Recommendation processing method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention discloses a recommendation processing method and device, equipment, a storage medium and a program product, at least relates to technologies such as artificial intelligence and the like, and not only improves content recommendation experience, but also obtains better advertisement revenue. The method comprises the following steps: obtaining a content fine arrangement result comprising a plurality of candidate contents sorted according to a target sequence, wherein the content fine arrangement result comprises at least one advertisement position and an advertisement fine arrangement result of at least one candidate advertisement; based on the plurality of candidate contents and the at least one advertisement space, generating M candidate sequences with different advertisement space ranks; determining a target advertisement placed on each advertisement position in each candidate sequence, and generating M advertisement rearrangement sequences based on a plurality of candidate contents in each candidate sequence and the target advertisement placed on each advertisement position; calculating a sequence score of each advertisement rearrangement sequence, and determining a target sequence from the M advertisement rearrangement sequences based on all the sequence scores; and recommending candidate contents and candidate advertisements in the target sequence according to the target sequence.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technologies, and particularly to a method, device, equipment, storage medium, and program product for recommendation processing. Background Art

[0002] With the rapid development of the Internet, contents such as videos, texts, and images, and even advertisements have become the main channels for people to obtain information and entertainment. In the scenario of recommended contents, a server usually composes a candidate sequence of multiple contents and advertisements in a certain order and sends it to a terminal device, so that a user can also browse advertisements interspersed during the process of browsing contents.

[0003] In the traditional method of recommendation processing, a content sorting model is usually used to determine a content list, and an advertisement sorting model is used to estimate advertisement features of advertisements to determine an advertisement list. After determining the content list and the advertisement list, a mixing process is performed on the content list and the advertisement list to determine all possible candidate sequences. Then, advertisement features in the advertisement list and content features in the content list are used to score all the candidate sequences, so as to determine a target sequence through the scores of the candidate sequences. In this way, the corresponding contents and advertisements are recommended to a user according to the target sequence.

[0004] However, in the current recommended manner, it is only determined at the mixing process stage which advertisement position each advertisement will finally be placed in. This will result in inaccurate advertisement features calculated before the candidate sequences are obtained by mixing, and further result in that the target sequence determined by using inaccurate advertisement features in the subsequent mixing process stage is not an optimal value. Therefore, when presenting contents and advertisements to a user according to the currently determined target sequence, it can neither bring a good content recommendation experience to the user nor obtain an optimal advertisement revenue. Summary of the Invention

[0005] The embodiments of the present application provide a method, device, equipment, storage medium, and program product for recommendation processing, which fully consider the influence of content context and the specific position of advertisement positions on candidate advertisements placed in different advertisement positions, can accurately estimate the target advertisement placed in each advertisement position, select an optimal target sequence, not only improve the content recommendation experience, but also obtain an optimal advertisement revenue, and is conducive to achieving a balance between content recommendation and advertisement promotion.

[0006] In a first aspect, an embodiment of the present application provides a method for recommendation processing. The method includes: obtaining a content fine ranking result and an advertisement fine ranking result, where the content fine ranking result includes multiple candidate contents sorted in a target order, and the advertisement fine ranking result includes at least one advertisement slot and at least one candidate advertisement; generating M candidate sequences with different advertisement slot rankings based on the multiple candidate contents and the at least one advertisement slot, where the arrangement order of the multiple candidate contents in each candidate sequence is the same as the target order, and each advertisement slot in each candidate sequence is used to pre-place multiple candidate advertisements, and M is an integer greater than or equal to 2; determining the target advertisement placed on each advertisement slot in each candidate sequence, and generating M advertisement re-ranking sequences based on the multiple candidate contents in each candidate sequence and the target advertisement placed on each advertisement slot, where the target advertisement is one of the multiple candidate advertisements; calculating the sequence score of each advertisement re-ranking sequence, and determining the target sequence from the M advertisement re-ranking sequences based on all the sequence scores; recommending the candidate contents and candidate advertisements in the target sequence according to the target sequence.

[0007] In a second aspect, an embodiment of the present application provides a recommendation processing device. The recommendation processing device includes an acquisition unit and a processing unit. Among them, the acquisition unit is used to obtain a content fine ranking result and an advertisement fine ranking result, where the content fine ranking result includes multiple candidate contents sorted in a target order, and the advertisement fine ranking result includes at least one advertisement slot and at least one candidate advertisement. The processing unit is used to generate M candidate sequences with different advertisement slot rankings based on the multiple candidate contents and the at least one advertisement slot, where the arrangement order of the multiple candidate contents in each candidate sequence is the same as the target order, and each advertisement slot in each candidate sequence is used to pre-place multiple candidate advertisements, and M is an integer greater than or equal to 2. The processing unit is used to determine the target advertisement placed on each advertisement slot in each candidate sequence, and generate M advertisement re-ranking sequences based on the multiple candidate contents in each candidate sequence and the target advertisement placed on each advertisement slot, where the target advertisement is one of the multiple candidate advertisements. The processing unit is used to calculate the sequence score of each advertisement re-ranking sequence, and determine the target sequence from the M advertisement re-ranking sequences based on all the sequence scores. The processing unit is used to recommend the candidate contents and candidate advertisements in the target sequence according to the target sequence.

[0008] In a possible design, in another implementation of the second aspect of the embodiments of the present application, the processing unit is configured to: for a first sequence, extract the content context features of the first sequence, where the first sequence is any one of M candidate sequences, and the content context features of the first sequence include the ad slot features of the ad slots in the first sequence and the content features of each candidate content; for the first ad slot in the first sequence, extract the ad features of each first candidate ad, where each first candidate ad is one of the multiple candidate ads pre-placed in the first ad slot, and the first ad slot is any one of at least one ad slot; based on the content context features of the candidate content and the ad features of each first candidate ad, calculate the cost per mille (CPM) revenue value of the corresponding first candidate ad; based on the CPM revenue values of all the first candidate ads, determine the target ad to be placed in the first ad slot from the multiple candidate ads pre-placed in the first ad slot.

[0009] In a possible design, in another implementation of the second aspect of the embodiments of the present application, the processing unit is configured to: perform feature fusion processing on the content context features of the candidate content and the ad features of a second candidate ad to obtain a first fusion feature, where the second candidate ad is any one of the first candidate ads; process the first fusion feature based on a preset expert network model to obtain the estimated conversion rate and the estimated click-through rate of the second candidate ad; calculate the product of the estimated conversion rate of the second candidate ad, the estimated click-through rate of the second candidate ad, and the ad bid of the second candidate ad to obtain the CPM revenue value of the second candidate ad.

[0010] In a possible design, in another implementation of the second aspect of the embodiments of the present application, the processing unit is configured to: determine the maximum CPM revenue value from the CPM revenue values of all the first candidate ads; determine the candidate ad corresponding to the maximum CPM revenue value as the target ad to be placed in the first ad slot.

[0011] In a possible design, in another implementation of the second aspect of the embodiments of the present application, the processing unit is configured to: for a second sequence, extract the content context features of the candidate content in the second sequence, where the second sequence is any one of M ad rearrangement sequences; for each ad slot in the second sequence, extract the ad features of the target ads placed in each ad slot; based on the content context features of the candidate content and the ad features of all the target ads in the second sequence, determine the sequence score of the second sequence.

[0012] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the processing unit is configured to: perform feature fusion processing on the content context features of the candidate content and the advertisement features of all target advertisements in the second sequence to obtain second fusion features; perform feature processing on the second fusion features based on a preset Gate network model to obtain promotion information of the candidate content, and perform feature processing on the second fusion features based on a preset expert network model to obtain the total advertisement revenue, where the promotion information is used to indicate the expected display situation when the candidate content is recommended for display, and the total advertisement revenue is used to indicate the expected revenue situation of all target advertisements in the second sequence; determine the sequence score of the second sequence based on the promotion information and the total advertisement revenue.

[0013] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the obtaining unit is further configured to: before determining the sequence score of the second sequence based on the promotion information and the total advertisement revenue, obtain a first weight and a second weight, where the first weight is used to indicate the content weight of the candidate content, and the second weight is used to indicate the weight of the revenue. The processing unit is configured to: weight the promotion information based on the first weight and weight the total advertisement revenue based on the second weight to obtain the weighted promotion information and the weighted total advertisement revenue respectively; sum the weighted promotion information and the weighted total advertisement revenue to obtain the sequence score of the second sequence.

[0014] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the processing unit is configured to: determine the maximum sequence score from all the sequence scores; determine the advertisement rearrangement sequence corresponding to the maximum sequence score as the target sequence.

[0015] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the processing unit is configured to: perform permutation and combination processing on all candidate contents and all advertisement positions to obtain N initial mixed sequences, where N is greater than or equal to M and N is an integer; select the initial mixed sequences that meet the preset sorting rule from the N initial mixed sequences to generate M candidate sequences with different advertisement position sortings.

[0016] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the preset sorting rule includes that there are at least T candidate contents between every two adjacent advertisement positions, where T is an integer greater than or equal to 1; the processing unit is configured to: calculate the first interval number in each initial mixed sequence, where the first interval number is used to indicate the number of candidate contents between every two adjacent advertisement positions in the corresponding initial mixed sequence; when the first interval number is greater than or equal to T, determine the initial mixed sequence corresponding to the first interval number as the candidate sequence.

[0017] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the obtaining unit is further configured to: before obtaining the content refined ranking result and the advertisement refined ranking result, obtain content browsing data and advertisement browsing data, where the content browsing data includes a plurality of content entries, and the advertisement browsing data includes a plurality of advertisement entries. The processing unit is configured to perform refined ranking processing on the plurality of content entries based on a content ranking model to obtain a content refined ranking result, and perform refined ranking processing on the plurality of advertisement entries based on an advertisement ranking model to obtain an advertisement refined ranking result, where the content entries include candidate content, and the advertisement entries include candidate advertisements.

[0018] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the processing unit is configured to: respectively extract the entry features of each content entry based on the content ranking model, and perform a scoring process on the entry features of each content entry to obtain the entry score corresponding to the content entry; sort the entry scores of all content entries based on the content ranking model to obtain a content entry ranking result; based on the content entries corresponding to the entry scores greater than a preset threshold in the content entry ranking result, obtain a content refined ranking result.

[0019] In a possible design, in another implementation manner of the second aspect of the embodiments of the present application, the processing unit is configured to: respectively extract the entry features of each advertisement entry based on the advertisement ranking model, and perform a scoring process on the entry features of each advertisement entry to obtain the entry score corresponding to the advertisement entry; sort the entry scores of all advertisement entries based on the advertisement ranking model to obtain an advertisement entry ranking result; based on the advertisement entries corresponding to the entry scores greater than a preset threshold in the advertisement entry ranking result and the corresponding advertisement positions, obtain an advertisement refined ranking result.

[0020] The third aspect of the embodiments of the present application provides a recommendation processing device, including: a memory, an input / output (I / O) interface, and a memory. The memory is used to store program instructions. The processor is configured to execute the program instructions in the memory to execute the recommendation processing method corresponding to the implementation manner of the first aspect above.

[0021] The fourth aspect of the embodiments of the present application provides a computer-readable storage medium, in which instructions are stored, and when the instructions run on a computer, the computer is caused to execute the method corresponding to the implementation manner of the first aspect above.

[0022] The fifth aspect of the embodiments of the present application provides a computer program product including instructions, and when the computer program product runs on a computer or a processor, the computer or the processor is caused to execute the method corresponding to the implementation manner of the first aspect above.

[0023] From the above technical solutions, it can be seen that the embodiments of the present application have the following advantages:

[0024] In the embodiments of the present application, since the content refined ranking result includes multiple candidate contents sorted in the target order, and the advertisement refined ranking result includes at least one advertisement space and at least one candidate advertisement. Then, after obtaining the content refined ranking result and the advertisement refined ranking result, based on the multiple candidate contents and at least one advertisement space, M candidate sequences with different advertisement space rankings are generated. Moreover, the arrangement order of the multiple candidate contents in each candidate sequence is the same as the target order, and each advertisement space in each candidate sequence is used to pre-place multiple candidate advertisements, where M is an integer greater than or equal to 2. By determining the target advertisement placed on each advertisement space in each candidate sequence, and generating M advertisement re-ranking sequences based on the multiple candidate contents in each candidate sequence and the target advertisement placed on each advertisement space. The described target advertisement is one of the multiple candidate advertisements. In this way, then calculate the sequence score of each advertisement re-ranking sequence, and determine the target sequence from the M advertisement re-ranking sequences based on all the sequence scores, and then recommend the candidate contents and candidate advertisements in the target sequence according to the target sequence. Through the above method, in the process of generating candidate sequences in the embodiments of the present application, the advertisement space and multiple candidate contents that meet the target order are taken into consideration, realizing the pre-fixing of the advertisement space in each candidate sequence and the candidate contents at each content position, and fully considering the influence of the specific position of the advertisement space and the content context on the candidate advertisements placed on different advertisement spaces. In this way, by re-estimating the pre-placement of different candidate advertisements for each advertisement space with a fixed position, it is ensured that the advertisement features of the target advertisement placed in each advertisement space are accurate, and thus the target sequence selected from the M different advertisement re-ranking sequences is also an optimal value, which not only brings a good content recommendation experience to the object, but also can obtain better advertisement revenue, and is conducive to achieving the balance between content recommendation and advertisement promotion. Description of the Drawings

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0026] Figure 1 Shows a schematic diagram of an application scenario for displaying advertisements and content;

[0027] Figure 2 Shows a schematic diagram of the recommendation process in the traditional solution;

[0028] Figure 3 Shows a schematic diagram of the mixed ranking processing framework in the traditional solution;

[0029] Figure 4Shows a schematic diagram of the system architecture provided by the embodiments of the present application;

[0030] Figure 5 Shows a flowchart of a method for recommendation processing provided by the embodiments of the present application;

[0031] Figure 6 Shows a schematic diagram of M candidate sequences provided by the embodiments of the present application;

[0032] Figure 7 Shows a schematic diagram of the structure for determining M advertisement rearrangement sequences provided by the embodiments of the present application;

[0033] Figure 8 Shows a schematic flowchart of sequence optimization provided by the embodiments of the present application;

[0034] Figure 9 Shows a schematic diagram of optional functional modules of a recommendation processing device provided by the embodiments of the present application;

[0035] Figure 10 Shows a schematic diagram of the optional hardware structure of a recommendation processing device provided by the embodiments of the present application. Detailed implementation manners

[0036] The embodiments of the present application provide a method, device, equipment, storage medium and program product for recommendation processing, which fully consider the influence of the content context and the specific position of the advertisement space on the candidate advertisements placed in different advertisement spaces, can accurately estimate the target advertisement placed in each advertisement space, select a better target sequence, not only improve the content recommendation experience, but also obtain better advertisement revenue, and is conducive to achieving the balance between content recommendation and advertisement promotion.

[0037] It can be understood that in the specific implementation manners of the present application, data related to user information and the like are involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0038] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0039] In the description, claims and the above-mentioned drawings of the present application, terms such as "first", "second", "third", "fourth", etc. (if any) are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0040] When an object requests access to different contents from a recommendation processing device through a terminal device, it may browse the advertisements interspersed in the contents. This is mainly because the recommendation processing device mixes multiple contents and advertisements in a certain order to generate a corresponding candidate sequence and sends the candidate sequence to the terminal device. After receiving the candidate sequence, the terminal device displays the contents and advertisements in order. In this way, through the terminal device, the object can not only view the contents but also view the advertisements.

[0041] For example, Figure 1 shows a schematic diagram of an application scenario for displaying advertisements and contents. As Figure 1 shown, this application scenario at least includes a recommendation processing device and a terminal device. Among them, the recommendation processing device obtains multiple candidate contents from content behavior data, such as candidate content 1 to candidate content X (X is an integer greater than or equal to 2), etc. And the recommendation processing device sorts these multiple candidate contents in a certain order, such as sorting from candidate content 1 to candidate content X, to form a content list. Similarly, the recommendation processing device obtains at least one candidate advertisement from advertisement behavior data, such as candidate advertisement 1 to candidate advertisement Y (Y is an integer greater than or equal to 1), etc. And the recommendation processing device sorts these multiple candidate advertisements in a certain order, such as sorting from candidate advertisement 1 to candidate advertisement Y, to form an advertisement list.

[0042] After generating the content list and the advertisement list, the recommendation processing device also mixes and arranges the content list and the advertisement list to obtain candidate sequences of different batches under the same session, such as candidate sequence 1, candidate sequence 2, and so on. Among them, for candidate sequence 1, the sorting of content and advertisement in the sequence is: candidate content 1 -> candidate advertisement 1 -> candidate content 2 -> candidate content 3 -> candidate content 4 -> candidate advertisement 2 -> candidate content 5, etc. For candidate sequence 2, the sorting of content and advertisement in the sequence is: candidate content 1 -> candidate advertisement 1 -> candidate content 2 -> candidate advertisement 2 -> candidate content 3 -> candidate content 4 -> candidate advertisement 3, etc. Further, the recommendation processing device sends different candidate sequences, such as candidate sequence 1 or candidate sequence 2, to the terminal device. The user continuously swipes down the content and advertisement in the candidate sequence through the terminal device until the entire candidate sequence is browsed.

[0043] For example, taking candidate sequence 1 as an example, its candidate content 1 is the video content of "The Ice and Snow World opens on January 1st", candidate advertisement 1 is the shopping link of "Spot Removing, Anti-Wrinkle and Whitening Cream", candidate content 2 is the video content of "Sports goods in sports events are deeply loved by the audience", and so on. At this time, through the terminal device, the user can, after viewing the video content of "The Ice and Snow World opens on January 1st", also click on the shopping link such as "https: / / s.click.tao.com / Spot Removing, Anti-Wrinkle and Whitening Cream" to view the specific content of "Spot Removing, Anti-Wrinkle and Whitening Cream". Subsequently, the user can also click to view the video content of "Sports goods in sports events are deeply loved by the audience". In this way, after the user browses the content and advertisement in candidate sequence 1, at this time, the recommendation processing device can calculate the total browsing duration of the user for the candidate content, etc., to reflect the user's browsing experience. And the recommendation processing device will also calculate the advertisement billing for the browsed candidate advertisement, etc., to calculate the total advertisement revenue for the candidate advertisement.

[0044] In the process of providing candidate sequences to the user, both the user's browsing experience for the content and the advertisement revenue need to be considered. Therefore, how to balance the user's browsing experience and the advertisement revenue requires ensuring that the advertisements and content in the candidate sequence are placed in reasonable positions.

[0045] In the traditional recommendation processing solution, usually, the candidate sequences are scored to determine the target sequence through the score. Exemplarily, Figure 2 shows the schematic diagram of the recommendation process in the traditional solution. As Figure 2As shown in the figure, the content behavior data of the object is used to extract content features such as the playing duration, likes, follows, and shares of each candidate content by means of a content sorting model, and the content features are scored, so as to determine a content list based on the content scores. For example, the content list includes candidate content A1 to candidate content A5. It should be noted that the candidate content A1 to candidate content A5 mentioned here can also be understood with reference to candidate content 1 to candidate content X in the foregoing Figure 1 , and the present application does not make specific limitations on the names of the candidate content.

[0046] Similarly, the advertisement features of the candidate advertisements are extracted by means of an advertisement sorting model, such as, but not limited to, one or more of the effective cost per mille (eCPM) value, advertisement playing features, and advertisement platform features. Among them, the advertisement playing features may include, but are not limited to, the fast-swipe rate, the completion rate, negative feedback, etc. The advertisement platform features may include, but are not limited to, the stay duration, the number of VVs, the session exit rate, etc., and the present application does not make limitations. Subsequently, the advertisement features are scored, so as to determine an advertisement list based on the advertisement scores. For example, the advertisement list includes candidate advertisement 1 and candidate advertisement 2. After determining the content list and the advertisement list, the content list and the advertisement list are mixed by means of a mixed sorting model to determine all possible candidate sequences, and a target sequence is determined from the candidate sequences. For example, the sorting between the content and the advertisement in the target sequence is: candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate advertisement 1 --> candidate content A4 --> candidate content A5 -> candidate advertisement 2.

[0047] Regarding how to determine the target sequence from the candidate sequences in the foregoing Figure 2 , the process can be understood with reference to the existing mixed sorting processing framework shown in Figure 3 . As Figure 3 shown, after the above Figure 2All possible candidate sequences determined by the mixed layout model include candidate sequence 1, candidate sequence 2, candidate sequence 3, candidate sequence 4, and candidate sequence 5, etc. Among them, for candidate sequence 1, the sorting between the candidate content and the candidate advertisement is: candidate content A1 --> candidate advertisement 1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> candidate advertisement 2 --> candidate content A5. For candidate sequence 2, the sorting between the candidate content and the candidate advertisement is: candidate content A1 --> candidate content A2 --> candidate advertisement 1 --> candidate content A3 --> candidate content A4 --> candidate content A5 --> candidate advertisement 2. For candidate sequence 3, the sorting between the candidate content and the candidate advertisement is: candidate content A1 --> candidate content A2 --> candidate advertisement 1 --> candidate content A3 --> candidate content A4 --> candidate content A5 --> candidate advertisement 2. For candidate sequence 4, the sorting between the candidate content and the candidate advertisement is: candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate advertisement 1 --> candidate content A4 --> candidate advertisement 2 --> candidate content A5. For candidate sequence 5, the sorting between the candidate content and the candidate advertisement is: candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate advertisement 1 --> candidate content A4 --> candidate advertisement 2 --> candidate content A5.

[0048] Thus, after generating the above-mentioned candidate sequences 1 to 5, corresponding content features are extracted for each content in each candidate sequence, as well as advertisement features such as the conversion rate and advertisement revenue of each advertisement. Then, the content features and advertisement features in each candidate sequence are used to score the corresponding candidate sequence, so as to determine the target sequence including content and advertisement through the scores of all candidate sequences. For example, the target sequence is candidate sequence 5. Then, according to this target sequence (i.e., Figure 3 the candidate sequence 5 shown), the corresponding content and advertisement are recommended to the target object.

[0049] However, in the process of determining the target sequence by using the above Figure 2 and Figure 3 methods, for each advertisement, which advertising position it will finally be placed in can only be determined in the mixed layout processing stage. This will result in the advertisement features calculated before the mixed layout to obtain the candidate sequence being inaccurate. As a result, the target sequence determined by using inaccurate advertisement features in the subsequent mixed layout processing stage is not an optimal value either. Therefore, when presenting content and advertisements to the object according to the currently determined target sequence, it can neither bring a good content recommendation experience to the object nor obtain an optimal advertisement revenue.

[0050] Therefore, to solve the above-mentioned technical problems, the embodiments of the present application provide a method for recommendation processing. Exemplarily, the method for recommendation processing provided by the present application can be applied to Figure 4 the system architecture shown in Figure 4 As shown, the system architecture at least includes a content recommendation engine, a mixed layout service module, and an advertisement re-ranking module. Optionally, the system architecture may further include a Mixer module, a model_server module, a datahub module, etc.

[0051] Among them, content browsing data and advertisement browsing data are first obtained from applications (APPs), etc., and then the corresponding content fine-ranking results and advertisement fine-ranking results are determined. Among the content fine-ranking results, there are multiple candidate contents sorted in the target order. For example, candidate content A1 to candidate content A5 sorted in sequence, etc. In the advertisement fine-ranking results, there are at least one advertisement slot and at least one candidate advertisement, such as advertisement slot 1 and advertisement slot 2, candidate advertisement a1, candidate advertisement a2, candidate advertisement b1, and candidate advertisement b2, etc.

[0052] In this way, M candidate sequences with different advertisement slot rankings are generated based on these multiple candidate contents and each advertisement slot, such as candidate sequence 1 to candidate sequence M, where M is an integer greater than or equal to 2. For example, the mixed layout service module is used to generate these M candidate sequences, and the specific generation process can be understood with reference to the content shown in Figure 6 below, and will not be elaborated here. It should be noted that for each candidate sequence, the arrangement order of the multiple candidate contents included in each candidate sequence is the same as the target order. For example, taking candidate sequence 1 as an example, even though advertisement slot 1 and advertisement slot 2 have been mixed with candidate content A1 to candidate content A5, the arrangement order of candidate content A1 to candidate content A5 in this candidate sequence 1 has not changed, and candidate content A1 to candidate content A5 are still arranged and displayed in sequence. In addition, for each advertisement slot in the present application, it can be used to pre-place multiple candidate advertisements. For example, taking advertisement slot 1 as an example, candidate advertisement a1, candidate advertisement a2, candidate advertisement b1, and candidate advertisement b2, etc. can be pre-placed, and will not be elaborated here.

[0053] After generating the M candidate sequences, in the embodiments of the present application, it is not necessary to directly score each candidate sequence as in the previous Figure 3 Instead, first use Figure 4The newly added advertisement rearrangement module completes the rearrangement process of the advertisements pre-placed in each advertising space among the M candidate sequences. Exemplarily, through the advertisement rearrangement module, it is possible to re-estimate the revenue value per thousand impressions of the candidate advertisements expected to be placed in each advertising space in each candidate sequence, etc., and then determine the target advertisement to be placed on each advertising space. Specifically, regarding how to determine the target advertisements placed on each advertising space, the calculation process can be understood with reference to the content described later Figure 7 and will not be elaborated here for now. Optionally, the advertisement rearrangement module can also be used to complete the billing process for each target advertisement.

[0054] In this way, based on the multiple candidate contents in each of the candidate sequences and the target advertisements placed on each advertising space, M advertisement rearrangement sequences are generated. For example, advertisement rearrangement sequence 1 to advertisement rearrangement sequence M. It should be noted that in each advertisement rearrangement sequence, it includes multiple candidate contents and the target advertisements on each advertising space. For example, taking advertisement rearrangement sequence 1 as an example, if it is determined that the target advertisement placed in advertising space 1 is candidate advertisement a1 and the target advertisement placed in advertising space 2 is candidate b2, then the sorting between the contents and advertisements in this advertisement rearrangement sequence 1 is: candidate advertisement a1 --> candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate advertisement b2 --> candidate content A4 --> candidate content A5.

[0055] After generating the M advertisement rearrangement sequences, calculate the sequence score of each advertisement rearrangement sequence, and determine the target sequence from the M advertisement rearrangement sequences based on all the sequence scores, such as advertisement rearrangement sequence 1. In this way, recommend the candidate contents and candidate advertisements according to the target sequence. For example, the target sequence can be fed back to Figure 4 the content recommendation engine shown to enable the content recommendation engine to recommend candidate contents and candidate advertisements according to this target sequence.

[0056] It can be seen that in the process of generating candidate sequences in the embodiments of the present application, full consideration is given to the advertising spaces and multiple candidate contents that meet the target order. On the basis of not affecting the recommendation order of candidate contents, the specific arrangement positions of the advertising spaces and the influence of the content context on the candidate advertisements placed on different advertising spaces are comprehensively considered. In this way, by re-estimating different candidate advertisements pre-placed in each advertising space with a fixed position, it is ensured that the advertisement characteristics, etc., of the target advertisements placed in each advertising space are accurate, and then the target sequence selected from the M different advertisement rearrangement sequences is also an optimal value, which not only brings a good content recommendation experience to the object, but also can obtain better advertising revenue, maximizing the balance between content recommendation and advertising promotion.

[0057] Optionally, the above Figure 4The shown datahub module can provide unified data access for the recommendation system, such as features, content browsing data, advertisement browsing data, object information, etc., and provide a unified object feature extraction interface to ensure the logical consistency of feature extraction both online and offline. For example, through the datahub module, corresponding object features can be extracted from the feature warehouse, etc.

[0058] The mentioned Mixer module can be understood as the central module of the recommendation system. Through the Mixer module, object information can be obtained according to, for example, content requests or advertisement requests. Or, through the Mixer module, fine ranking can be performed on advertisement entries, content entries, etc.; it can also recall advertisements and content and perform rough ranking, etc. In addition, through the Mixer module, advertisement information can be supplemented and recommendation request logs can be generated, such as exposure / click log tables, etc., to provide data support for the subsequent mixed service module to process the mixed ranking between candidate advertisements and candidate content, etc.

[0059] In addition, the above-mentioned mixed ranking service module can also generate a mixed ranking log table during the process of generating M candidate sequences, so as to provide data support for the subsequent advertisement re-ranking module to re-estimate the eCPM of candidate advertisements on different advertisement positions, etc.

[0060] Exemplarily, the above-mentioned recommendation processing method can also be applied in the field of artificial intelligence (AI). The recommendation processing method provided in the embodiments of this application can be implemented based on artificial intelligence. Artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence is also about researching the design principles and implementation methods of various intelligent machines to enable the machine to have the functions of perception, reasoning, and decision-making. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech technology, natural language processing technology, and machine learning / deep learning. In the embodiments of this application, the artificial intelligence technologies mainly involved include artificial neural networks, etc. in the above-mentioned machine learning (ML). In this application, content features, advertisement position features, advertisement features, etc. are extracted through a neural network model.

[0061] The recommendation processing method provided by this application can be applied to recommendation processing devices with data processing capabilities, such as servers. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (CDN), and big data and artificial intelligence platforms. This application does not make specific limitations.

[0062] In addition, the terminal devices mentioned in this application may include, but are not limited to, smartphones, desktop computers, laptop computers, tablet computers, smart speakers, in-vehicle devices, smart watches, wearable smart devices, intelligent voice interaction devices, smart home appliances, aircraft, etc.

[0063] It should be noted that the candidate content mentioned in the embodiments of this application may include, but is not limited to, short video content, long video content, text content, picture content, information content, commodities, products, items, etc. in different scenarios. This application does not make limitations. The candidate advertisements mentioned may also include, but are not limited to, video advertisements, in-feed advertisements, commodity advertisements, text link advertisements, self-media advertisements, public welfare advertisements, promotional advertisements, etc. This application does not make limitations.

[0064] Exemplarily, the embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc.

[0065] Combined with the above Figure 4 shown system architecture, the following introduces a recommendation processing method provided by the embodiments of this application with reference to the accompanying drawings. Figure 5 shows a flowchart of a recommendation processing method provided by the embodiments of this application. As Figure 5 shown, the recommendation processing method may include the following steps:

[0066] 501. Obtain the content fine ranking result and the advertisement fine ranking result. The content fine ranking result includes multiple candidate contents sorted in the target order, and the advertisement fine ranking result includes at least one advertisement position and at least one candidate advertisement.

[0067] In this example, if the target object wants to browse different contents, such as food videos, entertainment videos, or travel videos, etc., it can send a recommendation request to the recommendation processing device through the terminal device. After receiving the recommendation request, the recommendation processing device obtains the content fine ranking result and the advertisement fine ranking result. Or, the recommendation processing device can also directly obtain the content fine ranking result and the advertisement fine ranking result without receiving the recommendation request.

[0068] In the content fine-ranking result, there are multiple candidate contents sorted in the target order. For example, the content fine-ranking result may include candidate content A1 to candidate content A5 sorted in the target order. Specifically, the sorting of the candidate contents in the content fine-ranking result is: candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> candidate content A5.

[0069] In addition, in the advertisement fine-ranking result, there is at least one advertisement position and at least one candidate advertisement. For example, the advertisement fine-ranking result includes advertisement position 1, advertisement position 2, and also includes candidate advertisement a1, candidate advertisement a2, candidate advertisement b1, candidate advertisement b2, etc., which are not limited here.

[0070] It should be noted that the above-mentioned content fine-ranking result including candidate content A1 to candidate content A5, and the advertisement fine-ranking result including advertisement position 1, advertisement position 2, candidate advertisement a1, candidate advertisement a2, candidate advertisement b1, and candidate advertisement b2 are just an example. In actual applications, the content fine-ranking result may also include candidate contents sorted in other orders, and the advertisement fine-ranking result may also include other numbers of advertisement positions, candidate advertisements, etc., which are not limited in this application.

[0071] Exemplarily, for the process of how to obtain the content fine-ranking result and the advertisement fine-ranking result, it can be specifically understood in the following way, that is:

[0072] The recommendation processing device first obtains content browsing data and advertisement browsing data. The described content browsing data includes multiple content items, and the advertisement browsing data includes multiple advertisement items.

[0073] After obtaining the content browsing data, the recommendation processing device then performs fine-ranking processing on the multiple content items based on the content ranking model to obtain the content fine-ranking result. As a schematic description, the recommendation processing device extracts the item features of each content item based on the content ranking model and performs a scoring process on the item features of each content item to obtain the item score corresponding to the content item. After calculating the item scores of all content items, the recommendation processing device sorts the item scores of all content items based on the content ranking model to obtain the content item ranking result. Finally, the recommendation processing device obtains the content fine-ranking result based on the content items corresponding to the item scores greater than the preset threshold in the content item ranking result. More specifically, if the content item ranking result is obtained in the descending order of item scores, at this time, the recommendation processing device can also use the first P content items in the content item ranking result as the content fine-ranking result, where P is an integer greater than or equal to 2.

[0074] For example, there are multiple content items including Content Item 1 to Content Item 10. At this time, the respective item scores calculated for Content Item 1 to Content Item 10 are: 90 points, 88 points, 86 points, 89 points, 87 points, 83 points, 79 points, 60 points, 65 points, and 75 points. At this time, sorting in descending order according to the item scores, the content sorting result can be obtained as: Content Item 1 --> Content Item 4 --> Content Item 2 --> Content Item 5 --> Content Item 3 --> Content Item 6 --> Content Item 7 --> Content Item 10 --> Content Item 9 --> Content Item 8. If the preset threshold is 85 points, at this time, the refined content sorting result can be extracted as: Content Item 1 --> Content Item 4 --> Content Item 2 --> Content Item 5 --> Content Item 3. Or rather, the multiple candidate contents sorted in the target order in the refined content sorting result are: Content Item 1 (i.e., the aforementioned candidate content A1) --> Content Item 4 (i.e., the aforementioned candidate content A2) --> Content Item 2 (i.e., the aforementioned candidate content A3) --> Content Item 5 (i.e., the aforementioned candidate content A4) --> Content Item 3 (i.e., the aforementioned candidate content A5).

[0075] Similarly, after obtaining the advertisement browsing data, the recommendation processing device further performs refined ranking processing on multiple advertisement items based on the advertisement ranking model to obtain the advertisement refined ranking result. As a schematic description, the recommendation processing device extracts the item features of each advertisement item based on the advertisement ranking model and performs a scoring process on the item features of each advertisement item to obtain the item score corresponding to the advertisement item. After calculating the item scores of all advertisement items, the recommendation processing device sorts the item scores of all advertisement items based on the advertisement ranking model to obtain the advertisement item ranking result. Finally, the recommendation processing device obtains the advertisement refined ranking result based on the advertisement items and the corresponding advertisement positions corresponding to the item scores greater than the preset threshold in the advertisement item ranking result. More specifically, if the advertisement item ranking result is obtained in the descending order of the item scores, at this time, the recommendation processing device can also use the first Q advertisement items and these Q advertisement positions in the advertisement item ranking result as the advertisement refined ranking result, where Q is an integer greater than or equal to 1.

[0076] It should be noted that for the advertisement positions in the advertisement refined ranking result, there is no need to arrange them in the sorting order of the advertisement items obtained by the refined ranking process. It only needs to include the advertisement positions and the candidate advertisements in the advertisement refined ranking result.

[0077] 502. Generate M candidate sequences with different advertisement position rankings based on multiple candidate contents and at least one advertisement position. The arrangement order of the multiple candidate contents in each candidate sequence is the same as the target order. Each advertisement position in each candidate sequence is used to pre-place multiple candidate advertisements, where M is an integer greater than or equal to 2.

[0078] In this example, after obtaining the content fine ranking result and the advertisement fine ranking result, the recommendation processing device further generates M candidate advertisements with different advertisement slot rankings based on multiple candidate contents in the content fine ranking result and at least one advertisement slot in the advertisement fine ranking result.

[0079] Exemplarily, the recommendation processing device performs permutation and combination processing on all candidate contents in the content fine ranking result and all advertisement slots in the advertisement fine ranking result to obtain N initial mixed sequences. After obtaining these N initial mixed sequences, the recommendation processing device determines whether all of these N initial mixed sequences meet the preset ranking rule. It should be noted that the described preset ranking rule can be understood as the expected ranking condition between the content that the user wants to view and the advertisements. In this way, the recommendation processing device matches each initial mixed sequence among these N initial mixed sequences with the preset ranking rule, and selects the initial mixed sequences that meet the preset ranking rule from these N initial mixed sequences to generate M candidate sequences with different advertisement slots.

[0080] On the contrary, in the case where it is determined based on the preset ranking rule that among these N initial mixed sequences, there is an initial mixed sequence that does not meet the preset ranking rule, the corresponding initial mixed sequence that does not meet the preset ranking rule is deleted.

[0081] The described preset ranking rule may include, but is not limited to, that there are at least T candidate contents between every two adjacent advertisement slots, where T is an integer greater than or equal to 1.

[0082] As a schematic description, taking the case where there are at least T candidate contents between every two adjacent advertisement slots as the preset ranking rule as an example, in the process of selecting the initial mixed sequences that meet this preset ranking rule from the N initial mixed sequences, it can be understood with reference to the following method, that is:

[0083] The recommendation processing device calculates the first interval number in each initial mixed sequence. The described first interval number can be used to indicate the number of candidate contents between every two adjacent advertisement slots in the corresponding initial mixed sequence. In this way, after the recommendation processing device calculates the first interval number in each initial mixed sequence, it compares the first interval number with the value T. If the recommendation processing device compares that the first interval number is greater than or equal to T, it determines the initial mixed sequence corresponding to the first interval number as a candidate sequence. On the contrary, if the recommendation processing device compares that the first interval number is less than T, it deletes the corresponding initial mixed sequence.

[0084] For example, Figure 6 shows a schematic diagram of the M candidate sequences provided by the embodiments of the present application. As Figure 6As shown, taking the candidate content A1 to candidate content A5, ad space 1, and ad space 2 mentioned in the above step 501 as examples, since the arrangement order among candidate content A1 to candidate content A5 has not changed, and no candidate ads are inserted into ad space 1 and ad space 2, at this time, ad space 1 and ad space 2 can be not specifically distinguished. In this way, 30 initial mixed sequences can be obtained through permutation and combination.

[0085] Assume T = 3. At this time, for each initial mixed sequence among these 30 initial mixed sequences, the first interval number is greater than or equal to 3, and the sequences that are non-repetitive are as follows. Specifically, it can be understood with reference to the content shown in Table 1 below, that is:

[0086] Table 1

[0087]

[0088]

[0089] It can be seen from Table 1 above that there are a total of 6 initial mixed sequences that meet the condition of being greater than or equal to 3, that is, 6 candidate sequences are generated. Specifically, in candidate sequence 1, the sorting between the ad space and the candidate content is: ad space 1 --> candidate content A1 --> candidate content A2 --> candidate content A3 --> ad space 2 --> candidate content A4 --> candidate content A5. In candidate sequence 2, the sorting between the ad space and the candidate content is: ad space 1 --> candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> ad space 2 --> candidate content A5. In candidate sequence 3, the sorting between the ad space and the candidate content is: ad space 1 --> candidate content A1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> candidate content A5 --> ad space 2. In candidate sequence 4, the sorting between the ad space and the candidate content is: candidate content A1 --> ad space 1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> ad space 2 --> candidate content A5. In candidate sequence 5, the sorting between the ad space and the candidate content is: candidate content A1 --> ad space 1 --> candidate content A2 --> candidate content A3 --> candidate content A4 --> candidate content A5 --> ad space 2. In candidate sequence 6, the sorting between the ad space and the candidate content is: candidate content A1 --> candidate content A2 --> ad space 1 --> candidate content A3 --> candidate content A4 --> candidate content A5 --> ad space 2.

[0090] It should be noted that the above Figure 6Only the case where T = 3, candidate content A1 to candidate content A5, ad slot 1, and ad slot A2 are used as examples for illustration. In practical applications, for the value of T, as well as other numbers of candidate content and ad slots, it can be specifically determined according to the browsing needs of the object, and no limitation is made in this application.

[0091] It should be noted that the above only takes the preset sorting rule that there are at least T candidate content between every two adjacent ad slots as an example for illustration. In practical applications, the preset sorting rule may also include but is not limited to: the first display slot is not an ad slot, or the last display slot is not an ad slot, or an ad slot needs to be inserted between every two adjacent candidate content, etc., which will not be elaborated here. The described display slot is understood as the position for displaying candidate content or candidate ads.

[0092] By setting the above preset sorting rule, it is possible to customize the browsing order according to the browsing needs of the object. Moreover, it is also possible to eliminate the initial mixed sequence that does not conform to the rule and quickly locate a reasonable candidate sequence.

[0093] 503. Determine the target ad placed on each ad slot in each candidate sequence, and generate M ad rearrangement sequences based on the multiple candidate content in each candidate sequence and the target ad placed on each ad slot. The target ad is one of the multiple candidate ads.

[0094] In this example, from the content Figure 6 shown above, it can be seen that after the recommendation processing device generates M candidate sequences, since the ad slots in each of these M candidate sequences only fix their positions in the sequence and no candidate ads are inserted. Therefore, after generating these M candidate sequences, the recommendation processing device also needs to determine the target ad placed on each ad slot in each candidate sequence in order to determine the corresponding M ad rearrangement sequences. The described target ad is one of the multiple candidate ads.

[0095] Figure 7 shows the structural schematic diagram of determining M ad rearrangement sequences provided by an embodiment of the present application. As Figure 7 shown, after determining M candidate sequences through the foregoing step 502, each candidate sequence can be used as the input of the ad rearrangement module to determine the target ad placed on each ad slot in each candidate sequence through this ad rearrangement module, and then generate the corresponding M ad rearrangement sequences. The mentioned ad rearrangement module at least includes a concatembedding layer and a preset expert network (FC layers) model.

[0096] In addition, in the process of determining the target advertisement to be placed in each ad slot of each candidate sequence, since the determination process of the target advertisement to be placed in each ad slot of each candidate sequence is basically similar. Therefore, in the embodiment of the present application, only one candidate sequence (i.e., the first sequence mentioned later) among the M candidate sequences and any ad slot (i.e., the first ad slot in the first sequence mentioned later) are taken as examples for illustration. Specifically, it can be understood with reference to the content of the following steps S11 to S14, that is:

[0097] Step S11: For the first sequence, extract the content context features of the first sequence. The first sequence is any one of the M candidate sequences, and the content context features of the first sequence include the ad slot features of the ad slots in the first sequence and the content features of each candidate content.

[0098] In this example, the described first sequence is any one of the M candidate sequences, such as the candidate sequence 1 mentioned above Figure 6 For this first sequence, the recommendation processing device can extract the content context features of this first sequence. The described content context features of the first sequence at least include the ad slot features of the ad slots in the first sequence and the content features of each candidate content.

[0099] For example, taking the candidate sequence 1 above Figure 6 as the first sequence as an example, the recommendation processing device can extract the content features of candidate content A1, the content features of candidate content A2,..., and the content features of candidate content A5, and extract the ad slot features of ad slot 1 and ad slot 2 in this candidate sequence 1. Further, fuse the content features of candidate content A1 to candidate content A5, and the ad slot features of ad slot 1 and ad slot 2 to be used as the content context features of this candidate sequence 1.

[0100] It should be noted that in some examples, the mentioned ad slot features are also called slot features, and the specific name is not limited.

[0101] Step S12: For the first ad slot in the first sequence, extract the ad features of each first candidate advertisement respectively. Each first candidate advertisement is one of the multiple candidate advertisements pre-placed in the first ad slot, and the first ad slot is any one of at least one ad slot.

[0102] In this example, the first sequence may include at least one ad slot. Taking any ad slot (i.e., the first ad slot) as an example, if multiple candidate ads can be pre-placed on the first ad slot, the recommendation processing device can also extract the ad features of each first candidate ad on the first ad slot, such as, but not limited to, the attribute features of the first candidate ad, the scoring features in the ad fine-ranking processing stage, the predicted click-through rate (pCTR), the predicted conversion rate (pCVR), etc. calculated in the ad fine-ranking processing stage. The first candidate ad mentioned can be understood as one of the multiple candidate ads pre-placed on the first ad slot.

[0103] For example, taking the candidate sequence 1 shown above Figure 6 as the first sequence and the ad slot 1 in the candidate sequence 1 as the first ad slot. From the content shown in step 502 above, it can be known that candidate ads a1, a2, b1, b2, etc. can be pre-placed in the ad slot 1. At this time, the recommendation processing device needs to extract the ad features of the candidate ads a1, a2, b1, and b2 respectively.

[0104] Step S13: Calculate the revenue per mille (eCPM) of the corresponding first candidate ad based on the content context features of the candidate content and the ad features of each first candidate ad.

[0105] In this example, after the recommendation processing device extracts the content context features of the candidate content and the ad features of each first candidate ad in the first sequence, it also needs to calculate the revenue per mille (i.e., eCPM) of the corresponding first candidate ad based on the content context features and the ad features of each first candidate ad.

[0106] Optionally, in the process of calculating the eCPM of the first candidate ad, object features such as object name and object attribute can also be considered in combination on the basis of considering the content context features of the candidate content and the ad features of each first candidate ad. In this way, the accuracy of calculating the eCPM of the candidate ad is improved.

[0107] As a schematic description, since the calculation methods for the revenue per mille (eCPM) values of each first candidate advertisement are basically similar during the calculation process, in the embodiments of the present application, only any one of the first candidate advertisements (i.e., the second candidate advertisement mentioned later) is taken as an example for illustration. Additionally, since the revenue per mille value is equivalent to the product of the predicted conversion rate (pCVR), the predicted click-through rate (pCTR), and the advertisement bid. Or rather, the revenue per mille value satisfies the following formula, that is: eCPM = pCVR × pCTR × advertisement bid. Therefore, during the process of calculating the revenue per mille value of the second candidate advertisement, the recommendation processing device can first determine the predicted conversion rate, the predicted click-through rate, and the advertisement bid of this first candidate advertisement. Specifically, it can be understood with reference to the following method, that is:

[0108] The recommendation processing device first performs feature fusion processing on the content context feature of the candidate content and the advertisement feature of the second candidate advertisement through the fusion coding layer to obtain a first fusion feature. After obtaining the first fusion feature through fusion, the recommendation processing device processes the first fusion feature based on a preset expert network model to obtain the predicted conversion rate of the second candidate advertisement and the predicted click-through rate of the second candidate advertisement. The described preset expert network model can include, but is not limited to, FC layers, etc., which are not limited in the present application.

[0109] In addition, since the advertisement bid of the second candidate advertisement needs to be considered during the process of calculating the revenue per mille value of the second candidate advertisement. And for the advertisement bid of this second candidate advertisement, it is generally pre-given by the corresponding advertiser according to the business requirements for the second candidate advertisement. Therefore, the recommendation processing device also needs to obtain the advertisement bid of the second candidate advertisement from the terminal device corresponding to the advertiser.

[0110] In this way, the recommendation processing device then uses the predicted conversion rate of the second candidate advertisement, the predicted click-through rate of the second candidate advertisement, and the advertisement bid of the second candidate advertisement to calculate the product of the predicted conversion rate of the second candidate advertisement, the predicted click-through rate of the second candidate advertisement, and the advertisement bid of the second candidate advertisement, and obtains the revenue per mille value of the second candidate advertisement.

[0111] Step S14: Based on the revenue per mille values of all the first candidate advertisements, determine the target advertisement placed on the first advertising position from the multiple candidate advertisements pre-placed on the first advertising position.

[0112] In this example, after calculating the revenue per mille value of each first candidate advertisement, the recommendation processing device also determines the target advertisement placed on the first advertising position from the multiple candidate advertisements pre-placed on the first advertising position based on the revenue per mille values of all these first candidate advertisements.

[0113] As a schematic description, the recommendation processing device may determine the maximum revenue per mille (eCPM) value from the eCPM values of all the first candidate advertisements. After determining the maximum eCPM value, the recommendation processing device determines the candidate advertisement corresponding to the maximum eCPM value as the target advertisement to be placed in the first advertising slot.

[0114] For example, taking Figure 7 the advertisement slot 1 in the candidate sequence 1 shown as an example, candidate advertisements a1, a2, b1, and b2 can be pre-placed in the advertisement slot 1. If the candidate advertisements a1, a2, b1, and b2 are inserted into the advertisement slot a1 respectively, the eCPM a1 corresponding to the candidate advertisement a1 = 1000, the eCPM a2 corresponding to the candidate advertisement a2 = 800, the eCPM b1 corresponding to the candidate advertisement b1 = 700, and the eCPM b2 corresponding to the candidate advertisement b2 = 400.

[0115] Thus, by comparing eCPM a1, eCPM a2, eCPM b1, and eCPM b2, the maximum eCPM can be calculated as eCPM a1. At this time, the candidate advertisement a1 corresponding to eCPM a1 can be determined as the target advertisement to be placed in the advertisement slot 1 in the candidate sequence 1 (i.e., the advertisement rearrangement sequence 1 mentioned later).

[0116] In addition, for the advertisement slot 2 in the candidate sequence 1 and which candidate advertisements should be placed in the advertisement slots 1 and 2 in the other candidate sequences 2 (i.e., the advertisement rearrangement sequence 2 mentioned later) to candidate sequence 6 (i.e., the advertisement rearrangement sequence 6 mentioned later), the calculation process can be understood with reference to Figure 7 the process shown in the advertisement slot 1 in the candidate advertisement 1 shown. Details are not described here. For example, through the calculation process described in the foregoing steps S11 to S14, the candidate advertisement b2 can be used as the target advertisement to be placed in the advertisement slot 2 in the candidate sequence 1, and the candidate advertisements a2 and b1 can be used as the target advertisements to be placed in the advertisement slots 1 and 2 in the candidate sequence 6, etc.

[0117] In addition, after the recommendation processing device determines the target advertisement placed on each ad slot in each candidate sequence, it also generates M ad rearrangement sequences based on the multiple candidate contents in each candidate sequence and the target advertisement placed on each ad slot. That is to say, after the recommendation processing device determines the target advertisement placed on each ad slot in each candidate sequence, it also inserts the target advertisement into the corresponding ad slot in the corresponding candidate sequence. For example, for candidate sequence 1, the recommendation processing device inserts the target advertisement in ad slot 1 (i.e., candidate advertisement a1) into ad slot 1 in candidate sequence 1, and inserts the target advertisement in ad slot 2 (i.e., candidate advertisement b1) into ad slot 2 in candidate sequence 1. Similarly, the same insertion operation is performed for other candidate sequences to generate M ad rearrangement sequences, such as Figure 7 the ad rearrangement sequences 1 to ad rearrangement sequence 6 shown in

[0118] In some other alternative examples, after generating the M ad rearrangement sequences, the ad billing information of each target advertisement in each ad rearrangement sequence can also be calculated. Exemplarily, based on the calculated revenue per mille (eCPM) values of each candidate advertisement pre-placed in each ad slot, the sorting result of the candidate advertisements in each ad slot can be determined first. Further, for each ad slot, using the idea of second-price billing, based on the eCPM values of the candidate advertisements pre-placed in the corresponding ad slot and the corresponding sorting result of the candidate advertisements, the ad billing information of the target advertisement placed in the corresponding ad slot can be determined. Exemplarily, through the ad billing information, data support can be provided for the subsequent sequence selection process. That is to say, the ad billing information can also be used as a part of the ad features in the process of calculating the sequence score in subsequent step 504.

[0119] For example, taking ad slot 1 in the ad rearrangement sequence 1 shown in the foregoing Figure 7 as an example, at this time, since the calculated eCPM a1, eCPM a2, eCPM b1, and eCPM b2 are 1000, 800, 700, and 400 respectively. Through sorting, the corresponding sorting result of the candidate advertisements can be determined as candidate advertisement a1 -> candidate advertisement a2 -> candidate advertisement b1 -> candidate advertisement b2. At this time, using the idea of second-price billing, the reserve price of candidate advertisement a2 can be used as the ad billing information of the target advertisement (i.e., candidate advertisement a1).

[0120] In some alternative examples, if the ad rearrangement module or other components fail and the target advertisements placed on each ad slot cannot be estimated, the traditional mixed arrangement processing scheme can still be used as a disaster recovery fallback scheme.

[0121] 504. Calculate the sequence score of each advertisement rearrangement sequence, and determine the target sequence from the M advertisement rearrangement sequences based on all the sequence scores.

[0122] In this example, after generating the M advertisement rearrangement sequences, the recommendation processing device can also calculate the sequence score of each advertisement rearrangement sequence. Moreover, after calculating the sequence scores of each advertisement rearrangement sequence, the recommendation processing device determines the target sequence from the M advertisement rearrangement sequences based on all the sequence scores.

[0123] Figure 8 The flowchart of sequence optimization provided by the embodiment of the present application is shown. As Figure 8 shown, after determining the M advertisement rearrangement sequences through the foregoing step 503, each advertisement rearrangement sequence can be processed through the Figure 8 shown processing flow, such as through a preset Gate network model and a preset expert network model, etc., to determine the final target sequence. As a schematic description, in the process of calculating the sequence scores of each advertisement rearrangement sequence, the calculation methods are basically similar. Therefore, in the embodiment of the present application, only any one advertisement rearrangement sequence (i.e., the second sequence mentioned later) is taken as an example for illustration. Specifically, it can be understood with reference to the content described in the following steps S21 to S23, that is:

[0124] Step S21: For this second sequence, the recommendation processing device can extract the context features of the candidate content in the second sequence. For example, taking the advertisement rearrangement sequence 1 shown in the foregoing Figure 7 as the second sequence, the recommendation processing device can extract the content features of candidate content A1, the content features of candidate content A2,..., and the content features of candidate content A5, and then fuse the content features of candidate content A1 to candidate content A5 as the context features of the candidate content in the advertisement rearrangement sequence 1.

[0125] Step S22: For each advertisement position in the second sequence, extract the advertisement features of the target advertisement placed on each advertisement position.

[0126] In this example, in this second sequence, since the corresponding target advertisement has been inserted for each advertisement position in the second sequence. Then, the recommendation processing device can also extract the advertisement features of each target advertisement on each advertisement position in the second sequence. For example, the advertisement features of each target advertisement include, but are not limited to, one or more of the following features, that is: the attribute features of the target advertisement; the estimated click-through rate, estimated conversion rate, and revenue value per thousand impressions calculated in the advertisement rearrangement processing stage; the scoring features in the advertisement fine-tuning processing stage, the estimated click-through rate and estimated conversion rate calculated in the advertisement fine-tuning processing stage, etc., which are not limited here.

[0127] For example, taking the advertisement rearrangement sequence 1 shown above as the second sequence, the advertisement features of the target advertisement a1 placed on the advertisement space 1 and the advertisement features of the target advertisement b2 placed on the advertisement space 2 can be extracted. Figure 7 For example, taking the advertisement rearrangement sequence 1 shown above as the second sequence, the advertisement features of the target advertisement a1 placed on the advertisement space 1 and the advertisement features of the target advertisement b2 placed on the advertisement space 2 can be extracted.

[0128] Step S23: Determine the sequence score of the second sequence based on the content context features of the candidate content and the advertisement features of all target advertisements in the second sequence.

[0129] In this example, after the recommendation processing device extracts the content context features of the candidate content in the second sequence and the advertisement features of each target advertisement in the second sequence, it is also necessary to determine the sequence score of the second sequence based on the content context features and the advertisement features of all target advertisements.

[0130] Exemplarily, in the process of calculating the sequence score of the second sequence, the recommendation processing device can Figure 8 through the fusion coding layer shown, perform feature fusion processing on the content context features of the candidate content in the second sequence and the advertisement features of all target advertisements in the second sequence to obtain the second fusion feature. After obtaining the second fusion feature, the recommendation processing device uses Figure 8 the preset Gate network model shown to perform feature processing on the second fusion feature to obtain the promotion information of the candidate content. The described promotion information can be used to indicate the expected display situation when the candidate content is recommended for display. For example, the promotion information may include, but is not limited to, content playback duration, content click-through rate, content exposure count, etc., which will not be elaborated here. In addition, the number of models of the preset Gate network model is not limited in this application.

[0131] In addition, after obtaining the second fusion feature, the recommendation processing device also uses Figure 8 the preset expert network model shown to perform feature processing on the second fusion feature to obtain the total advertisement revenue. The described total advertisement revenue can be used to indicate the expected revenue situation of all target advertisements in the second sequence. In addition, the number of models of the preset expert network model is not limited in this application.

[0132] In this way, the recommendation processing device can determine the sequence score of the second sequence based on the promotion information and the total advertising revenue. As a schematic description, in the process of determining the sequence score of the second sequence based on the promotion information and the total advertising revenue, the recommendation processing device can first obtain the first weight and the second weight. Among them, the mentioned first weight is used to indicate the content weight of the candidate content. The described second weight is used to indicate the weight of the revenue. Further, the recommendation processing device weights the promotion information using the first weight and weights the total advertising revenue using the second weight, respectively obtaining the weighted promotion information and the weighted total advertising revenue. In this way, after completing the weighting process of the promotion information and the total advertising revenue, the weighted promotion information and the weighted total advertising revenue are summed to obtain the sequence score of the second sequence.

[0133] Thus, the recommendation processing device can repeatedly execute the above steps S21 to S23 to calculate the sequence score of each advertisement rearrangement sequence among these M advertisement rearrangement sequences. Exemplarily, after calculating the sequence score of each advertisement rearrangement sequence, the recommendation processing device can determine the target sequence from these M advertisement rearrangement sequences based on all the sequence scores. As a schematic description, the recommendation processing device can determine the maximum sequence score from all the sequence scores and determine the advertisement rearrangement sequence corresponding to the maximum sequence score as the target sequence.

[0134] For example, taking Figure 7 the advertisement rearrangement sequences 1 to 6 shown as an example, if the sequence scores of the advertisement rearrangement sequences 1 to 6 are calculated through the Figure 8 processing flow shown, which are: 89 points, 82 points, 65 points, 58 points, 75 points, and 68 points respectively. At this time, through comparison, it can be known that the maximum sequence score is 89 points. Thus, the advertisement rearrangement sequence 1 corresponding to this 89 points can be used as the target sequence.

[0135] Optionally, in the process of calculating each sequence score, the object features such as the object name and object attributes can also be considered in combination on the basis of considering the content context features of the candidate content and the advertisement features of each target advertisement. In the above way, the accuracy of calculating the sequence score is improved.

[0136] 505. Recommend the candidate content and candidate advertisements in the target sequence according to the target sequence.

[0137] In this example, after the recommendation processing device determines the target sequence, it can recommend the candidate content and candidate advertisements in the target sequence according to the sorting order in the target sequence. Exemplarily, the recommendation processing device can map the target sequence to a recommendation feedback message and send the recommendation feedback message to the terminal device to complete informing the terminal device of the candidate content and candidate advertisements in the target sequence. In this way, after receiving the recommendation feedback message, the terminal device performs a demapping process on it to obtain the target sequence in the recommendation feedback message. Further, the terminal device displays the corresponding candidate content and candidate advertisements to the target object according to the sorting order in the target sequence. Thus, the target object can view the corresponding candidate content and candidate advertisements in sequence.

[0138] Compared with the Figure 2 and Figure 3 traditional recommendation methods shown above, as can be seen from the content Figures 4 to 8 shown in this application, before the sequence optimization operation in the embodiments of this application, the newly added advertisement rearrangement module is first used to Figure 4 re-evaluate the candidate advertisements pre-placed on each advertisement position. And as can be seen from Figure 7 , during the process of re-evaluating using the advertisement rearrangement module, it is necessary to consider the advertisement features of each candidate advertisement and the content context features including content features and advertisement position features to re-evaluate the corresponding estimated conversion rate, estimated click-through rate, eCPM, etc., so that no matter which advertisement rearrangement sequence the target sequence determined in the subsequent mixed arrangement processing stage (such as Figure 8 ) is, it can ensure that the pCTR, eCPM, advertisement billing information, etc. of each advertisement rearrangement sequence are accurate. In this way, it can be ensured that the target sequence determined by using accurate advertisement features in the mixed arrangement processing stage is a better value, so that when presenting content and advertisements to an object according to the currently determined target sequence, a better content recommendation experience can be brought to the object, and accurate and better advertisement revenue can be obtained.

[0139] The above mainly introduces the solution provided by the embodiments of this application from the perspective of methods. It can be understood that in order to implement the above functions, it includes the corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should easily realize that, combining the modules and algorithm steps of each example described in the embodiments disclosed in this application, this application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the way of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution.

[0140] In the embodiments of the present application, the device can be divided into functional modules according to the above method examples. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. It should be noted that the division of modules in the embodiments of the present application is illustrative, only a logical function division, and there can be other division methods in actual implementation.

[0141] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other relevant parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of the overall module or unit that includes the function of the module or unit.

[0142] The recommended processing device in the embodiments of the present application will be described in detail below. Figure 9 It is a schematic diagram of optional functional modules of the recommended processing device provided in the embodiments of the present application. As Figure 9 shown, the recommended processing device may include an acquisition unit 901 and a processing unit 902.

[0143] Among them, the acquisition unit 901 is used to acquire the content fine ranking result and the advertisement fine ranking result. The content fine ranking result includes multiple candidate contents sorted in the target order, and the advertisement fine ranking result includes at least one advertisement position and at least one candidate advertisement. Specifically, reference can be made to the content described in step 501 above for understanding, and details are not elaborated here. Figure 5 For understanding, details are not elaborated here.

[0144] The processing unit 902 is used to generate M candidate sequences with different advertisement position rankings based on the multiple candidate contents and at least one advertisement position. The arrangement order of the multiple candidate contents in each candidate sequence is the same as the target order, and each advertisement position in each candidate sequence is used to pre-place multiple candidate advertisements, where M is an integer greater than or equal to 2. Specifically, reference can be made to the content described in step 502 above for understanding, and details are not elaborated here. Figure 5 For understanding, details are not elaborated here.

[0145] The processing unit 902 is used to determine the target advertisement placed on each advertisement position in each candidate sequence, and generate M advertisement re-ranking sequences based on the multiple candidate contents in each candidate sequence and the target advertisement placed on each advertisement position. The target advertisement is one of the multiple candidate advertisements. Specifically, reference can be made to the content described in step 503 above for understanding, and details are not elaborated here. Figure 5 For understanding, details are not elaborated here.

[0146] The processing unit 902 is configured to calculate the sequence score of each advertisement rearrangement sequence, and determine the target sequence from the M advertisement rearrangement sequences based on all the sequence scores. Specifically, reference may be made to the content described in step 504 above for understanding, and details are not elaborated here. Figure 5 For the content described in step 504 above, reference may be made for understanding, and details are not elaborated here.

[0147] The processing unit 902 is configured to recommend the candidate content and candidate advertisements in the target sequence according to the target sequence. Specifically, reference may be made to the content described in step 505 above for understanding, and details are not elaborated here. Figure 5 For the content described in step 505 above, reference may be made for understanding, and details are not elaborated here.

[0148] In some alternative examples, the processing unit 902 is configured to: for the first sequence, extract the content context features of the first sequence, where the first sequence is any one of the M candidate sequences, and the content context features of the first sequence include the advertisement slot features of the advertisement slots in the first sequence and the content features of each candidate content; for the first advertisement slot in the first sequence, extract the advertisement features of each first candidate advertisement respectively, where each first candidate advertisement is one of the multiple candidate advertisements pre-placed in the first advertisement slot, and the first advertisement slot is any one of at least one advertisement slot; based on the content context features of the candidate content and the advertisement features of each first candidate advertisement, calculate the thousand-impression revenue value of the corresponding first candidate advertisement; based on the thousand-impression revenue values of all the first candidate advertisements, determine the target advertisement placed in the first advertisement slot from the multiple candidate advertisements pre-placed in the first advertisement slot.

[0149] In some other alternative examples, the processing unit 902 is configured to: perform feature fusion processing on the content context features of the candidate content and the advertisement features of the second candidate advertisement to obtain the first fusion feature, where the second candidate advertisement is any one of the first candidate advertisements; process the first fusion feature based on a preset expert network model to obtain the estimated conversion rate and the estimated click-through rate of the second candidate advertisement; calculate the product of the estimated conversion rate of the second candidate advertisement, the estimated click-through rate of the second candidate advertisement, and the advertisement bid of the second candidate advertisement to obtain the thousand-impression revenue value of the second candidate advertisement.

[0150] In some other alternative examples, the processing unit 902 is configured to: determine the maximum thousand-impression revenue value from the thousand-impression revenue values of all the first candidate advertisements; determine the candidate advertisement corresponding to the maximum thousand-impression revenue value as the target advertisement placed in the first advertisement slot.

[0151] In some other alternative examples, the processing unit 902 is configured to: for the second sequence, extract the content context features of the candidate content in the second sequence, where the second sequence is any one of the M advertisement rearrangement sequences; for each advertisement slot in the second sequence, extract the advertisement features of the target advertisement placed on each advertisement slot; and determine the sequence score of the second sequence based on the content context features of the candidate content and the advertisement features of all the target advertisements in the second sequence.

[0152] In some other alternative examples, the processing unit 902 is configured to: perform feature fusion processing on the content context features of the candidate content and the advertisement features of all the target advertisements in the second sequence to obtain the second fusion feature; perform feature processing on the second fusion feature based on a preset Gate network model to obtain the promotion information of the candidate content, and perform feature processing on the second fusion feature based on a preset expert network model to obtain the total advertisement revenue, where the promotion information is used to indicate the expected display situation when the candidate content is recommended for display, and the total advertisement revenue is used to indicate the expected revenue situation of all the target advertisements in the second sequence; and determine the sequence score of the second sequence based on the promotion information and the total advertisement revenue.

[0153] In some other alternative examples, the obtaining unit 901 is further configured to: before determining the sequence score of the second sequence based on the promotion information and the total advertisement revenue, obtain a first weight and a second weight, where the first weight is used to indicate the content weight of the candidate content, and the second weight is used to indicate the weight of the revenue. The processing unit 902 is configured to: weight the promotion information based on the first weight and weight the total advertisement revenue based on the second weight to respectively obtain the weighted promotion information and the weighted total advertisement revenue; and sum the weighted promotion information and the weighted total advertisement revenue to obtain the sequence score of the second sequence.

[0154] In some other alternative examples, the processing unit 902 is configured to: determine the maximum sequence score from all the sequence scores; and determine the advertisement rearrangement sequence corresponding to the maximum sequence score as the target sequence.

[0155] In some other alternative examples, the processing unit 902 is configured to: perform permutation and combination processing on all the candidate contents and all the advertisement slots to obtain N initial mixed sequences, where N is greater than or equal to M and N is an integer; and select the initial mixed sequences that meet the preset sorting rule from the N initial mixed sequences to generate M candidate sequences with different advertisement slot orderings.

[0156] In some other alternative examples, the preset sorting rule includes that there are at least T candidate contents between every two adjacent advertisement spaces, where T is an integer greater than or equal to 1; the processing unit 902 is configured to: calculate the first interval number in each initial mixed sequence, where the first interval number is used to indicate the number of candidate contents between every two adjacent advertisement spaces in the corresponding initial mixed sequence; and when the first interval number is greater than or equal to T, determine the initial mixed sequence corresponding to the first interval number as a candidate sequence.

[0157] In some other alternative examples, the obtaining unit 901 is further configured to: before obtaining the content fine-ranking result and the advertisement fine-ranking result, obtain content browsing data and advertisement browsing data, where the content browsing data includes a plurality of content entries, and the advertisement browsing data includes a plurality of advertisement entries. The processing unit 902 is configured to

[0158] perform fine-ranking processing on the plurality of content entries based on a content ranking model to obtain a content fine-ranking result, and perform fine-ranking processing on the plurality of advertisement entries based on an advertisement ranking model to obtain an advertisement fine-ranking result, where the content entries include candidate contents, and the advertisement entries include candidate advertisements.

[0159] In some other alternative examples, the processing unit 902 is configured to: respectively extract the entry features of each content entry based on the content ranking model, and perform a scoring process on the entry features of each content entry to obtain the entry score of the corresponding content entry; sort the entry scores of all content entries based on the content ranking model to obtain a content entry sorting result; and obtain a content fine-ranking result based on the content entries corresponding to the entry scores greater than a preset threshold in the content entry sorting result.

[0160] In some other alternative examples, the processing unit 902 is configured to: respectively extract the entry features of each advertisement entry based on the advertisement ranking model, and perform a scoring process on the entry features of each advertisement entry to obtain the entry score of the corresponding advertisement entry; sort the entry scores of all advertisement entries based on the advertisement ranking model to obtain an advertisement entry sorting result; and obtain an advertisement fine-ranking result based on the advertisement entries and the corresponding advertisement spaces corresponding to the entry scores greater than a preset threshold in the advertisement entry sorting result.

[0161] The above describes the recommendation processing device in the embodiments of the present application from the perspective of modular functional entities. The following describes the recommendation processing device in the embodiments of the present application from the perspective of hardware processing. Figure 10 is a schematic structural diagram of a recommendation processing device provided by an embodiment of the present application. The recommendation processing device may vary greatly due to different configurations or performances, and may include but is not limited to Figure 9 the recommendation processing device shown. As Figure 10As shown, the recommendation processing device 300 can vary significantly due to differences in configuration or performance, and may include one or more central processing units (CPUs) 322 (e.g., one or more processors) and a memory 332, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 342 or data 344. Among them, the memory 332 and the storage media 330 can be transient storage or persistent storage. The programs stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations for the recommendation processing device. Further, the central processing unit 322 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the recommendation processing device 300. Exemplarily, the central processing unit 322 is used to execute the computer-executable instructions stored in the storage media 330, thereby implementing the recommendation processing method provided in the foregoing embodiments of the present application.

[0162] The recommendation processing device 300 may further include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, and so on.

[0163] Exemplarily, Figure 10 the central processing unit 322 in may cause the recommendation processing device to execute the method in the corresponding method embodiment by calling the computer-executable instructions stored in the memory 332. Figure 5

[0164] Specifically, Figure 9 the function / implementation process of the processing unit 902 in may be implemented by the central processing unit 322 in calling the computer-executable instructions stored in the memory 332. Figure 10 Figure 9 the function / implementation process of the acquisition unit 901 in may be implemented by the input / output interface 358 in. Figure 10

[0165] The steps performed by the recommendation processing device in the foregoing embodiments may be based on the structure of the recommendation processing device shown in this Figure 10

[0166] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the methods described in the foregoing embodiments are implemented.

[0167] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the methods described in the foregoing embodiments.

[0168] In the foregoing embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. Those skilled in the art can clearly understand that, for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above may refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0169] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods may be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other may be through some interfaces, indirect couplings or communication connections of devices or units, and may be in electrical, mechanical, or other forms.

[0170] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0171] A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, they generate, wholly or partly, a process or function in accordance with the embodiments of the present application. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be stored by a computer, or a data storage device such as a server or data center that includes one or more integrated available media. The available medium may be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as an SSD), etc.

[0172] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for recommendation processing, characterized in that, Including: Obtain the refined content ranking result and the refined advertisement ranking result. The refined content ranking result includes a plurality of candidate contents sorted in a target order, and the refined advertisement ranking result includes at least one advertisement slot and at least one candidate advertisement; Based on the plurality of candidate contents and the at least one advertisement slot, generate M candidate sequences with different advertisement slot rankings. The arrangement order of the plurality of candidate contents in each candidate sequence is the same as the target order, and each advertisement slot in each candidate sequence is used to pre-place a plurality of the candidate advertisements, where M is an integer greater than or equal to 2; Determine the target advertisement placed on each advertisement slot in each candidate sequence, and generate M advertisement re-ranking sequences based on the plurality of candidate contents in each candidate sequence and the target advertisement placed on each advertisement slot. The target advertisement is one of the plurality of candidate advertisements; Calculate the sequence score of each advertisement re-ranking sequence, and determine the target sequence from the M advertisement re-ranking sequences based on all the sequence scores; Recommend the candidate contents and candidate advertisements in the target sequence according to the target sequence.

2. The method according to claim 1, wherein Determine the target advertisement placed on each advertisement slot in each candidate sequence, including: For the first sequence, extract the content context feature of the first sequence. The first sequence is any one of the M candidate sequences, and the content context feature of the first sequence includes the advertisement slot feature of the advertisement slot in the first sequence and the content feature of each candidate content; For the first advertisement slot in the first sequence, extract the advertisement features of each first candidate advertisement. Each first candidate advertisement is one of the plurality of candidate advertisements pre-placed on the first advertisement slot, and the first advertisement slot is any one of the at least one advertisement slot; Based on the content context feature of the candidate content and the advertisement features of each first candidate advertisement, calculate the thousand-impression revenue value corresponding to each first candidate advertisement; Based on the thousand-impression revenue values of all the first candidate advertisements, determine the target advertisement placed on the first advertisement slot from the plurality of candidate advertisements pre-placed on the first advertisement slot.

3. The method according to claim 2, wherein Based on the content context feature of the candidate content and the advertisement features of each first candidate advertisement, calculate the thousand-impression revenue value corresponding to each first candidate advertisement, including: Perform feature fusion processing on the content context feature of the candidate content and the advertisement feature of the second candidate advertisement to obtain a first fusion feature. The second candidate advertisement is any one of the first candidate advertisements; Based on a preset expert network model, process the first fusion feature to obtain the estimated conversion rate of the second candidate advertisement and the estimated click-through rate of the second candidate advertisement; Calculate the product of the estimated conversion rate of the second candidate advertisement, the estimated click-through rate of the second candidate advertisement, and the advertisement bid of the second candidate advertisement to obtain the thousand-impression revenue value of the second candidate advertisement.

4. The method according to any one of claims 2 to 3, characterized in that Determining the target advertisement placed on the first ad slot from the multiple candidate advertisements pre-placed on the first ad slot based on the revenue per mille values of all the first candidate advertisements includes: Determining the maximum revenue per mille value from the revenue per mille values of all the first candidate advertisements; Determining the candidate advertisement corresponding to the maximum revenue per mille value as the target advertisement placed on the first ad slot.

5. The method according to any one of claims 1 to 3, characterized in that, Calculating the sequence score of each of the advertisement rearrangement sequences includes: For a second sequence, extracting the content context features of the candidate content in the second sequence, where the second sequence is any one of the M advertisement rearrangement sequences; For each ad slot in the second sequence, extracting the advertisement features of the target advertisements placed on each of the ad slots; Determining the sequence score of the second sequence based on the content context features of the candidate content and the advertisement features of all the target advertisements in the second sequence.

6. The method according to claim 5, wherein Determining the sequence score of the second sequence based on the content context features of the candidate content and the advertisement features of all the target advertisements in the second sequence includes: Performing feature fusion processing on the content context features of the candidate content and the advertisement features of all the target advertisements in the second sequence to obtain a second fusion feature; Performing feature processing on the second fusion feature based on a preset Gate network model to obtain the promotion information of the candidate content, and performing feature processing on the second fusion feature based on a preset expert network model to obtain the total advertisement revenue, where the promotion information is used to indicate the expected display situation when the candidate content is recommended for display, and the total advertisement revenue is used to indicate the expected revenue situation of all the target advertisements in the second sequence; Determining the sequence score of the second sequence based on the promotion information and the total advertisement revenue.

7. The method according to claim 6, characterized in that, Before determining the sequence score of the second sequence based on the promotion information and the total advertisement revenue, the method further includes: Obtaining a first weight value and a second weight value, where the first weight value is used to indicate the content weight of the candidate content, and the second weight value is used to indicate the weight of the revenue; Determining the sequence score of the second sequence based on the promotion information and the total advertisement revenue includes: Weighting the promotion information based on the first weight value and weighting the total advertisement revenue based on the second weight value to respectively obtain a weighted promotion information and a weighted total advertisement revenue; Summing the weighted promotion information and the weighted total advertisement revenue to obtain the sequence score of the second sequence.

8. The method according to any one of claims 1 to 3, characterized in that, Determining a target sequence from the M advertisement rearrangement sequences based on all the sequence scores includes: Determining the maximum sequence score from all the sequence scores; Determining the advertisement rearrangement sequence corresponding to the maximum sequence score as the target sequence.

9. The method according to any one of claims 1 to 3, characterized in that, Generating M candidate sequences with different ad slot orderings based on the multiple candidate contents and the at least one ad slot, including: Perform permutation and combination processing on all the candidate contents and all the ad slots to obtain N initial mixed sequences, where N is greater than or equal to M and N is an integer; Select initial mixed sequences that meet the preset sorting rules from the N initial mixed sequences to generate M candidate sequences with different ad slot sortings.

10. The method according to claim 9, wherein The preset sorting rule includes that there are at least T candidate contents between every two adjacent ad slots, and T is an integer greater than or equal to 1; Selecting initial mixed sequences that meet the preset sorting rules from the N initial mixed sequences to generate M candidate sequences with different ad slot sortings includes: Calculate the first interval number in each initial mixed sequence, and the first interval number is used to indicate the number of candidate contents between every two adjacent ad slots in the corresponding initial mixed sequence; When the first interval number is greater than or equal to T, determine the initial mixed sequence corresponding to the first interval number as a candidate sequence.

11. The method according to any one of claims 1 to 3, characterized in that, Before obtaining the content fine sorting result and the ad fine sorting result, the method further includes: Obtain content browsing data and ad browsing data, where the content browsing data includes multiple content items, and the ad browsing data includes multiple ad items; Obtaining the content fine sorting result and the ad fine sorting result includes: Perform fine sorting on the multiple content items based on a content sorting model to obtain the content fine sorting result, and perform fine sorting on the multiple ad items based on an ad sorting model to obtain the ad fine sorting result. The content items include the candidate contents, and the ad items include the candidate ads.

12. The method according to claim 11, wherein Performing fine sorting on the multiple content items based on a content sorting model to obtain the content fine sorting result includes: Extract the item features of each content item based on the content sorting model, and perform a scoring process on the item features of each content item to obtain the item score corresponding to the content item; Sort the item scores of all the content items based on the content sorting model to obtain a content item sorting result; Based on the content items corresponding to the item scores greater than a preset threshold in the content item sorting result, obtain the content fine sorting result.

13. The method according to claim 11, wherein Performing fine sorting on the multiple ad items based on an ad sorting model to obtain the ad fine sorting result includes: Extract the item features of each ad item based on the ad sorting model, and perform a scoring process on the item features of each ad item to obtain the item score corresponding to the ad item; Sort the item scores of all the ad items based on the ad sorting model to obtain an ad item sorting result; Based on the ad items and the corresponding ad slots corresponding to the item scores greater than a preset threshold in the ad item sorting result, obtain the ad fine sorting result.

14. A recommendation processing device, characterized in that, Includes: An acquisition unit for acquiring a content fine sorting result and an ad fine sorting result. The content fine sorting result includes multiple candidate contents sorted in a target order, and the ad fine sorting result includes at least one ad slot and at least one candidate ad; A processing unit, configured to generate M candidate sequences with different ad slot rankings based on the multiple pieces of candidate content and the at least one ad slot, wherein the arrangement order of the multiple pieces of candidate content in each candidate sequence is the same as the target order, and each ad slot in each candidate sequence is used to pre-place multiple pieces of candidate ads, and M is an integer greater than or equal to 2; The processing unit is configured to determine the target ad placed on each ad slot in each candidate sequence, and generate M ad rearrangement sequences based on the multiple pieces of candidate content in each candidate sequence and the target ad placed on each ad slot, wherein the target ad is one of the multiple pieces of candidate ads; The processing unit is configured to calculate the sequence score of each ad rearrangement sequence, and determine the target sequence from the M ad rearrangement sequences based on all the sequence scores; The processing unit is configured to recommend the candidate content and candidate ads in the target sequence according to the target sequence.

15. A recommended processing device, characterized in that, Comprising: An input / output interface, a processor, and a memory, wherein program instructions are stored in the memory; The processor is configured to execute the program instructions stored in the memory and execute the method according to any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes instructions, which, when running on a computer device, cause the computer device to execute the method according to any one of claims 1 to 13.

17. A computer program product, characterized in that, The computer program product includes instructions, which, when running on a computer device, cause the computer device to execute the method according to any one of claims 1 to 13.

Images