Training method of recall model, recall method, processing device and storage medium

Abstract

Embodiments of the present disclosure disclose a training method of a recall model, a recall method, a screening method and device of material samples, a processing device and a storage medium. The method comprises: obtaining at least two material samples labeled with predetermined feature labels; wherein different types of material samples contain different predetermined features; determining an initial recall model corresponding to the predetermined features based on the predetermined features contained in the material samples; inputting the material samples into the initial recall model corresponding to the predetermined features contained therein for training until a predetermined condition is met, to obtain a first trained recall model; and fusing the first recall models obtained by training at least two material samples to obtain a second recall model for predicting a second candidate path corresponding to at least one recommended material. When the recall model is used for content recommendation, it can better adapt to user needs, accurately recommend personalized content to users, and provide a better experience for users.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of content recommendation, and in particular to a training method of a recall model, a recall method, a screening method of material samples, a device, a processing apparatus and a storage medium. BACKGROUND

[0002] In the field of content recommendation, recall, as the first layer of the funnel structure of the recommendation system, needs to quickly screen out ten thousand level materials from the million or even ten million level material library to pass to the subsequent link. In the information flow scenario, compared with the video flow scenario, there are more types of materials. For example, in the micro-blog popular recommendation flow, in addition to videos, there are also text, long pictures, articles, voting and other material types, and the scene is more complex. The content recommendation method in the information flow scenario cannot adapt to the user demand, and brings bad experience to the user. SUMMARY

[0003] Therefore, the present disclosure discloses a training method of a recall model, a recall method, a screening method of material samples, a device, a processing apparatus and a storage medium, so as to better adapt to the user demand, accurately recommend personalized content to the user, and bring better experience to the user.

[0004] According to a first aspect of the present disclosure, a training method of a recall model is provided, the method is executed by a training end, and the method comprises:

[0005] Obtaining at least two material samples labeled with predetermined feature labels; wherein different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features and interest matching features;

[0006] Based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined;

[0007] Inputting the material samples into the initial recall model corresponding to the contained predetermined features for training until a predetermined condition is met, to obtain a first recall model after training; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined features;

[0008] Fusing the first recall models obtained by training at least two material samples to obtain a second recall model used to predict a second candidate path corresponding to at least one to-be-recommended material.

[0009] In one embodiment, the method further comprises:

[0010] Sending the model parameters of the second recall model after training to a training parameter server PS;

[0011] wherein the PS is configured to update and / or synchronize the model parameters.

[0012] In an embodiment, the recall model comprises an Expectation-Maximization (EM) model; the initial recall model is trained with the material sample input corresponding to the predetermined features until a predetermined condition is met, to obtain a trained first recall model, comprising:

[0013] controlling path parameters of the initial recall model to be unchanged, and training deep neural network (DNN) parameters of the initial recall model based on the material sample until the predetermined condition is met;

[0014] keeping the trained DNN parameters unchanged, and updating path parameters of the initial recall model until the predetermined condition is met, to obtain the trained first recall model.

[0015] In an embodiment, the method further comprises:

[0016] obtaining path parameters of the initial recall model from a training parameter server (PS).

[0017] In an embodiment, the method further comprises:

[0018] receiving information of a second key-value pair sent by a timing task end; the second key-value pair comprises a path corresponding to a material sample that is eliminated and a quantity of material samples under the path;

[0019] based on the path and the quantity indicated by the information of the second key-value pair, determining a penalty term that makes a material distribution corresponding to the first candidate path balanced;

[0020] in response to a loss function converging, determining that the predetermined condition is met; wherein the loss function comprises the penalty term.

[0021] According to a second aspect of the embodiments of the present disclosure, a screening method of a material sample is provided, the method is performed by a timing task end, and the method comprises:

[0022] obtaining a material identifier of a material sample from a material library; wherein the material sample in the material library is used for training of a recall model; different kinds of material samples comprise different predetermined features, and the predetermined features are one of the following: a click feature, an interaction feature, a time length feature, and an interest matching feature;

[0023] based on the material identifier, in a case where a first key-value pair matching the material identifier is obtained from a training parameter server (PS), determining a path included in the first key-value pair; wherein the first key-value pair comprises a material identifier of a material sample that is eliminated and a path corresponding to the material identifier;

[0024] determining the material sample in the path as a material sample eliminated in the material library;

[0025] sending information of the eliminated material sample to the training end and / or the service end.

[0026] In an embodiment, the method further comprises:

[0027] generating, based on the first key-value pair, a second key-value pair between the path and the number of material samples under the path;

[0028] sending information of the second key-value pair to the training end.

[0029] According to a third aspect of embodiments of the present disclosure, a recall method is provided, the method being performed by a service end, and the method comprising:

[0030] obtaining first request information containing target features, wherein the target features include at least two of the following predetermined features: click features, interaction features, time length features, and interest matching features;

[0031] based on the target features, predicting, by using a second recall model trained by any embodiment of the present disclosure, a second candidate path corresponding to at least one to-be-recommended material, to obtain a first recall result.

[0032] In an embodiment, the method further comprises:

[0033] receiving information of an eliminated material sample sent by a timing task end;

[0034] generating, based on the information of the eliminated material sample and first key-value pairs formed by material identifiers and corresponding paths, an inverted index of the first key-value pairs;

[0035] determining, based on the inverted index, to-be-recommended materials corresponding to the second candidate path.

[0036] In an embodiment, the method further comprises:

[0037] inputting to-be-recommended materials contained in the first recall result into an online deep model to obtain a second recall result;

[0038] The online deep model is configured to score and / or sort the to-be-recommended materials to filter out target materials for recommending to users.

[0039] According to a fourth aspect of embodiments of the present disclosure, a training device of a recall model is provided, the device comprising:

[0040] The acquisition module is configured to acquire at least two material samples labeled with predetermined feature labels; wherein different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features, and interest matching features;

[0041] The determination module is configured to determine an initial recall model corresponding to the predetermined feature based on the predetermined feature contained in the material sample.

[0042] The training module is configured to input the material sample into the initial recall model corresponding to the predetermined feature for training until a predetermined condition is met, to obtain a first trained recall model; wherein a prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one recommended material corresponding to the predetermined feature.

[0043] The fusion module is configured to fuse the first recall models obtained by training at least two material samples, to obtain a second recall model used to predict a second candidate path corresponding to at least one recommended material.

[0044] According to a fifth aspect of an embodiment of the present disclosure, a device for screening material samples is provided, and the device comprises:

[0045] The acquisition module is configured to acquire a material identifier of a material sample from a material library; wherein the material samples in the material library are used for training of a recall model; different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features, and interest matching features.

[0046] The determination module is configured to determine a path included in a first key-value pair based on the material identifier, in a case where the first key-value pair matching the material identifier is acquired from a training parameter server PS; wherein the first key-value pair includes a material identifier of an eliminated material sample and a path corresponding to the material identifier; and determine the material sample in the path as the eliminated material sample in the material library.

[0047] The sending module is configured to send information of the eliminated material sample to a training end and / or a service end.

[0048] According to a sixth aspect of an embodiment of the present disclosure, a recall device is provided, and the device comprises:

[0049] The acquisition module is configured to acquire first request information containing a target feature, wherein the target feature includes at least two of the following predetermined features: click features, interaction features, time length features, and interest matching features.

[0050] The prediction module is configured to predict, based on the target feature, a second candidate path corresponding to at least one to-be-recommended material by using the second recall model trained according to any one of claims 1 to 5, to obtain a first recall result.

[0051] According to a seventh aspect of embodiments of the present disclosure, a processing device is provided, and the processing device comprises:

[0052] a memory configured to store an executable program;

[0053] a processor configured to execute the executable program stored in the memory to implement the method according to any one of the embodiments of the present disclosure.

[0054] According to an eighth aspect of embodiments of the present disclosure, a computer storage medium is provided, and the computer storage medium stores an executable program, and the executable program is executed by a processor to implement the method according to any one of the embodiments of the present disclosure.

[0055] In embodiments of the present disclosure, at least two material samples labeled with predetermined features are obtained; different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features, and interest matching features; an initial recall model corresponding to the predetermined features is determined based on the predetermined features contained in the material samples; the material samples are input into the initial recall model corresponding to the predetermined features for training until a predetermined condition is met, to obtain a trained first recall model; the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material; and the first recall models trained by at least two types of material samples are fused to obtain a second recall model used to predict a second candidate path corresponding to at least one to-be-recommended material. Here, since different material samples are input into the initial recall model corresponding to the predetermined features for training until a predetermined condition is met, to obtain a trained first recall model, a plurality of first recall models trained for material samples containing different predetermined features can be obtained. Since the first recall models trained by at least two types of material samples are fused, the prediction results of different first recall models are used to jointly determine the second candidate path, which is essentially to fuse the prediction results of different first recall models. Compared with the way of obtaining prediction results by using a single recall model, the prediction accuracy and adaptability of the results can be improved. When content recommendation is performed by using the second recall model, the personalized content can be accurately recommended to the user, and the user can have a better experience. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1A flowchart of a training method of a recall model according to an exemplary embodiment.

[0057] Figure 2 A schematic diagram of an index structure according to an exemplary embodiment.

[0058] Figure 3 A schematic diagram of a forward propagation according to an exemplary embodiment.

[0059] Figure 4 A flowchart of a training method of a recall model according to an exemplary embodiment.

[0060] Figure 5 A flowchart of a recall method according to an exemplary embodiment.

[0061] Figure 6a A flowchart of a recall method according to an exemplary embodiment.

[0062] Figure 6b A flowchart of a recall method according to an exemplary embodiment.

[0063] Figure 6c A flowchart of a recall method according to an exemplary embodiment.

[0064] Figure 6d A flowchart of a recall method according to an exemplary embodiment.

[0065] Figure 7 A schematic diagram of a training apparatus of a recall model according to an exemplary embodiment.

[0066] Figure 8 A schematic diagram of a training apparatus of a recall model according to an exemplary embodiment.

[0067] Figure 9 A schematic diagram of a recall apparatus according to an exemplary embodiment. DETAILED DESCRIPTION

[0068] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application. All other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0069] In the following description, reference is made to "some embodiments", which describe only a subset of all possible embodiments, but which can be understood to be the same or different subsets of all possible embodiments, and which can be combined with each other, without conflict, unless otherwise indicated.

[0070] In the following description, the terms "first\second\third" are only to distinguish similar objects, and do not represent a specific order of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence, if permitted, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein.

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the application only, and is not intended to be limiting of the application.

[0072] In order to better understand the embodiments of the present disclosure, the scenarios in the related art are described below through exemplary embodiments:

[0073] In one embodiment, the deep model-based recall is to model the recall problem as a nearest neighbor search problem in a vector space, both the user and the item are represented by vectors, the item vectors are indexed offline, and the candidate item set is obtained by fuzzy nearest neighbor search method when online service. This nearest neighbor search (ANN, Approximately Nearest Neighbor Search) recall method is not an end-to-end recall method. With this method, since the user tower and the item tower are separated, the interaction is simple, the model expression ability is weak, and the user-item training sample cannot be directly optimized. The deep recall (Deep Retrieval) algorithm is a kind of deep retrieval model, which directly models the user-item sample, maps the item to the corresponding index by a method similar to clustering, and obtains the corresponding index by reasoning the user's information when online service. The items in the index cluster are selected as the recall results and transmitted to the subsequent link.

[0074] In an embodiment, the Deep Retrieval algorithm is mainly applied in the video stream recommendation scenario, and compared with the video stream scenario, the information stream scenario has more types of materials. Taking the microblog popular recommendation stream as an example, in addition to videos, there are also text and images, long texts, articles, voting and other material types, and the scene is more complex. In the video stream scenario, the positive samples used for training the model of the Deep Retrieval algorithm are videos that have been interacted by users, and this target is too broad in the information stream scenario. If it is directly used, it is not conducive to the mapping of the model and the user interest. At the same time, in the online learning streaming scenario, the Deep Retrieval algorithm does not explicitly propose how to eliminate expired materials, which will cause the material count in an index path to only increase and not decrease, so that the penalty term becomes larger and larger, so that the materials in the index path are less likely to be updated or not updated, which affects the model training and prediction effect.

[0075] As shown in Figure 1 The method provided in the embodiment of the present disclosure is a training method of a recall model, and the method is executed by a training end, and the method comprises the following steps:

[0076] Step S11, at least two material samples labeled by a feature label are obtained; wherein different types of the material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features and interest matching features;

[0077] Step S12, based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined;

[0078] Step S13, the material samples are input into the initial recall model corresponding to the predetermined features for training until a predetermined condition is met, and a first recall model after training is obtained; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined features;

[0079] Step S14, the first recall models obtained by training at least two material samples are fused to obtain a second recall model used to predict a second candidate path corresponding to at least one to-be-recommended material.

[0080] The training method of the recall model of the present disclosure can be applied in an electronic device, and the electronic device related to the present disclosure can be, but is not limited to, a computer, a mobile phone, a wearable device, a vehicle-mounted terminal, a roadside unit (RSU, Road Side Unit), a smart home terminal, an industrial sensing device and / or a medical device, etc.

[0081] It should be noted that the above method corresponds to the training process of the recall model, and the training process can be performed on a special device. After training the recall model, a trained recall model can be obtained, and the trained recall model can be transplanted to an electronic device for application of the recall model, and the trained recall model is applied. Of course, the training process of the recall model and the trained recall model can also be performed on the same electronic device, which is not limited here. When the training process of the recall model is performed on the special device, after the training of the recall model is completed, the trained recall model can be transplanted from the special device to the electronic device for applying the model.

[0082] In the embodiments of the present disclosure, the recall model can be applied to an information flow scenario. In the information flow scenario, material samples can be screened based on four different features. The material samples contain predetermined features, which can be one of the following: click features, interaction features, time length features, and interest matching features. The click feature can be: for non-video materials, it can be the feature generated by the user clicking the article, picture, comment, forwarding, liking, following and / or entering the personal homepage; for video materials, it can be the feature generated by the user clicking the video, liking, commenting or automatically playing the video for more than t seconds (for example, 3s). The interaction feature can be a sample that the user has interacted with (such as forwarding, commenting, and liking interaction operations) in the real exposure sample. The time length feature is a feature related to the time length. For example, a threshold is set for the time length, and samples with a user stay time longer than t seconds (for example, 4s) are regarded as positive samples. The interest matching feature can be: a feature determined based on the user N-level interest label and the material N-level interest label, for example, if the first-level and second-level interest labels of a user in a sample can match the first-level and second-level interest of a material, then the sample is regarded as a positive sample.

[0083] In an embodiment, at least two material samples labeled with predetermined feature labels are obtained; wherein different kinds of material samples contain different predetermined features, the predetermined features being one of the following: click feature, interaction feature, time length feature, and interest matching feature; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined; the material samples are input into the initial recall model corresponding to the predetermined features for training until a predetermined condition is met, to obtain a first trained recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one recommended material corresponding to the predetermined features; the first recall models trained by at least two kinds of material samples are fused to obtain a second recall model for predicting a second candidate path corresponding to at least one recommended material. A first request information is received. Based on the predetermined features contained in the first request information, the second recall model is used to predict the second candidate path, and at least one recommended material is included in the second candidate path. In an embodiment, the final recommended recommended material can be determined from the at least one recommended material included in the second candidate path.

[0084] In an embodiment, the recall model is a deep recall (DR) model.

[0085] It should be noted that in the information flow scenario, the DR model trained by the material sample containing different predetermined features will have different prediction tendencies. For example, see Table 1 for the performance indicators of the single DR model and the DR model fused by multiple single DR models. The DR model trained by the material sample containing the click feature (single click DR model) will have more competitive advantage in the ranking link, and the exposure ratio will be higher, but the final selected material style will be biased. The DR model trained by the material sample containing the interaction feature (single interaction DR model) will have a lower exposure ratio. In the recommendation scenario, in addition to pursuing the performance indicators, the ecological indicators such as user interest also need to be focused on. The model trained based on the sample containing the interest matching feature (single interest matching DR model) is from this point, which can greatly improve the interest matching degree of the user and the material. Considering the characteristics of the recommendation flow scenario and the balance of the recommendation performance indicators, ecological indicators, and style, the embodiments of the present application train four target DR models (four target fusion DR model) based on four kinds of material samples respectively, and fuse the screening results of the four DR models as the pre-selection material when online service, and then further screen and score the material through the deep model for fine screening, and the final recalled material is obtained by truncating according to the recall quota, and is sent to the subsequent ranking link. In this way, the user's demand can be more accurately adapted.

[0086] Table 1:

[0087]

[0088] In this embodiment of the disclosure, the DR model can map items to corresponding path indices, with each path index resembling a cluster, such as... Figure 2 As shown, a D×K dimensional matrix is ​​used as the index structure. The model has a total of D layers, each represented by a softmax fully connected layer (MLP). The model prediction requires D steps, with K choices at each step. At the end, there will be KD possible paths. Each path can represent a set containing multiple items, and each item can also be in the set of multiple paths simultaneously.

[0089] For example, the forward process of the model's training phase is as follows: Figure 3 As shown, each layer of the MLP outputs K probabilities, representing the probability of containing this node. The entire forward process is as follows: the user-side embedding is input into the first layer of the MLP, generating K-dimensional probabilities. Then, all embeddings along the path are concatenated and input into the d-th layer of the MLP. The final probability is the product of all probabilities along the path. The loss function uses the log-likelihood function, and π represents the mapping from item to KD paths.

[0090]

[0091] To model the ambiguity of an item, it can belong to multiple paths. The probability of an item belonging to multiple paths is the sum of the probabilities of belonging to each path, thus achieving multi-path expansion. The expanded objective function (or loss function) is:

[0092]

[0093] In an embodiment, at least two material samples with labeled predetermined feature tags are obtained; wherein different types of material samples contain different predetermined features, the predetermined features being one of the following: click feature, interaction feature, time length feature, and interest matching feature; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined; the material samples are input into the initial recall model corresponding to the contained predetermined features for training until a predetermined condition is met, obtaining a first trained recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined features; wherein, in response to the loss function converging, it is determined that the predetermined condition is met; wherein the loss function includes a penalty term for balancing the distribution of materials corresponding to the first path. The first recall model trained by at least two material samples is fused to obtain a second recall model for predicting a second candidate path corresponding to at least one to-be-recommended material. It should be noted that the first recall model trained by at least two material samples can include the prediction result of the first recall model trained by at least two material samples.

[0094] In an embodiment, the second key-value pair information sent by the timing task end is received; the second key-value pair includes the path corresponding to the eliminated material sample and the number of material samples under the path; based on the path and the number indicated by the second key-value pair information, a penalty term for balancing the distribution of materials corresponding to the first candidate path is determined; in response to the loss function converging, it is determined that the predetermined condition is met; wherein the loss function includes the penalty term.

[0095] In an embodiment, in order to avoid excessive concentration of items in the path, a penalty term is added to the loss function (objective function):

[0096]

[0097] That is:

[0098] Where |c| is the number of items in path c, and the f function can be freely defined. After adding the penalty term, the distribution of items in the path (for example, the first candidate path) can be more balanced.

[0099] In an embodiment, the recall model includes an expectation-maximization (EM) model.

[0100] In an embodiment, the recall model comprises an expectation maximization (EM) model. At least two material samples labeled with a predetermined feature label are obtained; wherein different types of material samples contain different predetermined features, and the predetermined features are one of the following: a click feature, an interaction feature, a time length feature, and an interest matching feature; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined. The path parameters of the initial recall model are controlled to be unchanged, and the DNN parameters of the initial recall model are trained based on the material samples until the predetermined condition is met. The DNN parameters are kept unchanged, and the path parameters of the initial recall model are updated until the predetermined condition is met, to obtain the first trained recall model; wherein the prediction result output by the first trained recall model is used to predict a first candidate path corresponding to at least one recommended material corresponding to the predetermined features; and the first trained recall models obtained by training at least two material samples are fused to obtain a second trained recall model used to predict a second candidate path corresponding to at least one recommended material.

[0101] Exemplarily, the initial recall model is an EM model, and the training process of the EM model comprises an E step and an M step. In the E step, the path π is fixed, and the neural network parameter θ is trained.

[0102] In the M step, the parameter θ is fixed, and the path π is updated. Q_str(θ, π) can be equivalently converted as:

[0103]

[0104] In an embodiment, the upper bound can be optimized as:

[0105] wherein,

[0106] indicates the number of times that the sample v appears in the training set. In an embodiment, s[v,c] can be set as

[0107] The final M step optimization target is simplified as:

[0108]

[0109] The function has no closed solution, and therefore, the coordinate descent method can be used for solving.

[0110] It should be noted that s[v,c] is referred to as a candidate path score, which records the score corresponding to the candidate path. The hidden path and the hidden score can be calculated in advance in the E step, and can be directly read in the M step.

[0111] In an embodiment, the EM model comprises an E-step module and an M-step module, which are executed alternately. After reading one batch of online samples, the E-step module is invoked first, which reads the path index corresponding to each material (mid) from the training parameter server PS, which will serve as the training label, according to the formula The loss is calculated to update the network parameters of the DNN layer, and based on the formula The hidden score and the corresponding hidden path are obtained and input into the M-step module. After obtaining the hidden path and the hidden score, the M-step module invokes the coordinate descent method to update the path parameters, and pushes the updated results to the training PS.

[0112] In an embodiment, the path parameters are obtained from the training parameter server PS, wherein the PS is used to update and / or synchronize the model parameters. At least two kinds of material samples with labeled predetermined feature labels are obtained; wherein different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, time length feature and interest matching feature; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined. The path parameters of the initial recall model are kept unchanged, and the DNN parameters of the initial recall model are trained based on the material samples until the predetermined condition is met. The DNN parameters are kept unchanged, and the path parameters of the initial recall model are updated until the predetermined condition is met, to obtain the trained first recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined feature; the first recall models trained by at least two kinds of material samples are fused to obtain a second recall model for predicting a second candidate path corresponding to at least one to-be-recommended material.

[0113] In an embodiment, at least two material samples with a predetermined feature label are obtained; wherein different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features, and interest matching features; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined. The path parameters of the initial recall model are controlled to be unchanged, and the DNN parameters of the initial recall model are trained based on the material samples until the predetermined condition is met. The DNN parameters are kept unchanged, and the path parameters of the initial recall model are updated until the predetermined condition is met, obtaining the first trained recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined features; the first recall models trained by at least two types of material samples are fused to obtain a second recall model used to predict a second candidate path corresponding to at least one to-be-recommended material. The model parameters of the trained second recall model are sent to a training parameter server PS; wherein the PS is used to update and / or synchronize the model parameters.

[0114] In an embodiment of the present disclosure, at least two material samples with a predetermined feature label are obtained; wherein different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, time length features, and interest matching features; based on the predetermined features contained in the material samples, an initial recall model corresponding to the predetermined features is determined; the material samples are input into the initial recall model corresponding to the contained predetermined features for training until a predetermined condition is met, obtaining a first trained recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one to-be-recommended material corresponding to the predetermined features; the first recall models trained by at least two types of material samples are fused to obtain a second recall model used to predict a second candidate path corresponding to at least one to-be-recommended material. Here, since different material samples are input into the initial recall model corresponding to the predetermined features for training until a predetermined condition is met, obtaining a first trained recall model, thus, a plurality of first trained recall models trained for material samples containing different predetermined features can be obtained, and since the first trained recall models trained by at least two types of material samples are fused, the prediction results of different first trained recall models are used to jointly determine a second candidate path, which is essentially to fuse the prediction results of different first trained recall models. Compared with the way of obtaining prediction results by using a single recall model, the prediction accuracy and the adaptability of the results can be improved, and when content recommendation is performed by using the second recall model, the personalized content can be accurately recommended to the user, and the user can have a better experience.

[0115] It should be noted that the method provided by the embodiments of the present disclosure can be executed alone or together with some methods in the embodiments of the present disclosure or some methods in related technologies.

[0116] As shown in Figure 4 The method provided in the embodiments of the present disclosure is executed by a timing task end, and the method comprises the following steps:

[0117] In step S41, the material identifier of the material sample is obtained from the material library; wherein the material samples in the material library are used for training of the recall model; different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, time length feature, and interest matching feature;

[0118] In step S42, based on the material identifier, if a first key-value pair matching the material identifier is obtained from the training parameter server PS, the path included in the first key-value pair is determined; wherein the first key-value pair includes the material identifier of the eliminated material sample and the path corresponding to the material identifier;

[0119] In step S43, it is determined that the material sample in the path is the eliminated material sample in the material library.

[0120] In one embodiment, the material identifier of the material sample is obtained from the material library; wherein the material samples in the material library are used for training of the recall model; different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, time length feature, and interest matching feature. Based on the material identifier, if a first key-value pair matching the material identifier is obtained from the training parameter server PS, the path included in the first key-value pair is determined; wherein the first key-value pair includes the material identifier of the eliminated material sample and the path corresponding to the material identifier. It is determined that the material sample in the path is the eliminated material sample in the material library. The information of the eliminated material sample is sent to the training end and / or the service end.

[0121] In one embodiment, the material identifier of the material sample is obtained from the material library; wherein the material samples in the material library are used for training of the recall model; different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, time length feature, and interest matching feature; based on the material identifier, a first key-value pair matching the material identifier is obtained from the training parameter server PS. Based on the first key-value pair, a second key-value pair between the path and the number of material samples under the path is generated; the information of the second key-value pair is sent to the training end.

[0122] It should be noted that in the information flow recommendation, the material library is updated in real time, which means that old materials will be eliminated at any time, and the eliminated materials may have been trained by the DR model and mapped to a path index cluster. If not handled, one is that it will affect the counting of the path cluster, that is, it will increase the penalty of the penalty term, and the formula Q pen (θ,π) will be smaller, which makes it more difficult for new materials to be mapped to the path, causing errors in the update of the path and affecting accuracy; two is that old materials will still be in the path cluster, which will cause a lack of service.

[0123] In an embodiment, the counting of the path is updated by a timing task module based on the global material library (timing task end), which independently runs on the training end and the service end and acts on the training end and the service end in real time.

[0124] In an embodiment, while the model is being trained, the timing task module will train the path to which the material id belongs on the PS (parameter server) at any time according to the ids (mid) of all materials in the material library. The PS stores key-value pairs in the form of <mid, path>, so that the paths corresponding to all materials can be obtained. By traversing all key-value pairs, the material set under each path and the number of materials in the path can be obtained in real time. The number is pushed to the training PS again to update the parameters for training, which makes the subsequent model training based on the real-time material library, achieving real-time dynamic adjustment.

[0125] In an embodiment, the generated <mid, path> is synchronized to the recall engine side as offline material, and the material set under each path is obtained by inverted index in the recall service, that is, the material set is also based on the real-time material library, and old materials will not be indexed. When a user request is received, the user-side features are formed into fid (feature hash id) through feature engineering, and the topn paths of the user are calculated by calling the online model to obtain the material set in the path, which is the recall result of this request.

[0126] It should be noted that those skilled in the art can understand that the method provided by the embodiments of the disclosure can be executed alone or together with some methods in some methods or related technologies in the embodiments of the disclosure.

[0127] As Figure 5 shown, a recall method is provided in the embodiments of the disclosure, which is executed by a service end, and the method comprises:

[0128] Step S51, obtaining first request information containing target features, wherein the target features include at least two of the following predetermined features: click features, interaction features, time length features, and interest matching features.

[0129] Step S52, predicting, based on the target features, a second candidate path corresponding to at least one material to be recommended by using the second recall model trained according to any embodiment of the present disclosure, to obtain a first recall result.

[0130] In one embodiment, first request information containing target features is obtained, wherein the target features include at least two of the following predetermined features: click features, interaction features, time length features, and interest matching features. Based on the target features, a second candidate path corresponding to at least one material to be recommended is predicted by using the second recall model trained according to any embodiment of the present disclosure, to obtain a first recall result. Information of eliminated material samples is obtained from a service parameter server PS; based on the information of the eliminated material samples and first key-value pairs formed by material identifiers and corresponding paths, an inverted index of the first key-value pairs is generated; and based on the inverted index, the material to be recommended corresponding to the second candidate path is determined.

[0131] In one embodiment, first request information containing target features is obtained, wherein the target features include at least two of the following predetermined features: click features, interaction features, time length features, and interest matching features. Based on the target features, a second candidate path corresponding to at least one material to be recommended is predicted by using the second recall model trained according to any embodiment of the present disclosure, to obtain a first recall result. The material contained in the first recall result is input into an online deep model to obtain a second recall result; wherein the online deep model is used to score and / or sort the material to screen out the material to be recommended.

[0132] In one embodiment, the server mainly relies on a recall engine, a service PS, and offline materials synchronized by a timing task. The recall engine processes user requests, performs feature engineering to form user-side feature ids (fids), calls a model on the service PS according to the fids to calculate topn paths of the user, and based on an inverted index result constructed by offline materials, coarsely screens out a material set. The material is combined to be further scored and sorted by a deep model, and after being truncated according to a recall quota, a final recall result (corresponding to the second recall result) is obtained.

[0133] It should be noted that those skilled in the art can understand that the method provided by the embodiments of the present disclosure can be executed alone or together with some methods in some methods or related technologies in the embodiments of the present disclosure.

[0134] To better understand the embodiments of this disclosure, the following exemplary embodiment will be used to further illustrate the embodiments of this disclosure:

[0135] like Figure 6a As shown in the embodiments of this disclosure, a recall method is provided, which is executed by the training end, and the method includes:

[0136] Training end:

[0137] Step 601: Based on the real exposure sample stream from Kafka, formulate corresponding training labels with clicks, interactions, duration, and interests as targets, and obtain material samples labeled with these labels for training the four models. The training method for each model is consistent from step 602 onwards.

[0138] Step 602: The model training module reads batch sample data from the Kafka sample stream.

[0139] Step 603: Based on this batch of sample data, perform the E-step calculation, retrieve the path parameters updated in the previous round from the training parameter server PS and keep the path parameters unchanged, based on the formula... Calculate the loss and perform backpropagation to update the DNN parameters, then push the DNN parameters onto the training PS.

[0140] Step 604: Perform another DNN forward operation to calculate the result according to the formula. Obtain the hidden score and the corresponding hidden path.

[0141] Step 605, the M-step module reads the hidden score and hidden path, based on the formula. The path is updated using the coordinate descent method, and the path is then pushed onto the training PS.

[0142] Step 606, repeat step 602.

[0143] like Figure 6b As shown in the embodiments of this disclosure, a method for screening material samples is provided. The method is executed by a timed task terminal, and the method includes:

[0144] Scheduled task client:

[0145] Step 611: Retrieve the IDs (mids) of all materials from the material library.

[0146] Step 612: Pull the data from the training PS based on mid.<mid,path> Key-value pairs, and count the number of times mid appears in each path, forming...<path,mid_cnt> Key-value pairs are pushed onto the training PS.

[0147] Step 613, based on the <mid, path> key-value pairs pulled from the training PS, form offline materials, and synchronize to the server.

[0148] As shown in Figure 6c , the present embodiment provides a recall method, which is executed by the server, and the method comprises:

[0149] The server comprises:

[0150] Step 621, the service PS synchronizes the training PS parameters in real time.

[0151] Step 622, the recall engine synchronizes the offline materials generated by the timing task, and establishes the inverted index of <path, mid> based on the <mid, path> key-value pairs.

[0152] Step 623, the recall engine processes the user request to form the user-side fid.

[0153] Step 624, according to the user-side fid request, the parameters of the four online models (click, interaction, duration, and interest, etc.) in the service PS are calculated to obtain the topn path index clusters (path) of each model for each user.

[0154] Step 625, based on the inverted index of <path, mid>, the candidate materials (mid) of each model for each user are obtained, and all of them are combined together as the preliminary screening result.

[0155] Step 626, the preliminary screened materials are pushed through the online deep model inference for further scoring and sorting, and the K materials are screened out as the final recall result according to the recall quota.

[0156] It should be noted that those skilled in the art can understand that the method provided by the present embodiment can be executed alone or together with some methods in the present embodiment or some methods in related technologies.

[0157] Please refer to Figure 6d , which shows a schematic diagram of the recall system.

[0158] As shown in Figure 7 , the present embodiment provides a recall model training device, which comprises:

[0159] The acquisition module 71 is configured to acquire at least two kinds of material samples labeled with predetermined feature labels; wherein different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, duration feature, and interest matching feature;

[0160] The determining module 72 is configured to determine an initial recall model corresponding to a predetermined feature based on the predetermined feature contained in the material sample;

[0161] The training module 73 is configured to train the initial recall model corresponding to the predetermined feature contained in the material sample until a predetermined condition is met, to obtain a trained first recall model; wherein a prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one recommended material corresponding to the predetermined feature;

[0162] The fusion module 74 is configured to fuse the first recall models trained by at least two material samples to obtain a second recall model used to predict a second candidate path corresponding to at least one recommended material.

[0163] It should be noted that those skilled in the art can understand that the method provided by the embodiments of the present disclosure can be executed alone or together with some methods in the embodiments of the present disclosure or some methods in related technologies.

[0164] As shown in Figure 8 , the embodiments of the present disclosure provide a screening device for material samples, which comprises:

[0165] The acquisition module 81 is configured to acquire a material identifier of a material sample from a material library; wherein the material samples in the material library are used for training of a recall model; different kinds of material samples contain different predetermined features, and the predetermined features are one of the following: click feature, interaction feature, time length feature and interest matching feature;

[0166] The determining module 82 is configured to determine a path included in a first key-value pair based on the material identifier, in a case that the first key-value pair matching the material identifier is acquired from a training parameter server PS; wherein the first key-value pair includes a material identifier of an eliminated material sample and a path corresponding to the material identifier; and determine the material sample in the path as the eliminated material sample in the material library;

[0167] The sending module 83 is configured to send information of the eliminated material sample to a training end and / or a service end.

[0168] It should be noted that those skilled in the art can understand that the method provided by the embodiments of the present disclosure can be executed alone or together with some methods in the embodiments of the present disclosure or some methods in related technologies.

[0169] As shown in Figure 9 , the embodiments of the present disclosure provide a recall device, which comprises:

[0170] The acquisition module 91 is configured to acquire first request information containing a target feature, wherein the target feature includes at least two of the following predetermined features: a click feature, an interaction feature, a time length feature, and an interest matching feature.

[0171] The prediction module 92 is configured to predict, based on the target feature, a second candidate path corresponding to at least one to-be-recommended material by using the second recall model trained according to any one of claims 1 to 5, to obtain a first recall result.

[0172] It should be noted that those skilled in the art can understand that the method provided by the embodiments of the present disclosure can be executed alone or together with some methods in some related technologies or some methods in the embodiments of the present disclosure.

[0173] The embodiments of the present disclosure provide a processing device, which comprises:

[0174] A memory for storing executable programs;

[0175] A processor for executing the executable programs stored in the memory to implement the method according to any one of the embodiments of the present disclosure.

[0176] It can be understood that the memory can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a ferromagnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM). The magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example but not limitation, many forms of RAM can be used, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), sync link dynamic random access memory (SLDRAM), and direct rambus random access memory (DRRAM).The memory described in the embodiments of the present application is intended to include, but not limited to, these and any other suitable types of memory.

[0177] The method disclosed in the present application can be applied to or implemented by the processor. The processor can be an integrated circuit chip with processing capability. In the implementation process, the steps of the method for speech conversion can be completed by the integrated logic circuit or the instruction of software form in the processor. The processor mentioned above can be a general processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The processor can implement or execute the disclosed methods, steps and logic block diagrams in the present application. The general processor can be a microprocessor or any conventional processor, etc. In combination with the steps of the method disclosed in the present application, the hardware decoding processor can be directly embodied to complete the execution, or the hardware and software modules in the decoding processor can be combined to complete the execution. The software module can be located in a storage medium, which is located in a memory. The processor reads the information in the memory and combines the hardware to complete the steps of the method for speech conversion provided in the embodiments of the present application.

[0178] The present application further provides a computer storage medium, which stores an executable program. When the executable program is executed by a processor, the method according to any one of the embodiments of the present application is implemented. Specifically, the computer storage medium can be a computer readable storage medium, such as a memory for storing a computer program, which can be executed by a processor of a processing device to complete the steps of the method according to the embodiments of the present application. The computer readable storage medium can be a ROM, a PROM, an EPROM, an EEPROM, a flash memory, a magnetic surface memory, an optical disc, or a CD-ROM, etc.

[0179] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A training method for a recall model, characterized in that, The method is executed by the training end, and the method includes: Obtain at least two types of material samples labeled with predetermined feature tags; wherein, different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, duration features, and interest matching features, and the material types of the material samples include videos, text and images, pictures, articles, or polls; Based on the predetermined features contained in the material sample, an initial recall model corresponding to the predetermined features is determined; The material sample is input into the initial recall model corresponding to the predetermined features it contains and trained until the predetermined conditions are met to obtain the trained first recall model; wherein, the prediction result output by the first recall model is used to predict the first candidate path corresponding to at least one recommended material corresponding to the predetermined features. A first recall model trained by fusing at least two material samples is used to obtain a second recall model for predicting a second candidate path corresponding to at least one material to be recommended. The method further includes: Receive information about a second key-value pair sent by the timed task terminal; the second key-value pair includes the path corresponding to the eliminated material sample and the number of material samples under the path; Based on the path and quantity indicated by the information of the second key-value pair, a penalty term is determined to make the material distribution corresponding to the first candidate path balanced. In response to the convergence of the loss function, it is determined that the predetermined conditions are met; wherein the loss function includes the penalty term.

2. The method according to claim 1, characterized in that, The method further includes: The trained parameters of the second recall model are sent to the training parameter server PS. The PS is used to update and / or synchronize the model parameters.

3. The method according to claim 1, characterized in that, The recall model includes an expected maximum (EM) model; the material samples are input into an initial recall model corresponding to the predetermined features they contain and trained until predetermined conditions are met, resulting in a trained first recall model, including: Keeping the path parameters of the initial recall model unchanged, the deep neural network (DNN) parameters of the initial recall model are trained based on the material samples until the predetermined conditions are met; Keeping the trained DNN parameters unchanged, the path parameters of the initial recall model are updated until the predetermined conditions are met, thus obtaining the trained first recall model.

4. The method according to claim 3, characterized in that, The method further includes: Obtain the path parameters of the initial recall model from the training parameter server PS.

5. A method for screening material samples, characterized in that, The method is executed by a scheduled task terminal, and the method includes: Obtain material identifiers for material samples from the material library; wherein, the material samples in the material library are used for training the recall model; different types of material samples contain different predetermined features, the predetermined features being one of the following: click features, interaction features, duration features, and interest matching features, and the material types of the material samples include videos, text and images, pictures, articles, or polls; Based on the material identifier, if a first key-value pair matching the material identifier is obtained from the training parameter server PS, the path included in the first key-value pair is determined; wherein, the first key-value pair includes the material identifier of the eliminated material sample and the path corresponding to the material identifier; The material samples in the path are identified as discarded material samples in the material library; Sending information about discarded material samples to the training end and / or the server; the method further includes: Based on the first key-value pair, a second key-value pair is generated between the path and the number of material samples under the path; Send the information of the second key-value pair to the training end.

6. A recall method, characterized in that, The method is executed by the server, and the method includes: Obtain first request information containing target features, wherein the target features include at least two of the following predetermined features: click features, interaction features, duration features, and interest matching features; Based on the target features, the second recall model trained using any one of claims 1 to 4 is used to predict the second candidate path corresponding to at least one recommended material, and the first recall result is obtained.

7. The method according to claim 6, characterized in that, The method further includes: Obtain information on the eliminated material samples from the service parameter server PS; Based on the information of the eliminated material samples, and the first key-value pair formed by the material identifier and the corresponding path, an inverted index of the first key-value pair is generated. The recommended material corresponding to the second candidate path is determined based on the inverted index.

8. The method according to claim 7, characterized in that, The method further includes: Input the materials to be recommended contained in the first recall result into the online deep model to obtain the second recall result; The online deep model is used to score and / or rank the materials to be recommended in order to filter out target materials for recommendation to users.

9. A training device for a recall model, characterized in that, The device includes: The acquisition module is configured to acquire at least two types of material samples labeled with predetermined feature tags; wherein, different types of material samples contain different predetermined features, and the predetermined features are one of the following: click features, interaction features, duration features, and interest matching features, and the material types of the material samples include videos, text and images, pictures, articles, or polls; The determination module is configured to determine an initial recall model corresponding to the predetermined features contained in the material sample. The training module is configured to input the material sample into an initial recall model corresponding to the predetermined features it contains and train it until the predetermined conditions are met, thereby obtaining a trained first recall model; wherein the prediction result output by the first recall model is used to predict a first candidate path corresponding to at least one recommended material corresponding to the predetermined features. The fusion module is configured to fuse a first recall model trained from at least two material samples to obtain a second recall model for predicting a second candidate path corresponding to at least one material to be recommended. The training module is configured to receive information of a second key-value pair sent by a timed task terminal; the second key-value pair includes the path corresponding to the eliminated material sample and the number of material samples under the path; based on the path and the number indicated by the information of the second key-value pair, a penalty term is determined to make the material distribution corresponding to the first candidate path balanced; in response to the convergence of the loss function, it is determined that the predetermined condition is met; wherein, the loss function includes the penalty term.

10. A recall device, characterized in that, The device includes: The acquisition module is configured to: acquire first request information containing target features, wherein the target features include at least two of the following predetermined features: click features, interaction features, duration features, and interest matching features; The prediction module is configured to: based on the target features, use a second recall model trained according to any one of claims 1 to 4 to predict a second candidate path corresponding to at least one material to be recommended, and obtain a first recall result.

11. A processing apparatus, characterized in that, The processing equipment includes: Memory, used to store executable programs; A processor, when executing an executable program stored in the memory, implements the method as described in any one of claims 1 to 4, 5, or 6 to 8.

12. A computer storage medium, characterized in that, The computer storage medium stores an executable program, which, when executed by a processor, implements the method as described in any one of claims 1 to 4, 5, or 6 to 8.

Images