Method and device for video recommendation and video recommendation model training, equipment and medium

Abstract

The embodiment of the present disclosure relates to a method for training a video recommendation model, comprising: obtaining a feature sample training set and a feature sample test set according to video-related features and user-related features; training a preset model according to the feature sample training set to obtain a prediction model, predicting the feature sample test set through the prediction model to obtain a first-class feature prediction index value set; predicting the feature sample test set through an evaluation model to obtain a second-class feature prediction index value set; comparing N second prediction index values in the second-class feature prediction index value set with corresponding first prediction index values in sequence, screening the second-class features with second prediction index values smaller than the corresponding first prediction index values in the second-class feature prediction index value set to form a feature set; training a video recommendation model by using the feature set, and determining feature identifiers of training features according to the feature set with the second prediction index values smaller than the first prediction index values; and improving the accuracy of the video recommendation model recommendation.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of big data, and particularly relates to a video recommendation method and a video recommendation model training method, a device, equipment and a medium. BACKGROUND

[0002] For video recommendation, video-related features and user-related features need to be collected. However, these features have various dimensions and are uneven, which makes it difficult to train a relatively accurate and suitable video recommendation model. The dimension of a feature is related to the expression ability of the feature, the applicable model and the complexity of calculation. If the dimension of a feature is too high, the calculation amount will increase, thereby reducing the recommendation accuracy and the calculation efficiency of the video recommendation model.

[0003] Therefore, there is a need for a device capable of improving the recommendation accuracy and the calculation efficiency of a video recommendation model. SUMMARY

[0004] The present disclosure discloses a video recommendation method and a video recommendation model training method, a device, equipment and a medium.

[0005] The first aspect of the present disclosure provides a video recommendation model training method, which comprises: obtaining a feature sample training set and a feature sample test set according to video-related features and user-related features; wherein the video-related features at least include attribute features of a video and window statistical features of the video; the user-related features at least include user attribute features, user behavior features and user preference features for a video; training a preset model to obtain a prediction model according to the feature sample training set, and predicting the feature sample test set through the prediction model to obtain a first feature prediction index value set; wherein the first feature prediction index value set includes first prediction index values of N features; m evaluation models are used to respectively predict the feature sample test set to obtain m second feature prediction index value sets; wherein each second feature prediction index value set includes second prediction index values of N features; the N second prediction index values in the mth second feature prediction index value set are compared with the first prediction index values corresponding to the serial numbers in sequence, and the second features corresponding to the serial numbers in the mth second feature prediction index value set whose second prediction index values are smaller than the first prediction index values are selected to form an mth feature set; a video recommendation model is trained using the m feature sets, and a feature identifier of a training feature is determined according to the feature set whose second prediction index value is smaller than the first prediction index value; and the feature identifier is used to determine the features used for video recommendation by the video recommendation model.

[0006] Optionally, the method further comprises: performing m-class preprocessing on the feature sample training set, training a preset model on the preprocessed feature sample training set, and obtaining m evaluation models.

[0007] Optionally, the m-class preprocessing on the feature sample training set and the training of the preset model on the preprocessed feature sample training set to obtain m evaluation models at least include one of the following: performing random ordering processing on the features in the feature sample training set, training the preset model on the feature sample training set after the random ordering processing, and obtaining a first evaluation model; performing zero setting processing on the features in the feature sample training set, training the preset model on the feature sample training set after the zero setting processing, and obtaining a second evaluation model; determining the average values between the features in the feature sample training set and performing average value assignment processing on the features, training the preset model on the feature sample training set after the average value assignment processing, and obtaining a third evaluation model.

[0008] Optionally, the screening of the second-class features with a second prediction index value smaller than the corresponding sequence number of the first prediction index value in the second prediction index value set of the mth second-class feature prediction index value set to form the mth feature set at least includes one of the following: screening the second-class features with a second prediction index value smaller than the corresponding sequence number of the first prediction index value in the second prediction index value set of the first second-class feature prediction index value set to form a first feature set; screening the second-class features with a second prediction index value smaller than the corresponding sequence number of the first prediction index value in the second prediction index value set of the second second-class feature prediction index value set to form a second feature set; and screening the second-class features with a second prediction index value smaller than the corresponding sequence number of the first prediction index value in the second prediction index value set of the third second-class feature prediction index value set to form a third feature set.

[0009] Optionally, the user attribute feature at least includes one of the following: an age feature of the user; a gender feature of the user; a region feature of the user; a language feature of the user; and a device used by the user and / or a network used by the user.

[0010] Optionally, the user behavior feature at least includes one of the following: a video address feature watched by the user; a history record feature containing time information of the video watched by the user; and a behavior feature of the user not clicking on a pushed video; wherein the user behavior feature is used to determine a user portrait score feature, and the user portrait score feature is used for personalized video recommendation for the user.

[0011] Optionally, the preference feature of the user for the video comprises at least one of the following: a video type feature watched or purchased by the user; a video type feature ignored by the user; a video type feature liked by the user; a video type feature disliked by the user; a video type feature actively commented or collected by the user; and a video type feature negatively commented by the user.

[0012] Optionally, the attribute feature of the video comprises at least one of the following: a cover feature of the video; a title feature of the video; a type identification feature of the video; a producer feature of the video; and a language feature of the video.

[0013] Optionally, the window statistic feature of the video comprises at least one of the following: an exposure rate feature of the video in a statistical time window; a click rate feature of the video in the statistical time window; a play quantity feature of the video in the statistical time window; a play completion rate feature of the video in the statistical time window; a play progress feature of each play of the video in the statistical time window; and a performance feature of the video on a relay platform in the statistical time window, wherein the performance feature of the video on the relay platform comprises at least a play quantity feature of video relay.

[0014] A second aspect of the present disclosure discloses a method for video recommendation, comprising: obtaining a target user feature of a target user and a target video feature of a target video according to feature identification; determining a target video to be recommended to the target user from candidate videos according to the target user feature, the target video feature, and a video recommendation model trained by the video recommendation model training method provided in the first aspect; and sending video information of the target video to a user account of the target user.

[0015] The third aspect of the present disclosure provides a device for training a video recommendation model, the device comprising: a first obtaining module configured to obtain a feature sample training set and a feature sample test set according to video-related features and user-related features; wherein the video-related features comprise at least attribute features of a video and window statistical features of the video; the user-related features comprise at least user attribute features, user behavior features and user preference features for a video; a first determining module configured to train a preset model according to the feature sample training set to obtain a prediction model, and predict the feature sample test set through the prediction model to obtain a first set of feature prediction indicator values; wherein the first set of feature prediction indicator values comprises first prediction indicator values of N features; a second determining module configured to predict the feature sample test set through m evaluation models respectively to obtain m second sets of feature prediction indicator values; wherein each of the second sets of feature prediction indicator values comprises second prediction indicator values of the N features; a comparing and screening module configured to compare the N second prediction indicator values in the mth second set of feature prediction indicator values with corresponding first prediction indicator values in sequence, screen second features with second prediction indicator values smaller than the corresponding first prediction indicator values in the mth second set of feature prediction indicator values, and form an mth feature set; and a training module configured to train a video recommendation model using the m feature sets, and determine feature identifiers of training features according to the feature sets with the second prediction indicator values smaller than the first prediction indicator values; the feature identifiers are used to determine features for video recommendation using the video recommendation model.

[0016] Optionally, the device further comprises a third determining module configured to perform m types of preprocessing on the user feature sample training set, train a preset model using the preprocessed user feature sample training set to obtain m evaluation models.

[0017] Optionally, the third determining module is configured to perform at least one of the following: perform random ordering processing on features in the user feature sample training set, train the preset model using the feature sample training set after the random ordering processing to obtain a first evaluation model; perform zeroing processing on the features in the feature sample training set, train the preset model using the feature sample training set after the zeroing processing to obtain a second evaluation model; determine average values between the features in the feature sample training set and perform average value assignment processing on the features, train the preset model using the feature sample training set after the average value assignment processing to obtain a third evaluation model.

[0018] Optionally, the comparing and screening module is configured to at least one of: screen the second-class features with second prediction index values less than the first prediction index values of the corresponding sequence numbers in the second prediction index values of the first set of second-class feature prediction index values to form a first feature set; screen the second-class features with second prediction index values less than the first prediction index values of the corresponding sequence numbers in the second prediction index values of the second set of second-class feature prediction index values to form a second feature set; and screen the second-class features with second prediction index values less than the first prediction index values of the corresponding sequence numbers in the second prediction index values of the third set of second-class feature prediction index values to form a third feature set.

[0019] Optionally, the user attribute feature comprises at least one of: an age feature of the user; a gender feature of the user; a region feature of the user; a language feature of the user; and a device and / or network feature used by the user.

[0020] Optionally, the user behavior feature comprises at least one of: a video address feature watched by the user; a history record feature with time information of the video watched by the user; and a behavior feature of the user not clicking the pushed video.

[0021] Optionally, the user preference feature for the video comprises at least one of: a video type feature watched or purchased by the user; a video type feature ignored by the user; a video type feature liked by the user; a video type feature disliked by the user; a video type feature actively commented or collected by the user; and a video type feature negatively commented by the user.

[0022] Optionally, the video attribute feature comprises at least one of: a cover feature of the video; a title feature of the video; a window feature of the video; a type identification feature of the video; a producer feature of the video; and a language feature of the video.

[0023] Optionally, the window statistical feature of the video comprises at least one of: a click rate feature of the video within a statistical duration; a play count feature of the video within the statistical duration; a play completion rate feature of the video within the statistical duration; a play progress feature of each play of the video within the statistical duration; and a performance feature of the video on a retransmission platform within the statistical duration, wherein the performance feature of the video on the retransmission platform comprises at least a play count feature of the video retransmission.

[0024] The fourth aspect of the present disclosure provides a device for video recommendation, the device comprising: a second acquisition module configured to acquire target user features of a target user according to a feature identifier; a fourth determination module configured to determine a target video to be recommended to the target user from candidate videos according to the target user features and a video recommendation model trained by the video recommendation model training method provided in the first aspect; and a sending module configured to send video information of the target video to a user account of the target user.

[0025] The fifth aspect of the present disclosure provides an electronic device, comprising: a memory for storing processor-executable instructions; and a processor connected to the memory; wherein the processor is configured to execute the method for video recommendation model training provided in the first aspect and / or the method for video recommendation provided in the second aspect.

[0026] The sixth aspect of the present disclosure provides a non-transitory computer-readable storage medium, the computer-readable storage medium storing computer-executable instructions, the computer-executable instructions being executed by a processor to implement the method for video recommendation model training provided in the first aspect and / or the method for video recommendation provided in the second aspect.

[0027] The method for video recommendation model training provided by the embodiments of the present disclosure comprises: predicting m second-class feature prediction index value sets by respectively predicting the feature sample test set by m evaluation models; comparing N second prediction index values in the mth second-class feature prediction index value set with first prediction index values corresponding to the serial numbers in sequence, and screening second-class features smaller than the first prediction index values corresponding to the serial numbers in the second prediction index values in the mth second-class feature prediction index value set to form the mth feature set; compared with the feature importance in the prior art, the embodiments of the present disclosure can individually evaluate and screen features meeting the importance requirements to form a feature set, thereby improving the accuracy of the video recommendation model and reducing the training cost.

[0028] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0029] The accompanying drawings incorporated in and forming a part of the specification illustrate an embodiment consistent with the present disclosure and serve to explain the principles of the present disclosure together with the specification.

[0030] Figure 1 is a flowchart of the method for video recommendation model training according to an exemplary embodiment of the present disclosure;

[0031] Figure 2is a flowchart of a method for training a video recommendation model according to an example embodiment of the present disclosure;

[0032] Figure 3 is a flowchart of a method for training a video recommendation model according to an example embodiment of the present disclosure;

[0033] Figure 4 is a flowchart of a method for training a video recommendation model according to an example embodiment of the present disclosure;

[0034] Figure 5 is a flowchart of a method for training a video recommendation model according to an example embodiment of the present disclosure;

[0035] Figure 6 is a flowchart of a method for training a video recommendation model according to an example embodiment of the present disclosure;

[0036] Figure 7 is a structural diagram of an apparatus for training a video recommendation model according to an example embodiment of the present disclosure;

[0037] Figure 8 is a structural diagram of an apparatus for training a video recommendation model according to an example embodiment of the present disclosure;

[0038] Figure 9 is a structural diagram of an apparatus for training a video recommendation model according to an example embodiment of the present disclosure. DETAILED DESCRIPTION

[0039] The example embodiments will be described in detail herein with reference to the drawings. When the following description refers to arrangements in the drawings, identical numbers on different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following example embodiments do not represent all implementations consistent with the present disclosure. Instead, they are merely examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the attached application files.

[0040] In conjunction with Figure 1 As shown, the example embodiments of the present disclosure provide a method for training a video recommendation model, the method comprising:

[0041] In step S101, a feature sample training set and a feature sample test set are obtained according to video-related features and user-related features; wherein the video-related features at least include attribute features of the video and window statistical features of the video; and the user-related features at least include user attribute features, user behavior features, and user preference features for the video;

[0042] Step S102, training a preset model according to the feature sample training set to obtain a prediction model, and predicting the feature sample test set through the prediction model to obtain a first-class feature prediction index value set; wherein the first-class feature prediction index value set includes first prediction index values of N features;

[0043] Step S103, predicting the feature sample test set through m evaluation models respectively to obtain m second-class feature prediction index value sets; wherein each second-class feature prediction index value set includes second prediction index values of N features;

[0044] Step S104, comparing the N second prediction index values in the mth second-class feature prediction index value set with the first prediction index values of the corresponding serial numbers in sequence, screening the second-class features whose second prediction index values are smaller than the first prediction index values of the corresponding serial numbers in the mth second-class feature prediction index value set to form the mth feature set;

[0045] Step S105, training a video recommendation model using the m feature sets, and determining feature identifiers of training features according to the feature sets whose second prediction index values are smaller than the first prediction index values; the feature identifiers are used to determine the features used for video recommendation by the video recommendation model.

[0046] In the embodiments of the present disclosure, the video-related features are features related to the video itself, such as attribute features of the video, and features related to video playback, such as window statistical features of the video.

[0047] Here, the attribute features of the video are features related to the playback content of the video. The video attribute features include but are not limited to cover features, title features, and type identifier features. The window statistical features of the video include but are not limited to click rate features, play count features, and play completion rate features.

[0048] In some embodiments, the user-related features are features related to the attributes, behaviors, and preferences of the user for the video. For example, the features related to the attributes of the user include but are not limited to age features of the user, gender features of the user, regional features of the user, and language features of the user. The behavior features of the user can be related behavior features of the user watching the video, including but not limited to video address features watched by the user; history record features of the user watching the video containing time information; and behavior features of the user not clicking on the pushed video.

[0049] Here, the video-related features and the user-related features can be obtained from offline data of the video application software. Thus, the online playback of the video will not be affected.

[0050] In some embodiments, with respect to the step S101, a feature sample training set A1 is generated by program code according to video-related features and user-related features, and a feature sample test set B1 conforming to independent and identical distribution is generated at the same time. Here, the format of the feature sample training set A1 is a format supporting training in tf_trainer. tf_trainer is a training tool of Python software.

[0051] In some embodiments, with respect to the step S102, the preset model can be a DeepFM model. Here, the DeepFM model is a new neural network model generated by integrating a Factor Machine (FM) model and a Deep Neural Network (DNN) model, the FM model learns the interaction of low-order features, and the DNN learns the interaction of high-order features.

[0052] Here, the prediction model can be a Base model. The prediction model is used to predict the first-class feature prediction indicator values of the feature sample test set. The first prediction indicator values of N features in the first-class feature prediction indicator value set are represented by {base auc1, base auc2, …, base aucN}. After the N test features in the feature sample test set are input into the prediction model, the corresponding prediction indicator values base aucN are respectively predicted. Here, auc is an indicator representing the importance of features.

[0053] In some embodiments, with respect to the step S103, the evaluation model is a DeepFM model trained by m feature sample training sets.

[0054] Here, by predicting the feature sample test set by m evaluation models, m second-class feature prediction indicator value sets can be obtained. For example, the first second-class feature prediction indicator value set can be represented by {auc11, auc12, …, auc1N}, and the mth second-class feature prediction indicator value set can be represented by {aucm1, aucm2, …, aucmN}.

[0055] In some embodiments, with respect to the step S104, N second prediction indicator values in each of the m second-type feature prediction indicator value sets are compared with the first prediction indicator value of the corresponding sequence number. That is, the Nth prediction indicator value in the second-type feature prediction indicator value set {aucm1, aucm2, …, aucmN} is compared with the first-type feature prediction indicator value set {baseauc1, baseauc2, …, baseaucN} according to the corresponding sequence number N. The second-type feature of the corresponding sequence number that is smaller than the corresponding first prediction indicator value is used to form the mth feature set.

[0056] Here, the m and N are positive integers. For example, the m can be 1, 2, and 3. The N can be a positive integer with a value of 1 to 100, 1000, or 10000, etc.

[0057] For example, if auc11 is smaller than baseauc1, the first feature in the feature test set B1 corresponding to the sequence number 1 in auc11 is used to form the first feature of the first feature set. If aucmN is smaller than baseaucN, the Nth feature in the feature test set B1 corresponding to the sequence number N in aucmN is used to form the Nth feature of the mth feature set. Similarly, the above process is repeated.

[0058] In this way, by comparing one by one, the features with required importance can be selected as the features that can be used to train the video recommendation model.

[0059] In some embodiments, with respect to the step S105, the video recommendation model is trained by the m feature sets, and a video recommendation model that meets the importance requirement can be obtained. Therefore, the accuracy of the video recommendation model is higher.

[0060] Here, the feature identifier is used to identify the feature whose second prediction indicator value is smaller than the first prediction indicator value. Therefore, when training the video recommendation model, the feature identifier is used to filter the features used by the video recommendation model to perform video recommendation.

[0061] In this way, the recommendation accuracy of the video recommendation model is higher, and the demand of the user is better met.

[0062] In some embodiments, the video recommendation model is used for the recommendation of non-incentive video. The video recommendation model of the embodiments of the present disclosure is not used for the recommendation of incentive video.

[0063] Here, the incentive issuing type video recommendation refers to a video recommendation mode in which a user can obtain an incentive benefit by watching a recommended video. The incentive can include a red packet, a shopping voucher, a shopping discount, points, or game coins, etc. Due to the incentive issuing type video recommendation, watching is encouraged by issuing an incentive, and the screening accuracy of video recommendation is relaxed to some extent. For example, the FM model can be used alone, or the base model or the Shuffle model can be used alone for video recommendation for the incentive issuing type video recommendation.

[0064] For example, a red packet or game coins can be obtained by watching a video advertisement of a short video software or a video website.

[0065] For another example, the user can obtain an incentive benefit by watching the video of the incentive distribution type video recommendation.

[0066] The above-mentioned incentive issuing type video is mostly not recommended based on user-related features and video-related features, and the personalized needs of the user are not considered. The video recommendation model provided by the present disclosure needs to consider the personalized needs of the user, and therefore the video recommendation model of the embodiment of the present disclosure is not used for the above-mentioned incentive issuing type video.

[0067] Here, the non-incentive issuing type video refers to a video that the user cannot obtain an economic benefit by watching, but the non-incentive issuing type video can meet the personalized needs of the user, such as the need for rest, personal interests and hobbies, and artistic learning, etc. The video recommendation model provided by the embodiment of the present disclosure can be trained according to user-related features and video-related features, so as to actually meet the personalized needs of the user.

[0068] Therefore, the video recommendation model of the embodiment of the present disclosure is applied to the above-mentioned non-incentive issuing type video, and accurately recommends a video meeting the needs and expectations of the user to the user.

[0069] As shown in Figure 2 The method further includes the following steps:

[0070] In step S106, the m-class pre-processing is performed on the feature sample training set, the pre-processed feature sample training set is used to train a preset model, and m evaluation models are obtained.

[0071] In some embodiments, the difference between the evaluation model and the prediction model is that the feature sample training set used to train the evaluation model is pre-processed before the preset model is trained to obtain the evaluation model. Different types of pre-processing can obtain different types of evaluation models, and m evaluation models are obtained by training the DeepFM model with m feature sample training sets obtained by m types of pre-processing.

[0072] Here, the preprocessing for class m includes, but is not limited to, setting features to 0, averaging features, and randomly sorting features.

[0073] Thus, by performing different preprocessing steps on the feature sample training set, different preprocessed feature sample training sets can be obtained. This allows for the training of different evaluation models, making the assessment of feature importance more diverse.

[0074] like Figure 3 As shown in this embodiment, step S106 involves performing m-class preprocessing on the feature sample training set, and then training a preset model with the preprocessed feature sample training set to obtain m evaluation models, including at least one of the following:

[0075] Step S1061: Randomly sort the features in the feature sample training set, and train the preset model with the feature sample training set after random sorting to obtain the first evaluation model.

[0076] Step S1062: Zero out the features in the feature sample training set, and train the preset model with the zeroed feature sample training set to obtain the second evaluation model.

[0077] Step S1063: Determine the average value among the features in the feature sample training set and assign the average value to the features. Train the preset model with the feature sample training set after the average value assignment to obtain the third evaluation model.

[0078] In this embodiment of the disclosure, regarding step S1061, regarding the random sorting (shuffle) of the features in the feature sample training set, shuffle is a random function in Python software used to randomly sort all features in the feature sample training set.

[0079] Here, when the number of feature samples in the feature sample training set is large, the feature sample training set can be divided into batches, and then the feature samples in each batch can be randomly sorted. For example, random sorting can be performed on a certain number of feature samples. Specifically, a feature sample training set can be divided into multiple batches of equal size, for example, each batch of 10 feature samples can be a training set, and each batch of 10 feature samples can be randomly sorted.

[0080] Thus, batch preprocessing can improve processing efficiency and scalability.

[0081] Here, the first evaluation model, obtained by training the preset model with the feature sample training set after random sorting, will predict the features in the feature sample test set and obtain the first set of second-class feature prediction index values.

[0082] In some embodiments, regarding step S1062, setting the features in the feature sample training set to zero means filling the feature values ​​with zeros.

[0083] Here, the second evaluation model, obtained by training the preset model using the feature sample training set after setting it to 0, will predict the features in the feature sample test set, thus obtaining a second set of second-class feature prediction index values.

[0084] In some embodiments, regarding step S1063, determining the average value among the features in the feature sample training set can be done by taking the average value of features in different batches, and then taking the average value again based on the different average values ​​of features in different batches, and performing average value processing on each feature.

[0085] Here, the second evaluation model, obtained by training the preset model using the feature sample training set after averaging, will predict the features in the feature sample test set, resulting in a third set of second-class feature prediction index values.

[0086] Thus, by using different types of preprocessing in the above embodiments, the evaluation types become more diversified, and the predicted index values ​​of the second type of features predicted by the evaluation model also become more diverse, making the evaluation of the importance of features more comprehensive and accurate.

[0087] like Figure 4 As shown in this embodiment, step S104 involves selecting second-class features with corresponding indices smaller than the first prediction index value from the second prediction index values ​​in the m-th second-class feature prediction index value set to form the m-th feature set, which includes at least one of the following:

[0088] Step S1041: Select the second type of features with corresponding numbers less than the first prediction index value from the second prediction index value of the first second type of feature prediction index value set to form the first feature set;

[0089] Step S1042: Select the second type of features with corresponding numbers less than the first prediction index value from the second prediction index value of the second second type of feature prediction index value set to form the second feature set;

[0090] Step S1043: Select the second type features with corresponding serial numbers less than the first prediction index value from the second prediction index values ​​in the third set of second type feature prediction index values ​​to form the third feature set.

[0091] In the embodiments of the present disclosure, if the second prediction indicator value in the second set of feature prediction indicator values is greater than or equal to the first prediction indicator value, it indicates that the second feature corresponding to the sequence number is not important or has a negative impact. If the second prediction value in the second set of feature prediction indicator values is less than the first prediction value, the second feature corresponding to the sequence number meets the importance requirement.

[0092] In some embodiments, regarding steps S1041, S1042 and S1043, the smaller the second prediction indicator value is than the first prediction indicator value, the greater the importance of the feature is. When the number of features in the sample test set is large, for example, tens of thousands to tens of millions of features, a part of the second features with greater importance, which are smaller than the first prediction indicator value, can be selected to form the feature set. There is no need to select all the second features smaller than the first prediction indicator value to form the feature set.

[0093] For example, the second features can be sorted in descending order of importance according to the difference between the second prediction indicator value and the first prediction indicator value, wherein the greater the difference is, the greater the importance is. According to the importance sorting of the second features and the predetermined number, the second features with the predetermined number in the front of the sorting are selected to form the feature set.

[0094] In this way, the efficiency of the overall training video recommendation model and the availability of the second features in the feature set can be improved.

[0095] In the embodiments of the present disclosure, the user attribute features include at least one of the following:

[0096] An age feature of the user;

[0097] A gender feature of the user;

[0098] A region feature of the user;

[0099] A language feature of the user;

[0100] A device used by the user and / or a network used by the user.

[0101] In the embodiments of the present disclosure, the region feature of the user can be determined according to the geographical location set by the user in the video software. The region feature of the user can also be determined according to the positioning information of the user. The region feature of the user includes but is not limited to the following: a country feature to which the user belongs, a province feature to which the user belongs, and a city, street, etc. feature to which the user belongs. The region feature of the user here can be the region feature where the user is currently located.

[0102] Thus, the user can be recommended with relevant videos of the corresponding region according to the regional characteristics of the user. For example, when the user travels to a certain region, the video application software can recommend videos of relevant content such as food, landscape, weather, and news.

[0103] In some embodiments, the network characteristics used by the user can be determined through the device used by the user. The device characteristics used by the user include but are not limited to the characteristics of the user using a smart phone or a tablet computer. The network characteristics used by the user include but are not limited to the characteristics of the user using 5G, 4G, and a local area network.

[0104] Thus, the user can be recommended with videos of different resolutions according to the device used by the user and / or the network characteristics used by the user. The better the network environment, the higher the resolution.

[0105] In some embodiments, video content of different age groups can be recommended according to the age characteristics of the user. For example, age groups that prefer to watch policy news, age groups that prefer to watch sports news, and age groups that prefer to watch entertainment news.

[0106] Thus, the user can be recommended with different types of video content according to the age characteristics of the user.

[0107] In some embodiments, different gender-biased videos can be recommended to the user through the video recommendation model according to the gender characteristics of the user. For example, different clothing videos can be recommended according to different genders.

[0108] Thus, important user attribute characteristics are filtered from the user attribute characteristics for training the video recommendation model, which can enable the video recommendation model to recommend different videos according to the user attributes.

[0109] In the embodiments of the present disclosure, the behavior characteristics of the user include at least one of the following:

[0110] Video address characteristics watched by the user;

[0111] History record characteristics of the user watching videos containing time information;

[0112] Behavior characteristics of the user not clicking on the pushed videos;

[0113] The behavior characteristics of the user are used to determine the portrait score characteristics of the user, and the portrait score characteristics of the user are used for personalized video recommendation for the user.

[0114] In some embodiments, the user has different behavior characteristics for the recommended videos on the recommendation page when using the video application software.

[0115] For example, for part of the videos on the recommendation page, the user does not click to watch, and for another part of the videos, the user clicks to watch. The video recommendation model trained according to the behavior characteristics of the user not clicking to push the video can not recommend the video with the same or similar type again.

[0116] For another example, the video recommendation model trained according to the video address characteristics of the video watched by the user can recommend the video with the same address identifier next time. For example, the video with the same author or website output.

[0117] For another example, the video recommendation model trained according to the history record characteristics of the user watching the video containing time information can recommend the video similar to the history record according to the time distribution when recommending.

[0118] In some embodiments, the behavior characteristics of the user can also include user search behavior characteristics. Thus, the video recommendation model trained according to the user search behavior characteristics can recommend the video with different search frequencies to the user.

[0119] In some embodiments, the scoring process according to the comprehensive behavior characteristics of the user obtains the portrait score characteristics of the user. Thus, the video recommendation model trained according to the portrait score characteristics of the user can make personalized video recommendation to the user.

[0120] Thus, the personalized recommendation of the video recommendation model is more in line with the individuality of the user.

[0121] In the embodiments of the present disclosure, the preference characteristics of the user for the video include at least one of the following:

[0122] The video type characteristics watched or purchased by the user;

[0123] The video type characteristics ignored by the user;

[0124] The video type characteristics liked by the user;

[0125] The video type characteristics disliked by the user;

[0126] The video type characteristics actively commented or collected by the user;

[0127] The video type characteristics passively commented by the user.

[0128] In some embodiments, the user can take the behavior of watching or purchasing part of the videos. The video recommendation model trained according to the video type characteristics watched or purchased by the user can recommend more similar or same type videos to the user. Thus, the conversion rate of the video is improved.

[0129] In some embodiments, for the video type features that the user ignores, for example, the user clicks the operation of not being interested in the video, the number of times of recommending the user the video type can be reduced.

[0130] In some embodiments, after the user watches some videos, the user can like, actively comment or collect, and then according to the video type features involved in these preferences, for example, the video type features that the user likes, the video type features that the user actively comments or collects, the recommendation of the user about the videos of this type can be increased.

[0131] In some other embodiments, after the user watches some videos, the user can click the operation of not being interested in the video, and then according to the video type features involved in these preferences, for example, the video type features that the user likes, the video type features that the user actively comments or collects, the recommendation of the user about the videos of this type can be reduced.

[0132] In this way, the video preference features of the user are used to train the video recommendation model, which can increase the degree of fitting of the video recommendation model to the personal interests of the user, so as to more accurately recommend videos to the user.

[0133] In the embodiments of the present disclosure, the attribute features of the video at least include one of the following:

[0134] the cover feature of the video;

[0135] the title feature of the video;

[0136] the type identification feature of the video;

[0137] the producer feature of the video;

[0138] the language feature of the video.

[0139] In some embodiments, the cover feature and the title feature of the video directly affect whether the user clicks to watch the video. The type identification feature and the producer feature of the video indirectly affect the search and the behavior of watching the video of the user.

[0140] In one embodiment, for the language feature of the video, some users have the demand of learning different languages, and the user can click the language type that is often learned. For the user who does not have the demand of learning different languages, the video in the language that meets the user's language can be recommended to the user.

[0141] For example, according to the language type that the user can involve, the video of the same language type can be recommended. For example, the user often uses Chinese and English, and the video with Chinese and / or English content can be recommended. Or the user often uses Chinese and often learns Spanish, and the video with Chinese and / or Spanish language content can be recommended. Therefore, the embodiments of the present disclosure are not limited to the above examples.

[0142] In the embodiments of the present disclosure, the window statistical features of the video at least include one of the following:

[0143] an exposure rate feature of the video in a statistical time window;

[0144] a click rate feature of the video in the statistical time window;

[0145] a play quantity feature of the video in the statistical time window;

[0146] a play completion rate feature of the video in the statistical time window;

[0147] a play progress feature of each play of the video in the statistical time window;

[0148] a performance feature of the video on a relay platform in the statistical time window; wherein the performance feature of the video on the relay platform at least includes a play quantity feature of a video relay.

[0149] In the embodiments of the present disclosure, the statistical time window refers to a statistical time length. The statistical time window includes but is not limited to 3 days, 4 days, 5 days, 6 days, and 7 days.

[0150] In some embodiments, the exposure rate indicates the frequency of the video being seen by a user. The click rate refers to the ratio of the number of times of being clicked to the number of times of being displayed. The play quantity refers to the number of times of playing the video. The play completion rate refers to the ratio of the number of times of the video being completely played to the play quantity. The play progress of each play refers to the time from the start time of the video to the time when the user stops watching the video.

[0151] In some embodiments, the relay platform refers to the performance feature of the video produced in a first video application software or platform played on a third-party video application software or platform, which includes but is not limited to the play quantity feature of the video relay, the click rate feature of the video relay, and the play completion rate of the video.

[0152] In this way, the video recommendation model can be better trained by the window statistical features of the video, so as to obtain more accurate video recommendations of the video recommendation model.

[0153] In combination with Figure 5 As shown in the drawings, in the embodiments of the present disclosure, a method for video recommendation is provided, which includes:

[0154] In step S201, target user features of a target user and target video features of a target video are obtained according to feature identifiers.

[0155] Step S202, determining a target video to be recommended to the target user from the candidate videos according to the target user features, the target video features, and a video recommendation model trained by the video recommendation model training method provided in the above embodiments.

[0156] Step S203, sending video information of the target video to a user account of the target user.

[0157] In the embodiments of the present disclosure, regarding the step S201, the feature identifier identifies the target user features and the target video features. Therefore, the target user features of the target user and the target video features of the target video can be determined according to the feature identifier.

[0158] In some embodiments, regarding the step S202, the video recommendation model predicts the target user features and the target video features, so that the target video to be recommended to the target user can be determined from the candidate videos.

[0159] Here, the target video and the target user can be a user or a video of an online video application software or platform.

[0160] In some embodiments, regarding the step S203, the video information of the target video includes but is not limited to video cover information, video title information, video address information, video content information, video producer information, video play quantity, and like information. Sending the video information of the target video to the user account of the target user can make the target video be recommended to the user, and the user can quickly select according to the video information.

[0161] In this way, the video recommendation model trained by the offline data can accurately determine the target video to be recommended to the target user online.

[0162] In the field of machine learning, feature engineering is the most important. A widely circulated sentence in the industry is: "data and features determine the upper limit of machine learning, and models and algorithms are only used to approach this upper limit". Through feature engineering, features are extracted from raw data to the maximum extent for algorithms and models. And feature engineering includes feature preprocessing, feature selection, and feature evaluation.

[0163] Feature evaluation is to verify the necessity of using the selected features for the model. Through evaluation, we find out the most useful feature subset, eliminate irrelevant or redundant features, thereby reducing the number of features, improving the accuracy of the model, and reducing the running time.

[0164] Current feature evaluation schemes and their shortcomings:

[0165] 1. Existing feature evaluation refers to the overall evaluation of the generated features, which occurs after feature selection and feature encoding, before the final quantitative analysis, and needs to be linearly evaluated from various aspects such as feature coverage, feature dimension, qualitative analysis, etc.

[0166] 2. Coverage refers to the proportion of videos and users that can successfully generate features in the total videos and users. The coverage is naturally different for different feature extraction and generation methods.

[0167] 3. In actual engineering implementation, the dimension of the feature is a very important consideration factor, because the dimension of the feature is related to the expression ability of the feature, the applicable model and the complexity of the calculation. If the dimension of the feature is too low, the expression ability of the feature is obviously limited. If the dimension of the feature is too high, not only the calculation amount will be upgraded, but also the problem of dimension disaster is easily caused. Therefore, the dimension of the feature needs to be adapted to the calculation framework and the data size.

[0168] 4. Qualitative analysis, for interpretable features, sampling-based qualitative analysis is very useful, such as using video tags to generate features, and each dimension is interpretable.

[0169] 5. Quantitative analysis, feature is poured into the model, and quantitative analysis is performed on the result. For a trained model, what is important is to evaluate its generalization ability.

[0170] The embodiment of the present disclosure is to evaluate each feature separately, and respectively use the methods of setting the feature to 0 (Zero), using the mean (Mean), and shuffling (Shuffle) to process the feature.

[0171] Then, by training the model, it is compared whether the loss value (loss) or the index value (auc) changes before and after, and if there is no change, it means that the feature is indeed not important and can be eliminated. The whole evaluation process is efficient and the reliability is improved.

[0172] In combination with Figure 6 As shown in the figure, the embodiment of the present disclosure includes the following contents:

[0173] (1) Generate a sample set tfrecord Sample Set A (the current sample set is in the format supporting training in tf_trainer) containing N features through program code as a training set, and simultaneously generate a corresponding test set Sample Set B conforming to independent and identical distribution; here, the sample set tfrecord Sample Set A can be expressed by the feature sample training set in the above embodiment, and the test set Sample Set B can be expressed by the feature sample test set in the above embodiment;

[0174] (2) Train the Base model through the network structure model of DeepFM by the tfrecord Sample Set A, and then predict and evaluate the test set Sample Set B by the Base model to obtain base auc;

[0175] (3) Process each batch, train a new model by sequentially randomly sorting the feature values of N features, and predict the test set Sample Set B by the models corresponding to the features processed in different ways to obtain corresponding auc1...N. Here, processing by batch can provide processing efficiency, and the scalability can be improved without processing from the sample;

[0176] (4) Perform the processing of filling 0 and taking the mean value according to the step 3, obtain the models corresponding to the features processed, and then evaluate the same test set to finally obtain corresponding auc1...N;

[0177] (5) Put the auc obtained by evaluating each feature by the three ways together for comprehensive comparison. If the auc evaluated by the three ways is negative relative to base auc, the feature can be filtered out. When the model is trained too much, the features that have a greater impact on auc can be screened, and the availability and efficiency of the features are improved as a whole.

[0178] In combination with Figure 7 As shown in the figure, in the embodiment of the disclosure, a device 300 for training a video recommendation model is provided, and the device 300 comprises:

[0179] A first obtaining module 301 is configured to obtain a feature sample training set and a feature sample test set according to video-related features and user-related features. The video-related features at least include attribute features of a video and window statistical features of the video. The user-related features at least include user attribute features, user behavior features, and user preference features for a video.

[0180] A first determining module 302 is configured to train a preset model to obtain a prediction model according to the feature sample training set, and predict the feature sample test set by the prediction model to obtain a first-class feature prediction index value set. The first-class feature prediction index value set includes first prediction index values of N features.

[0181] A second determining module 303 is configured to predict the feature sample test set by m evaluation models respectively to obtain m second-class feature prediction index value sets. Each second-class feature prediction index value set includes second prediction index values of N features.

[0182] The comparison and screening module 304 is configured to compare the N second prediction indicator values in the mth second characteristic prediction indicator value set with the first prediction indicator values in corresponding sequences one by one, and screen the second characteristics in the second prediction indicator values in the mth second characteristic prediction indicator value set that are smaller than the first prediction indicator values in corresponding sequences to form an mth characteristic set.

[0183] The training module 305 is configured to train a video recommendation model by using the m characteristic sets, and determine a feature identifier of a training feature according to the characteristic set in which the second prediction indicator value is smaller than the first prediction indicator value; the feature identifier is used to determine a feature used for video recommendation by using the video recommendation model.

[0184] In combination with Figure 8 As shown in the figure, in the embodiment of the present disclosure, the device further comprises:

[0185] The third determination module 306 is configured to perform m-class preprocessing on the user feature sample training set, train a preset model by using the user feature sample training set after the preprocessing, and obtain m evaluation models.

[0186] In the embodiment of the present disclosure, the third determination module 306 is at least configured as one of the following:

[0187] The third determination module 306 is configured to perform random ordering processing on the features in the user feature sample training set, train the preset model by using the feature sample training set after the random ordering processing, and obtain a first evaluation model.

[0188] The third determination module 306 is configured to perform zero processing on the features in the user feature sample training set, train the preset model by using the feature sample training set after the zero processing, and obtain a second evaluation model.

[0189] The third determination module 306 is configured to determine the average values between the features in the user feature sample training set, perform average value assignment processing on the features, train the preset model by using the feature sample training set after the average value assignment processing, and obtain a third evaluation model.

[0190] In the embodiment of the present disclosure, the comparison and screening module 304 is at least configured as one of the following:

[0191] The comparison and screening module 304 is configured to screen the second characteristics in the second prediction indicator values in the first second characteristic prediction indicator value set that are smaller than the first prediction indicator values in corresponding sequences to form a first characteristic set.

[0192] The comparison and screening module 304 is configured to screen the second characteristics in the second prediction indicator values in the second second characteristic prediction indicator value set that are smaller than the first characteristic prediction indicator values in corresponding sequences to form a second characteristic set.

[0193] Filtering, in a second prediction index value of a third second-type feature prediction index value set, a second-type feature smaller than a corresponding serial number of the first-type feature prediction index value, to form a third feature set.

[0194] In the embodiments of the present disclosure, the user attribute feature at least includes one of the following:

[0195] An age feature of the user;

[0196] A gender feature of the user;

[0197] A region feature of the user;

[0198] A language feature of the user;

[0199] A device and / or network feature used by the user.

[0200] In the embodiments of the present disclosure, the behavior feature of the user at least includes one of the following:

[0201] A video address feature watched by the user;

[0202] A history record feature of the user watching the video containing time information;

[0203] A behavior feature of the user not clicking a pushed video;

[0204] The behavior feature of the user is used to determine a user portrait score feature, and the user portrait score feature is used for personalized video recommendation for the user.

[0205] In the embodiments of the present disclosure, the user preference feature for the video at least includes one of the following:

[0206] A video type feature watched or purchased by the user;

[0207] A video type feature ignored by the user;

[0208] A video type feature liked by the user;

[0209] A video type feature disliked by the user;

[0210] A video type feature actively commented or collected by the user;

[0211] A video type feature passively commented by the user.

[0212] In the embodiments of the present disclosure, the attribute feature of the video at least includes one of the following:

[0213] A cover feature of the video;

[0214] A title feature of the video;

[0215] A window feature of the video;

[0216] a type identification feature of the video;

[0217] a producer feature of the video;

[0218] a language feature of the video.

[0219] In the embodiments of the present disclosure, the window statistical feature of the video at least includes one of the following:

[0220] a click rate feature of the video within a statistical duration;

[0221] a play count feature of the video within a statistical duration;

[0222] a play completion rate feature of the video within a statistical duration;

[0223] a play progress feature of each play of the video within a statistical duration;

[0224] a performance feature of the video on a relay platform within a statistical duration; wherein the performance feature of the video on the relay platform at least includes a play count feature of video relay.

[0225] In combination with Figure 9 the present disclosure, an apparatus 400 for video recommendation is provided, and the apparatus includes:

[0226] a second acquisition module 401 configured to acquire a target user feature of a target user according to a feature identifier;

[0227] a fourth determination module 402 configured to determine a target video to be recommended to the target user from candidate videos according to the target user feature and a video recommendation model trained by the video recommendation model training method;

[0228] a sending module 403 configured to send video information of the target video to a user account of the target user.

[0229] The embodiments of the present disclosure provide an electronic device, which includes:

[0230] a memory for storing processor-executable instructions;

[0231] a processor connected with the memory;

[0232] The processor is configured to execute the method for training the video recommendation model provided in any of the foregoing technical solutions.

[0233] The processor can include various types of storage media, which is a non-transitory computer storage medium and can continue to store information thereon after the electronic device is powered off.

[0234] The processor can be connected with the memory through a bus or the like for reading an executable program stored on the memory, for example, capable of executing the method of the foregoing one or more technical solutions.

[0235] An embodiment of the present disclosure provides a structure of an electronic device. The electronic device includes a processing component further comprising one or more processors, and a memory resource represented by a memory for storing instructions executable by the processing component, for example, an application program. The application program stored in the memory can include one or more than one module each corresponding to a set of instructions. In addition, the processing component is configured to execute the instructions to perform any method of the foregoing method applied to the electronic device, for example, the method of the foregoing one or more technical solutions.

[0236] The electronic device can also include a power supply component configured to perform power management of the electronic device, a wired or wireless network interface configured to connect the electronic device to a network, and an input / output (I / O) interface. The electronic device can operate based on an operating system stored in the memory, for example, Windows Server TM, Mac OS X TM, Unix TM, Linux TM, FreeBSD TM or the like.

[0237] An embodiment of the present disclosure provides a non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by a processor of a computer, the computer can execute the method of training the video recommendation model of the foregoing one or more technical solutions.

[0238] Other embodiments of the present disclosure will be apparent to those skilled in the art with the consideration of the specification and practice of the disclosed application. The present disclosure is intended to cover any variations, uses or adaptive changes of the present disclosure following the general principles of the present disclosure and including common knowledge or conventional technical means in the art not disclosed in the present disclosure. The specification and examples are only regarded as exemplary, and the true scope and spirit of the present disclosure are indicated by the appended application documents.

[0239] It should be understood that the present disclosure is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended application documents.

Claims

1. A method for training a video recommendation model, the method comprising: The method comprises: According to the video-related features and the user-related features, a feature sample training set and a feature sample test set are obtained; wherein the video-related features at least include attribute features of the video and window statistical features of the video; the user-related features at least include user attribute features, user behavior features and user preference features of the video; A preset model is trained according to the feature sample training set to obtain a prediction model, and the feature sample test set is predicted by the prediction model to obtain a first-class feature prediction index value set; wherein the first-class feature prediction index value set includes first prediction index values of N features; The feature sample test set is predicted by m evaluation models respectively to obtain m second-class feature prediction index value sets; wherein each second-class feature prediction index value set includes second prediction index values of N features; N second prediction index values in the mth second-class feature prediction index value set are compared with the first prediction index values of the corresponding serial numbers in sequence, and the second-class features with the second prediction index values smaller than the first prediction index values of the corresponding serial numbers in the mth second-class feature prediction index value set are screened to form the mth feature set; A video recommendation model is trained by using the m feature sets, and the feature identification of the training feature is determined according to the feature set with the second prediction index value smaller than the first prediction index value; the feature identification is used to determine the features used for video recommendation by the video recommendation model.

2. The method of Claim 1, wherein, The method further comprises: The feature sample training set is preprocessed in m classes, and the feature sample training set after the preprocessing is used to train a preset model to obtain m evaluation models.

3. The method of Claim 2, wherein, The preprocessing of the feature sample training set in m classes and the training of the preset model by using the feature sample training set after the preprocessing to obtain m evaluation models at least include one of the following: The features in the feature sample training set are randomly sorted, and the feature sample training set after the random sorting is used to train the preset model to obtain a first evaluation model; The features in the feature sample training set are set to zero, and the feature sample training set after the zero setting is used to train the preset model to obtain a second evaluation model; The average value between the features in the feature sample training set is determined, and the average value is assigned to the features, and the feature sample training set after the assignment of the average value is used to train the preset model to obtain a third evaluation model.

4. The method of Claim 1, wherein, The screening of the second-class features with the second prediction index values smaller than the first prediction index values of the corresponding serial numbers in the mth second-class feature prediction index value set to form the mth feature set at least includes one of the following: The second-class features with the second prediction index values smaller than the first prediction index values of the corresponding serial numbers in the first second-class feature prediction index value set are screened to form the first feature set; The second-class features with the second prediction index values smaller than the first prediction index values of the corresponding serial numbers in the second second-class feature prediction index value set are screened to form the second feature set; Filtering the second-class features with a second prediction index value less than the corresponding sequence number of the first prediction index value in the second prediction index value of the third second-class feature prediction index value set to form a third feature set.

5. The method of training a video recommendation model according to claim 1, wherein, The user attribute feature includes at least one of the following: An age feature of the user; A gender feature of the user; A region feature of the user; A language feature of the user; A device used by the user and / or a network used by the user feature.

6. The method of training a video recommendation model according to claim 1, wherein, The user behavior feature includes at least one of the following: A video address feature watched by the user; A history record feature of the user watching a video containing time information; A behavior feature of the user not clicking a pushed video; The user behavior feature is used to determine a user portrait score feature, and the user portrait score feature is used for personalized video recommendation for the user.

7. The method of training a video recommendation model according to claim 1, wherein, The user preference feature for a video includes at least one of the following: A video type feature watched or purchased by the user; A video type feature ignored by the user; A video type feature liked by the user; A video type feature disliked by the user; A video type feature actively commented or collected by the user; A video type feature negatively commented by the user.

8. The method of training a video recommendation model according to claim 1, wherein, The video attribute feature includes at least one of the following: A cover feature of the video; A title feature of the video; A type identification feature of the video; A producer feature of the video; A language feature of the video.

9. The method of training a video recommendation model according to claim 1, wherein, The video window statistical feature includes at least one of the following: An exposure rate feature of the video in a statistical time window; A click rate feature of the video in the statistical time window; A play quantity feature of the video in the statistical time window; A play completion rate feature of the video in the statistical time window; A play progress feature of each play of the video in the statistical time window; A performance feature of the video on a relay platform in the statistical time window; wherein the performance feature of the video on the relay platform includes at least a video relay play quantity feature.

10. A method of video recommendation, the method comprising: The method includes: According to the feature identifier, obtaining a target user feature of a target user and a target video feature of a target video; According to the target user feature, the target video feature, and the video recommendation model trained in any one of claims 1 to 9, determining a target video to be recommended to the target user from the alternative videos; Sending video information of the target video to a user account of the target user.

11. An apparatus for video recommendation model training, the apparatus comprising: The device includes: A first acquisition module configured to obtain a feature sample training set and a feature sample test set according to video-related features and user-related features; wherein the video-related features include at least video attribute features and video window statistical features; and the user-related features include at least user attribute features, user behavior features, and user preference features for a video; A first determination module configured to train a preset model to obtain a prediction model according to the feature sample training set, and to predict the feature sample test set through the prediction model to obtain a first-class feature prediction index value set; wherein the first-class feature prediction index value set includes first prediction index values of N features. The second determining module is configured to predict the feature sample test set by using the m evaluation models respectively to obtain m second-class feature prediction index value sets; each of the second-class feature prediction index value sets includes second prediction index values of N features; The comparison and screening module is configured to compare the N second prediction index values in the mth second-class feature prediction index value set with the first prediction index values of the corresponding serial numbers in sequence, and screen the second-class features with the second prediction index values smaller than the first prediction index values of the corresponding serial numbers in the mth second-class feature prediction index value set to form the mth feature set. The training module is configured to train the video recommendation model by using the m feature sets, and determine feature identifiers of training features according to the feature sets with the second prediction index values smaller than the first prediction index values.

12. The apparatus for video recommendation model training according to claim 11, wherein, The device further comprises: The third determining module is configured to perform m-class preprocessing on the user feature sample training set, train a preset model by using the preprocessed user feature sample training set, and obtain m evaluation models.

13. An apparatus for video recommendation, the apparatus comprising: The device comprises: The second obtaining module is configured to obtain target user features of a target user according to the feature identifiers. The fourth determining module is configured to determine a target video to be recommended to the target user from the candidate videos according to the target user features and the video recommendation model trained in any one of claims 1 to 9. The sending module is configured to send video information of the target video to a user account of the target user.

14. An electronic device, comprising: It comprises: A memory for storing processor-executable instructions; A processor connected with the memory; The processor is configured to execute the method for training the video recommendation model and / or the method for video recommendation provided in any one of claims 1 to 9.

15. A non-transitory computer-readable storage medium, comprising: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by the processor to implement the method for training the video recommendation model and / or the method for video recommendation provided in any one of claims 1 to 9. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by the processor to implement the method for training the video recommendation model and / or the method for video recommendation provided in any one of claims 1 to 9.

Images