Data processing method and device, data recall method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a data processing method and device, a data recall method and device, electronic equipment and a storage medium, and the data processing method comprises the steps: obtaining first consumption data of a target account in a reference recommendation scene and to-be-recalled data in a to-be-recommended scene; determining feature information of the first consumption data and the to-be-recalled data; and calculating the similarity between the feature information of the first consumption data and the feature information of the to-be-recalled data, and determining the to-be-recalled data associated with the first consumption data based on the sequence of the similarity between the feature information of the first consumption data and the feature information of the to-be-recalled data. According to the data processing method and device, the data recall method and device, the electronic equipment and the storage medium provided by the embodiment of the invention, the recommendation precision to the user is improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a data processing and recall method, device, electronic device and storage medium. Background Art

[0002] Different users may prefer to consume in different scenarios. For example, user 1 prefers to consume in scenario 1, user 2 prefers to consume in scenario 2, and so on. However, in actual applications, there are also cases where users who prefer to consume in one scenario consume in another scenario. It is not difficult to understand that if a user prefers to consume in one scenario, the consumption time in that scenario is generally longer, while the consumption time in other scenarios is generally shorter.

[0003] If scenarios with shorter user consumption time are understood as weak consumption scenarios, weak consumption scenarios cannot well capture user interest preferences due to less consumption data, which will result in lower accuracy of recommendations to users. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a data processing and recall method, device, electronic device and storage medium to improve the accuracy of recommendations to users. The specific technical solution is as follows:

[0005] In a first aspect, a data processing method is provided, comprising:

[0006] Obtain the first consumption data of the target account in the reference recommendation scenario and the to-be-recalled data in the to-be-recommended scenario;

[0007] Determining characteristic information of the first consumption data and the data to be recalled;

[0008] The similarity between the feature information of the first consumption data and the feature information of the data to be recalled is calculated, and the data to be recalled associated with the first consumption data is determined based on the order of the similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

[0009] Optionally, the data to be recalled includes: second consumption data of the target account in the scenario to be recommended;

[0010] The determining of the characteristic information of the first consumption data and the data to be recalled includes:

[0011] determining characteristic information of the first consumption data and the second consumption data;

[0012] The calculating the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, and determining the target recalled data from the data to be recalled based on the order of the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, includes:

[0013] The similarity between the feature information of the first consumption data and the feature information of the second consumption data is calculated, and based on the order of the similarity between the feature information of the first consumption data and the feature information of the second consumption data, target recall data is determined from the second consumption data.

[0014] Optionally, the feature information includes embedded features;

[0015] The determining the characteristic information of the first consumption data and the second consumption data includes:

[0016] sorting the first consumption data and the second consumption data according to consumption time;

[0017] Take every two adjacent consumption data as a sample pair and calculate the corresponding weight of each sample pair;

[0018] Taking each consumption data in each sample pair as the center point data, all consumption data that form a sample pair with the center point data are aggregated into a combination as a neighbor node combination;

[0019] For each center point data, the consumption data in the combination of the center point data and the neighboring nodes is used as a training sample;

[0020] The training sample is used as the input of the graph attention model GAT, the feature vector of the graph attention layer in the graph attention model GAT is adjusted, and when the graph attention model GAT outputs the center point data, the feature vector of the graph attention layer in the graph attention model GAT is used as the embedded feature of the center point data.

[0021] Optionally, after taking each consumption data in each sample pair as the center point data and aggregating all consumption data constituting a sample pair with the center point data into a combination as a neighbor node combination, the method further includes:

[0022] Selecting target neighbor data from the neighbor node combination based on the number of consumption data in the neighbor node combination and / or the weight of sample pairs corresponding to the center point data and the consumption data in the neighbor node combination;

[0023] For each center point data, the center point data and the target neighbor data are used as test samples;

[0024] After taking the training sample as the input of the graph attention model GAT and adjusting the feature vector of the graph attention layer in the graph attention model GAT, the method further includes:

[0025] The test sample is used as the input of the graph attention model GAT, the feature vector of the graph attention layer in the graph attention model GAT is adjusted, and when the graph attention model GAT outputs the center point data, the feature vector of the graph attention layer in the graph attention model GAT is used as the embedded feature of the center point data.

[0026] Optionally, determining the target recall data from the data to be recalled based on the order of similarity between the feature information of the first consumption data and the feature information of the data to be recalled includes:

[0027] Based on the order of similarity between the characteristic information of the first consumption data and the characteristic information of the to-be-recalled data, inverted data is generated with the identifier of the first consumption data as a keyword and the target recall data as a value corresponding to the keyword.

[0028] Optionally, the reference recommended scene is a long video scene, and the scene to be recommended is a short video scene.

[0029] In a second aspect, a data recall method is provided, comprising:

[0030] In response to a data recall request sent by a client, determining consumption data of a target account in a reference recommendation scenario, wherein the data recall request is used to request data in a scenario to be recommended;

[0031] Search for the data to be recalled associated with the consumption data, and use the data to be recalled associated with the consumption data as the first target recalled data of the target account in the scenario to be recommended, wherein the data to be recalled associated with the consumption data is obtained according to the data processing method described in any one of the first aspects.

[0032] Optionally, determining the consumption data of the target account in the reference recommendation scenario includes:

[0033] Determine user preference data of the target account in a reference recommendation scenario, where the user preference data is data of interest to the user in the consumption data.

[0034] Optionally, the scenario to be recommended includes a scenario in which the consumption time of the target account is less than a preset time or a scenario in which the consumption data is less than a preset amount;

[0035] The determining of the consumption data of the target account in the reference recommendation scenario includes:

[0036] Determine the scenario with the longest consumption time of the target account as a reference recommendation scenario;

[0037] Determine the consumption data of the target account in the scenario with the longest consumption duration.

[0038] Optionally, after searching for the to-be-recalled data associated with the consumption data, the method further includes:

[0039] Determine matching data in the data in the to-be-recommended scenario that matches the attributes of the to-be-recalled data, and use the matching data as second target recall data of the target account in the to-be-recommended scenario;

[0040] and / or,

[0041] Matching data in the data under the to-be-recommended scenario that matches the attributes of the consumption data of the target account under the reference recommendation scenario is determined, and the matching data is used as third target recall data of the target account under the to-be-recommended scenario.

[0042] In a third aspect, a data processing device is provided, comprising:

[0043] An acquisition module, used to acquire the first consumption data of the target account in the reference recommendation scenario and the to-be-recalled data in the to-be-recommended scenario;

[0044] A first determining module, used to determine characteristic information of the first consumption data and the data to be recalled;

[0045] The second determination module is used to calculate the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, and determine the data to be recalled associated with the first consumption data based on the order of the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled.

[0046] In a fourth aspect, a data recall device is provided, comprising:

[0047] A third determination module, configured to determine the consumption data of the target account in the reference recommendation scenario in response to a data recall request sent by the client, wherein the data recall request is used to request data in the scenario to be recommended;

[0048] A search module is used to search for the data to be recalled associated with the consumption data, and use the data to be recalled associated with the consumption data as the first target recalled data of the target account in the scenario to be recommended, wherein the data to be recalled associated with the consumption data is obtained according to the data processing device described in the third aspect.

[0049] In a fifth aspect, an electronic device is provided, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0050] Memory, used to store computer programs;

[0051] The processor is used to implement the method steps described in any one of the first aspect or the second aspect when executing the program stored in the memory.

[0052] In another aspect of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any of the above-mentioned data processing methods or data recall methods is implemented.

[0053] In another aspect of the present invention, a computer program product comprising instructions is provided, which, when executed on a computer, enables the computer to execute any of the above-mentioned data processing methods or data recall methods.

[0054] In the embodiment of the present invention, by determining the characteristic information of the consumption data of the target account in the reference recommendation scenario and the characteristic information of the data to be recommended in the to-be-recommended scenario, and based on the order of similarity between the characteristic information of the consumption data and the characteristic information of the data to be recommended, and based on the order of similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, the data to be recalled associated with the first consumption data is determined, that is, the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario is established, so that in the to-be-recommended scenario, the recalled data can be determined based on the consumption data of the user in the reference recommendation scenario, through the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario, so that the recommendation to the user in the to-be-recommended scenario utilizes the data of the user in the reference recommendation scenario, thereby improving the accuracy of the recommendation to the user. Even if the user has relatively little consumption data in the to-be-recommended scenario, the user's interest preference can be well obtained, thereby improving the accuracy of the recommendation to the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0056] Figure 1 is a flow chart of a data processing method in an embodiment of the present invention;

[0057] Figure 2 for Figure 1 A flow chart of S102;

[0058] Figure 3 for Figure 1 Another flow chart of S102;

[0059] Figure 4 is a flow chart of a data recall method in an embodiment of the present invention;

[0060] Figure 5 A schematic diagram of a data processing method in a specific scenario according to an embodiment of the present invention;

[0061] Figure 6 Another schematic diagram of the data processing method in a specific scenario according to an embodiment of the present invention;

[0062] Figure 7 This is another schematic diagram of the data processing method in a specific scenario according to an embodiment of the present invention;

[0063] Figure 8 A schematic diagram of a specific embodiment of the present invention;

[0064] Fig. 9 A schematic diagram of a data recall method in a specific scenario according to an embodiment of the present invention;

[0065] Fig.10 A schematic diagram of triggering a data recall request in an embodiment of the present invention;

[0066] Fig.11 A schematic diagram of the structure of a data processing device in an embodiment of the present invention;

[0067] Fig.12 It is a structural schematic diagram of a data recall device in an embodiment of the present invention;

[0068] Fig.13 Schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0070] Reference Figure 1 , an embodiment of the present invention provides a data processing method, including:

[0071] S101, obtaining first consumption data of a target account in a reference recommendation scenario and to-be-recalled data in a to-be-recommended scenario;

[0072] S102, determining characteristic information of the first consumption data and the data to be recalled;

[0073] S103, calculating the similarity between the feature information of the first consumption data and the feature information of the data to be recalled, and determining the data to be recalled associated with the first consumption data based on the order of the similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

[0074] In the embodiment of the present invention, by determining the characteristic information of the consumption data of the target account in the reference recommendation scenario and the characteristic information of the data to be recommended in the to-be-recommended scenario, and based on the order of similarity between the characteristic information of the consumption data and the characteristic information of the data to be recommended, and based on the order of similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, the data to be recalled associated with the first consumption data is determined, that is, the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario is established, so that in the to-be-recommended scenario, the recalled data can be determined based on the consumption data of the user in the reference recommendation scenario, through the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario, so that the recommendation to the user in the to-be-recommended scenario utilizes the data of the user in the reference recommendation scenario, thereby improving the accuracy of the recommendation to the user. Even if the user has relatively little consumption data in the to-be-recommended scenario, the user's interest preference can be well obtained, thereby improving the accuracy of the recommendation to the user.

[0075] In the embodiment of the present invention, consumption data of a target account in one scenario is used to influence data recall of the target account in another scenario.

[0076] The data processing method provided by the embodiment of the present invention can be applied to a server, etc.

[0077] In S101 , the scene to be recommended and the reference recommended scene may be different scenes.

[0078] For example, if a scenario involves data recall, you can execute Figure 1 In the data mathematical method shown, in this process, the scene can be used as a scene to be recommended, and another scene different from the scene can be used as a reference recommendation scene.

[0079] For example, the reference recommendation scene is a long video scene, and the scene to be recommended is a short video scene.

[0080] In one possible implementation, the scenario to be recommended may be a scenario where the target account has less consumption data and a shorter consumption time. The reference recommendation scenario may be a scenario where the target account has more consumption data and a longer consumption time.

[0081] For example, in a scenario involving data recall, before recalling data for the target account in this scenario, you can execute Figure 1The data mathematical method shown in the figure can detect the historical consumption records of the target account in the scenario during the process. If it is detected that the target account has less consumption data and / or a shorter consumption time in the scenario, such as the consumption data is less than a preset number and / or the consumption time is less than a preset time, the scenario will be used as a scenario to be recommended, wherein the preset number and preset time can be determined based on actual needs or experience.

[0082] Furthermore, the historical consumption records of the target account in other scenarios different from the scenario may be detected. If it is detected that the target account has more consumption data and longer consumption time in the other scenarios, such as more consumption data in other scenarios than the consumption data in the above-mentioned scenario, and / or the consumption time in other scenarios is longer than the consumption time in the above-mentioned scenario, then the other scenarios may be used as reference recommendation scenarios. Among them, if it is detected that the target account has more consumption data and / or longer consumption time in multiple other scenarios, such as more consumption data in multiple other scenarios than the consumption data in the above-mentioned scenario to be recommended, and / or the consumption time in multiple other scenarios is longer than the consumption time in the above-mentioned scenario to be recommended, then one or more may be randomly selected from the multiple other scenarios as reference recommendation scenarios; or, the scenario with the most consumption data and / or the longest consumption time may be selected from the multiple other scenarios as the reference recommendation scenario; or, the scenario with consumption data greater than a preset amount and / or consumption time greater than a preset time may be selected from the multiple other scenarios as the reference recommendation scenario.

[0083] The "first" in the first consumption data is only used to distinguish the consumption data used subsequently. The first consumption data here is actually the consumption data of the target account in the reference recommendation scenario. It can also be understood as the historical behavior data of the target account in the reference recommendation scenario.

[0084] The consumption data of the target account in the reference recommendation scenario can also be understood as the consumption data of the user using the target account in the reference recommendation scenario.

[0085] The data to be recalled in the scenario to be recommended may be all the data in the scenario to be recommended, or may be the consumption data of the target account in the recommendation scenario.

[0086] The data in different scenarios can be in any form, such as pictures, videos, audio, text, etc.

[0087] In S102, any method of characterizing data characteristic information may be used to determine the characteristic information of the first consumption data and the data to be recalled, which is not limited in the embodiment of the present invention.

[0088] For example, the first consumption data and the data to be recalled may be encoded respectively, and the encoded data may be used as feature information corresponding to the first consumption data and the data to be recalled respectively, wherein the encoding method may be selected according to actual needs and the like.

[0089] For another example, a model may be pre-trained respectively, and the model is used to determine the characteristic information of the data. Thus, the first consumption data and the data to be recalled may be input into the model, and the characteristic information corresponding to the first consumption data and the data to be recalled may be obtained through the model.

[0090] In S103, the similarity between the first consumption data and the data to be recalled is represented by the similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

[0091] The data to be recalled associated with the first consumption data may be determined based on the order of similarity between the feature information of the first consumption data and the feature information of the data to be recalled, which may include: selecting the data to be recalled corresponding to the feature information of one or more data to be recalled that are ranked first in the order of similarity from high to low as the data to be recalled associated with the first consumption data. Alternatively, selecting the data to be recalled corresponding to the feature information of one or more data to be recalled that are ranked last in the order of similarity from low to high as the data to be recalled associated with the first consumption data.

[0092] The first consumption data may be one or more, and the to-be-recalled data associated with each first consumption data may be obtained respectively.

[0093] Determining the data to be recalled associated with the first consumption data may also be understood as establishing an association relationship between the first consumption data and the data to be recalled.

[0094] Furthermore, after determining the data to be recalled associated with the first consumption data, or in other words, establishing the association relationship between the first consumption data and the data to be recalled, the data to be recalled associated with the first consumption data can be saved, or in other words, saving the association relationship between the first consumption data and the data to be recalled, so that in the data recall stage, the data to be recalled associated with the first consumption data can be found.

[0095] S103 may include: generating inverted data using the identifier of the first consumption data as a keyword and the target recall data as a value corresponding to the keyword based on the order of similarity between the feature information of the first consumption data and the feature information of the to-be-recalled data.

[0096] The above process of determining the to-be-recalled data associated with the first consumption data may also be understood as a process of generating inverted data using the identifier of the consumption data of the reference recommendation scenario as a keyword and the target recall data of the to-be-recommended scenario as a value corresponding to the keyword.

[0097] In an optional embodiment, the data to be recalled in the scenario to be recommended obtained in S101 includes: second consumption data of the target account in the scenario to be recommended.

[0098] The "second" in the second consumption data is only used to distinguish the consumption data of the target account in the aforementioned reference recommendation scenario and the consumption data that may be used subsequently. The second consumption data here is actually the consumption data of the target account in the scenario to be recommended, which can also be understood as the historical behavior data of the target account in the scenario to be recommended, etc.

[0099] In the case where the acquired data to be recalled in the scenario to be recommended includes: second consumption data of the target account in the scenario to be recommended, S102 includes: determining feature information of the first consumption data and the second consumption data;

[0100] S103 includes: calculating the similarity between the feature information of the first consumption data and the feature information of the second consumption data, and determining the target recall data from the second consumption data based on the order of the similarity between the feature information of the first consumption data and the feature information of the second consumption data.

[0101] In an optional embodiment, the feature information includes embedded features.

[0102] The data to be recalled in the scenario to be recommended obtained in S101 includes: in the case of the second consumption data of the target account in the scenario to be recommended, determining the characteristic information of the first consumption data and the second consumption data, such as Figure 2 As shown, including:

[0103] S201, sorting the first consumption data and the second consumption data according to consumption time;

[0104] For example, the order is from nearest to the current time to the farthest one or from farthest to nearest one.

[0105] S202, taking every two adjacent consumption data as a sample pair, and calculating the weight corresponding to each sample pair;

[0106] According to the above S201 and S202, the consumption data corresponding to each account can obtain a corresponding sample pair, and each sample pair is processed separately.

[0107] S203, taking each consumption data in each sample pair as the center point data, aggregating all consumption data that form a sample pair with the center point data into a combination as a neighbor node combination;

[0108] S204, for each center point data, using the center point data and the consumption data in the combination of neighboring nodes as training samples;

[0109] S205, using the training sample as the input of the Graph Attention Network (GAT), and adjusting the feature vector of the graph attention layer in the GAT.

[0110] Furthermore, when the output of GAT is center point data, the feature vector of the graph attention layer in GAT is used as the embedded feature of the center point data.

[0111] It can be understood that in the embodiment of the present invention, GAT is used to determine the embedded features of consumption data, and other consumption data are used in the process of determining a consumption data. In this way, the accuracy of the determined feature information can be improved, so that the feature information can more accurately represent the consumption data.

[0112] For example, 1) the consumption data of the same target account in different scenarios are sorted according to the consumption time, and every two adjacent consumption data in the sorted list are regarded as co-occurrence sample pairs.

[0113] 2) Taking each consumption data ID in each pair in 1) as the key, aggregate all consumption data that can form a pair with it into a list, that is, take the current consumption data ID as the center point of the graph, and aggregate the consumption data IDs of all neighboring points connected to it.

[0114] 3) For each center point in 2) and its neighboring nodes, the pair is used as a training sample;

[0115] 4) Train the GAT graph model to obtain the embedding representation of all consumption data.

[0116] That is, the training samples are used as the input of GAT, the feature vector of the graph attention layer in GAT is adjusted, and when the output of GAT is the center point consumption data, the feature vector of the graph attention layer in GAT is used as the embedded feature of the center point consumption data.

[0117] In an optional embodiment, each consumption data in each sample pair is used as the center point data, and all consumption data that form a sample pair with the center point data are aggregated into a combination as a neighbor node combination, such as Figure 3 As shown, it also includes:

[0118] S301, selecting target neighbor data from the neighbor node combination based on the number of consumption data in the neighbor node combination and / or the weight of the sample pair corresponding to the center point data and the consumption data in the neighbor node combination;

[0119] S302, for each center point data, the center point data and the target neighbor data are used as test samples;

[0120] S303, after taking the training sample as the input of GAT and adjusting the feature vector of the graph attention layer in GAT, taking the test sample as the input of the graph attention model GAT, adjusting the feature vector of the graph attention layer in the graph attention model GAT, and when the output of the graph attention model GAT is the center point data, using the feature vector of the graph attention layer in the graph attention model GAT as the embedded feature of the center point data.

[0121] Specifically, in the process of S202 taking every two adjacent consumption data as a sample pair, the weight corresponding to each sample pair can be calculated. Then, based on the weight of the sample pair corresponding to the center point consumption data and the consumption data in the neighbor node combination, the target neighbor consumption data can be selected from the neighbor node combination, and for each center point consumption data, the center point consumption data and the target neighbor consumption data can be used as test samples.

[0122] The target neighbor consumption data is selected from the neighbor node combination based on the number of consumption data in the neighbor node combination; for each center point consumption data, the center point consumption data and the target neighbor consumption data are used as test samples. This process can also be understood as randomly sampling the center points whose neighbor point lists connected to each center point consumption data in the above step 2) are larger than a certain threshold to generate test samples.

[0123] In the embodiment of the present invention, during the training of the GAT graph model, a test sample may be selected, and the GAT graph model may be further trained based on the test sample to further improve the accuracy of the obtained features.

[0124] The embodiment of the present invention is Figure 1 The illustrated embodiment can obtain the association relationship between consumption data of one or more accounts in a reference recommendation scenario and data to be recalled in a to-be-recommended scenario. Thus, when data is to be recalled in the to-be-recommended scenario, the data can be recalled based on the association relationship.

[0125] like Figure 4 As shown, an embodiment of the present invention provides a data recall method, including:

[0126] S401, in response to a data recall request sent by a client, determining consumption data of a target account in a reference recommendation scenario, wherein the data recall request is used to request data in a scenario to be recommended;

[0127] S402, searching for the to-be-recalled data associated with the consumption data, and using the to-be-recalled data associated with the consumption data as the first target recall data of the target account in the to-be-recommended scenario, wherein the to-be-recalled data associated with the consumption data is based on the above Figure 1 It is obtained by the data processing method of the illustrated embodiment.

[0128] In the embodiment of the present invention, by determining the characteristic information of the consumption data of the target account in the reference recommendation scenario and the characteristic information of the data to be recommended in the to-be-recommended scenario, and based on the order of similarity between the characteristic information of the consumption data and the characteristic information of the data to be recommended, and based on the order of similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, the data to be recalled associated with the first consumption data is determined, that is, the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario is established, so that in the to-be-recommended scenario, the recalled data can be determined based on the consumption data of the user in the reference recommendation scenario, through the association relationship between the consumption data of the user in the reference recommendation scenario and the data to be recommended in the to-be-recommended scenario, so that the recommendation to the user in the to-be-recommended scenario utilizes the data of the user in the reference recommendation scenario, thereby improving the accuracy of the recommendation to the user. Even if the user has relatively little consumption data in the to-be-recommended scenario, the user's interest preference can be well obtained, thereby improving the accuracy of the recommendation to the user.

[0129] The data recall method provided in the embodiment of the present invention can be applied to a server, etc.

[0130] In S401, when the user wants to obtain data in the recommended scenario, a data recall request can be triggered through the client. The data recall request can carry the user's account information. In this way, the server receives the data recall request sent by the client and can determine the target account based on the account information, and further determine the consumption data of the target account in the reference recommendation scenario.

[0131] The data recall request may be sent when switching from other scenarios to the scenario to be recommended, or it may be sent when requesting consumption in the scenario to be recommended again. Both are possible.

[0132] In an optional embodiment, after receiving a data recall request sent by a client, the target account can be determined based on the account information carried in the data recall request. Afterwards, it can be determined whether the target account is a user preferred by the scenario to be recommended. If not, it can be understood that it is a user preferred by other scenarios, or the user corresponding to the target account has not consumed in the scenario to be recommended before or has less consumption data in the scenario to be recommended. Then, the data recall method provided by the embodiment of the present invention can be further executed to determine the consumption data of the target account in the reference recommendation scenario. The data recall request is used to request data in the scenario to be recommended; and the data to be recalled associated with the consumption data is searched, and the data to be recalled associated with the consumption data is used as the first target recall data of the target account in the scenario to be recommended.

[0133] Among them, each account can be marked in advance based on whether the user corresponding to the target account has consumed in the scenario to be recommended before or the amount of consumption data in the scenario to be recommended. If the user corresponding to the target account has not consumed in the scenario to be recommended before or has less consumption data in the scenario to be recommended, the target account will be marked as a user with preference for other scenarios; if the user corresponding to the target account has consumed in the scenario to be recommended before or has more consumption data in the scenario to be recommended, the target account will be marked as a user with preference for the scenario to be recommended.

[0134] Specifically, the reference recommendation scene, the scene to be recommended, the relationship between the reference recommendation scene and the scene to be recommended, and the consumption data corresponding to the reference recommendation scene and the scene to be recommended have been described in detail in the above embodiments and will not be repeated here.

[0135] In an optional embodiment, determining the consumption data of the target account in the reference recommendation scenario in S401 includes:

[0136] Determine the user preference data of the target account in the reference recommendation scenario. The user preference data is the user's interested data in the consumption data.

[0137] Among them, the data of interest include data with more consumption times and longer consumption duration.

[0138] For example, the consumption data of the target account in the reference recommendation scenario can be integrated and counted to determine the number of consumptions and / or the length of consumption for each consumption data, so that data with a large number of consumptions and / or a long length of consumption can be selected. Among them, a consumption number greater than a preset consumption number can be understood as a large number of consumptions, a consumption length greater than a preset consumption length can be understood as a long consumption length, and so on.

[0139] The above method is only an exemplary description of determining user preference data. In the embodiment of the present invention, the method of determining preference data in the relevant technology can be used to determine the user preference data of the target account in the reference recommendation scenario, and this is not limited in the embodiment of the present invention.

[0140] In an optional embodiment, the scenario to be recommended includes a scenario in which the consumption time of the target account is less than a preset time or a scenario in which the consumption data is less than a preset amount;

[0141] Determine the consumption data of the target account in the reference recommendation scenario, including: determine the scenario with the longest consumption time of the target account as the reference recommendation scenario; determine the consumption data of the target account in the scenario with the longest consumption time.

[0142] Specifically, in the process of recalling data in the scenario to be recommended, if the consumption time of the target account in the scenario to be recommended is less than the preset time or the consumption data is less than the preset number, the scenario with the longest consumption time of the target account can be first determined as the reference recommendation scenario, and then the data of the target account in the reference recommendation scenario is determined, that is, the consumption data of the target account in the scenario with the longest consumption time is determined.

[0143] In this way, when a user spends less time in one scenario, the consumption data of the user in another scenario with a longer consumption time can be used to recall the data in the scenario with a shorter consumption time. Therefore, even if the user has less consumption data in the scenario to be recommended, the user's interest preference can be well obtained, which improves the accuracy of recommendation to the user.

[0144] Above Figure 1 The data to be recalled associated with the first consumption data obtained in the embodiment, or the association between the first consumption data and the data to be recalled, can be understood as a recall source. Searching for the data to be recalled associated with the consumption data in S402 can also be understood as searching for a recall source.

[0145] If the above Figure 1 In the illustrated embodiment, the acquired data to be recalled in the scenario to be recommended includes all data in the scenario to be recommended, and what is determined is the association relationship between the consumption data of the target account in the reference recommendation scenario and the data associated with all the data, and all the data may include data that has not been consumed by the target account before. Figure 4 During the data recall process of the illustrated embodiment, the associated data to be recalled that is searched using the consumption data of the target account in the reference recommendation scenario may include data that has not been consumed before. In this way, even when the target account recalls data for the first time in the to-be-recommended scenario, the data recall method provided in the embodiment of the present invention may be used to perform data recall.

[0146] In an optional embodiment, after searching for the to-be-recalled data associated with the consumption data, the method further includes:

[0147] Determine matching data in the data in the to-be-recommended scenario that matches the attributes of the to-be-recalled data, and use the matching data as second target recall data of the target account in the to-be-recommended scenario;

[0148] and / or,

[0149] Matching data in the data under the to-be-recommended scenario that matches the attributes of the consumption data of the target account under the reference recommendation scenario is determined, and the matching data is used as the third target recall data of the target account under the to-be-recommended scenario.

[0150] Among them, the attributes can be theme, style, etc.

[0151] Specifically, in the actual application process, the attributes of each data can be determined separately, for example, attribute tags can be marked, and then attribute matching can be performed to determine the matching data in the data to be recommended scenario that matches the attributes of the data to be recalled and / or determine the matching data in the data to be recommended scenario that matches the attributes of the consumption data of the target account in the reference recommendation scenario.

[0152] In the embodiment of the present invention, in the data recall process, in order to enrich the recall results and better meet the user's preferences, in addition to using the above Figure 1 The association relationship obtained in the illustrated embodiment determines the target recall data of the target account in the scenario to be recommended. The embodiment of the present invention can further determine the matching data with attribute matching based on the consumption data of the target account in the reference recommendation scenario and the data to be recalled associated with the consumption data of the target account in the reference recommendation scenario.

[0153] After obtaining the target recall data in S402, sorting and recommendation can be performed based on the target recall data. Specifically, the sorting and recommendation can adopt the methods in the related art, and the embodiment of the present invention is not limited to this.

[0154] In order to more clearly illustrate the data processing method and the data recall method provided by the embodiments of the present invention, a detailed description is given below by taking the reference recommendation scene as a long video scene and the scene to be recommended as a short video scene as an example.

[0155] In applications (Applications, Apps) that play long videos and short videos, different users have different preferences for long videos and short videos. At the same time, there are also cases where users who prefer long videos (i.e., users who prefer long videos) browse and consume short video information flow scenarios. Consumption may include operations such as clicking and sharing.

[0156] There are essential differences between short videos and long videos in terms of length, production difficulty and cost, viewing methods and scenarios, content and form, and audience needs and preferences. Different forms and contents meet different needs of the audience and have different application values ​​in different scenarios.

[0157] For example, long videos and short videos can be distinguished based on time and length. For example, short videos generally refer to video content with a shorter duration, usually within a few minutes; while long videos refer to video content with a longer duration, usually several hours or longer.

[0158] In order to better recommend short videos to users who prefer long videos, and thereby attract users who prefer long videos to effectively consume short video information flow scenarios, etc., an embodiment of the present invention provides a data processing method and a data recall method.

[0159] The data processing method provided in the embodiment of the present invention can be applied to a server, etc. The server can be a server corresponding to a video player.

[0160] In an embodiment of the present invention, an association relationship between long videos and short videos can be established first. Then, when a long video preferring user enters a short video information flow scenario, short video content that the long video preferring user may be interested in can be pulled based on the long videos in the long video preferring user's historical interests and the pre-established association relationship between long videos and short videos.

[0161] In order to enable a long video preference user to pull short video content that may be of interest to the long video preference user based on the long videos in the long video preference user's historical interests and the pre-established association relationship between long videos and short videos when the long video preference user enters the short video information flow scenario, an association relationship between long videos and short videos can be first established in an embodiment of the present invention, and the process of pre-establishing the association relationship between long videos and short videos can also be understood as a process of pre-generating a recall source.

[0162] The following is a description of the process of establishing the association relationship between the long video and the short video in the embodiment of the present invention. Figure 1 Specific applications of the illustrated embodiments.

[0163] like Figure 5 As shown, this may include:

[0164] S501, obtaining the first consumed video of the target account in the long video scenario and the to-be-recalled video in the short video scenario;

[0165] S502, determining feature information of the first consumed video of the target account in the long video scenario and the video to be recalled in the short video scenario;

[0166] S503, calculate the similarity between the feature information of the first consumption video of the target account in the long video scenario and the feature information of the video to be recalled in the short video scenario, and determine the video to be recalled in the short video scenario associated with the first consumption video of the target account in the long video scenario based on the order of the similarity between the feature information of the first consumption video of the target account in the long video scenario and the feature information of the video to be recalled in the short video scenario.

[0167] In one embodiment, referring to Figure 6 , the recall source is obtained through the following steps:

[0168] S601, obtaining consumption data of at least one account, wherein for each user in the at least one account, the consumption data includes at least one long video and at least one short video.

[0169] Users make purchases through their accounts, and the account’s consumption data can also be understood as the user’s consumption data.

[0170] Consumption data may include viewing behavior data, etc. Specifically, the viewing behavior data of users in all scenarios in the APP can be accessed, including long videos and short videos. The videos watched by each user are collected as a list for sample collection, that is, a viewing video list is obtained for each user, and the viewing video list contains the video identifiers of the videos watched by the user in the historical time. This process can also be understood as a sample collection process.

[0171] S602, determining feature information of all videos including long videos and short videos included in consumption data of at least one user;

[0172] The embodiment of the present invention does not limit the method of determining the feature information of the video.

[0173] S603, for each long video in all videos, calculate the similarity between the feature information of the long video and the feature information of each short video in all videos, and determine the short videos associated with the long video based on the order of the similarity between the feature information of the long video and the feature information of each short video in all videos, and use the short videos associated with each long video as recall sources.

[0174] Wherein, determining the short videos associated with the long video based on the order of similarity between the feature information of the long video and the feature information of each short video in all videos may include:

[0175] The short videos associated with the long video are determined according to the similarity between the feature information of the long video and the feature information of each short video in all videos in order from high to low or from low to high.

[0176] If the similarity between the feature information of the long video and the feature information of each short video in all videos is in descending order, a preset number of short videos ranked first are selected as short videos associated with the long video;

[0177] If the similarity between the feature information of the long video and the feature information of each short video in all videos is ordered from low to high, a preset number of short videos ranked later are selected as short videos associated with the long video.

[0178] In one implementation, the feature information may include an embedding feature.

[0179] According to the embeddings of all videos obtained in S602, the embedding data of long videos and short videos are distinguished, and then the long video embedding is used as the key and the short video embedding is used as the value to calculate the video similarity. The short videos associated with the long videos are established in the order of similarity, which can also be understood as generating a recall source.

[0180] Among them, in S603, determining the short video associated with the long video based on the order of similarity between the feature information of the long video and the feature information of each short video in all videos may include: generating inverted data with the video identifier of the long video as a keyword and the video identifier of the short video associated with the long video as the value corresponding to the keyword based on the order of similarity between the feature information of the long video and the feature information of each short video in all videos.

[0181] For example, generate inverted data with long video id as key and short video id as value. Inverted data example: a represents the id of the long video "The Legend of Zhen Huan", and b, c, d, etc. are the ids of short videos related to "The Legend of Zhen Huan".

[0182] In one possible implementation, GAT can be used for training to obtain the embedding features of the video.

[0183] like Figure 7 As shown, S602 may include:

[0184] S701, for each account, sorting the videos included in the consumption data of the account according to viewing time;

[0185] For example, in order from near to far from the current time or in order from far to near from the current time.

[0186] S702, taking every two adjacent videos as a sample pair;

[0187] S703, taking each video in each sample pair as a center point video, aggregating all videos forming a sample pair with the center point video into a combination as a neighbor node combination;

[0188] According to the above S701 and S702, the video corresponding to each account can obtain a corresponding sample pair, and each sample pair is processed separately.

[0189] S704, for each center point video, taking the center point video and the video in the combination of neighboring nodes as training samples;

[0190] S705, taking the training sample as the input of GAT, adjusting the feature vector of the graph attention layer in GAT, and taking the feature vector of the graph attention layer in GAT as the embedding feature of the center point video when the output of GAT is the center point video.

[0191] For example, A) sorting the videos in the watched video list of the same account according to the viewing time, and after sorting, every two adjacent videos in the list are regarded as co-occurrence sample pairs.

[0192] B) Taking each video ID in each pair in A as the key, aggregate all videos that can form a pair with it into a list, that is, take the current video ID as the center point of the graph and aggregate the video IDs of all neighboring points connected to it.

[0193] C) For each pair of center point in B and its neighboring nodes, use them as training samples;

[0194] D) Train the GAT graph model to obtain embedding representations of all videos.

[0195] That is, the training samples are used as the input of the graph attention model GAT, the feature vector of the graph attention layer in the graph attention model GAT is adjusted, and when the output of the graph attention model GAT is a center point video, the feature vector of the graph attention layer in the graph attention model GAT is used as the embedded feature of the center point video.

[0196] In one possible implementation, during the training of the GAT graph model, a test sample may be selected, and the GAT graph model may be further trained based on the test sample to improve the accuracy of the obtained features.

[0197] Specifically, after S703, the method may further include: selecting a target neighbor video from the neighbor node combination based on the number of videos in the neighbor node combination and / or the weight of the sample pairs corresponding to the center point video and the videos in the neighbor node combination; for each center point video, using the center point video and the target neighbor video as test samples;

[0198] Furthermore, after taking the training sample as the input of GAT and adjusting the feature vector of the graph attention layer in GAT, it may also include: taking the test sample as the input of GAT, adjusting the feature vector of the graph attention layer in GAT, and when the GAT outputs the center point video, using the feature vector of the graph attention layer in GAT as the embedded feature of the center point video.

[0199] Specifically, in the process of treating every two adjacent videos as a sample pair in S702, the weight corresponding to each sample pair can be calculated. Then, based on the weights of the sample pairs corresponding to the center point video and the videos in the neighbor node combination, the target neighbor video can be selected from the neighbor node combination, and for each center point video, the center point video and the target neighbor video can be used as test samples.

[0200] A target neighbor video is selected from the neighbor node combination based on the number of videos in the neighbor node combination; for each center point video, the center point video and the target neighbor video are used as test samples. This process can also be understood as randomly sampling the center points whose neighbor point list size connected to each center point video in the above step B is greater than a certain threshold to generate test samples.

[0201] In an embodiment of the present invention, an association relationship between long videos and short videos can be established. Specifically, the viewing behaviors of long videos and short videos can be integrated for joint modeling, and inverted data with long videos as keys and short videos as values ​​can be generated to construct a recall source. In this way, when a long video preference user enters the short video information flow scene, in the recommended recall stage, the short video content that the user may be interested in is pulled according to the long video ID interest in his historical interests. In this way, the interest acceptance of the long video preference user in the recall stage is achieved, which can improve the accuracy of recommending short videos to the long video preference user. In addition, the consumption indicators of the long video preference user in the short video information flow scene can be improved. Better acceptance of the long video preference user entering the short video information flow recommendation scene is achieved, which improves the user experience of this part of users.

[0202] Figure 8 A schematic diagram of a specific embodiment provided for an embodiment of the present invention is shown below in conjunction with Figure 8 The embodiment of the present invention pre-establishes an association relationship between long videos and short videos, so that when a user with a long video preference enters a short video information flow scene, in the recall phase, the short video content that the user may be interested in is pulled out and described in detail based on the long video ID interest in his historical interests.

[0203] In general, in the embodiment of the present invention, training samples are constructed by collecting the long video and short video viewing behaviors of users in all scenes in the APP, and GAT is used for training to establish the association relationship between long videos and short videos, thereby obtaining inverted data with long videos as keys and short video lists as values, and realizing the long video interest acceptance of users who prefer long videos. It can be understood that the embodiment of the present invention realizes a method for accepting long video preference users based on long and short video joint modeling.

[0204] Specifically include:

[0205] 1) Sample collection:

[0206] That is, collect samples of users watching long videos & short videos.

[0207] Specifically, the viewing behavior data of users in all scenarios in the APP can be accessed, including long videos and short videos. The videos watched by each user can be used as a list for sample collection.

[0208] 2) GAT graph model training:

[0209] That is, GAT is used to obtain feature information of each collected video.

[0210] Specifically, the embedding representations of all videos may be obtained by adopting steps A) to D) in the above embodiment.

[0211] 3) Inverse generation of long videos to short videos:

[0212] According to the embedding data of all videos obtained in 2), the embedding data of long videos and short videos are distinguished, and then the long video embedding is used as the key and the short video embedding is used as the value to calculate the video similarity, and the inverted data with the long video id as the key and the short video id as the value is generated in the order of similarity.

[0213] Inverse data example: a represents the ID of the long video "The Legend of Zhen Huan", and b, c, d... are the IDs of short videos related to "The Legend of Zhen Huan".

[0214] Conduct training to establish the association between long videos and short videos, so as to obtain inverted data with long videos as keys and short video lists as values.

[0215] 4) Use of inverted data:

[0216] When a user who prefers long videos enters the short video information flow scenario, in the recommendation recall phase, the short video content that he may be interested in is pulled based on the long video ID in his historical interests.

[0217] For example, the long video interests id1, id2 and id3 of the long video preference user are determined. It can also be understood that the video identifiers of the long videos that the long video preference user is interested in include id1, id2 and id3. Then, the video identifiers of the short videos corresponding to id1, id2 and id3 can be found from the recalled source inverted data.

[0218] For example, if a user who prefers long videos has an interest in "Empresses in the Palace", and the corresponding video id is a, then videos b, c, d, etc. can be recalled in the recall stage to inherit their long video interests, and then enter the subsequent stages such as fine ranking. For example, in the recommendation system, in the fine ranking stage, the videos obtained in the recall stage are ranked by the ranking model, and recommendations are made based on the ranked videos.

[0219] The following is a description of the process of extracting short video content that may be of interest to long video preference users based on the long videos in the long video preference user's historical interests and the pre-established association relationship between long videos and short videos. Figure 4 The data method shown is applicable to specific applications.

[0220] Reference Fig. 9 , which may include:

[0221] S901, in response to a data recall request sent by a client, determining the consumed videos of a target account in a long video scenario, where the data recall request is used to request videos in a short video scenario;

[0222] For example, in response to a long video preference user entering a short video information flow scenario, a target long video that the long video preference user is interested in is determined.

[0223] Long video preference users can also be understood as users who prefer long videos. In one implementation, long video preference users can represent users whose consumption of long videos meets preset conditions, where the preset conditions can be determined by actual needs or experience, for example, browsing long videos within a historical time range meets a preset duration, and / or consumption of long videos meets a preset number, etc.

[0224] In one possible implementation, in response to a long video preference user entering a short video information stream scenario, the method may include: receiving a request from the long video preference user for a short video information stream.

[0225] This request is the data recall request mentioned above.

[0226] Wherein, receiving a request from a user who prefers long videos for a short video information stream may include: receiving a trigger request from a user who prefers long videos for an entry option on a short video page. Fig.10 As shown, when the video player is opened, a page is displayed, and the page may include multiple options: Option 1, Option 2, and Option 3, etc. Different options correspond to different categories, among which Option 2 may correspond to a short video page, that is, after clicking Option 2, the video player displays a short video page, and at least one short video is displayed in the short video page. At the same time, operation options for the short video, such as sharing, etc., may be displayed in the window displaying the short video content.

[0227] Specifically, after receiving a trigger request for the entry option of the short video page, it can be determined whether the user sending the trigger request is a long video preference user, such as determining whether the consumption of long videos by the user sending the trigger request within a historical time range meets preset conditions. If so, it means that a request for a short video information stream from a long video preference user has been received.

[0228] The long videos that the long video preference users are interested in can be determined by determining the long video identifiers of the long videos that the long video preference users are interested in.

[0229] In one possible implementation, the video identifiers of long videos that the long video preference user is interested in may be determined from historical consumption data.

[0230] Specifically, the user's behavior of watching long videos in historical time can be recorded, and the long videos that the user is interested in can be counted. For example, the long videos that the user browses and / or consumes that meet preset conditions can be regarded as the long videos that the user is interested in, and the user ID of the user and the video ID of the long video that the user is interested in can be recorded accordingly to obtain historical consumption data. In this way, after receiving the request for the short video information flow from the long video preference user, the video ID of the corresponding long video can be found by using the user ID of the long video preference user.

[0231] S902, searching for the to-be-recalled videos in the short video scenario associated with the consumption videos of the target account in the long video scenario, and taking the to-be-recalled videos in the short video scenario associated with the consumption data of the target account in the long video scenario as the first target recalled videos of the target account in the short video scenario.

[0232] Among them, the videos to be recalled associated with the consumed videos are based on the above Figure 5 It is obtained by the data processing method of the illustrated embodiment.

[0233] It can be understood that the short video corresponding to the target long video is searched from the above-mentioned recall source, and the short video corresponding to the target long video is used as the target recall video, so as to recommend it to the long video preference users based on the target recall video, wherein the recall source includes short videos respectively associated with multiple long videos.

[0234] In the embodiment of the present invention, corresponding short videos can be recommended to long video preference users based on the long videos that the long video preference users are interested in, which can improve the accuracy of recommending short videos to long video preference users.

[0235] Furthermore, the consumption indicators and user experience of users who prefer long videos in short video information flow scenarios can be improved.

[0236] The short videos respectively associated with multiple long videos can be determined in advance and used as recall sources. For example, the corresponding relationship between the video identifiers of short videos respectively associated with the video identifiers of multiple long videos can be determined in advance and saved. In this way, after receiving the request for the short video information flow from the long video preference user and determining the long videos that the long video preference user is interested in, the short videos corresponding to the long videos that the long video preference user is interested in can be found from the recall source.

[0237] In an embodiment of the present invention, short videos corresponding to long videos that long video preference users are interested in are used as target recall videos, so as to be recommended to long video preference users based on the target recall videos in a subsequent stage.

[0238] In the embodiment of the present invention, short videos corresponding to long videos that the long video preference user is interested in are searched from the recall source, and the short videos corresponding to the long videos that the long video preference user is interested in are used as the target recall videos, that is, short videos that the long video preference user may be interested in are determined based on the long videos that the long video preference user is interested in. It can also be understood that the interest of the long video preference user is taken over, and the interest of the long video preference user in the long video is taken over in the process of recommending short videos to the long video preference user.

[0239] In the process of recommending short videos to users who prefer long videos, the industry generally adds users' long video interest features or viewing sequence features in the fine ranking model stage. However, the problem with this is that if the interest in long videos is not taken over in the recall stage, it will be difficult for the work in the fine ranking model stage to be effective.

[0240] When users who prefer long videos enter the short video information flow scene, whether the interests of these users can be captured is directly related to whether these users can effectively consume on the short video information flow page.

[0241] The embodiment of the present invention realizes the interest acceptance of users who prefer long videos in the recommendation recall stage, and can improve the accuracy of recommending short videos to users who prefer long videos. In addition, it can improve the consumption index of users who prefer long videos in the short video information flow scene, and improve the user experience of these users.

[0242] In the embodiment of the present invention, an association relationship between long videos and short videos can be established based on the consumption data of at least one user (generating a recall source). When a long video preference user enters a short video information flow recommendation scenario, the short video corresponding to the long video that the long video preference user is interested in can be searched based on the association relationship established in the above manner. In this way, a better connection is achieved for the long video preference user to enter the short video information flow recommendation scenario, further improving the accuracy of recommending short videos to long video preference users, improving the consumption indicators of long video preference users in the short video information flow scenario, and improving the user experience of these users.

[0243] In one implementable method, when the long video identifier of the long video that the long video preference user is interested in is determined to determine the long video that the long video preference user is interested in, the short video corresponding to the long video that the long video preference user is interested in is searched from the recall source, and the short video corresponding to the long video that the long video preference user is interested in is used as the target recall video, including: searching from the recall source for the short video identifier of the short video corresponding to the long video identifier, and using the short video corresponding to the short video identifier as the target recall video.

[0244] The recommendation process generally includes a recall phase and a sorting phase. In the video recommendation process, the recall phase first recalls some videos from a large number of videos, and the sorting phase further sorts the videos obtained in the recall phase based on user characteristics and / or video characteristics such as user long video interest characteristics or viewing sequence characteristics, and then recommends videos to users based on the sorting results. Among them, the sorting phase can specifically include a rough sorting phase and a fine sorting phase. The rough sorting phase and the fine sorting phase are essentially sorting, the difference is that the sorting basis is different, and the fine sorting phase can also be understood as continuing to obtain a more accurate sorting result based on more or more accurate user characteristics and / or video characteristics on the results obtained in the rough sorting phase.

[0245] When a user who prefers long videos enters the short video information flow scene, in the recommended recall stage, the short videos that the user may be interested in are pulled according to the long video id (identifier) ​​in his historical interests. For example, it is predetermined that the long video with video id a is associated with short videos with video ids b, c, and d. If a user who prefers long videos has an interest in "The Legend of Zhen Huan" in his historical interests, and the corresponding video id is a, then videos b, c, d, etc. can be recalled in the recall stage, taking over the long video interests of the long video preference user, and then entering the subsequent stages such as fine sorting.

[0246] Corresponding to the data processing method provided in the above embodiment, the embodiment of the present invention further provides a data processing device, such as Fig.11 As shown, including:

[0247] An acquisition module 1101 is used to acquire the first consumption data of the target account in the reference recommendation scenario and the to-be-recalled data in the to-be-recommended scenario;

[0248] A first determining module 1102, used to determine characteristic information of the first consumption data and the data to be recalled;

[0249] The second determination module 1103 is used to calculate the similarity between the feature information of the first consumption data and the feature information of the data to be recalled, and determine the data to be recalled associated with the first consumption data based on the order of the similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

[0250] Optionally, the data to be recalled includes: second consumption data of the target account in the scenario to be recommended;

[0251] A first determining module 1102 is specifically used to determine characteristic information of the first consumption data and the second consumption data;

[0252] The second determination module 1103 is specifically used to calculate the similarity between the feature information of the first consumption data and the feature information of the second consumption data, and determine the target recall data from the second consumption data based on the order of the similarity between the feature information of the first consumption data and the feature information of the second consumption data.

[0253] Optionally, the feature information includes embedded features;

[0254] The first determination module 1102 is specifically used to: sort the first consumption data and the second consumption data according to the consumption time; take every two adjacent consumption data as a sample pair, and calculate the weight corresponding to each sample pair; take each consumption data in each sample pair as the center point data, and aggregate all the consumption data that form the sample pair with the center point data into a combination as a neighbor node combination; for each center point data, take the consumption data in the combination of the center point data and the neighbor node as a training sample; take the training sample as the input of the graph attention model GAT, adjust the feature vector of the graph attention layer in the graph attention model GAT, and when the output of the graph attention model GAT is the center point data, use the feature vector of the graph attention layer in the graph attention model GAT as the embedded feature of the center point data.

[0255] Optionally, the first determination module 1102 is specifically used to: after taking each consumption data in each sample pair as the center point data, aggregating all consumption data that form sample pairs with the center point data into a combination as a neighbor node combination, and then selecting target neighbor data from the neighbor node combination based on the number of consumption data in the neighbor node combination and / or the weight of the sample pairs corresponding to the center point data and the consumption data in the neighbor node combination; for each center point data, taking the center point data and the target neighbor data as test samples; after taking the training sample as the input of the graph attention model GAT and adjusting the feature vector of the graph attention layer in the graph attention model GAT, taking the test sample as the input of the graph attention model GAT, adjusting the feature vector of the graph attention layer in the graph attention model GAT, and when the output of the graph attention model GAT is the center point data, taking the feature vector of the graph attention layer in the graph attention model GAT as the embedded feature of the center point data.

[0256] Optionally, the second determination module 1103 is specifically configured to generate inverted data using the identifier of the first consumption data as a keyword and the target recall data as a value corresponding to the keyword based on the order of similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

[0257] Optionally, the reference recommendation scene is a long video scene, and the scene to be recommended is a short video scene.

[0258] Corresponding to the data recall method provided in the above embodiment, the embodiment of the present invention provides a data recall device, such as Fig.12 As shown, including:

[0259] The third determination module 1201 is used to determine the consumption data of the target account in the reference recommendation scenario in response to the data recall request sent by the client, where the data recall request is used to request the data in the scenario to be recommended;

[0260] The search module 1202 is used to search for the data to be recalled associated with the consumption data, and use the data to be recalled associated with the consumption data as the first target recalled data of the target account in the scenario to be recommended, wherein the data to be recalled associated with the consumption data is obtained according to the data processing device of the above embodiment.

[0261] Optionally, the third determination module 1201 is specifically used to determine user preference data of the target account in a reference recommendation scenario, where the user preference data is the user's interested data in the consumption data.

[0262] Optionally, the scenario to be recommended includes a scenario in which the consumption time of the target account is less than a preset time or a scenario in which the consumption data is less than a preset amount;

[0263] The third determination module 1201 is specifically used to determine the scenario with the longest consumption time of the target account as a reference recommendation scenario; and determine the consumption data of the target account in the scenario with the longest consumption time.

[0264] Optionally, the device further comprises:

[0265] The fourth determination module 1203 is used to determine, after searching for the data to be recalled associated with the consumption data, matching data in the data under the scenario to be recommended that matches the attributes of the data to be recalled, and use the matching data as the second target recall data of the target account in the scenario to be recommended; and / or determine matching data in the data under the scenario to be recommended that matches the attributes of the consumption data of the target account in the reference recommendation scenario, and use the matching data as the third target recall data of the target account in the scenario to be recommended.

[0266] The embodiment of the present invention further provides an electronic device, such as Fig.13As shown, it includes a processor 1301, a communication interface 1302, a memory 1303 and a communication bus 1304, wherein the processor 1301, the communication interface 1302, and the memory 1303 communicate with each other through the communication bus 1304;

[0267] Memory 1303, used for storing computer programs;

[0268] The processor 1301 is used to implement the method steps of the data processing method or the data recall method in the above-mentioned embodiment when executing the program stored in the memory 1303.

[0269] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0270] The communication interface is used for communication between the above terminal and other devices.

[0271] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0272] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0273] In another embodiment provided by the present invention, a computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, the data processing method or data recall method described in any one of the above embodiments is implemented.

[0274] In another embodiment provided by the present invention, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute the data processing method or data recall method described in any one of the above embodiments.

[0275] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0276] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0277] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, electronic device, computer-readable storage medium, and computer program product embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0278] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A data processing method, characterized in that: include: Obtain the first consumption data of the target account in the reference recommendation scenario and the to-be-recalled data in the to-be-recommended scenario; Determining characteristic information of the first consumption data and the data to be recalled; The similarity between the feature information of the first consumption data and the feature information of the data to be recalled is calculated, and the data to be recalled associated with the first consumption data is determined based on the order of the similarity between the feature information of the first consumption data and the feature information of the data to be recalled.

2. The method according to claim 1, characterized in that: The data to be recalled includes: second consumption data of the target account in the scenario to be recommended; The determining of the characteristic information of the first consumption data and the data to be recalled includes: determining characteristic information of the first consumption data and the second consumption data; The calculating the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, and determining the target recalled data from the data to be recalled based on the order of the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, includes: The similarity between the feature information of the first consumption data and the feature information of the second consumption data is calculated, and based on the order of the similarity between the feature information of the first consumption data and the feature information of the second consumption data, target recall data is determined from the second consumption data.

3. The method according to claim 2, characterized in that The feature information includes embedded features; The determining the characteristic information of the first consumption data and the second consumption data includes: sorting the first consumption data and the second consumption data according to consumption time; Take every two adjacent consumption data as a sample pair and calculate the corresponding weight of each sample pair; Taking each consumption data in each sample pair as the center point data, all consumption data that form a sample pair with the center point data are aggregated into a combination as a neighbor node combination; For each center point data, the consumption data in the combination of the center point data and the neighboring nodes is used as a training sample; The training sample is used as the input of the graph attention model GAT, the feature vector of the graph attention layer in the graph attention model GAT is adjusted, and when the graph attention model GAT outputs the center point data, the feature vector of the graph attention layer in the graph attention model GAT is used as the embedded feature of the center point data.

4. The method according to claim 3, characterized in that After taking each consumption data in each sample pair as the center point data and aggregating all consumption data constituting a sample pair with the center point data into a combination as a neighbor node combination, the method further includes: Selecting target neighbor data from the neighbor node combination based on the number of consumption data in the neighbor node combination and / or the weight of sample pairs corresponding to the center point data and the consumption data in the neighbor node combination; For each center point data, the center point data and the target neighbor data are used as test samples; After taking the training sample as the input of the graph attention model GAT and adjusting the feature vector of the graph attention layer in the graph attention model GAT, the method further includes: The test sample is used as the input of the graph attention model GAT, the feature vector of the graph attention layer in the graph attention model GAT is adjusted, and when the graph attention model GAT outputs the center point data, the feature vector of the graph attention layer in the graph attention model GAT is used as the embedded feature of the center point data.

5. The method according to claim 1, characterized in that The determining target recall data from the data to be recalled based on the order of similarity between the feature information of the first consumption data and the feature information of the data to be recalled comprises: Based on the order of similarity between the characteristic information of the first consumption data and the characteristic information of the to-be-recalled data, inverted data is generated with the identifier of the first consumption data as a keyword and the target recall data as a value corresponding to the keyword.

6. The method according to any one of claims 1 to 5, characterized in that: The reference recommendation scene is a long video scene, and the scene to be recommended is a short video scene.

7. A data recall method, characterized in that: include: In response to a data recall request sent by a client, determining consumption data of a target account in a reference recommendation scenario, wherein the data recall request is used to request data in a scenario to be recommended; Search for the data to be recalled associated with the consumption data, and use the data to be recalled associated with the consumption data as the first target recalled data of the target account in the scenario to be recommended, wherein the data to be recalled associated with the consumption data is obtained according to the data processing method according to any one of claims 1 to 6.

8. The method according to claim 7, characterized in that The determining of the consumption data of the target account in the reference recommendation scenario includes: Determine user preference data of the target account in a reference recommendation scenario, where the user preference data is data of interest to the user in the consumption data.

9. The method according to claim 7, characterized in that: The scenarios to be recommended include scenarios where the target account's consumption duration is less than a preset duration or where the consumption data is less than a preset amount; The determining of the consumption data of the target account in the reference recommendation scenario includes: Determine the scenario with the longest consumption time of the target account as a reference recommendation scenario; Determine the consumption data of the target account in the scenario with the longest consumption duration.

10. The method according to claim 7, characterized in that After searching for the to-be-recalled data associated with the consumption data, the method further includes: Determine matching data in the data in the to-be-recommended scenario that matches the attributes of the to-be-recalled data, and use the matching data as second target recall data of the target account in the to-be-recommended scenario; and / or, Matching data in the data under the to-be-recommended scenario that matches the attributes of the consumption data of the target account under the reference recommendation scenario is determined, and the matching data is used as third target recall data of the target account under the to-be-recommended scenario.

11. A data processing device, characterized in that: include: An acquisition module, used to acquire the first consumption data of the target account in the reference recommendation scenario and the to-be-recalled data in the to-be-recommended scenario; A first determining module, used to determine characteristic information of the first consumption data and the data to be recalled; The second determination module is used to calculate the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled, and determine the data to be recalled associated with the first consumption data based on the order of the similarity between the characteristic information of the first consumption data and the characteristic information of the data to be recalled.

12. A data recall device, characterized in that: include: A third determination module, configured to determine the consumption data of the target account in the reference recommendation scenario in response to a data recall request sent by the client, wherein the data recall request is used to request data in the scenario to be recommended; A search module is used to search for the data to be recalled associated with the consumption data, and use the data to be recalled associated with the consumption data as the first target recalled data of the target account in the scenario to be recommended, wherein the data to be recalled associated with the consumption data is obtained according to the data processing device according to claim 11.

13. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, for implementing the method steps described in any one of claims 1 to 10 when executing a program stored in a memory.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 10 are implemented.

Images