Object recommendation method, electronic equipment and storage medium

Abstract

The invention provides an object recommendation method, electronic equipment and a storage medium. The method comprises the following steps: acquiring a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; the historical behavior is a forward feedback behavior aiming at the behavior object; determining a first historical behavior from the plurality of historical behaviors, and determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining a correlation degree between the comprehensive object features and reference behavior objects corresponding to historical behaviors except the first historical behavior, and determining a target behavior object based on the correlation degree; fusing the comprehensive object features with object features of the target behavior object to obtain fused features; and based on the fusion features, determining a recommendation object corresponding to the user account from the alternative objects. According to the mode, the data value of a plurality of forward feedback behaviors can be brought into full play, and the recommendation accuracy and individuation degree are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a recommendation method of an object, an electronic device and a storage medium. BACKGROUND

[0002] In various recommendation scenarios such as video recommendation, music recommendation, e-commerce commodity recommendation, etc., in order to achieve accurate recommendation and prediction of user interest preferences, different types of user behavior sequences are often analyzed and modeled. Taking the video recommendation scenario as an example, the user behavior sequence includes a like video sequence, a comment video sequence, and a watch-to-end video sequence, etc. The like video sequence contains all the video names that the user actively likes; the comment video sequence contains the video names that the user publishes a good comment on; and the watch-to-end video sequence contains the video names that the user watches completely. These different behavior sequences reflect the user's interest tendency from different dimensions.

[0003] In related technologies, the like video sequence, the comment video sequence and the watch-to-end video sequence are simply connected to form a comprehensive sequence, which is then input into a neural network model for overall modeling to obtain a recommended video. This method cannot fully utilize the comprehensive value of multi-sequence data, and has certain limitations in recommendation accuracy and prediction accuracy, and cannot meet the user's demand for high-quality personalized recommendation. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a recommendation method of an object, an electronic device and a storage medium, so as to fully utilize the data value of multiple historical behaviors, improve the recommendation accuracy and prediction accuracy of the object, and meet the user's demand for high-quality personalized recommendation.

[0005] In a first aspect, an embodiment of the present application provides a recommendation method of an object, the method comprising: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate an object acted on by the historical behavior; and the historical behavior is a positive feedback behavior of the user account to the behavior object; determining a first historical behavior from the plurality of historical behaviors, determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining a correlation degree between the comprehensive object feature and a reference behavior object corresponding to a historical behavior other than the first historical behavior, and determining a target behavior object from the reference behavior object based on the correlation degree; fusing the comprehensive object feature and an object feature of the target behavior object to obtain a fused feature; and determining a recommended object corresponding to the user account from a plurality of candidate objects based on the fused feature.

[0006] In a second aspect, an embodiment of the present application provides a recommendation device for an object, the device comprising: a first obtaining module configured to obtain a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate an object acted on by the historical behavior; the historical behavior is a positive feedback behavior for the behavior object; a first determining module configured to determine a first historical behavior from the plurality of historical behaviors; determine a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; a second determining module configured to determine a correlation degree between the comprehensive object feature and reference behavior objects corresponding to historical behaviors other than the first historical behavior, and determine a target behavior object from the reference behavior objects based on the correlation degree; and a first fusing module configured to fuse the comprehensive object feature and an object feature of the target behavior object to obtain a fused feature, and determine a recommended object corresponding to the user account from candidate objects based on the fused feature.

[0007] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, the memory storing machine executable instructions capable of being executed by the processor, and the processor executes the machine executable instructions to implement the above-mentioned recommendation method for an object.

[0008] In a fourth aspect, an embodiment of the present application provides a storage medium, the storage medium storing machine executable instructions, and the machine executable instructions, when invoked and executed by a processor, cause the processor to implement the above-mentioned recommendation method for an object.

[0009] The embodiments of the present application bring the following beneficial effects: The above-mentioned recommendation method for an object, electronic device and storage medium, the method comprising: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate an object acted on by the historical behavior; the historical behavior is a positive feedback behavior for the behavior object; determining a first historical behavior from the plurality of historical behaviors, and determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining a correlation degree between the comprehensive object feature and reference behavior objects corresponding to historical behaviors other than the first historical behavior, and determining a target behavior object from the reference behavior objects based on the correlation degree; fusing the comprehensive object feature and an object feature of the target behavior object to obtain a fused feature; and determining a recommended object corresponding to the user account from candidate objects based on the fused feature.

[0010] In the manner, the core preference of the first historical behavior of the user is used to realize directional feature fusion and generate a fusion feature, and then based on the fusion feature, a recommended object corresponding to the user account is determined from the candidate objects. In the feature fusion process, the first historical behavior can be selected as the behavior with the highest indication strength of the user's preference, and the comprehensive object feature of the first historical behavior reflects the core preference of the user embodied by the behavior. Then, based on the comprehensive object feature, the behavior objects corresponding to the historical behaviors other than the first historical behavior are filtered out, which are highly matched with the core preference corresponding to the first historical behavior. The comprehensive object feature is fused with the object features of the filtered behavior objects to obtain the fusion feature. The fusion feature not only retains the core preference reflected by the first historical behavior, but also expands the dimension of the user's related interest, avoiding that the user's preference represented by the fusion feature is too single. The manner can fully play the data value of multiple positive feedback behaviors, effectively improve the matching degree of the recommended object and the real demand of the user, and improve the accuracy and personalization degree of the recommendation.

[0011] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application will be realized and attained by the structure particularly pointed out in the description and claims.

[0012] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0014] Figure 1 A flow chart of an object recommendation method provided by an embodiment of the present application is shown in the figure. Figure 2 A flow chart of another object recommendation method provided by an embodiment of the present application is shown in the figure. Figure 3 A structure schematic diagram of an object recommendation device provided by an embodiment of the present application is shown in the figure. Figure 4 A structure schematic diagram of an electronic device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0015] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0016] Based on this, the embodiments of the present application provide a recommended method of an object, an electronic device and a storage medium, which can be applied to various recommendation scenarios such as video recommendation, music recommendation, e-commerce commodity recommendation, etc.

[0017] In order to facilitate the understanding of the present embodiment, first, a recommended method of an object disclosed by the embodiments of the present application will be described in detail, as shown in the following figure: Figure 1 The method comprises the following steps: Step S102, obtaining a plurality of historical behaviors of a user account and a behavior object corresponding to the historical behaviors; wherein the behavior object is used to indicate the object of the historical behavior; and the historical behavior is a positive feedback behavior for the behavior object.

[0018] The historical behavior is a positive feedback behavior actively initiated by the user account in different scenarios. The positive feedback behavior refers to the behavior of showing recognition and love to the target object through a specific behavior. Such behavior can reflect the preference tendency or interest point of the user. The positive feedback behaviors in different scenarios have different forms, for example: the "heart" behavior and the "complete playback" behavior for a song in a music scenario; the "watching to the end", "liking" and "collecting" behaviors for a video in a video scenario; and the "adding to cart" and "ordering" behaviors for a commodity in an e-commerce scenario. It should be noted that the indication strength of the user's preference for different types of historical behaviors is obviously different: taking the music scenario as an example, the "heart" behavior is the behavior of the user actively marking a highly loved song, and the preference indication strength thereof is higher than that of the "complete playback" behavior, and can more accurately reflect the core preference of the user for the song; similarly, the preference indication strength of the "ordering" behavior in the e-commerce scenario is also much higher than that of the "adding to cart" behavior, and can represent the real purchase demand of the user.

[0019] The behavior object is used to indicate the object of the historical behavior. The behavior object is a unique identifier indicating the object of the historical behavior, such as a song name in a music scenario, a video name in a video scenario, a commodity name in an e-commerce scenario, etc. The object of the historical behavior can be accurately obtained through the behavior object, and the accurate association of each historical behavior with the corresponding object is ensured.

[0020] Here, a plurality of historical behaviors of a user account and a behavior object corresponding to each historical behavior are obtained.

[0021] Step S104, determining a first historical behavior from the plurality of historical behaviors, determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior.

[0022] The first historical behavior can be a behavior with the highest indication strength of user preference among the plurality of historical behaviors. For example, in a music scenario, if there are two types of historical behaviors of "heart" and "complete play" in the user account, the "heart" behavior is a behavior of actively marking a highly favorite song, and the indication strength of the "heart" behavior is higher than that of the "complete play" behavior. Therefore, the "heart" behavior can be the first historical behavior. It should be noted that the first behavior object corresponding to the first historical behavior includes at least one.

[0023] The object feature can be in the form of a feature vector. In actual implementation, the behavior object can be input into a pre-trained feature extraction model, core semantic information of the behavior object corresponding to the historical behavior object is extracted through the model, for example, in a music scenario, the style of a song, the theme of a song, in an e-commerce scenario, the category of a product, the function attribute, and finally the object feature vector of the behavior object is output, and the vector is used as the object feature of the behavior object. It should be noted that the feature dimensions of the object feature of the behavior object corresponding to each historical behavior need to be consistent, that is, no matter which type of historical behavior the behavior object belongs to, the number of dimensions of the object feature and the arrangement position of each dimension in the vector need to be completely consistent. For example, in a music scenario, if the feature vector dimension of song A corresponding to the "heart" behavior is set to "[style, rhythm, lyrics emotion] 3 dimensions, the feature vector of song B corresponding to the "complete play" behavior must also include "style, rhythm, lyrics emotion" 3 dimensions, and the arrangement position of each dimension in all vectors needs to be completely consistent. This requirement for consistency of dimensions is the basis for subsequent implementation of multi-behavior object feature operation.

[0024] The comprehensive object feature of the first historical behavior is used to indicate the overall feature of the behavior object corresponding to the first historical behavior, and to reflect the core user preference behind the type of historical behavior. If the first behavior object corresponding to the first historical behavior is one, the object feature of the first behavior object is used as the comprehensive object feature. If the first behavior object corresponding to the first historical behavior includes a plurality of objects, preferably, the average value of the vector can be obtained by averaging the object feature vectors of the plurality of first behavior objects, and the obtained vector average value is used as the comprehensive object feature.

[0025] In other manners, different generation manners can be adopted to generate the comprehensive object feature according to actual scene requirements and data characteristics. For example, if key information in the feature needs to be highlighted, a max-pooling manner can be adopted to take the maximum value of each dimension of the object feature vectors of the plurality of first behavior objects in the dimension, so as to strengthen the preference feature of the user in the core dimension, and finally generate the comprehensive object feature. If the weight needs to be dynamically adjusted in combination with the user interaction difference, an attention mechanism can be introduced, the object features of different first behavior objects are assigned with corresponding weights according to the correlation degree between the different first behavior objects and the user preference, and then the comprehensive object feature is generated based on the weighted feature vectors, so that the overall feature is more in line with the real preference of the user.

[0026] In this step, the first historical behavior is determined from the plurality of historical behaviors, the comprehensive object feature of the first historical behavior is determined according to the object features of the plurality of first behavior objects corresponding to the first historical behavior, the core preference of the user behind this type of historical behavior is reflected, and the common feature of the user in this type of behavior is highlighted.

[0027] In step S106, the correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior is determined, and the target behavior object is determined from the reference behavior object based on the correlation degree.

[0028] The correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior can be represented by the feature distance between the comprehensive object feature and the object feature of the reference behavior object. The smaller the feature distance, the higher the correlation degree, which means that the reference behavior object and the behavior object corresponding to the first historical behavior are more consistent in core attributes such as music style, lyrics theme, product category, functional attribute, etc., and are more in line with the core preference of the user reflected by the first historical behavior.

[0029] Specifically, the form of the feature vector can be further explained. Since the comprehensive object feature and the object feature of the reference behavior object are both in the form of a feature vector. For example, in the music scene, the comprehensive object feature vector is [0.8 (ballad style), 0.3 (fast rhythm), 0.9 (healing lyrics)], for each historical behavior other than the first historical behavior, the object feature vector corresponding to each reference behavior object corresponding to the historical behavior is calculated with the comprehensive object feature vector by using common algorithms such as Euclidean distance and cosine distance to calculate the vector distance between the two vectors, and the calculated vector distance is determined as the correlation degree. Then, find the specified number of reference behavior objects with the smallest vector distance value from the comprehensive object feature vector, and take them as the target behavior object. For example, in the music scene listed above, the smaller the vector distance value, the higher the consistency of the reference behavior object and the behavior object corresponding to the first historical behavior in the dimensions of "style, rhythm, lyrics emotion".

[0030] In step S108, the integrated object feature is fused with the object feature of the target behavior object to obtain a fused feature; and based on the fused feature, a recommended object corresponding to the user account is determined from the candidate objects.

[0031] Here, the object feature is in the form of a feature vector. If the target behavior object is one, a preset weight parameter can be called according to the scene requirement, and the integrated object feature vector and the object feature vector of the target behavior object are processed by dimension-by-dimension weighted summation to obtain the fused feature.

[0032] For example, in a music scene, the integrated feature vector is [0.8 (lyric), 0.3 (slow rhythm)], the target feature vector is [0.75 (lyric), 0.25 (slow rhythm)], the weight parameter corresponding to the integrated feature vector is 0.7, and the weight parameter corresponding to the target feature vector is 0.3. Then, the dimension value of “lyric” after fusion is 0.8*0.7+0.75*0.3=0.785, the dimension value of “slow rhythm” after fusion is 0.3*0.7+0.25*0.3=0.285, and the fused feature vector is [0.785, 0.285].

[0033] If the target behavior object is multiple, the object feature vectors of the multiple target behavior objects can be subjected to vector average operation to obtain an average vector. Then, according to a preset weight parameter, the integrated object feature vector is multiplied by the corresponding weight parameter, the average vector is multiplied by the corresponding weight parameter, and the two products are added to obtain a fused feature vector. The fused feature vector is determined as the fused feature.

[0034] Further, the candidate objects are also input into the pre-trained feature extraction model to extract the candidate object features corresponding to the candidate objects, i.e., candidate object feature vectors, and the feature dimensions of the candidate object feature vectors are completely consistent with those of the fused feature vector. Then, for each candidate object, the candidate object feature vector and the fused feature vector are spliced in a preset order, such as the fused feature vector first and the candidate object feature vector second, to form a first joint feature containing both the fused feature vector and the candidate object feature vector. The first joint feature is input into a preset scoring function, which is used to quantitatively evaluate the matching degree of the candidate object feature vector and the fused feature vector. For example, the scoring function can be designed based on the Sigmoid function: first, linearly weight each dimension of the first joint feature to obtain a linear combination result; then, input the result into the Sigmoid function to map it to the interval [0, 1], and finally output a recommendation score. The closer the score is to 1, the higher the matching degree of the candidate object feature vector and the fused feature vector, and the more suitable the candidate object is as a recommended object. Finally, the recommended object corresponding to the user account is determined from the candidate objects according to the output recommendation score.

[0035] In this way, the comprehensive object feature is fused with the object feature of the target behavior object to obtain a fused feature. In the process of obtaining the fused feature, the object feature highly matched with the comprehensive object feature of the first historical behavior is accurately screened out as a supplement, so that the generated fused feature not only retains the core preference direction of the user reflected by the first historical behavior, but also expands the dimension of the related interests of the user through the supplement feature, forming a more comprehensive and accurate preference representation.

[0036] This way effectively mines the cross-relation information between the behavior objects corresponding to different historical behaviors of the user account, improving the understanding of the user behavior patterns and interest preferences. In the music recommendation scenario, by analyzing the playlists corresponding to the "heart" behavior and the "complete playback" behavior, the common characteristics of songs that are both "hearted" and "completely played" are mined, which can more accurately find the types of songs that the user likes and completely listens to, so as to recommend music works that better meet the user's interests, and improve the user's satisfaction and participation in the recommendation results. In the e-commerce field, by analyzing the product lists corresponding to the "add to cart" and "order" behaviors, the user's potential purchase of the product after browsing can be accurately grasped, the conversion rate of product recommendation can be improved, and the sales growth of the e-commerce platform can be promoted.

[0037] The above object recommendation method includes: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate: the object of the historical behavior; the historical behavior is a positive feedback behavior for the behavior object; determining a first historical behavior from the plurality of historical behaviors, determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining the correlation degree between the comprehensive object feature and a reference behavior object corresponding to a historical behavior other than the first historical behavior, determining a target behavior object from the reference behavior object based on the correlation degree; fusing the comprehensive object feature with the object feature of the target behavior object to obtain a fused feature; and determining a recommended object corresponding to the user account from the candidate objects based on the fused feature.

[0038] In this mode, the core preference of the first historical behavior of the user is used to realize directional feature fusion and generate a fused feature, and then based on the fused feature, a recommended object corresponding to the user account is determined from the candidate objects. In the feature fusion process, the first historical behavior can be selected as the behavior with the highest indication of user preference, and the comprehensive object feature of the first historical behavior reflects the core preference of the user embodied by the behavior; and based on the comprehensive object feature, behavior objects highly matching the core preference of the first historical behavior are selected from behavior objects corresponding to historical behaviors other than the first historical behavior, and the comprehensive object feature is fused with the object features of the selected behavior objects to obtain the fused feature. The fused feature not only retains the core preference reflected by the first historical behavior, but also expands the dimension of the user's related interests, avoiding that the user preference represented by the fused feature is too single. This mode can fully utilize the data value of multiple positive feedback behaviors, effectively improve the matching degree of the recommended object and the real needs of the user, and improve the accuracy and personalization of the recommendation.

[0039] In one mode, before determining the comprehensive object feature of the first historical behavior, the behavior object corresponding to each historical behavior is input into a pre-trained feature extraction model, and the object feature of the behavior object is output.

[0040] Here, for each historical behavior, the behavior object corresponding to the historical behavior is input into a pre-trained feature extraction model, and the object corresponding to the historical behavior of the behavior object is pre-processed by the model Embedding, and finally the object feature of the behavior object is output. The object feature is in the form of a feature vector.

[0041] The feature extraction model is pre-trained based on a large-scale corpus or user historical behavior data to construct training samples, and is trained using a word vector training algorithm. During training, the algorithm maps the object corresponding to the historical behavior of each behavior object into a low-dimensional dense vector space to form a corresponding Embedding vector. Here, objects with similar semantics or similar behavior patterns are closer in the vector space. For example, the vector distance of two songs of the same lyrical style is significantly smaller than that of songs of different styles.

[0042] The following embodiments provide specific implementation modes for determining the comprehensive object feature of the first historical behavior.

[0043] In one mode, a plurality of first behavior objects corresponding to the first historical behavior are determined, and a first average feature corresponding to the object features of the plurality of first behavior objects is determined as the comprehensive object feature.

[0044] The first historical behavior can be a behavior with the highest indication intensity of user preference among multiple historical behaviors. For example, in a music scenario, if a user account has both "heart" and "complete play" historical behaviors, the "heart" behavior can be the first historical behavior.

[0045] First, a plurality of first behavior objects corresponding to the first historical behavior are determined. Since the object features are in the form of feature vectors, the object feature vector corresponding to each first behavior object needs to be obtained first. The feature dimensions of these vectors are completely consistent. For example, each vector contains three dimensions of "style, rhythm, and lyrics emotion", and the arrangement positions of each dimension in the vector are completely consistent.

[0046] Then, the object feature vectors corresponding to the plurality of first behavior objects are subjected to average value operation between vectors. That is, for each fixed dimension, the values of all vectors in this dimension are summed and divided by the total number of first behavior objects to obtain the average value of this dimension. The average values of all dimensions are integrated according to the original dimension arrangement order to form a new vector, which is the first average feature, and is determined as the comprehensive object feature corresponding to the first historical behavior.

[0047] The comprehensive object feature summarizes the overall preference of the user for the first historical behavior. The comprehensive object feature is obtained by averaging the features, which weakens the noise influence of a single behavior object. In addition, by calculating the average value by dimension, the common features of all first behavior objects can be accurately extracted, and the preference tendency of the user on the first historical behavior reflected by the comprehensive object feature is more reliable.

[0048] In one way, the correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior can be represented by the feature distance between the comprehensive object feature and the object feature of the reference behavior object: Specifically, the historical behavior other than the first historical behavior is determined, the object feature of the reference behavior object corresponding to the historical behavior other than the first historical behavior is obtained, the feature distance between the comprehensive object feature and the object feature of the reference behavior object is determined, and the feature distance between the comprehensive object feature and the object feature of the reference behavior object is determined as the correlation degree.

[0049] The feature vector can be further explained in detail. For example, in a music scene, the comprehensive object feature vector is [0.8 (lyricism), 0.3 (fast rhythm), 0.9 (healing lyrics)], and for each historical behavior other than the first historical behavior, the object feature vector corresponding to each reference behavior object corresponding to the historical behavior is calculated by using a commonly used algorithm such as Euclidean distance or cosine distance to calculate the vector distance between the comprehensive object feature vector and the object feature vector, and then the vector distance value is determined to be the correlation degree. Then, the specified number of reference behavior objects with the smallest vector distance value are found as the target behavior objects.

[0050] Further, the target behavior objects are determined from the reference behavior objects based on the feature distance between the comprehensive object feature and the object feature of the reference behavior objects. The target behavior objects have a specified number, and the feature distance between the object feature of the target behavior objects and the comprehensive object feature is not greater than the feature distance between the object feature of the reference behavior objects other than the target behavior objects and the comprehensive object feature.

[0051] That is, the vector distance between the feature vector of all reference behavior objects and the comprehensive object feature vector is calculated to obtain the distance value corresponding to each reference behavior object. Then, all reference behavior objects are sorted in order of "distance value from small to large". Finally, the first specified number of reference behavior objects are selected from the sorting result to determine the target behavior objects.

[0052] In addition to determining the correlation degree between the comprehensive object feature and the reference behavior objects by calculating the vector distance, the correlation degree can also be determined by using the Pearson correlation coefficient and the Gaussian kernel function. Specifically, the Pearson correlation coefficient measures the linear correlation trend between the comprehensive object feature vector and the object feature vector of the reference behavior objects to represent the correlation degree between them. The Pearson correlation coefficient has a value range of [-1, 1], and the closer it is to 1, the more consistent the trend of the two vectors in each feature dimension, and the higher the correlation degree. The closer it is to -1, the more opposite the trend, and the lower the correlation degree. Close to 0, there is no obvious linear correlation. The Gaussian kernel function maps the comprehensive object feature vector and the object feature vector of the reference behavior objects to a high-dimensional space to capture the non-linear correlation that is difficult to detect in the original low-dimensional space, and then represents the correlation degree by the kernel function value. The kernel function value has a value range of (0, 1], and the closer it is to 1, the higher the similarity of the two vectors in the high-dimensional space, and the stronger the correlation degree. The closer it is to 0, the weaker the correlation degree.

[0053] The following embodiments provide specific implementation methods for obtaining fused features.

[0054] In one mode, the target behavior object includes multiple, the second average feature corresponding to the object feature of the multiple target behavior objects is calculated; based on the preset weight parameter, the integrated object feature and the second average feature are fused to obtain the fusion feature.

[0055] Specifically, the average vector of the object feature vectors of the multiple target behavior objects can be obtained as the second average feature according to the mode of the feature vector. Then, the integrated object feature vector is multiplied by the corresponding weight parameter, the vector corresponding to the second average feature is multiplied by the corresponding weight parameter, and the two product results are added to obtain the fusion feature vector. The fusion feature vector is determined as the fusion feature.

[0056] For example, in a music scene, the feature vectors of two target behavior objects are [0.2, 0.2, 0.8] and [0.8, 0.4, 0.6] respectively, and the feature dimensions are style, rhythm, and lyrics emotion in turn. The average value of the style dimension is (0.2+0.8) / 2=0.5, the average value of the rhythm dimension is (0.2+0.4) / 2=0.3, and the average value of the lyrics emotion dimension is (0.8+0.6) / 2=0.7. Finally, the average vector [0.5, 0.3, 0.5] is obtained, which is taken as the second average feature.

[0057] Then, two groups of weight parameters are preset according to the business scene requirements. For example, if it is required to give priority to preserving the user's core preferences reflected by the integrated object feature, the weight parameter of the integrated object feature vector can be set to 0.6, and the weight parameter of the vector corresponding to the second average feature can be set to 0.4. Then, the multiplication operation of the vector and the weight parameter is performed respectively, the value of each dimension of the integrated object feature vector is multiplied by the corresponding weight parameter to obtain an integrated feature weighted vector, and the value of each dimension of the vector corresponding to the second average feature is multiplied by the corresponding weight parameter to obtain an average feature weighted vector. Finally, the values of the corresponding dimensions of the two weighted vectors are added to obtain the values of the dimensions of the fusion feature vector, and the integrated vector is the final fusion feature vector, which is determined as the fusion feature.

[0058] For example, the integrated object feature vector is [0.8, 0.3, 0.7], the weight parameter is 0.6, the average vector is [0.5, 0.3, 0.5], and the weight parameter is 0.4. The integrated feature weighted vector is [0.48, 0.18, 0.42], the average feature weighted vector is [0.2, 0.12, 0.2], and the fusion feature vector is [0.68, 0.3, 0.62], which is the fusion feature.

[0059] In another way, the target behavior objects can be classified by a clustering algorithm such as K-Means, and target behavior objects with similar features are classified into different categories. Then, representative core behavior objects are selected from each category. Subsequently, the average features of the core behavior objects are calculated. Finally, the average features are fused with the comprehensive object features to generate a fusion feature that not only retains the core preferences reflected by the first historical behavior but also takes into account the diversity of the target behavior.

[0060] The following embodiments provide specific implementation modes of determining the recommended objects corresponding to the user account.

[0061] The candidate object features of the candidate objects are obtained, wherein the feature dimensions of the candidate object features are consistent with the fusion features. For each candidate object, the object features of the candidate object are spliced with the fusion features to obtain first joint features. The first joint features are processed by a preset scoring function to obtain a recommendation score corresponding to the candidate object. Based on the recommendation score, the recommended objects corresponding to the user account are determined from the candidate objects.

[0062] The object features are in the form of feature vectors. The feature dimensions of the object features of the candidate objects and the fusion feature vectors are consistent, that is, the number of dimensions and the arrangement positions of each dimension in the vector of the object feature vector of the candidate object and the fusion feature vector must be completely consistent. The behavior objects corresponding to the candidate objects and the historical behavior can be input into the same pre-trained feature extraction model to extract the object features.

[0063] After the object feature vectors of the candidate objects are extracted, the fusion feature vector and the object feature vector of the candidate object are spliced in a preset order to form a first joint feature vector containing two types of feature information. For example, the fusion feature vector is [0.7, 0.2, 0.8], and the object feature vector of a candidate object is [0.8, 0.3, 0.8]. After splicing in the order of “fusion feature vector first and candidate object feature vector second”, the first joint feature vector is [0.7, 0.2, 0.8, 0.8, 0.3, 0.8].

[0064] Then, the first joint feature vector corresponding to each candidate object is input into a preset scoring function such as a Sigmoid function for processing, and the recommendation score corresponding to the candidate object is output. The core function of the scoring function is to quantitatively evaluate the matching degree of the candidate object features and the fusion features, and the higher the score, the more suitable the candidate object is for the user's preferences.

[0065] Finally, the recommendation scores of all candidate objects are sorted in descending order of the recommendation scores, and the top several candidate objects with high recommendation scores are selected from the sorting result as the recommended objects corresponding to the user account.

[0066] In this mode, by effectively mining the cross-relation information between the behavior objects corresponding to different historical behaviors of the user account, the understanding ability of the user behavior mode and interest preference is improved. In the music recommendation scene, by analyzing the playlists corresponding to the "heart" behavior and the "complete playback" behavior, the common characteristics of the songs that are both "heart" and "complete playback" are mined, the song types that the user likes and completely listens to can be more accurately found, so that the music works that are more consistent with the interests of the user are recommended, and the satisfaction and participation of the user to the recommendation result are improved. In the e-commerce field, by analyzing the product lists corresponding to the "add to cart" and "order" behaviors, the products that the user is likely to purchase after browsing can be accurately grasped, the conversion rate of product recommendation is improved, and the sales growth of the e-commerce platform is promoted.

[0067] Next, a recommendation method of another object disclosed by the embodiment of the application is described in detail, as shown in the following Figure 2 The method comprises the following steps: In step S202, a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors are obtained; the behavior object is used to indicate the object of the historical behavior; and the historical behavior is a positive feedback behavior of the user account to the behavior object.

[0068] This step is the same as step S201, and will not be described here.

[0069] In step S204, a first historical behavior is determined from the plurality of historical behaviors, and a comprehensive object feature of the first historical behavior is determined based on an object feature of a first behavior object corresponding to the first historical behavior.

[0070] This step is the same as step S104, and the above-mentioned first historical behavior can be a behavior with the highest indication intensity of user preference in the plurality of historical behaviors. The first historical behavior is determined from the plurality of historical behaviors, and the comprehensive object feature of the first historical behavior is determined according to the object feature of the first behavior object corresponding to the first historical behavior, so as to reflect the core preference of the user behind this type of historical behavior and highlight the common characteristics of the user in this type of behavior.

[0071] In step S206, the correlation degree between the comprehensive object feature and a reference behavior object corresponding to the historical behavior other than the first historical behavior is determined, and the target behavior object is determined from the reference behavior object based on the correlation degree.

[0072] The step S106 can be used to represent the correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior by integrating the feature distance between the object feature of the comprehensive object and the object feature of the reference behavior object. The smaller the feature distance is, the higher the compatibility of the reference behavior object with the behavior object corresponding to the first historical behavior is in core attributes such as the style of music, the theme of lyrics, the category of goods, the functional attributes and the like, and the more the reference behavior object conforms to the user core preference reflected by the first historical behavior. The vector distance value calculated is determined as the correlation degree, and then a specified number of reference behavior objects with the smallest vector distance value from the comprehensive object feature vector are found out and used as the target behavior objects.

[0073] In step S208, the comprehensive object feature is fused with the object feature of the target behavior object to obtain a fused feature.

[0074] If the target behavior objects are multiple, the object feature vectors of the multiple target behavior objects can be subjected to vector average operation to obtain an average vector. Then, the comprehensive object feature vector is multiplied by the corresponding weight parameter, the average vector is multiplied by the corresponding weight parameter, and the two product results are added to obtain a fused feature vector. The fused feature vector is determined as the fused feature.

[0075] Finally, the generated fused feature not only retains the user core preference direction reflected by the first historical behavior, but also expands the dimension of the user related interest through the supplementary feature, thereby forming a more comprehensive and accurate preference representation.

[0076] In step S210, a user multi-source feature is obtained, the fused feature is fused with the user multi-source feature based on a preset fusion weight parameter to obtain a multi-source fused feature, and a recommended object is determined from the candidate objects based on the multi-source fused feature. The user multi-source feature includes a user portrait feature and a time feature of historical behavior.

[0077] The user multi-source feature includes the user portrait feature and the time feature of historical behavior. The user portrait feature can be extracted from user portrait information by a portrait feature extraction model trained in advance, and reflects the identity attribute of the user. The user portrait information can include the age of the user account, the gender of the user account, the occupation of the user and the like.

[0078] The time feature of historical behavior can be extracted from time data of historical behavior by a time feature extraction model trained in advance. Specifically, the time feature of historical behavior can be a dynamic feature extracted based on time nodes, time intervals and time frequencies of historical behavior, and represents the behavior rule of the user in the time dimension. The time data of historical behavior can be the time stamp when the user performs the historical behavior, the frequency of historical behavior and the like.

[0079] The user profile features and the temporal features of historical behavior can be represented by user profile feature vectors and temporal feature vectors of historical behavior, respectively. Specifically, based on the business scenario requirements, fusion weight parameters are preset, and the fusion feature vectors corresponding to the fusion features, user profile feature vectors, and temporal feature vectors of historical behavior are fused to generate multi-source fusion feature vectors, which are then used as multi-source fusion features.

[0080] After extracting the object feature vectors of the candidate objects, the multi-source fusion feature vectors and the object feature vectors of the candidate objects are concatenated in a preset order. The input is then processed by a preset scoring function such as the Sigmoid function, and the recommended score corresponding to the candidate object is output. The higher the score, the more the candidate object matches the user's preferences.

[0081] Finally, the recommendation scores of all candidates are sorted from highest to lowest, and the top-ranked candidates are selected as the recommended candidates for the user accounts.

[0082] This approach first anchors users' core behavioral preferences by fusing features. Based on these core preferences, user profile features and historical behavior time features are added. This allows the multi-source fusion features to not only anchor core needs but also take into account user attributes and dynamic behavioral patterns, achieving a more comprehensive and in-depth representation of user behavior and interests. Through multi-source fusion features, recommended objects are determined from the candidate objects, which are more in line with users' real needs and usage scenarios, effectively improving the accuracy and personalization of recommendations.

[0083] Specifically, user profile information is acquired and input into a pre-trained profile feature extraction model to output user profile features; the number of feature dimensions of the user profile features is consistent with the number of feature dimensions of the fused features; historical behavior time data is acquired and input into a pre-trained time feature extraction model to output historical behavior time features; the number of feature dimensions of the historical behavior time features is consistent with the number of feature dimensions of the fused features, facilitating direct subsequent fusion operations with the fused features.

[0084] The aforementioned user profile information may include: user age, gender, region, occupation, and other information provided during account registration or identified by the system. The aforementioned historical behavior time data may include: timestamps of the user's historical actions and the frequency of those actions.

[0085] The user portrait information is input into a pre-trained portrait feature extraction model, and a user portrait feature is output, which is represented by a user portrait feature vector; the time data is input into a pre-trained time feature extraction model, and a time feature of a historical behavior is output, which is represented by a time feature vector of the historical behavior. The dimension number of the final output vector of the user portrait feature vector or the time feature vector of the historical behavior needs to be completely consistent with the dimension number of the fusion feature vector, so as to facilitate subsequent fusion operation with the fusion feature vector and generate a multi-source fusion feature vector.

[0086] Corresponding to the method embodiment, referring to Figure 3 A schematic diagram of a recommended object device is shown, and the device comprises: The first acquisition module 302 is configured to acquire a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; the behavior object is used to indicate an object acted on by the historical behavior; and the historical behavior is a positive feedback behavior of the user account to the behavior object. The first determination module 304 is configured to determine a first historical behavior from the plurality of historical behaviors; and determine a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior. The second determination module 306 is configured to determine a correlation degree between the comprehensive object feature and a reference behavior object corresponding to a historical behavior other than the first historical behavior; and determine a target behavior object from the reference behavior object based on the correlation degree. The first fusion module 308 is configured to fuse the comprehensive object feature and an object feature of the target behavior object to obtain a fusion feature; and determine a recommended object corresponding to the user account from the candidate objects based on the fusion feature.

[0087] In this way, the directional feature fusion is realized around the core preference of the first historical behavior of the user, and the fusion feature is generated, and then the recommended object corresponding to the user account is determined from the candidate objects based on the fusion feature. In the feature fusion process, the first historical behavior can preferentially select a behavior with the highest indication strength of user preference, and the comprehensive object feature of the first historical behavior reflects the core preference of the user embodied by the behavior; and then, based on the comprehensive object feature, a behavior object highly matched with the core preference corresponding to the first historical behavior is selected from the behavior objects corresponding to the historical behaviors other than the first historical behavior, and the comprehensive object feature is fused with the object feature of the selected behavior object to obtain the fusion feature. The fusion feature not only retains the core preference reflected by the first historical behavior, but also expands the dimension of the user's related interest, avoiding that the user preference represented by the fusion feature is too single. This way can fully play the data value of the plurality of positive feedback behaviors, effectively improve the matching degree of the recommended object and the real demand of the user, and improve the accuracy and personalization degree of the recommendation.

[0088] The first determining module is further configured to determine a plurality of first behavior objects corresponding to the first historical behavior, determine a first average feature corresponding to object features of the plurality of first behavior objects, and determine the first average feature as the comprehensive object feature.

[0089] The second determining module is further configured to determine a historical behavior other than the first historical behavior, acquire object features of a reference behavior object corresponding to the historical behavior other than the first historical behavior, determine a feature distance between the comprehensive object feature and the object features of the reference behavior object, and determine the feature distance between the comprehensive object feature and the object features of the reference behavior object as the correlation degree.

[0090] The second determining module is further configured to determine a target behavior object from the reference behavior object based on the feature distance between the comprehensive object feature and the object features of the reference behavior object, wherein the target behavior object has a specified number, and the feature distance between the object features of the target behavior object and the comprehensive object feature is not greater than the feature distance between the object features of a reference behavior object other than the target behavior object and the comprehensive object feature.

[0091] The target behavior object includes a plurality of target behavior objects, and the first fusion module is configured to calculate a second average feature corresponding to object features of the plurality of target behavior objects, and fuse the comprehensive object feature and the second average feature based on a preset weight parameter to obtain a fused feature.

[0092] The first fusion module is configured to acquire an object feature of a candidate object, wherein the object feature of the candidate object is consistent with the fused feature in feature dimension, splice the object feature of the candidate object and the fused feature to obtain a first joint feature for each candidate object, process the first joint feature by using a preset scoring function to obtain a recommendation score corresponding to the candidate object, and determine a recommended object corresponding to the user account from the candidate object based on the recommendation score.

[0093] The device further includes a second fusion module configured to acquire user multi-source features, wherein the user multi-source features include user portrait features and time features of the historical behavior, fuse the fused feature and the user multi-source features based on a preset fusion weight parameter to obtain multi-source fused features, and determine the recommended object from the candidate object based on the multi-source fused features.

[0094] The device further includes a second acquisition module configured to acquire user portrait information, input the user portrait information into a pre-trained portrait feature extraction model, and output user portrait features; wherein the number of feature dimensions of the user portrait features is consistent with the number of feature dimensions of the fusion features; acquire time data of the historical behaviors, input the time data into a pre-trained time feature extraction model, and output time features of the historical behaviors; wherein the number of feature dimensions of the time features of the historical behaviors is consistent with the number of feature dimensions of the fusion features.

[0095] The embodiment also provides an electronic device including a processor and a memory, the memory storing machine executable instructions capable of being executed by the processor, and the processor executing the machine executable instructions to implement the object recommendation method. The electronic device can be a server or a terminal device.

[0096] Referring to Figure 4 The electronic device includes a processor 100 and a memory 101, the memory 101 storing machine executable instructions capable of being executed by the processor 100, and the processor 100 executing the machine executable instructions to implement the object recommendation method.

[0097] Further, Figure 4 The electronic device shown further includes a bus 102 and a communication interface 103, and the processor 100, the communication interface 103 and the memory 101 are connected through the bus 102.

[0098] The memory 101 can include a high-speed random access memory (RAM) and can also include a non-volatile memory such as at least one disk memory. The communication between the system network element and at least one other network element is realized through at least one communication interface 103 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 102 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one bidirectional arrow is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0099] The processor 100 can be an integrated circuit chip with processing capability. In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor 100 or the instruction in the form of software. The processor 100 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block disclosed in the embodiment of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiment of the present application can be directly embodied as a hardware code processor to execute, or be executed by a combination of hardware and software modules in the code processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium in the art. The storage medium is located in the memory 101, and the processor 100 reads the information in the memory 101, and combines the hardware to complete the steps of the method of the above embodiment.

[0100] The processor in the above electronic device, by executing machine executable instructions, can realize the following operations of the above object recommendation method: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate: the object of the historical behavior; the historical behavior is a positive feedback behavior of the user account to the behavior object; determining a first historical behavior from the plurality of historical behaviors; determining the comprehensive object feature of the first historical behavior based on the object feature of the first behavior object corresponding to the first historical behavior; determining the correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior, and determining the target behavior object from the reference behavior object based on the correlation degree; fuse the comprehensive object feature and the object feature of the target behavior object to obtain the fusion feature; and determining the recommended object corresponding to the user account from the candidate object based on the fusion feature.

[0101] In the manner, the core preference of the first historical behavior of the user is used to realize the directional feature fusion and generate the fused feature, and then the recommendation object corresponding to the user account is determined from the candidate objects based on the fused feature. In the feature fusion process, the first historical behavior can be selected as the behavior with the highest preference indication strength, and the comprehensive object feature of the first historical behavior reflects the core preference of the user embodied by the behavior. Then, based on the comprehensive object feature, the behavior objects highly matching the core preference of the first historical behavior are selected from the behavior objects corresponding to the historical behaviors other than the first historical behavior, and the comprehensive object feature is fused with the object features of the selected behavior objects to obtain the fused feature. The fused feature not only retains the core preference reflected by the first historical behavior, but also expands the dimension of the user's related interests, avoiding that the user preference represented by the fused feature is too single. The manner can fully utilize the data value of multiple positive feedback behaviors, effectively improve the matching degree of the recommendation object and the real demand of the user, and improve the accuracy and personalization degree of the recommendation.

[0102] The processor in the electronic device can realize the following operations of the object recommendation method by executing the machine executable instructions: inputting the behavior object corresponding to each historical behavior into the pre-trained feature extraction model, and outputting the object feature of the behavior object.

[0103] The processor in the electronic device can realize the following operations of the object recommendation method by executing the machine executable instructions: determining a plurality of first behavior objects corresponding to the first historical behavior; determining a first average feature corresponding to the object features of the plurality of first behavior objects, and determining the first average feature as the comprehensive object feature.

[0104] The processor in the electronic device can realize the following operations of the object recommendation method by executing the machine executable instructions: determining a historical behavior other than the first historical behavior, obtaining the object feature of a reference behavior object corresponding to the historical behavior other than the first historical behavior; determining a feature distance between the comprehensive object feature and the object feature of the reference behavior object, and determining the feature distance between the comprehensive object feature and the object feature of the reference behavior object as the correlation degree.

[0105] The processor in the electronic device can realize the following operations of the object recommendation method by executing the machine executable instructions: determining a target behavior object from the reference behavior objects based on the feature distance between the comprehensive object feature and the object feature of the reference behavior object; wherein, the target behavior object has a specified number; and the feature distance between the object feature of the target behavior object and the comprehensive object feature is not greater than the feature distance between the object feature of the reference behavior object other than the target behavior object and the comprehensive object feature.

[0106] The aforementioned target behavior objects include multiple objects; the processor in the aforementioned electronic device can perform the following operations of the recommendation method for the aforementioned objects by executing machine-executable instructions: calculating the second average feature corresponding to the object features of the multiple target behavior objects; and performing fusion processing on the comprehensive object features and the second average feature based on preset weight parameters to obtain fused features.

[0107] The processor in the aforementioned electronic device, by executing machine-executable instructions, can perform the following operations of the recommendation method for the aforementioned objects: obtaining candidate object features; wherein the candidate object features and the fused features have the same feature dimension; for each candidate object, concatenating the candidate object's object features with the fused features to obtain a first joint feature; processing the first joint feature through a preset scoring function to obtain a recommendation score corresponding to the candidate object; and based on the recommendation score, determining the recommended object corresponding to the user account from the candidate objects.

[0108] The processor in the aforementioned electronic device can execute machine-executable instructions to perform the following operations of the recommendation method for the aforementioned object: acquiring multi-source user features; wherein, the multi-source user features include: user profile features and time features of historical behavior; based on preset fusion weight parameters, fusing the fusion features with the multi-source user features to obtain multi-source fusion features; and based on the multi-source fusion features, determining the recommended object from the candidate objects.

[0109] The processor in the aforementioned electronic device, by executing machine-executable instructions, can implement the following operations of the recommendation method for the aforementioned object: acquiring user profile information, inputting the user profile information into a pre-trained profile feature extraction model, and outputting user profile features; wherein the number of feature dimensions of the user profile features is consistent with the number of feature dimensions of the fused features; acquiring historical behavior time data, inputting the time data into a pre-trained time feature extraction model, and outputting historical behavior time features; wherein the number of feature dimensions of the historical behavior time features is consistent with the number of feature dimensions of the fused features.

[0110] This embodiment also provides a storage medium storing machine-executable instructions. When the machine-executable instructions are invoked and executed by a processor, the machine-executable instructions cause the processor to implement the recommended method of the above-mentioned object.

[0111] The machine-executable instructions stored in the storage medium can realize the following operations in the object recommendation method: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior object is used to indicate an object acted on by the historical behavior; the historical behavior is a positive feedback behavior of the user account to the behavior object; determining a first historical behavior from the plurality of historical behaviors; determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining a correlation degree between the comprehensive object feature and a reference behavior object corresponding to a historical behavior other than the first historical behavior, and determining a target behavior object from the reference behavior object based on the correlation degree; fusing the comprehensive object feature and an object feature of the target behavior object to obtain a fused feature; and determining a recommended object corresponding to the user account from the candidate objects based on the fused feature.

[0112] In this way, the core preference of the first historical behavior of the user is used to realize directional feature fusion and generate a fused feature, and then the recommended object corresponding to the user account is determined from the candidate objects based on the fused feature. In the feature fusion process, the first historical behavior can be used to select a behavior with the highest indication strength of user preference, and the comprehensive object feature of the first historical behavior reflects the core preference of the user embodied by the behavior. Then, based on the comprehensive object feature, behavior objects highly matching the core preference corresponding to the first historical behavior are selected from behavior objects corresponding to historical behaviors other than the first historical behavior, and the comprehensive object feature is fused with the object features of the selected behavior objects to obtain the fused feature. This way not only retains the core preference reflected by the first historical behavior, but also expands the dimension of user-related interests, avoiding the user preference represented by the fused feature being too single. This way can fully utilize the data value of multiple positive feedback behaviors, effectively improve the matching degree of the recommended object and the real needs of the user, and improve the accuracy and personalization of the recommendation.

[0113] The machine-executable instructions stored in the storage medium can realize the following operations in the object recommendation method: inputting the behavior object corresponding to the historical behavior into the pre-trained feature extraction model for each historical behavior, and outputting the object feature of the behavior object.

[0114] The machine-executable instructions stored in the storage medium can realize the following operations in the object recommendation method: determining a plurality of first behavior objects corresponding to the first historical behavior; determining a first average feature corresponding to the object features of the plurality of first behavior objects, and determining the first average feature as the comprehensive object feature.

[0115] The machine-executable instructions stored in the storage medium can implement the following operation in the object recommendation method: determining a historical behavior other than the first historical behavior, and obtaining an object feature of a reference behavior object corresponding to the historical behavior other than the first historical behavior; determining a feature distance between the integrated object feature and the object feature of the reference behavior object, and determining the feature distance between the integrated object feature and the object feature of the reference behavior object as the correlation degree.

[0116] The machine-executable instructions stored in the storage medium can implement the following operation in the object recommendation method: determining a target behavior object from the reference behavior object based on the feature distance between the integrated object feature and the object feature of the reference behavior object; the target behavior object has a specified number; and the feature distance between the object feature of the target behavior object and the integrated object feature is not greater than the feature distance between the object feature of a reference behavior object other than the target behavior object and the integrated object feature.

[0117] The target behavior object includes a plurality of target behavior objects, and the machine-executable instructions stored in the storage medium can implement the following operation in the object recommendation method: calculating a second average feature corresponding to the object features of the plurality of target behavior objects; and performing fusion processing on the integrated object feature and the second average feature based on a preset weight parameter to obtain a fusion feature.

[0118] The machine-executable instructions stored in the storage medium can implement the following operation in the object recommendation method: obtaining a candidate object feature of a candidate object; the candidate object feature has a same feature dimension as the fusion feature; splicing the object feature of the candidate object and the fusion feature to obtain a first joint feature for each candidate object; processing the first joint feature by using a preset scoring function to obtain a recommendation score corresponding to the candidate object; and determining a recommended object corresponding to the user account from the candidate objects based on the recommendation score.

[0119] The machine-executable instructions stored in the storage medium can implement the following operation in the object recommendation method: obtaining a user multi-source feature; the user multi-source feature includes a user portrait feature and a time feature of a historical behavior; fusing the fusion feature and the user multi-source feature based on a preset fusion weight parameter to obtain a multi-source fusion feature; and determining the recommended object from the candidate objects based on the multi-source fusion feature.

[0120] The machine-executable instructions stored in the storage medium can be executed to implement the following operations in the object recommendation method: obtaining user portrait information, inputting the user portrait information into a pre-trained portrait feature extraction model, and outputting user portrait features; wherein the number of feature dimensions of the user portrait features is consistent with the number of feature dimensions of the fusion features; obtaining time data of historical behaviors, inputting the time data into a pre-trained time feature extraction model, and outputting time features of the historical behaviors; wherein the number of feature dimensions of the time features of the historical behaviors is consistent with the number of feature dimensions of the fusion features.

[0121] The object recommendation method, the electronic device and the computer program product of the storage medium provided by the embodiments of the present application include a computer readable storage medium storing program codes, the instructions included in the program codes can be used to execute the method in the foregoing method embodiments, and specific implementation can be referred to the method embodiments, which will not be described here.

[0122] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0123] In addition, in the description of the embodiments of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection" and "connection" should be understood in a broad sense, for example, can be fixedly connected, can also be detachably connected, or integrally connected; can be mechanically connected, can also be electrically connected; can be directly connected, can also be indirectly connected through an intermediate medium, and can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0124] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the technical solutions that essentially contribute to the prior art can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0125] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0126] Finally, it should be noted that the above embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, and are not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent substitutions for some technical features; and these modifications, changes or substitutions do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A recommendation method of an object, characterized by, The method comprises: obtaining a plurality of historical behaviors of a user account and behavior objects corresponding to the historical behaviors; wherein the behavior objects are used to indicate objects acted on by the historical behaviors; the historical behaviors are positive feedback behaviors of the user account for the behavior objects; determining a first historical behavior from the plurality of historical behaviors, determining a comprehensive object feature of the first historical behavior based on an object feature of a first behavior object corresponding to the first historical behavior; determining a correlation degree between the comprehensive object feature and a reference behavior object corresponding to a historical behavior other than the first historical behavior, and determining a target behavior object from the reference behavior object based on the correlation degree; fusing the comprehensive object feature and an object feature of the target behavior object to obtain a fused feature, and determining a recommended object corresponding to the user account from candidate objects based on the fused feature.

2. The method of claim 1, wherein, The first behavior object corresponding to the first historical behavior comprises a plurality of first behavior objects; the step of determining the comprehensive object feature of the first historical behavior based on the object feature of the first behavior object corresponding to the first historical behavior comprises: determining a plurality of first behavior objects corresponding to the first historical behavior; determining a first average feature corresponding to the object features of the plurality of first behavior objects, and determining the first average feature as the comprehensive object feature.

3. The method of claim 1, wherein, The step of determining the correlation degree between the comprehensive object feature and the reference behavior object corresponding to the historical behavior other than the first historical behavior comprises: determining a historical behavior other than the first historical behavior, and obtaining an object feature of a reference behavior object corresponding to the historical behavior other than the first historical behavior; determining a feature distance between the comprehensive object feature and the object feature of the reference behavior object, and determining the feature distance between the comprehensive object feature and the object feature of the reference behavior object as the correlation degree.

4. The method of claim 3, wherein, The step of determining a target behavior object from the reference behavior object based on the correlation degree comprises: determining a target behavior object from the reference behavior object based on the feature distance between the comprehensive object feature and the object feature of the reference behavior object; wherein the target behavior object has a specified number; the feature distance between the object feature of the target behavior object and the comprehensive object feature is not greater than the feature distance between the object feature of a reference behavior object other than the target behavior object and the comprehensive object feature.

5. The method of claim 1, wherein, The target behavior object comprises a plurality of target behavior objects; the step of fusing the comprehensive object feature and the object feature of the target behavior object to obtain a fused feature comprises: calculating a second average feature corresponding to the object features of the plurality of target behavior objects; fusing the comprehensive object feature and the second average feature based on a preset weight parameter to obtain a fused feature.

6. The method of claim 1, wherein, The step of determining a recommended object corresponding to the user account from candidate objects based on the fused feature comprises: obtaining a candidate object feature of the candidate object; wherein the candidate object feature is consistent with the feature dimension of the fused feature; For each candidate object, the object feature of the candidate object is spliced with the fusion feature to obtain a first joint feature; the first joint feature is processed by a preset scoring function to obtain a recommendation score corresponding to the candidate object; Based on the recommendation score, a recommended object corresponding to the user account is determined from the candidate objects.

7. The method of claim 1, wherein, After the step of obtaining the fusion feature, the method further comprises: Obtaining user multi-source features; wherein the user multi-source features include user portrait features and time features of the historical behaviors; Based on a preset fusion weight parameter, the fusion feature is fused with the user multi-source features to obtain multi-source fusion features; Based on the multi-source fusion features, a recommended object is determined from the candidate objects.

8. The method of claim 7, wherein, The step of obtaining the user multi-source features comprises: Obtaining user portrait information, inputting the user portrait information into a pre-trained portrait feature extraction model, and outputting the user portrait features; wherein the number of feature dimensions of the user portrait features is consistent with the number of feature dimensions of the fusion features; Obtaining time data of the historical behaviors, inputting the time data into a pre-trained time feature extraction model, and outputting the time features of the historical behaviors; wherein the number of feature dimensions of the time features of the historical behaviors is consistent with the number of feature dimensions of the fusion features.

9. An electronic device, comprising: A processor and a memory are included, the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the object recommendation method of any one of claims 1-8.

10. A storage medium, characterized by The storage medium stores machine executable instructions, and when the machine executable instructions are called and executed by the processor, the machine executable instructions cause the processor to implement the object recommendation method of any one of claims 1-8.

Images