Recommended method, device, electronic equipment and storage medium

Abstract

The present disclosure provides a recommendation method and device, electronic equipment and storage medium, wherein the method comprises: determining input feature information according to feature information of a candidate recommendation object and user feature information; encoding the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoding feature information; determining first time recommendation feature information according to the first time encoding feature information; determining second time recommendation feature information according to the input feature information and the first time encoding feature information; determining a recommendation probability of the candidate recommendation object according to the first time recommendation feature information and the second time recommendation feature information; and determining a target recommendation object from the candidate objects and performing recommendation according to the recommendation probability. Thus, the e-commerce can be accurately recommended according to the first time recommendation feature information and the second time recommendation feature information.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and in particular, to a recommendation method and device, an electronic device, and a storage medium. BACKGROUND

[0002] With the development of Internet technology, e-commerce is everywhere in our daily life and work. E-commerce, commonly known as e-commerce, usually refers to a new business operation mode in which buyers and sellers conduct various commercial activities without meeting each other in the global business trade activities in the open network environment of the Internet, based on the browser / server application mode, to realize online shopping of consumers, online transactions and online electronic payment between merchants, and various business activities, transaction activities, financial activities and related comprehensive service activities. Due to the explosive growth of available information on the network, personalized recommendation is a basic strategy to give consumers a better user experience.

[0003] At present, in the e-commerce recommendation scene, due to the complexity of the e-commerce business itself, and as the business development time elapses, it is difficult to capture long-term user historical behavior information for recommendation. SUMMARY

[0004] The present disclosure aims to at least solve one of the technical problems in the related art to some extent.

[0005] To this end, the present disclosure proposes a recommendation method, device, electronic device and storage medium to determine first time recommendation feature information according to first time encoding features, and determine second time recommendation feature information according to the first encoding features and input information, and then determine the recommendation probability of the candidate recommendation object according to the first time recommendation feature information and the second time recommendation feature information, and determine the target recommendation object from the candidate object according to the recommendation probability to perform recommendation. Therefore, according to the loading of the short-term behavior information of the user, the long-term user historical behavior information can be effectively determined, and then according to the long-term historical behavior information of the user and the short-term behavior information of the user, the e-commerce can be accurately recommended.

[0006] In a first aspect, an embodiment of the present disclosure provides a recommendation method, including: obtaining candidate recommendation objects and user feature information; determining input feature information according to feature information of the candidate recommendation objects and the user feature information; encoding the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information; determining first time recommendation feature information according to the first time encoded feature information; determining second time recommendation feature information according to the input feature information and the first time encoded feature information; determining a recommendation probability of the candidate recommendation objects according to the first time recommendation feature information and the second time recommendation feature information; and determining a target recommendation object from the candidate objects and performing recommendation according to the recommendation probability.

[0007] In a second aspect, an embodiment of the present disclosure provides a recommendation device, including: an obtaining module, configured to obtain candidate recommendation objects and user feature information; a first determining module, configured to determine input feature information according to feature information of the candidate recommendation objects and the user feature information; an encoding module, configured to encode the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information; a second determining module, configured to determine first time recommendation feature information according to the first time encoded feature information, and determine second time recommendation feature information according to the input feature information and the first time encoded feature information; a third determining module, configured to determine a recommendation probability of the candidate recommendation objects according to the first time recommendation feature information and the second time recommendation feature information; and a recommendation module, configured to determine a target recommendation object from the candidate objects and perform recommendation according to the recommendation probability.

[0008] In a third aspect, an embodiment of the present disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the method in the first aspect of the present disclosure is implemented.

[0009] In a fourth aspect, an embodiment of the present disclosure provides a non-transitory computer-readable storage medium, having a computer program stored thereon, and when the computer program is executed by a processor, the method in the first aspect of the present disclosure is implemented.

[0010] In a fifth aspect, an embodiment of the present disclosure provides a computer program product, including a computer program, and when the computer program is executed by a processor, the method in the first aspect of the present disclosure is implemented.

[0011] The technical solution provided by the embodiments of the present disclosure has the following beneficial effects:

[0012] The technical solution of the present disclosure determines the first-time recommendation feature information according to the first-time coding feature, determines the second-time recommendation feature information according to the first coding feature and the input information, further determines the recommendation probability of the candidate recommendation object according to the first-time recommendation feature information and the second-time recommendation feature information, and determines the target recommendation object from the candidate object according to the recommendation probability to perform recommendation. Therefore, by loading and saving the first-time coding feature information and converting the first-time coding feature information, the second-time recommendation feature information can be effectively determined, so that the user short-term behavior information can be loaded and saved and converted into user long-term behavior information, and further, according to the user short-term behavior information and the long-term behavior information, the e-commerce can be accurately recommended.

[0013] The additional aspects and advantages of the present disclosure will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0014] The above and / or additional aspects and advantages of the present disclosure will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0015] Figure 1 A flowchart of a recommendation method provided by an embodiment of the present disclosure;

[0016] Figure 2 A flowchart of another recommendation method provided by an embodiment of the present disclosure;

[0017] Figure 3 A flowchart of another recommendation method provided by an embodiment of the present disclosure;

[0018] Figure 4 A flowchart of another recommendation method provided by an embodiment of the present disclosure;

[0019] Figure 5 A structural diagram of a multi-layer deep learning network model provided by an embodiment of the present disclosure;

[0020] Figure 6 A structural diagram of a recommendation device provided by an embodiment of the present disclosure;

[0021] Figure 7 A block diagram of an exemplary electronic device suitable for implementing an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0022] Embodiments of the present disclosure are described below in detail with reference to the accompanying drawings, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present disclosure, and cannot be understood as a limitation of the present disclosure.

[0023] Due to the complexity of the business of e-commerce itself, there are great technical challenges for e-commerce recommendation. The main technical challenges are: 1) huge amount of data, as the business development time goes by, the amount of data for many years can reach a very large number, which cannot be solved by increasing computing resources; 2) due to the time being relatively long, it is difficult to effectively learn the long-term user behavior information a long time ago; 3) even if the long-period user behavior rule is learned, it is easy to be disturbed by short-term noise; 4) short-term memory information is easy to change, and there is much noise, and there is no effective information to judge whether the information is noise; 5) it is difficult to identify and store long-short period information.

[0024] In view of at least one of the above problems, the present disclosure provides a recommendation method, device, electronic equipment and storage medium.

[0025] The recommendation method, device, electronic equipment and storage medium of the embodiments of the present disclosure are described below with reference to the accompanying drawings.

[0026] Figure 1 A flowchart of a recommendation method provided by the embodiments of the present disclosure is shown. It should be noted that the recommendation method of the embodiments of the present disclosure can be applied to the recommendation device of the embodiments of the present disclosure, and the device can be configured in an electronic equipment. The electronic equipment can be a mobile terminal, such as a mobile phone, a tablet computer, a personal digital assistant, and other hardware devices with various operating systems.

[0027] As shown in Figure 1 The recommendation method can include the following steps:

[0028] Step 101, obtaining candidate recommendation objects and user feature information.

[0029] In the embodiments of the present disclosure, the candidate recommendation objects can be goods, e-commerce and other recommendable objects. The user feature information can include user attribute feature information and user historical behavior feature information; for example, the user attribute feature information is user name, age, gender, hobby, etc. The user historical behavior feature information can include the historical behavior of the user to the candidate recommendation image within a set time, for example, the candidate recommendation object is e-commerce A, and the user historical behavior can be the browsing rate and click rate of the user to e-commerce A in the past week; for example, the candidate recommendation object is goods A, and the user historical behavior can be the click rate, browsing rate or purchase rate of the user to goods A in the past week.

[0030] In the embodiments of the present disclosure, the user attribute feature information can be determined by the user registration on the relevant page, and the user historical behavior feature information can be determined by the statistics of the user behavior within a set time.

[0031] In step 102, the input feature information is determined according to the feature information of the candidate recommended object and the user feature information.

[0032] In the embodiments of the present disclosure, the feature extraction can be performed on the candidate recommended object to determine the feature information of the candidate object. Furthermore, in order to better associate the candidate recommended object and the user feature information, the corresponding high-order feature can be determined according to the candidate recommended object and the user feature information, and the high-order feature is taken as the input feature.

[0033] In step 103, the input information is encoded by using the encoding layer of the first time recurrent neural network to obtain at least one first time encoding feature information.

[0034] Furthermore, the input information can be encoded by using the encoding layer of the first time recurrent neural network to obtain at least one first time encoding feature, wherein the first time recurrent neural network can memorize and process the time feature vector sequence within the first time. The encoding layer of the first time recurrent neural network can include a first time positive correlation sub-encoder, a first time negative correlation sub-encoder and a first time all sub-encoder, and correspondingly, the first time encoding feature can include a first time positive correlation sub-encoding feature, a first time negative correlation sub-encoding feature and a first time all sub-encoding feature.

[0035] For example, the first time is one week, the first time recurrent neural network can memorize and process the time feature vector sequence within the time, the first time recurrent neural network can be called a short-term information memory network, and the memory encoder (encoder) of the short-term information memory network is used to encode the input feature to obtain the first time encoding feature. In addition, the first time recurrent neural network can also perform classification prediction according to the input feature, for example, the classification can include the purchase preference of the user to the candidate object, such as like or dislike.

[0036] In step 104, the first time recommendation feature information is determined according to the first time encoding feature information.

[0037] In the embodiments of the present disclosure, the first-time coding features can include: first-time positive correlation sub-coding features and first-time negative correlation sub-coding features. The first-time positive correlation sub-coding features can be coding features with high correlation between the user and the candidate object, and the first-time negative correlation sub-coding features can be coding features with low or zero correlation between the user and the candidate object. For example, the user is a young woman, the candidate object is a cosmetic product, and the correlation between the user and the candidate object is high (positive correlation). The user and the corresponding candidate object can correspond to the first-time positive correlation sub-coding features. For another example, the user is an old man, the candidate object is a cosmetic product, and the correlation between the user and the candidate object is low or zero (negative correlation). The user and the corresponding candidate object can correspond to the second-time negative correlation sub-coding features.

[0038] In order to improve the accuracy of the first-time positive correlation sub-coding features and the first-time negative correlation sub-coding features, the first-time coding features can further include first-time all sub-coding features. The first-time all sub-coding features of the user and the corresponding candidate object can be determined according to the correlation (positive correlation or negative correlation) between the user and the candidate object.

[0039] Further, the first-time positive correlation sub-coding features, the first-time negative correlation sub-coding features and the first-time all sub-coding features are weighted and summed by using a multi-head attention network mechanism, and the weighted sum result is taken as the first-time recommendation feature information.

[0040] In step 105, the second-time recommendation feature information is determined according to the input feature information and the first-time coding feature information.

[0041] Further, in order to improve the accuracy of the recommendation, after the first-time recommendation feature information is determined, the first-time coding feature information corresponding to the short-term behavior of the user can be converted into the second-time recommendation feature information (multi-dimensional long-term behavior of the user).

[0042] Optionally, the input features are encoded and converted by using a multi-head attention network to obtain the encoded and converted feature information corresponding to the input feature information. The intermediate coding feature information corresponding to the first-time coding feature information is generated according to the encoded and converted feature information and the first-time coding feature information. Further, the intermediate coding features are stored and coded by using the dimensions of the full connection layers in the second-time recurrent neural network to obtain the second-time recommendation feature information.

[0043] In step 106, the recommendation probability of the candidate object is determined according to the first-time recommendation feature information and the second-time recommendation feature information.

[0044] In the embodiments of the present disclosure, the first-time recommendation feature information and the second-time recommendation feature information are spliced to obtain a shared recommendation feature sequence, and then the recommendation probability of the candidate recommendation object is determined according to the first-time feature sequence corresponding to the first-time recommendation feature information, the second-time feature sequence corresponding to the second-time recommendation feature information, and the shared feature sequence corresponding to the shared recommendation feature sequence.

[0045] In step 107, the target recommendation object is determined from the candidate recommendation objects according to the recommendation probability, and the target recommendation object is recommended.

[0046] Further, when the number of candidate recommendation objects is one, if the recommendation probability of the candidate recommendation object is greater than a set probability threshold, the candidate recommendation object is recommended; when the number of candidate recommendation objects is at least two, the recommendation probabilities corresponding to the at least two candidate recommendation objects are compared to determine the target recommendation object from the candidate recommendation objects, and the target recommendation object is recommended. For example, the candidate recommendation object with the highest recommendation probability is taken as the target recommendation object and is recommended.

[0047] To sum up, the candidate recommendation objects and the user feature information are obtained, the input feature information is determined according to the feature information of the candidate recommendation objects and the user feature information, the input information is encoded by using the encoding layer of the first-time recurrent neural network to obtain at least one first-time encoding feature information, the first-time recommendation feature information is determined according to the first-time encoding feature information, the second-time recommendation feature information is determined according to the input feature information and the first-time encoding feature information, the recommendation probability of the candidate recommendation object is determined according to the first-time recommendation feature information and the second-time recommendation feature information, and the target recommendation object is determined from the candidate recommendation objects according to the recommendation probability and is recommended. Thus, by loading the saved first-time encoding feature information and converting the first-time encoding feature information, the second-time recommendation feature information can be effectively determined. In this way, by loading the saved user short-term behavior information and converting the user short-term behavior information into user long-term behavior information, the e-commerce can be accurately recommended according to the user short-term behavior information and the user long-term behavior information.

[0048] In order to clearly illustrate how the second-time recommendation feature information is obtained in the above-mentioned embodiments of the present disclosure, the present disclosure further provides a recommendation method.

[0049] Figure 2 A flowchart of another recommendation method provided by the embodiments of the present disclosure.

[0050] As shown in Figure 2 , the recommendation method can include the following steps:

[0051] In step 201, the candidate recommendation objects and the user feature information are obtained.

[0052] Step 202, determining the input feature information according to the characteristic information of the candidate recommended object and the user characteristic information.

[0053] Step 203, encoding the input feature information by using an encoding layer of the first time recurrent neural network to obtain at least one first time encoded feature information.

[0054] In the embodiments of the present disclosure, it can be expressed as the following formula:

[0055] Wherein, W i is the matrix corresponding to the encoding layer of the first time recurrent neural network, X is the input feature information, d is the normalization factor, (W i X) T is the transpose matrix corresponding to the dot product of the matrix W i and the input feature information, encoder i is the first time encoded feature information. The encoding layer of the first time recurrent neural network can include a first time positive correlation sub-encoder, a first time negative correlation sub-encoder and a first time all sub-encoder, and accordingly, encoder i may be a first time positive correlation sub-encoding feature encoder all , a first time negative correlation sub-encoding feature encoder neg or a first time all sub-encoding feature encoder pos .

[0056] In order to obtain more accurate encoded feature information, in the embodiments of the present disclosure, after obtaining at least one first time encoded feature information, the category probability value corresponding to the first time encoded feature information can be determined according to the first time encoded feature information, and the first time recurrent neural network is modified according to the category probability value and the corresponding set auxiliary loss function. It can be expressed as the following formula:

[0057] pref_p i = sigmoid (encoder i *V), wherein the first time encoded feature information encoder i may be encoder all , encoder neg or encoder pos , and accordingly, the category probability value pref_p i corresponding to the first time encoded feature information may be pref_p all , pref_p neg or pref_p pos .

[0058] Further, the loss all = -(y log(pref_p all + (1 - y) log(1 - pref_p all )) ;

[0059] the loss neg = (y - 1) log(1 - pref_p neg ) ;

[0060] the loss pos = -y log(pref_p pos ) ;

[0061] wherein y is a coefficient of each set auxiliary loss function, and the loss all is less than or equal to a first set threshold, such as the loss all = 0, it can be determined that the first-time encoding feature information encoder all has an accuracy higher than a first accuracy threshold; when the loss all is greater than the first set threshold, the first-time positive correlation sub-encoder of the encoding layer of the first neural network can be modified; when the loss neg is less than or equal to a second set threshold, such as the loss neg = 0, it can be determined that the first-time encoding feature information encoder neg has an accuracy higher than a second accuracy threshold, and the loss neg is greater than the set threshold; when the loss neg is greater than the second set threshold, the first-time negative correlation sub-encoder of the encoding layer of the first neural network can be modified; when the loss pos is less than or equal to a third set threshold, such as the loss pos = 0, it can be determined that the first-time encoding feature information loss pos has an accuracy higher than a third accuracy threshold; and when the loss pos is greater than the third set threshold, the first-time all sub-encoders of the encoding layer of the first neural network can be modified.

[0062] Step 204, determining first-time recommended feature information according to the first-time encoding feature information.

[0063] Step 205, using a conversion network to encode and convert the input feature information to obtain the encoding conversion feature information corresponding to the input feature information.

[0064] As an example, a multi-head attention network mechanism can be used to encode and convert the input feature information to obtain the encoding conversion feature information corresponding to the input feature information.

[0065] In step 206, the intermediate coding feature information corresponding to the first-time coding feature information is determined according to the coding conversion feature information and the first-time coding feature information.

[0066] Optionally, the weight corresponding to the first-time coding feature information is determined according to the coding conversion feature information and the first-time coding feature information; and the intermediate coding feature information corresponding to the first-time coding feature information is determined according to the first-time coding feature information and the weight corresponding to the first-time coding feature information.

[0067] That is, the dot product can be performed according to the transpose matrix corresponding to the coding conversion feature information and the first-time coding feature information, the dot product result is normalized, the normalized result is taken as the weight corresponding to the first-time coding feature information, the first-time coding feature information is multiplied by the weight corresponding to the first-time coding feature, and the result is taken as the intermediate coding feature information corresponding to the first-time coding feature information. This can be expressed as the following formula:

[0068] input_transfor=MultiHead(high_feat);

[0069] discriminator_pi=sigmoid(input_transfor*encoder i T );

[0070] trans_encoder i =discriminator_p i ·encoder i ;

[0071] Wherein, high_feat can represent the input feature, input_transfor can represent the coding conversion feature information, encoder i may represent each first-time coding feature information, encoder i T may represent the transpose matrix corresponding to each first-time coding feature information, discriminator_p i may represent the weight corresponding to the first-time coding feature information, trans_encoder i is the intermediate coding feature information corresponding to each first-time coding feature information. In the embodiment of the present disclosure, the first-time coding feature information can include the first-time positive correlation sub-coding feature, the first-time negative correlation sub-coding feature, and the first-time all sub-coding feature.

[0072] In step 207, the storage encoding feature information corresponding to the dimension information of each full connection layer in the second time recurrent neural network is determined according to the dimension information of each full connection layer in the second time recurrent neural network and the intermediate encoding feature information.

[0073] Further, in order to obtain the data information of different dimensions in the first time encoding feature information, in the embodiment of the present disclosure, the intermediate encoding feature is stored and memorized by using the second time recurrent neural network, and as an example, the intermediate encoding feature information can be memorized and stored according to the dimension information of each full connection layer in the second time recurrent neural network, so that the storage encoding feature information corresponding to the dimension information of each full connection layer can be determined.

[0074] In step 208, the second time recommendation feature information is generated according to the storage encoding feature information.

[0075] Further, the storage encoding feature information corresponding to the dimension information of each layer full connection layer can be spliced, and the spliced feature information can be used as the second time recommendation feature information. This can be expressed as the following formula:

[0076] long_term_store j =(trans_encoder all +trans_encoder neg +trans_encoder pos )*W j ;

[0077] long_term_info=Concat(long_term_store1,...,long_term_store n );

[0078] Wherein, long_term_store j represents the storage encoding feature information corresponding to the jth layer full connection layer, and long_term_info is the second time recommendation feature information.

[0079] In step 209, the recommendation probability of the candidate recommendation object is determined according to the first time recommendation feature information and the second time recommendation feature information.

[0080] In step 210, the target recommendation object is determined from the candidate recommendation object according to the recommendation probability and is recommended.

[0081] The execution process of steps 201 to 202, steps 209 to 210 can refer to the execution process of any embodiment of the present disclosure, which will not be repeated here.

[0082] In summary, the conversion network is used to encode and convert the input feature information to obtain the encoding conversion feature information corresponding to the input feature information; the intermediate encoding feature information corresponding to the first time encoding feature information is determined according to the encoding conversion feature information and the first time encoding feature information; the storage encoding feature information corresponding to the dimension information of each full connection layer in the second time recurrent neural network is determined according to the dimension information of each full connection layer in the second time recurrent neural network and the intermediate encoding feature information; and the second time recommendation feature information is generated according to the storage encoding feature information. In this way, by loading and saving the first time encoding feature information and converting the first time encoding feature information, the second time recommendation feature information can be effectively determined, so that the user short-term behavior information can be loaded and saved, and the user long-term behavior information can be obtained by converting the user short-term behavior information.

[0083] To clearly illustrate how the recommendation probability of the candidate recommendation object is determined according to the first time recommendation feature information and the second time recommendation feature information in the above-mentioned embodiments of the present disclosure, the present disclosure further provides a recommendation method.

[0084] Figure 3 A flowchart of another recommendation method provided by the embodiments of the present disclosure.

[0085] As shown in Figure 3 , the recommendation method can include the following steps:

[0086] Step 301, obtaining candidate recommendation objects and user feature information.

[0087] Step 302, determining input feature information according to the feature information of the candidate recommendation objects and the user feature information.

[0088] Step 303, encoding the input information by using the encoding layer of the first time recurrent neural network to obtain at least one first time encoding feature information.

[0089] Step 304, determining the first time recommendation feature information according to the first time encoding feature information.

[0090] Step 305, determining the second time recommendation feature information according to the input feature information and the first time encoding feature information.

[0091] Step 306, determining the shared recommendation feature information according to the first time recommendation feature information and the second time recommendation feature information.

[0092] As an example, the first time recommendation feature information and the second time recommendation feature information can be spliced, and the splicing result can be used as the shared recommendation feature information.

[0093] In step 307, the first neural network is used to perform feature extraction on the first-time recommendation feature information to obtain a first-time feature sequence.

[0094] To reduce the interference of noise and improve the accuracy of the recommendation result, in the embodiments of the present disclosure, the first neural network can be used to perform feature extraction on the first-time recommendation feature information to obtain a first-time feature sequence corresponding to the first-time recommendation feature.

[0095] In step 308, the second neural network is used to perform feature extraction on the second-time recommendation feature information to obtain a second-time feature sequence.

[0096] Similarly, the second neural network can be used to perform feature extraction on the second-time recommendation feature information to obtain a second-time feature sequence corresponding to the second-time recommendation feature information.

[0097] In step 309, the third neural network is used to perform feature extraction on the shared recommendation feature information to obtain a shared feature sequence.

[0098] In the embodiments of the present disclosure, the third neural network is used to perform feature extraction on the shared recommendation feature information to obtain a shared feature sequence corresponding to the shared recommendation feature information.

[0099] In step 310, the recommendation probability of the candidate object matched with each historical behavior in the user historical behavior feature information is determined according to the first-time feature sequence, the second-time feature sequence and the shared feature sequence.

[0100] Optionally, the weight corresponding to each historical behavior in the user historical behavior feature information is determined; the first feature sequence matched with each historical behavior is determined according to the first-time feature sequence, the second-time feature sequence and / or the shared feature sequence; the second feature sequence matched with each historical behavior is determined according to the feature sequence matched with each historical behavior and the weight of the corresponding historical behavior; the second feature sequence is input into the multi-layer neural network to obtain the recommendation probability value corresponding to each historical behavior; and the recommendation probability values corresponding to each historical behavior are weighted and summed according to the set weight to determine the recommendation probability of the candidate object matched with each historical behavior.

[0101] That is, weights corresponding to each historical behavior in the user historical behavior feature information can be determined according to the gating network, and a first feature sequence matching each historical behavior can be determined according to the first time feature sequence, the second time feature sequence and / or the shared feature sequence. Then, the first feature sequence matching each historical behavior is multiplied by the weight of the corresponding historical behavior, and the multiplication result is taken as a second feature sequence matching each historical behavior. The second feature sequence is input into the multi-layer neural network, and the multi-layer neural network can output a recommendation probability value (confidence) corresponding to each historical behavior. Then, according to a set weight, the probability values corresponding to each historical behavior are weighted and summed, and the weighted sum result is taken as a recommendation probability of the candidate object matching each historical behavior.

[0102] As an example, it can be expressed as the following formula:

[0103] gate k (x)=softmax(W k x);

[0104]

[0105] cgc_output i (x)=gate k (x)S k (x);

[0106] task i (x)=mlp i (cgc_output i (x));

[0107]

[0108] wherein k is a historical behavior (such as a click), W k is a gating matrix corresponding to the historical behavior, x is a first time feature sequence, a second time feature sequence or a shared feature sequence, is a jthneural network matrix corresponding to a feature sequence matching the historical behavior, is a high-order feature obtained by feature extraction of x using different neural networks, S k (x) is a first feature sequence matching the historical behavior (for example, the historical behavior is a click, and the first feature sequence matching the click can be a feature sequence obtained by combining a high-order feature obtained by neural network extraction of the first time feature sequence and a high-order feature obtained by neural network extraction of the second time feature sequence), cgc_output i (x) is a second feature sequence matching each historical behavior, and task i(x) is a recommended probability value corresponding to each historical behavior, and the mlp_output is a recommended probability of the candidate object matched with each historical behavior.

[0109] Step 311, determining a target recommended object from the candidate recommended objects according to the recommended probability and performing recommendation.

[0110] Steps 301 to 305 and step 311 can refer to the execution process of any embodiment of the present disclosure, which will not be repeated here.

[0111] In summary, by determining the shared recommendation feature information according to the first-time recommendation feature information and the second-time recommendation feature information, performing feature extraction on the first-time recommendation feature information by using the first neural network to obtain the first-time feature sequence, performing feature extraction on the second-time recommendation feature information by using the second neural network to obtain the second-time feature sequence, performing feature extraction on the shared recommendation feature information by using the third neural network to obtain the shared feature sequence, and determining the recommended probability of the candidate object matched with each historical behavior in the user historical behavior feature information according to the first-time feature sequence, the second-time feature sequence and the shared feature sequence, the recommended probability of the candidate recommended object can be determined according to the first-time recommendation feature information and the second-time recommendation feature information, and then the e-commerce can be accurately recommended according to the recommended probability of the candidate recommended object.

[0112] In order to clearly illustrate how the embodiments of the present disclosure determine the input feature information, the present disclosure further provides a recommendation method.

[0113] Figure 4 A flowchart of another recommendation method provided by the embodiments of the present disclosure.

[0114] As shown in the figure, the recommendation method can include the following steps: Figure 4

[0115] Step 401, obtaining a candidate recommended object and user feature information, wherein the user feature information includes user attribute feature information and user historical behavior feature information.

[0116] Step 402, sorting the user historical behavior feature information according to time to obtain user serialized feature information.

[0117] In the embodiments of the present disclosure, the user historical behavior feature information can be sorted according to time, and the sorting result is used as the user serialized feature information.

[0118] Step 403, performing feature statistics on the user serialized feature information to obtain user statistical feature information corresponding to the user serialized feature information.

[0119] ​Further, the user serialized feature information is subjected to feature statistics, for example, each feature in the user serialized feature information can be subjected to statistics, and the user statistical feature information can be determined according to the statistics result.

[0120] At step 404, the user serialized feature information is subjected to vectorization to obtain the vectorized user serialized feature information.

[0121] Further, the user serialized feature information is subjected to vectorization to obtain the vectorized user serialized feature information, for example, which can be expressed as the following formula:

[0122] embed_input s =embedding(inputs s );

[0123] Wherein, inputs s is the user serialized feature information, and embed_input s is the vectorized user serialized feature information.

[0124] At step 405, the multi-head attention network mechanism is adopted to extract features from the vectorized user serialized feature information to obtain the first feature information.

[0125] In order to associate the features between the user serialized feature information, the multi-head attention network mechanism can be adopted to extract features from the user serialized feature information to obtain the first feature information. For example, which can be expressed as the following formula: multi_feat=MultiHead(embed_input s ), wherein, embed_input s is the vectorized user serialized feature information, and multi_feat is the first feature information.

[0126] At step 406, the user attribute feature information, the feature information of the candidate recommended object and the user statistical feature information are spliced to obtain the first spliced feature information.

[0127] In the embodiments of the present disclosure, it can be expressed as the following formula: first spliced feature information=input u +input p +input a ;

[0128] Wherein, input u is the user attribute feature information, input p is the feature information of the candidate recommended object, and input a is the user statistical feature information.

[0129] In step 407, the first spliced feature information is subjected to feature extraction by using a multi-layer neural network to obtain second feature information.

[0130] In the embodiments of the present disclosure, it can be expressed as the following formula: conti_feat = Mlp(input u +input p +input a ), wherein input u is the user attribute feature information, input p is the feature information of the candidate recommended object, and input a is the user statistical feature information.

[0131] In step 408, the first spliced feature information and the vectorized user serialized feature information are spliced to obtain second spliced feature information.

[0132] In the embodiments of the present disclosure, it can be expressed as the following formula: input all = input u +input p +input a +embed_input s , wherein input u +input p +input a is the first spliced feature information, input all is the second spliced feature information, and embed_input s is the vectorized user serialized feature information.

[0133] In step 409, input features are generated according to the first feature information, the second feature information and the second spliced feature information.

[0134] In the embodiments of the present disclosure, it can be expressed as the following formula: high_feat = LayerNorm(multi_feat + conti_feat + input all ), wherein multi_feat is the first feature information, conti_feat is the second feature information, and input all is the second spliced feature information.

[0135] In step 410, the input information is encoded by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information.

[0136] In step 411, the first time recommended feature information is determined according to the first time encoded feature information.

[0137] Step 412: Determine the second-time recommended feature information based on the input feature information and the first-time encoded feature information.

[0138] Step 413: Determine the recommendation probability of the candidate recommendation object based on the first-time recommendation feature information and the second-time recommendation feature information.

[0139] Step 414: Based on the recommendation probability, determine the target recommendation object from the candidate recommendation objects and make a recommendation.

[0140] In summary, by acquiring candidate recommendation objects and user feature information, including user attribute features and user historical behavior features; sorting the user historical behavior features by time to obtain user serialization feature information; performing feature statistics on the user serialization feature information to obtain corresponding user statistical feature information; vectorizing the user serialization feature information to obtain vectorized user serialization feature information; employing a multi-head attention network mechanism to extract features from the vectorized user serialization feature information to obtain first feature information; concatenating the user attribute features, candidate recommendation object features, and user statistical feature information to obtain first concatenated feature information; concatenating the first concatenated feature information and the vectorized user serialization feature information to obtain second concatenated feature information; and generating input features based on the first feature information, second feature information, and second concatenated feature information. Therefore, based on the candidate recommendation object features and user feature information, the input feature information can be accurately determined.

[0141] To illustrate the embodiments of this disclosure more clearly, examples are given below.

[0142] For example, such as Figure 5 As shown, the following can be used Figure 5 The multi-layer deep learning network model shown implements the recommendation method of this disclosure embodiment. The multi-layer deep learning network model may include: 1. The bottom layer is a high-order information extraction layer, which is responsible for extracting high-order feature information from the input; 2. The middle layer is a memory network layer, including a first recurrent neural network (short-term memory network), a transformation network (memory transformation network), and a second recurrent neural network (long-term memory network); 3. The top two neural networks (expert networks) may include: a first neural network (short-term memory expert network), a second neural network (long-term memory expert network), and a third neural network (shared expert network); 4. The top layer is a task-specific tower, which is responsible for learning task-specific feature information and generating recommendation probability values ​​for each task based on the task-specific feature information.

[0143] The specific process can be as follows: User feature information includes user historical behavior feature information and user attribute feature information. Sorting the user historical behavior feature information by time yields user serialized feature information. This user serialized feature information is then vectorized (e.g., ...). Figure 5 (embedding), and employs a multi-head attention mechanism (e.g., Figure 5 Multi-Head Attention (in this context) extracts features from the vectorized user sequence information to obtain the first feature information. Then, a multi-layer neural network (e.g., ...) is used. Figure 5 The Multilayer Perceptron (MLP) extracts features from the concatenated features of user attribute features, candidate recommendation object features, and user statistical features to obtain second feature information. Then, the first concatenated feature information and the vectorized user serialization feature information are concatenated to obtain the second concatenated feature information. The first feature information, the second feature information, and the second concatenated feature information are normalized to obtain the input features.

[0144] Furthermore, a first-time recurrent neural network (e.g., Figure 5The encoding layer of the first time recurrent neural network encodes the input feature information to obtain at least one first time encoded feature information. The encoding layer of the first time recurrent neural network can include a first time positive information sub-encoder, a first time negative information sub-encoder, and a first time all information sub-encoder. Correspondingly, the first time encoded feature can include a first time positive information sub-encoding feature, a first time negative information sub-encoding feature, and a first time all information sub-encoding feature. Then, the first time positive information sub-encoding feature, the first time negative information sub-encoding feature, and the first time all information sub-encoding feature are weighted and summed by using a multi-head attention network mechanism, and the weighted and summed result is taken as first time recommendation feature information (short term higher order features). Further, the input feature is encoded and converted by using a conversion network (hippocampus) to obtain encoding and conversion feature information corresponding to the input feature information, and according to the encoding and conversion feature information and the first time encoded feature information, intermediate encoding feature information corresponding to the first time encoded feature information is generated. Then, the intermediate encoding feature is stored and encoded by using the dimensions of each fully connected layer in the second time recurrent neural network (long term memory network) to obtain second time recommendation feature information (long term higher order features). The first time recommendation feature information and the second time recommendation feature information are spliced to obtain a shared recommendation feature sequence. Then, according to the first time feature sequence corresponding to the first time recommendation feature information, the second time feature sequence corresponding to the second time recommendation feature information, and the shared feature sequence corresponding to the shared recommendation feature sequence, the recommendation probability of the candidate recommended object is determined.

[0145] In Figure 5 , the first time recommendation feature information (short_term_info) is feature extracted by using a first neural network (short term memory expert network, such as Figure 5 In Figure 5The first neural network, the second neural network and the third neural network output are controlled by a gating network, and the weight control result is input into a multi-layer neural network. The multi-layer neural network can output a recommendation probability value corresponding to each historical behavior. Then, according to a set weight, the probability values corresponding to each historical behavior are weighted and summed, and the weighted sum result is taken as a recommendation probability of a candidate object matched with each historical behavior.

[0146] The recommendation method provided by the embodiments of the present disclosure comprises the following steps: obtaining candidate recommendation objects and user feature information; determining input feature information according to the feature information of the candidate recommendation objects and the user feature information; encoding the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoding feature information; determining first time recommendation feature information according to the first time encoding feature information; determining second time recommendation feature information according to the input feature information and the first time encoding feature information; determining a recommendation probability of the candidate recommendation objects according to the first time recommendation feature information and the second time recommendation feature information; and determining a target recommendation object from the candidate objects and performing recommendation according to the recommendation probability. The method determines the first time recommendation feature information according to the first time encoding feature, determines the second time recommendation feature information according to the first encoding feature and the input information, then determines the recommendation probability of the candidate recommendation objects according to the first time recommendation feature information and the second time recommendation feature information, and determines the target recommendation object from the candidate objects according to the recommendation probability to perform recommendation. Therefore, according to the short-term behavior information of the user, the long-term historical behavior information of the user can be effectively determined, and then according to the long-term historical behavior information of the user and the short-term behavior information of the user, the e-commerce can be accurately recommended.

[0147] The recommendation method provided by the embodiments of the present disclosure corresponds to the recommendation method provided by the above Figures 1 to 5 The recommendation method provided by the embodiments of the present disclosure corresponds to the recommendation method provided by the above Figures 1 to 5 The recommendation method provided by the embodiments of the present disclosure corresponds to the recommendation method provided by the above

[0148] Figure 6 A structural schematic diagram of a recommendation device provided by the embodiments of the present disclosure.

[0149] As Figure 6As shown, the recommendation device 600 can include an acquisition module 610, a first determination module 620, an encoding module 630, a second determination module 640, a third determination module 650, a fourth determination module 660, and a recommendation module 670.

[0150] The acquisition module 610 is configured to acquire candidate recommendation objects and user feature information. The first determination module 620 is configured to determine input feature information according to feature information of the candidate recommendation objects and the user feature information. The encoding module 630 is configured to encode the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information. The second determination module 640 is configured to determine first time recommendation feature information according to the first time encoded feature information. The third determination module 650 is configured to determine second time recommendation feature information according to the input feature information and the first time encoded feature information. The fourth determination module 660 is configured to determine a recommendation probability of the candidate recommendation objects according to the first time recommendation feature information and the second time recommendation feature information. The recommendation module 670 is configured to determine a target recommendation object from the candidate objects according to the recommendation probability and perform recommendation.

[0151] As a possible implementation manner of the embodiments of the present disclosure, the third determination module is configured to: encode and convert the input feature information by using a conversion network to obtain encoded and converted feature information corresponding to the input feature information; determine intermediate encoded feature information corresponding to the first time encoded feature information according to the encoded and converted feature information and the first time encoded feature information; determine stored encoded feature information corresponding to the dimension information of each full connection layer in the second time recurrent neural network according to the dimension information of each full connection layer in the second time recurrent neural network and the intermediate encoded feature information; and generate the second time recommendation feature information according to the stored encoded feature information.

[0152] As a possible implementation manner of the embodiments of the present disclosure, the third determination module is further configured to: determine a weight corresponding to the first time encoded feature information according to the encoded and converted feature information and the first time encoded feature information; and determine the intermediate encoded feature information corresponding to the first time encoded feature information according to the first time encoded feature information and the weight corresponding to the first time encoded feature information.

[0153] As a possible implementation manner of the embodiment of the present disclosure, the fourth determining module is configured to: determine shared recommendation feature information according to the first-time recommendation feature information and the second-time recommendation feature information; perform feature extraction on the first-time recommendation feature information by using a first neural network to obtain a first-time feature sequence; perform feature extraction on the second-time recommendation feature information by using a second neural network to obtain a second-time feature sequence; perform feature extraction on the shared recommendation feature information by using a third neural network to obtain a shared feature sequence; and determine a recommendation probability of a candidate object matched with each historical behavior in the user historical behavior feature information according to the first-time feature sequence, the second-time feature sequence, and the shared feature sequence.

[0154] As a possible implementation manner of the embodiment of the present disclosure, the fourth determining module is further configured to: determine a weight corresponding to each historical behavior in the user historical behavior feature information; determine a first feature sequence matched with each historical behavior from the first-time feature sequence, the second-time feature sequence, and the shared feature sequence; determine a second feature sequence matched with each historical behavior according to the first feature sequence matched with each historical behavior and the weight of the corresponding historical behavior; determine a probability value corresponding to each historical behavior according to the feature sequence corresponding to each historical behavior; and determine a recommendation probability of a candidate object matched with each historical behavior by performing weighted summation on the probability value corresponding to each historical behavior according to a set weight.

[0155] As a possible implementation manner of the embodiment of the present disclosure, the user feature information includes user historical behavior feature information; and the recommendation device further includes a sorting module and a statistical module.

[0156] The sorting module is configured to sort the user historical behavior feature information according to time to obtain user serialized feature information; and the statistical module is configured to perform feature statistics on the user serialized feature information to obtain user statistical feature information corresponding to the user serialized feature information.

[0157] As a possible implementation manner of the embodiment of the present disclosure, the user feature information comprises user attribute feature information, the first determining module is configured to: vectorize the user sequential feature information to obtain vectorized user sequential feature information; perform feature extraction on the vectorized user sequential feature information by using a multi-head attention network mechanism to obtain first feature information; splice the user attribute feature information, the feature information of the candidate recommended object and the user statistical feature information to obtain first spliced feature information; perform feature extraction on the first spliced feature information by using a multi-layer neural network to obtain second feature information; splice the first spliced feature information and the vectorized user sequential feature information to obtain second spliced feature information; and generate input feature information according to the first feature information, the second feature information and the second spliced feature information.

[0158] As a possible implementation manner of the embodiment of the present disclosure, the recommendation device further comprises a fifth determining module and a correction module.

[0159] The fifth determining module is configured to determine a category probability value corresponding to the first time encoded feature information according to the first time encoded feature information, and the correction module is configured to correct the first time recurrent neural network according to the category probability value and a corresponding set auxiliary loss function.

[0160] The recommendation device of the embodiment of the present disclosure acquires candidate recommended objects and user feature information, determines input feature information according to the feature information of the candidate recommended objects and the user feature information, encodes the input information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information, determines first time recommendation feature information according to the first time encoded feature information, determines second time recommendation feature information according to the input feature information and the first time encoded feature information, determines a recommendation probability of the candidate recommended objects according to the first time recommendation feature information and the second time recommendation feature information, and determines a target recommended object from the candidate objects according to the recommendation probability and performs recommendation. The device can determine the first time recommendation feature information according to the first time encoded feature, determine the second time recommendation feature information according to the first encoded feature and the input information, and then determine the recommendation probability of the candidate recommended objects according to the first time recommendation feature information and the second time recommendation feature information, and determine the target recommended object from the candidate objects according to the recommendation probability to perform recommendation. Thus, according to the loaded short-term behavior information of the user, the long-term user historical behavior information can be effectively determined, and then according to the long-term historical behavior information of the user and the short-term behavior information of the user, the e-commerce can be accurately recommended.

[0161] To achieve the above-mentioned embodiments, the embodiments of the present disclosure further provide an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the recommendation method according to any one of the preceding method embodiments.

[0162] To achieve the above-mentioned embodiments, the embodiments of the present disclosure further provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the recommendation method according to any one of the preceding method embodiments.

[0163] To achieve the above-mentioned embodiments, the embodiments of the present disclosure further provide a computer program product, wherein the instructions in the computer program product are executable by a processor to implement the recommendation method according to any one of the preceding method embodiments.

[0164] Figure 7 A block diagram of an exemplary electronic device suitable for use in implementing embodiments of the present disclosure is shown. The electronic device can be a server, or it can be a client. Figure 7 The electronic device 700 shown is merely an example and should not be taken as limiting the functionality or use of embodiments of the present disclosure.

[0165] As Figure 7 shown, the electronic device 700 includes:

[0166] a memory 710 and a processor 720, a bus 730 connecting different components (including the memory 710 and the processor 720), the memory 710 storing a computer program, and the processor 720 executing the program to implement the recommendation method according to embodiments of the present disclosure.

[0167] The bus 730 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor or local bus using any of a variety of bus architectures. By way of example, these architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0168] The electronic device 700 typically includes a variety of electronic device readable media. These media can be any available media that is accessible by the electronic device 700 and includes both volatile and non-volatile media, removable and non-removable media.

[0169] The storage 760 can also include a separate memory unit, such as a flash memory card, a floppy disk drive, a magnetic hard disk drive, or a tape drive, among others. The storage 760 can also include an external storage drive accessible via one or more external storage interfaces such as a USB, Firewire, or IEEE 1394 interface, among others. The storage 760 can also include a storage controller capable of accessing multiple storage devices and / or formats. Figure 7 Not shown are, typically, a "hard disk drive", commonly referred to as a "hard drive", or "hard drive". While not specifically shown, such can provide Figure 7 read / write capability of a removable nonvolatile magnetic medium, such as a "floppy disk" or a "flexible disk", and can be accessed via one or more disk drives. In these cases the disk drive can be connected to the system bus 730 by one or more disk drive interfaces. The storage 710 can include one or more program products, comprising related data and / or code, for example, being configured to be executed by the processor 710.

[0170] Program / utility 780, having a set (at least one) of program modules 770, can be stored in memory 710 by way of example, and not limitation, including an operating system, one or more application programs, other program modules, and program data, each of which

[0171] The electronic device 700 can also communicate with one or more external devices 790 such as a keyboard or pointing device, a display 791, etc.; one or more devices that enable a user to interact with the electronic device 700; and / or one or more devices that enable the electronic device 700 to communicate with one or more other computing devices. Such communication can be via an input / output (I / O) interface 792. Similarly, the electronic device 700 can communicate with one or more networks such as one or more local area networks (LANs), wide area networks (WANs), telecommunication networks, and / or the Internet via the network adapter 793. As Figure 7 illustrated, the network adapter 793 can be communicatively coupled to the other components of the electronic device 700 via the bus 730. It will be appreciated that the network adapter 793 can be implemented as part of the processor 710. Alternatively, the network adapter 793 can be implemented as a separate and distinct component within the electronic device 700. Figure 7 It is to be appreciated that the electronic device 700 can be a specially configured server, desktop computer, laptop computer, notebook computer, netbook computer, netpad, tablet computer, cellular radiotelephone, smart phone, personal digital assistant (PDA), or other portable communication device, among others. It is to be appreciated that the electronic device 700 can be configured to operate in a distributed computing environment, such as a cloud computing environment. It is to be appreciated that the electronic device 700 can be configured to operate in a virtualized computing environment, such as a virtual machine environment. It is to be appreciated that the electronic device 700 can be configured to operate in a virtualized computing environment, such as a virtual machine environment.

[0172] The processor 720 performs various function applications and data processing by running programs stored in the memory 710.

[0173] It should be noted that the implementation process and technical principles of the electronic device of the present embodiment are described in the foregoing description of the recommended method of the present disclosure, and will not be repeated here.

[0174] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0175] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present disclosure, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0176] Any process or method descriptions in the flowchart or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing the specified logic functions or processes, and the preferred embodiments of the present disclosure include additional implementations in which the order of execution is not necessarily the same as that shown or discussed, including the implementation of functions according to the involved functions in a substantially simultaneous manner, or in reverse order, which should be understood by those skilled in the art to which the embodiments of the present disclosure belong.

[0177] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of instructions to implement logic functions, and can be embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, processor- containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be a machine-readable storage device (e.g., magnetic, optical or other) a machine-readable storage diskette (e.g., floppy disk, optical disk, CD- ROM, etc.), a machine- readable storage card (e.g., PCMCIA card, etc.), a machine-readable storage tape (e.g., magnetic tape, optical tape, etc.), a machine-readable storage medium (e.g., RAM, ROM, etc.), a machine-readable signal (e.g., electrical, optical, etc.), a machine-readable medium (e.g., carrier wave, etc.) or any other suitable medium or means of embodying the program. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a RAM, a ROM, an EPROM, a FLASH memory card, an optical fiber, and a portable compact disc read-only memory (CD-ROM). Additionally, the computer-readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and stored in a computer memory.

[0178] It should be understood that portions of the present disclosure can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, the various steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, or their combinations can be used: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0179] Those skilled in the art can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing related hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0180] In addition, each functional unit in each embodiment of the present disclosure can be integrated in one processing module, or each unit can exist physically separately, or two or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. When the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0181] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present disclosure have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present disclosure, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present disclosure.

Claims

1. A recommendation method characterized by comprising: The method comprises: obtaining candidate recommended objects and user feature information; determining input feature information according to the feature information of the candidate recommended objects and the user feature information; encoding the input feature information by using an encoding layer of a first time recurrent neural network to obtain at least one first time encoded feature information; determining first time recommended feature information according to the first time encoded feature information; determining second time recommended feature information according to the input feature information and the first time encoded feature information; determining a recommendation probability of the candidate recommended objects according to the first time recommended feature information and the second time recommended feature information; determining a target recommended object from the candidate recommended objects and performing recommendation according to the recommendation probability.

2. The method of claim 1, wherein, The method of determining second time recommended feature information according to the input feature information and the first time encoded feature information comprises: encoding and converting the input feature information by using a conversion network to obtain encoding and conversion feature information corresponding to the input feature information; determining intermediate encoded feature information corresponding to the first time encoded feature information according to the encoding and conversion feature information and the first time encoded feature information; determining storage encoded feature information corresponding to the dimension information of each fully connected layer in a second time recurrent neural network according to the dimension information of each fully connected layer in the second time recurrent neural network and the intermediate encoded feature information; generating second time recommended feature information according to the storage encoded feature information.

3. The method of claim 2, wherein, The method of determining intermediate encoded feature information corresponding to the first time encoded feature information according to the encoding and conversion feature information and the first time encoded feature information comprises: determining a weight corresponding to the first time encoded feature information according to the encoding and conversion feature information and the first time encoded feature information; determining intermediate encoded feature information corresponding to the first time encoded feature information according to the first time encoded feature information and the weight corresponding to the first time encoded feature information.

4. The method of claim 1, wherein, The method of determining a recommendation probability of the candidate recommended objects according to the first time recommended feature information and the second time recommended feature information comprises: determining shared recommended feature information according to the first time recommended feature information and the second time recommended feature information; extracting features of the first time recommended feature information by using a first neural network to obtain a first time feature sequence; extracting features of the second time recommended feature information by using a second neural network to obtain a second time feature sequence; extracting features of the shared recommended feature information by using a third neural network to obtain a shared feature sequence; determining a recommendation probability of a candidate object matching each historical behavior in user historical behavior feature information according to the first time feature sequence, the second time feature sequence and the shared feature sequence.

5. The method of claim 4, wherein, The method of determining a recommendation probability of a candidate object matching each historical behavior in user historical behavior feature information according to the first time feature sequence, the second time feature sequence and the shared feature sequence comprises: Determine the weight corresponding to each historical behavior in the user historical behavior feature information; According to the first time feature sequence, the second time feature sequence and / or the shared feature sequence, determine the first feature sequence matched with each historical behavior; According to the first feature sequence matched with each historical behavior and the weight of the corresponding historical behavior, determine the second feature sequence matched with each historical behavior; Input the second feature sequence into the multi-layer neural network to obtain the recommendation probability value corresponding to each historical behavior; According to the set weight, weight sum of the recommendation probability value corresponding to each historical behavior is determined to determine the recommendation probability of the candidate object matched with each historical behavior.

6. The method of claim 1, wherein, The user feature information includes: user historical behavior feature information; According to the feature information of the candidate recommendation object and the user feature information, before determining the input feature information, it further includes: Sort the user historical behavior feature information according to time to obtain user serialized feature information; Feature statistics are performed on the user serialized feature information to obtain user statistical feature information corresponding to the user serialized feature information.

7. The method of claim 6, wherein, The user feature information includes: user attribute feature information, and the input feature information is determined according to the feature information of the candidate recommendation object and the user feature information, including: Vectorize the user serialized feature information to obtain the vectorized user serialized feature information; Using multi-head attention network mechanism, feature extraction is performed on the vectorized user serialized feature information to obtain first feature information; Splicing is performed on the user attribute feature information, the feature information of the candidate recommendation object and the user statistical feature information to obtain first splicing feature information; Using multi-layer neural network, feature extraction is performed on the first splicing feature information to obtain second feature information; The first splicing feature information and the vectorized user serialized feature information are spliced to obtain second splicing feature information; According to the first feature information, the second feature information and the second splicing feature information, the input feature is generated.

8. The method according to any one of claims 1 to 7, characterized in that, After the encoding layer of the first time recurrent neural network is used to encode the input feature information to obtain at least one first time encoding feature information, it further includes: According to the first time encoding feature information, determine the category probability value corresponding to the first time encoding feature information; According to the category probability value and the corresponding set auxiliary loss function, the first time recurrent neural network is corrected.

9. A recommendation apparatus characterized by comprising: Including: The acquisition module is used for acquiring candidate recommendation objects and user feature information; The first determination module is used for determining the input feature information according to the feature information of the candidate recommendation object and the user feature information; The encoding module is used for encoding the input feature information by using the encoding layer of the first time recurrent neural network to obtain at least one first time encoding feature information; The second determination module is used for determining the first time recommendation feature information according to the first time encoding feature information; The third determining module is configured to determine second-time recommendation feature information according to the input feature information and the first-time encoded feature information. The fourth determining module is configured to determine a recommendation probability of the candidate recommendation object according to the first-time recommendation feature information and the second-time recommendation feature information. The recommendation module is configured to determine a target recommendation object from the candidate recommendation object according to the recommendation probability and perform recommendation.

10. The apparatus of claim 9, wherein, The third determining module is configured to: perform encoding conversion on the input feature information by using a conversion network to obtain encoded conversion feature information corresponding to the input feature information; determine intermediate encoded feature information corresponding to the first-time encoded feature information according to the encoded conversion feature information and the first-time encoded feature information; determine storage encoded feature information corresponding to the dimension information of each full connection layer in the second-time recurrent neural network according to the dimension information of each full connection layer in the second-time recurrent neural network and the intermediate encoded feature information; generate second-time recommendation feature information according to the storage encoded feature information.

11. The apparatus of claim 10, wherein, The third determining module is further configured to: determine a weight corresponding to the first-time encoded feature information according to the encoded conversion feature information and the first-time encoded feature information; determine intermediate encoded feature information corresponding to the first-time encoded feature information according to the first-time encoded feature information and the weight corresponding to the first-time encoded feature information.

12. The apparatus of claim 9, wherein, The fourth determining module is configured to: determine shared recommendation feature information according to the first-time recommendation feature information and the second-time recommendation feature information; perform feature extraction on the first-time recommendation feature information by using a first neural network to obtain a first-time feature sequence; perform feature extraction on the second-time recommendation feature information by using a second neural network to obtain a second-time feature sequence; perform feature extraction on the shared recommendation feature information by using a third neural network to obtain a shared feature sequence; determine a recommendation probability of a candidate object matched with each historical behavior in the user historical behavior feature information according to the first-time feature sequence, the second-time feature sequence and the shared feature sequence.

13. The apparatus of claim 12, wherein, The fourth determining module is further configured to: determine a weight corresponding to each historical behavior in the user historical behavior feature information; determine a first feature sequence matched with each historical behavior according to the first-time feature sequence, the second-time feature sequence and / or the shared feature sequence; determine a second feature sequence matched with each historical behavior according to the first feature sequence matched with each historical behavior and the weight of the corresponding historical behavior; input the second feature sequence into a multi-layer neural network to obtain a recommendation probability value corresponding to each historical behavior; weight and sum the recommendation probability values corresponding to each historical behavior according to a set weight to determine a recommendation probability of a candidate object matched with each historical behavior.

14. The apparatus of claim 9, wherein, The user feature information includes user historical behavior feature information; and the apparatus further includes: The sorting module is configured to sort the user historical behavior feature information according to time to obtain user serialized feature information. The statistical module is configured to perform feature statistics on the user serialized feature information to obtain user statistical feature information corresponding to the user serialized feature information.

15. The apparatus of claim 14, wherein, The user feature information includes user attribute feature information, and the first determining module is configured to: vectorize the user serialized feature information to obtain vectorized user serialized feature information; perform feature extraction on the vectorized user serialized feature information by using a multi-head attention network mechanism to obtain first feature information; splice the user attribute feature information, the feature information of the candidate recommended object, and the user statistical feature information to obtain first spliced feature information; perform feature extraction on the first spliced feature information by using a multi-layer neural network to obtain second feature information; splice the first spliced feature information and the vectorized user serialized feature information to obtain second spliced feature information; generate the input feature according to the first feature information, the second feature information, and the second spliced feature information.

16. The apparatus of any one of claims 9-15, wherein, The apparatus further includes: The fifth determining module is configured to determine a category probability value corresponding to the first time coding feature information according to the first time coding feature information. The correction module is configured to correct the first time recurrent neural network according to the category probability value and a corresponding set auxiliary loss function.

17. An electronic device, comprising: A computer program product comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of claims 1-8 when executing the computer program.

18. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program product comprises a computer program which, when executed by a processor, implements the method of any one of claims 1-8.

19. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-8.

Images