponding to different preset targets respectively according to the embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, and the output layer being used to output prediction results of the feature data pair corresponding to different preset targets respectively according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers; training the neural network model using the plurality of feature data pairs and the plurality of label results of the plurality of feature data pairs corresponding to different preset targets respectively to obtain the recommendation model.

2. The method of claim 1, wherein, Each of the two groups of deep neural network layers comprises a plurality of deep neural network structures, and the number of the deep neural network structures comprised is the same as the number of the preset targets.

3. The method of claim 1, wherein, The outputting, by each of the two groups of deep neural network layers, of the feature vectors corresponding to different preset targets for the same feature data according to the embedding vectors corresponding to different preset targets output by the embedding layer for the feature data comprises: For each group of deep neural network layers, determining the preset target corresponding to each deep neural network structure in the group of deep neural network layers; The embedding vectors corresponding to different preset targets are respectively input into the corresponding deep neural network structures to obtain the feature vectors corresponding to different preset targets output by each deep neural network structure.

4. The method of claim 1, wherein, The outputting, by each of the two groups of deep neural network layers, of the feature vectors corresponding to different preset targets for the same feature data according to the embedding vectors corresponding to different preset targets output by the embedding layer for the feature data comprises: For each group of deep neural network layers, the embedding vectors corresponding to different preset targets are spliced; The spliced results are respectively input into each deep neural network structure in the group of deep neural network layers to obtain the feature vectors corresponding to different preset targets output by each deep neural network structure.

5. The method of claim 1, wherein, The outputting, by the output layer, of the prediction results of the feature data pairs corresponding to different preset targets according to the feature vectors corresponding to different preset targets output by the two groups of deep neural network layers comprises: For each preset target, the feature vector corresponding to the preset target is obtained; According to the feature vector, the prediction result of the feature data pair corresponding to the preset target is obtained.

6. The method of claim 1, wherein, The training of the neural network model using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain the recommendation model comprises: The plurality of feature data pairs are respectively input into the neural network model to obtain the prediction results of the feature data pairs corresponding to different preset targets output by the neural network model for each feature data pair; According to the prediction results and the labeled results of each feature data corresponding to different preset targets, a loss function value is calculated; According to the calculated loss function value, the parameters in the neural network model are adjusted until the neural network model converges, and the recommendation model is obtained.

7. The method of claim 1, wherein, The training of the neural network model using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain the recommendation model comprises: The neural network model is trained using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain a first neural network model, and each deep neural network structure in each group of deep neural network layers in the neural network model is used to output the feature vectors corresponding to different preset targets according to the embedding vectors corresponding to different preset targets; The first neural network model layer is trained using the plurality of feature data and label results corresponding to different preset targets respectively corresponding to the plurality of feature pairs, to obtain the recommendation model, each deep neural network structure of each group of deep neural network layers in the first neural network model being used to output feature vectors corresponding to different preset targets according to embedding vectors corresponding to all preset targets.

8. A recommendation apparatus, comprising: an acquisition unit configured to acquire target feature data and feature data of at least one candidate promotion information; a generation unit configured to generate at least one candidate feature data pair according to the target feature data and the feature data of the at least one candidate promotion information, different candidate feature data pairs corresponding to different candidate promotion information; a prediction unit configured to obtain prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively; a recommendation unit configured to determine a recommendation score of the at least one candidate promotion information according to the prediction results corresponding to different preset targets respectively, and determine a target promotion information from the at least one candidate promotion information according to the recommendation score; wherein, when determining the recommendation score of the at least one candidate promotion information according to the prediction results corresponding to different preset targets respectively, the recommendation unit is specifically configured to: for each candidate feature data pair, add or average the prediction results corresponding to different preset targets respectively; use the addition result or the average result as the recommendation score of the candidate promotion information corresponding to the candidate feature data pair; wherein, when obtaining the prediction results of the at least one candidate feature data pair corresponding to different preset targets respectively, the prediction unit is specifically configured to: input the at least one candidate feature data pair into a recommendation model to obtain the prediction results output by the recommendation model for each candidate feature data pair corresponding to different preset targets respectively; further comprising a training unit configured to train the recommendation model in the following manner: acquire training data, the training data containing a plurality of feature data pairs and label results of the plurality of feature data pairs corresponding to different preset targets respectively, each feature data pair containing first feature data and second feature data, the feature data containing input end feature data and promotion information feature data, the input end feature data being a model of an input end, a name, a gender, an age, and a profession of a user corresponding to the input end, the promotion information feature data being a title and an industry of promotion information, the preset target containing at least two of a click rate target, a conversion rate target, a complete play rate target, a deep conversion rate target, and a delay estimation target. The neural network model comprises an embedding layer, two sets of deep neural network layers and an output layer, the embedding layer is configured to output embedding vectors corresponding to different preset targets for a first feature data pair and a second feature data pair according to an input feature data pair, the two sets of deep neural network layers correspond to the first feature data and the second feature data respectively, each set of deep neural network layers is configured to output feature vectors corresponding to different preset targets for the feature data pair according to embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, and the output layer is configured to output prediction results corresponding to different preset targets for the feature data pair according to feature vectors corresponding to different preset targets output by the two sets of deep neural network layers. The neural network model is trained by using the plurality of feature data pairs and the plurality of labeled results corresponding to different preset targets for the plurality of feature data pairs, and a recommendation model is obtained.

9. The apparatus of claim 8, wherein, Each set of deep neural network layers comprises a plurality of deep neural network structures, and the number of the deep neural network structures is the same as the number of the preset targets.

10. The apparatus of claim 8, wherein, When each set of deep neural network layers outputs feature vectors corresponding to different preset targets for the feature data pair according to embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, the following steps are performed: For each set of deep neural network layers, the preset target corresponding to each deep neural network structure in the set of deep neural network layers is determined. The embedding vectors corresponding to different preset targets are input into the corresponding deep neural network structures respectively, and feature vectors corresponding to different preset targets output by each deep neural network structure are obtained.

11. The apparatus of claim 8, wherein, When each set of deep neural network layers outputs feature vectors corresponding to different preset targets for the feature data pair according to embedding vectors corresponding to different preset targets output by the embedding layer for the same feature data, the following steps are performed: For each set of deep neural network layers, the embedding vectors corresponding to different preset targets are spliced. The spliced results are input into each deep neural network structure in the set of deep neural network layers respectively, and feature vectors corresponding to different preset targets output by each deep neural network structure are obtained.

12. The apparatus of claim 8, wherein, When the output layer outputs prediction results corresponding to different preset targets for the feature data pair according to feature vectors corresponding to different preset targets output by the two sets of deep neural network layers, the following steps are performed: For each preset target, the feature vector corresponding to the preset target is obtained. According to the feature vector, the prediction result corresponding to the preset target for the feature data pair is obtained.

13. The apparatus of claim 8, wherein, When the neural network model is trained by using the plurality of feature data pairs and the plurality of labeled results corresponding to different preset targets for the plurality of feature data pairs, and the recommendation model is obtained, the following steps are performed: The plurality of feature data pairs are respectively input into the neural network model, and prediction results of the neural network model for each feature data pair are obtained, the prediction results corresponding to different preset targets respectively; According to the prediction results and the labeled results of each feature data corresponding to different preset targets, a loss function value is calculated; According to the calculated loss function value, parameters in the neural network model are adjusted until the neural network model converges, and the recommendation model is obtained.

14. The apparatus of claim 8, wherein, When the training unit trains the neural network model using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain the recommendation model, the following is specifically performed: The neural network model is trained using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain a first neural network model, and each deep neural network structure of each group of deep neural network layers in the neural network model is used to output a feature vector corresponding to different preset targets according to an embedding vector corresponding to different preset targets; The first neural network model is trained using the plurality of feature data pairs and the labeled results of the plurality of feature data pairs corresponding to different preset targets to obtain the recommendation model, and each deep neural network structure of each group of deep neural network layers in the first neural network model is used to output a feature vector corresponding to different preset targets according to an embedding vector corresponding to all preset targets.

15. An electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

16. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1-7.

17. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-7.

Images