Novelty item recall model generation method, recall method, device and system

Abstract

The present disclosure proposes a method for generating a novelty item recall model, a recall method, a device and a system, and relates to the field of computers. The generation method includes: constructing multiple novelty item recall tasks; obtaining a novelty sample corresponding to each novelty item recall task to obtain multiple novelty samples of multiple novelty item recall tasks; using multiple novelty samples to train a multi-task dual-tower model to generate a novelty item recall model. This improves the recall effect and improves the novelty of the recommendation while satisfying the recommendation accuracy.

Description

Technical Field

[0001] The present disclosure relates to the field of computers, and in particular to a method and device for generating a novelty item recall model, a novelty item recall method and device, and a recommendation system. Background Art

[0002] Recommendation systems are widely used in various scenarios to match users with information. In e-commerce recommendation systems, after a user makes a request, the recommendation system selects some items from a large number of items and displays them to the user, providing personalized services to the user. Some recommendation algorithms estimate user interests based on the user's historical click behavior and make recommendations. This is prone to the problem of "what content is clicked, what content is recommended" homogeneity of recommended content. Although the recommended content is accurate, it is easy to make the recommended content too monotonous, causing users to become aesthetically fatigued and lose their sense of freshness in the recommended content. Summary of the invention

[0003] The disclosed embodiments propose a scheme for generating a novelty item recall model and a scheme for recalling novelty items based on the recall model, thereby improving the recall effect and enhancing the novelty of the recommendation while satisfying the accuracy of the recommendation.

[0004] Some embodiments of the present disclosure provide a method for generating a novelty item recall model, including:

[0005] Construct multiple novel object recall tasks;

[0006] Obtaining a novel sample corresponding to each novel item recall task to obtain multiple novel samples for multiple novel item recall tasks;

[0007] Use a variety of novel samples to train the multi-task dual-tower model to generate a novel item recall model.

[0008] Among them, the multi-task dual-tower model includes an item tower, a user tower and an output module. The item tower is configured to process item features in novelty samples and output corresponding representation vectors of the item features. The user tower is shared by the multiple novelty item recall tasks and is configured to process user features in each novelty sample and output corresponding representation vectors of the user features respectively. The output module is configured to calculate the similarity between the output of the item tower and the output of the user tower, and output it as a predicted value.

[0009] In some embodiments, training the multi-task dual-tower model using a plurality of novel samples includes:

[0010] Calculate the similarity between the representation vector corresponding to the user feature in each novelty sample output by the user tower and the representation vector corresponding to the item feature in each novelty sample output by the item tower as the prediction value of each novelty sample, and calculate the loss of the novelty item recall task corresponding to each novelty sample according to the prediction value and label value of each novelty sample;

[0011] Determine the total loss based on the losses in each novelty item recall task;

[0012] The parameters of the multi-task dual-tower model are updated according to the total loss so as to train the multi-task dual-tower model.

[0013] In some embodiments, the item tower includes an input module and an item feature learning module, and the user tower includes the input module and the user feature learning module, wherein:

[0014] The input module is configured to generate a first item characterization vector corresponding to an item feature or a first user characterization vector corresponding to a user feature;

[0015] The item feature learning module is configured to perform dimensionality reduction processing on the first item representation vector to obtain a second item representation vector;

[0016] The user feature learning module is configured to perform dimensionality reduction processing on the first user representation vector to obtain a second user representation vector.

[0017] In some embodiments, the user feature learning module includes a user portrait feature learning unit and a user behavior feature learning unit.

[0018] The user portrait feature learning unit is configured to perform mean processing on the first user portrait representation vector in the first user representation vector to obtain the second user portrait representation vector in the second user representation vector;

[0019] The user behavior feature learning unit is configured to perform sequence feature learning on a first user behavior representation vector in the first user representation vector to obtain a second user behavior representation vector in the second user representation vector.

[0020] In some embodiments, the input module is configured to generate a first item image representation vector corresponding to the item image feature; the item feature learning module is configured to perform mean processing on the first item image representation vector to obtain a second item image representation vector.

[0021] In some embodiments, it also includes:

[0022] The recall ability of the trained multi-task dual-tower model for novel items is evaluated using evaluation indicators, and the multi-task dual-tower model that passes the evaluation is used as the novel item recall model.

[0023] The evaluation indicator is used to characterize the probability that a first probability of predicting a novel positive sample is greater than a second probability of predicting a novel negative sample.

[0024] In some embodiments, each novelty sample is constructed based on business or based on implicit semantics. Constructing novelty samples based on business includes: taking samples corresponding to the category of items affected by the user's current preset behavior and the category that the user's preset behavior has not acted on in the past n days as novelty positive samples, negatively sampling samples corresponding to items that the user's preset behavior has not acted on as novelty negative samples, or taking samples corresponding to the category of items affected by the user's current preset behavior and the category that the user's preset behavior has not acted on in the past n days as novelty negative samples, where n is a preset value;

[0025] Constructing novel samples based on latent semantics includes: taking samples corresponding to items affected by the user's current preset behavior and whose novelty is greater than a certain threshold as novelty positive samples, negatively sampling samples corresponding to items not affected by the user's preset behavior as novelty negative samples, or taking samples corresponding to items affected by the user's current preset behavior and whose novelty is not greater than a certain threshold as novelty negative samples.

[0026] In some embodiments, the novelty is represented by the minimum distance between the first characterization vector of the item affected by the user's current preset behavior and each second characterization vector of each item affected by the user's historical preset behavior.

[0027] In some embodiments, the preset behavior includes at least one of a click, an exposure, or an order.

[0028] In some embodiments, the category includes at least one of different levels of categories or product words.

[0029] Some embodiments of the present disclosure provide a novelty item recall method, including:

[0030] The item tower of the novel item recall model is used to process the features of the candidate items to obtain a representation vector of the features of the candidate items;

[0031] The target user features are processed using the user tower of the novelty item recall model to obtain a representation vector of the target user features;

[0032] According to the similarity between the representation vector of the target user's features and the representation vector of the candidate item's features, the recalled items are selected from the candidate items.

[0033] In some embodiments, the target user characteristics include the target user's latest behavior characteristics and portrait characteristics acquired in real time in response to the user's recall request.

[0034] In some embodiments, selecting the recalled items from the candidate items includes: selecting a preset number of recalled items with the greatest similarity from the candidate items.

[0035] Some embodiments of the present disclosure provide a device for generating a novelty item recall model, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute the method for generating the novelty item recall model of each embodiment based on instructions stored in the memory.

[0036] Some embodiments of the present disclosure provide a device for generating a novelty item recall model, including:

[0037] A task construction module, configured to construct a plurality of novelty item recall tasks;

[0038] A sample acquisition module is configured to acquire a novelty sample corresponding to each novelty item recall task to obtain a plurality of novelty samples for a plurality of novelty item recall tasks;

[0039] The model generation module is configured to train the multi-task dual-tower model using a variety of novel samples to generate a novel item recall model.

[0040] Among them, the multi-task dual-tower model includes an item tower and a user tower. The item tower is configured to process the item features in the novelty sample and output the corresponding representation vector of the item features. The user tower is shared by the multiple novelty item recall tasks and is configured to process the user features in each novelty sample and output the corresponding representation vector of the user features.

[0041] Some embodiments of the present disclosure provide a novelty item recall device, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute the novelty item recall method of each embodiment based on instructions stored in the memory.

[0042] Some embodiments of the present disclosure provide a novelty item recall device, including:

[0043] An item vector processing module is configured to process the candidate item features using the item tower of the novel item recall model to obtain a representation vector of the candidate item features;

[0044] A user vector processing module is configured to process the target user features using the user tower of the novelty item recall model to obtain a representation vector of the target user features;

[0045] The item recall module is configured to select a recall item from the candidate items according to the similarity between the representation vector of the target user feature and the representation vector of the candidate item feature.

[0046] Some embodiments of the present disclosure provide a recommendation system, including: a novelty item recall device.

[0047] Some embodiments of the present disclosure provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for generating a novelty item recall model or the steps of the novelty item recall method of each embodiment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. The present disclosure can be more clearly understood according to the following detailed description with reference to the drawings.

[0049] Obviously, the drawings described below are only some embodiments of the present disclosure, and a person skilled in the art can obtain other drawings based on these drawings without creative work.

[0050] Figure 1 A schematic diagram of a recommendation link of an e-commerce recommendation system according to some embodiments of the present disclosure is shown.

[0051] Figure 2 A schematic diagram illustrating a multi-mission dual-tower model of some embodiments of the present disclosure.

[0052] Figure 3 A schematic diagram showing a method for generating a novelty item recall model based on a multi-task dual-tower model according to some embodiments of the present disclosure.

[0053] Figure 4 A schematic diagram showing a method for recalling novelty items based on a novelty item recall model according to some embodiments of the present disclosure.

[0054] Figure 5 A schematic diagram showing a device for generating a novelty item recall model according to some embodiments of the present disclosure.

[0055] Figure 6 A schematic diagram showing a device for generating a novelty item recall model according to some embodiments of the present disclosure.

[0056] Figure 7 A schematic diagram showing a novelty item recall device according to some embodiments of the present disclosure.

[0057] Figure 8 A schematic diagram showing a novelty item recall device according to some embodiments of the present disclosure.

[0058] Fig. 9 A schematic diagram showing a recommendation system according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure.

[0060] Unless otherwise specified, descriptions such as “first” and “second” in the present disclosure are used to distinguish different objects and are not used to indicate meanings such as size or time sequence.

[0061] Figure 1 A schematic diagram of a recommendation link of an e-commerce recommendation system according to some embodiments of the present disclosure is shown.

[0062] like Figure 1 As shown in the figure, in the e-commerce recommendation system, after the user sends a request, the e-commerce recommendation system selects K items from the massive item library to display to the user, providing personalized services for the user. The recommendation link generally includes multiple stages such as recall, rough sorting, fine sorting, and re-ranking. Recall is to quickly retrieve items that the user is potentially interested in from the massive item library and narrow the scope of items, for example, narrowing the candidate items of the order of 100 million to the order of 10,000, so that the user can obtain the recommended information with a delay of milliseconds, and then hand it over to the subsequent sorting link. Rough sorting is to filter the recalled items to reduce the number of items entering the fine sorting, for example, narrowing the candidate items of the order of 10,000 to the order of 1,000 / 100. Fine sorting is to score the items one by one, generally based on user characteristics, item characteristics, and the interaction characteristics between users and items, calculate the fine sorting score, and then sort them in descending order according to the score to generate a sequence of Top-N items, for example, narrowing the candidate items of the order of 1,000 / 100 to the order of 100 / 10. Re-ranking is to re-rank the sequence of Top-N items generated by the fine sorting to generate a Top-K item sequence, which is directly presented to the user as the final recommendation result of the sorting system. Among them, the sorting methods such as rough sorting, fine sorting, and re-ranking can refer to the existing technology. The present disclosure mainly improves the recall method, and proposes a scheme for generating a novelty item recall model and a scheme for recalling novelty items based on the recall model, aiming to improve the recall effect, so that the novelty of the recommendation is improved while satisfying the accuracy of the recommendation. Through the novelty recall scheme, it is hoped that the recommended pages with good novelty will be displayed to users, and the user's interest range will be explored and expanded from the user's interactive feedback, so as to promote the user's sense of surprise and enhance the browseability of the recommended page. The improvement of novelty item exposure can drive the diversification of user clicks, expand the user's perception of the range of interest, and drive the update and iteration of the entire recall model, solve the serious problem of homogeneity of recommended content, and achieve the purpose of improving the recommended category structure and ecological health, and improving the quality of recommendations.

[0063] The disclosed embodiment proposes a solution for generating a novelty item recall model based on a multi-task dual-tower model. The multi-task dual-tower model is first described below.

[0064] Figure 2 A schematic diagram illustrating a multi-mission dual-tower model of some embodiments of the present disclosure.

[0065] like Figure 2 As shown, the multi-task dual-tower model includes an item tower, a user tower and an output module. The item tower is configured to process the item features in the novelty sample and output the corresponding characterization vector of the item features. The user tower is shared by the multiple novelty item recall tasks, each novelty item recall task corresponds to a novelty sample, and the user tower is configured to process the user features in each novelty sample and output the corresponding characterization vector of the user features. The output module is configured to calculate the similarity between the output of the item tower and the output of the user tower, and output it as a predicted value. Among them, the similarity can be calculated, for example, by inner product or cosine similarity. By calculating the similarity, the user is closer to the positive example item in the vector space and farther away from the negative example item in the vector space.

[0066] The item tower includes an input module and an item feature learning module shared with the user tower. The input module is configured to generate a first item representation vector corresponding to the item feature (i.e., the input of the item tower) or a first user representation vector corresponding to the user feature (i.e., the input of the user tower), and each representation vector may be, for example, an embedding vector. The item feature learning module is configured to perform dimensionality reduction processing on the first item representation vector to obtain a second item representation vector. In some embodiments, the item feature is, for example, an item image feature, and the item image feature includes, for example: the category to which the item belongs, the gender or age preference to which the item is adapted, etc. Then the input module is configured to generate a first item image representation vector corresponding to the item image feature; and the item feature learning module is configured to perform mean processing on the first item image representation vector to obtain a second item image representation vector.

[0067] The user tower includes an input module and a user feature learning module shared with the item tower. The input module is configured to generate a first item representation vector corresponding to the item feature (i.e., the input of the item tower) or a first user representation vector corresponding to the user feature (i.e., the input of the user tower), and each representation vector may be, for example, an embedding vector. The user feature learning module is configured to perform dimensionality reduction processing on the first user representation vector to obtain a second user representation vector.

[0068] In some embodiments, user features include, for example, user portrait features and user behavior features. User portrait features include, for example, user gender, user age, etc. User behavior features include, for example, user click behavior, user order behavior, user exposure behavior, and other behavioral features. Taking user click behavior features as an example, for example, they include: a sequence of 150 items that the user has recently clicked in the past 60 days, a sequence of categories to which the user has recently clicked 150 items in the past 60 days, etc. Then the user feature learning module includes a user portrait feature learning unit and a user behavior feature learning unit. The user portrait feature learning unit is configured to perform mean processing on the first user portrait representation vector in the first user representation vector to obtain the second user portrait representation vector in the second user representation vector. The user behavior feature learning unit is configured to perform sequence feature learning on the first user behavior representation vector in the first user representation vector to obtain the second user behavior representation vector in the second user representation vector. The user behavior feature learning unit is, for example, a Transfromer model based on user sequence modeling or a MIND (Multi-Interest Network with Dynamic Routing) model that learns to reflect multiple points of interest of users.

[0069] From a horizontal perspective, the multi-task dual-tower model includes a discrete input layer, a feature learning layer, and an output layer. The discrete input layer is used to generate a representation vector corresponding to the input feature. The discrete input layer is, for example, an embedding layer. The discrete input layer includes, for example, an input module. The feature learning layer is used to perform dimensionality reduction processing on the representation vector output by the discrete input layer. The feature learning layer includes, for example, an item feature learning module and a user feature learning module. The output layer is used to output the final prediction result based on the representation vector obtained by the feature learning layer. The output layer includes, for example, an output module.

[0070] In the multi-task dual-tower model, the input module is shared by the user tower and the item tower, and the user feature learning module is shared by multiple novelty item recall tasks. Therefore, the input module and the user feature learning module belong to the shared layer. The item feature learning module is independent and belongs to the independent layer.

[0071] Figure 3 A schematic diagram showing a method for generating a novelty item recall model based on a multi-task dual-tower model according to some embodiments of the present disclosure.

[0072] like Figure 3 As shown, the method for generating a novel item recall model based on a multi-task dual-tower model includes the following steps.

[0073] At step 310, a plurality of novelty item recall tasks are constructed.

[0074] Novelty needs to be defined as novelty at the user-perceivable level, but the perceivable level is difficult to quantify. The present disclosure explores novelty that is similar to user perception from multiple dimensions.

[0075] Novelty is defined based on the business. Business is divided into time dimension and category dimension. For example, items in a category (such as first-level category / second-level category / third-level category / product word) for which the user has no preset behavior (such as exposure / click / order) within a period of time (such as one month / one quarter / one year / history) are novel to the user.

[0076] Novelty is defined based on implicit semantics. The minimum distance between the first characterization vector of the item affected by the user's current preset behavior and the second characterization vectors of the items affected by the user's historical preset behavior is used to characterize novelty. Items with a novelty greater than a certain threshold are novel to the user. The characterization vector of the item is obtained, for example, based on methods such as word2vec, and the characterization vector is, for example, an embedding vector.

[0077] For example, the minimum distance between the representation vector of the item currently clicked by the user and the representation vectors of the items previously clicked by the user is taken as the novelty. The formula for novelty is as follows:

[0078] Novelty

[0079] =min{D(tgt, clicked1), D(tgt, clicked2)...D(tgt, clicked k-1 ), D(tgt, clicked k )}

[0080] Among them, Novelty represents novelty, min represents the minimum value, D(tgt, clicked k ) represents the characterization vector tgt of the item currently clicked by the user and the characterization vector clicked of the kth item clicked by the user in history k The distance between clicked1, clicked2, clicked k-1 , clicked k Represents the representation vector of each item clicked by the user in history.

[0081] The calculation formula of the distance between two representation vectors is as follows:

[0082]

[0083] Where D(X, Y) represents the distance between vector X and vector Y, x i represents the i-th element of vector X, y irepresents the i-th element of vector Y, and n represents the dimension of vector X and vector Y.

[0084] In step 320, a novelty sample corresponding to each novelty item recall task is obtained to obtain a plurality of novelty samples for a plurality of novelty item recall tasks.

[0085] According to the definition of novelty, each novel sample is constructed based on business or based on latent semantics.

[0086] Constructing novelty samples based on business includes: taking samples corresponding to the category of items affected by the user's current preset behavior and the category that the user's preset behavior has not acted on in the past n days as novelty positive samples, negatively sampling samples corresponding to items that the user's preset behavior has not acted on as novelty negative samples, or taking samples corresponding to the category of items affected by the user's current preset behavior and the category that the user's preset behavior has not acted on in the past n days as novelty negative samples, where n is a preset value. The preset behavior includes at least one of clicks, exposures, or orders. Categories include at least one of categories or product words at different levels.

[0087] Constructing novelty samples based on latent semantics includes: taking samples corresponding to items affected by the user's current preset behavior and samples with a novelty greater than a certain threshold as novelty positive samples, negatively sampling samples corresponding to items not affected by the user's preset behavior as novelty negative samples, or taking samples corresponding to items affected by the user's current preset behavior and samples with a novelty not greater than a certain threshold as novelty negative samples. The novelty is characterized by the minimum distance between the first characterization vector of the item affected by the user's current preset behavior and the second characterization vectors of each item affected by the user's historical preset behavior. The preset behavior includes at least one of click, exposure, or order placement.

[0088] Here are some novelty samples. For example, the sample corresponding to the category of the item currently clicked by the user that has not been clicked in the past n days is regarded as a novelty positive sample, recorded as "n days-category novel click sample", such as "60 days-cid1 novel click sample" means that the category cid1 of the item currently clicked by the user is a category that the user has not clicked in 60 days. Negative sampling is performed on the samples that the user did not click as novelty negative samples, or the non-"60 days-cid1 novel click samples" in the user click samples are regarded as novelty negative samples.

[0089] Since there are multiple definitions of novelty dimensions, without loss of generality, we take the following four dimensions of novelty as an example to describe multiple novel item recall tasks (referred to as “tasks”) and the corresponding multiple novel samples, among which novel positive samples are referred to as “positive samples” and novel negative samples are referred to as “negative samples”. Task 1: 60-day-cid1 novel click task, take 60-day-cid1 novel click samples as positive samples, negative samples of samples not clicked by the current user and other click samples (i.e., non-60-day-cid1 novel click samples in click samples) as negative samples; Task 2: 60-day-cid2 novel click task, take 60-day-cid2 novel click samples as positive samples, negative samples of samples not clicked by the current user and other click samples (i.e., non-60-day-cid2 novel click samples in click samples) as negative samples; Task 3: 60-day-cid3 novel click task, take 60-day-cid3 novel click samples as positive samples, negative samples of samples not clicked by the current user and other click samples (i.e., non-60-day-cid3 novel click samples in click samples) as negative samples; Task 4: Latent semantic novel click task, take samples that are clicked and whose novelty is greater than a threshold (m) as positive samples, negative samples of samples not clicked by the current user and other click samples (i.e., clicked and whose novelty is not greater than the threshold) as negative samples. Where, cid1, cid2, cid3 represent different categories respectively.

[0090] In step 330 , the multi-task dual-tower model is trained using a variety of novelty samples to generate a novelty item recall model.

[0091] Looking back at historical novelty sample data for example for 60 days, randomly selecting for example 3% of the samples as the evaluation set, and the rest as the training set, the training data is input into the multi-task dual-tower model to train the multi-task dual-tower model.

[0092] Using a variety of novel samples to train the multi-task dual-tower model includes: for example, using the inner product, calculating the corresponding representation vector X of the user feature in each novel sample output by the user tower = {x i}The representation vector Y corresponding to the item features in each novelty sample output by the item tower = {y i The similarity between As the predicted value of each novelty sample For example, a cross entropy loss function is used to calculate the predicted value of each novel sample. And the label value y calculates the loss of the novel item recall task corresponding to each novel sample According to the loss of each novel item recall task j j Determine the total loss L = ∑ j w j *lossj , the weight coefficient w of each novelty item recall task j j It can be set that according to the total loss, the total loss is minimized and the parameters of the multi-task dual-tower model are updated so as to train the multi-task dual-tower model.

[0093] In step 340, the recall capability of the trained multi-task dual-tower model for novel items is evaluated using evaluation indicators, and the multi-task dual-tower model that passes the evaluation is used as a novel item recall model.

[0094] The evaluation index is used to characterize the probability that the first probability of predicting a novel positive sample is greater than the second probability of predicting a novel negative sample. The evaluation index is, for example, AUC (area under the curve), which represents the area under the ROC (receiver operating characteristic curve).

[0095]

[0096]

[0097] Among them, P 正样本 represents the first probability of predicting a novel positive sample, P 负样本 represents the second probability of predicting a novel negative sample, M and N represent the number of novel positive samples and novel negative samples, respectively. Among them, 1, 0.5, and 0 in Formula I are only examples and are not limited to these example values, as long as they meet the size relationship between the three presented so far.

[0098] Therefore, the multi-task dual-tower model is trained and evaluated through novel samples, and finally a novel item recall model is generated.

[0099] Figure 4 A schematic diagram showing a method for recalling novelty items based on a novelty item recall model according to some embodiments of the present disclosure.

[0100] like Figure 4 As shown, the method for recalling novel items based on the novel item recall model includes the following steps.

[0101] In step 410, the candidate item features are processed using the item tower of the novel item recall model to obtain a representation vector of the candidate item features.

[0102] For a large number of candidate items, the item tower can be used to process offline to obtain the representation vectors of all candidate items, and store them in an ANN (Approximate Nearest Neighbor) retrieval system, such as vector retrieval FAISS, for query.

[0103] In step 420, the target user features are processed using the user tower of the novelty item recall model to obtain a representation vector of the target user features.

[0104] The target user features may be, for example, the latest behavior features and portrait features of the target user acquired in real time in response to a user's recall request. The target user features are processed in real time using a user tower to obtain the latest representation vector of the target user.

[0105] In step 430, a recalled item is selected from the candidate items according to the similarity between the characterization vector of the target user feature and the characterization vector of the candidate item feature.

[0106] The similarity may be calculated, for example, by inner product or cosine similarity. Selecting the recalled items from the candidate items includes: selecting a preset number of recalled items with the greatest similarity from the candidate items.

[0107] Therefore, novel items are recalled based on the novel item recall model to improve the recall effect, thereby improving the novelty of the recommendation while satisfying the recommendation accuracy.

[0108] Figure 5 A schematic diagram showing a device for generating a novelty item recall model according to some embodiments of the present disclosure.

[0109] like Figure 5 As shown, the device 500 for generating a novelty item recall model includes:

[0110] A task construction module 510 is configured to construct a plurality of novelty item recall tasks;

[0111] The sample acquisition module 520 is configured to acquire a novelty sample corresponding to each novelty item recall task to obtain a plurality of novelty samples for a plurality of novelty item recall tasks;

[0112] The model generation module 530 is configured to train the multi-task dual-tower model using a variety of novelty samples to generate a novelty item recall model, wherein the multi-task dual-tower model includes an item tower and a user tower, the item tower is configured to process item features in the novelty samples and output a corresponding representation vector of the item features, the user tower is shared by the multiple novelty item recall tasks, and is configured to process user features in each novelty sample and output a corresponding representation vector of the user features.

[0113] The model generation module 530 is specifically configured to calculate the similarity between the representation vector corresponding to the user features in each novelty sample output by the user tower and the representation vector corresponding to the item features in each novelty sample output by the item tower as the prediction value of each novelty sample, calculate the loss of the novelty item recall task corresponding to each novelty sample according to the prediction value and label value of each novelty sample; determine the total loss according to the losses of each novelty item recall task; and update the parameters of the multi-task dual-tower model according to the total loss so as to train the multi-task dual-tower model.

[0114] The model generation module 530 is also configured to use an evaluation indicator to evaluate the recall ability of the trained multi-task dual-tower model for novel items, and use the multi-task dual-tower model that has passed the evaluation as a novel item recall model, wherein the evaluation indicator is used to characterize the probability that a first probability of predicting a novel positive sample is greater than a second probability of predicting a novel negative sample.

[0115] Figure 6 A schematic diagram showing a device for generating a novelty item recall model according to some embodiments of the present disclosure.

[0116] like Figure 6 As shown, the device 600 for generating a novelty item recall model includes: a memory 610 and a processor 620 coupled to the memory 610, and the processor 620 is configured to execute the method for generating a novelty item recall model in any of the aforementioned embodiments based on instructions stored in the memory 610.

[0117] The memory may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.

[0118] The processor may be implemented as a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistors and other discrete hardware components.

[0119] The device 600 may further include an input / output interface 630 , a network interface 640 , a storage interface 650 , etc. These interfaces 630 , 640 , 650 , the memory 610 , and the processor 620 may be connected via a bus 660 , for example.

[0120] Among them, the input and output interface provides a connection interface for input and output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface provides a connection interface for various networked devices. The storage interface provides a connection interface for external storage devices such as SD cards and USB flash drives. The bus can use any bus structure among a variety of bus structures. For example, the bus structure includes but is not limited to the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MCA) bus, and the Peripheral Component Interconnect (PCI) bus.

[0121] Figure 7 A schematic diagram showing a novelty item recall device according to some embodiments of the present disclosure.

[0122] like Figure 7 As shown, the novelty item recall device 700 includes:

[0123] The item vector processing module 710 is configured to process the candidate item features using the item tower of the novel item recall model to obtain a representation vector of the candidate item features.

[0124] The user vector processing module 720 is configured to process the target user characteristics using the user tower of the novelty item recall model to obtain a representation vector of the target user characteristics; wherein the target user characteristics include the latest behavior characteristics and portrait characteristics of the target user obtained in real time in response to the user's recall request.

[0125] The item recall module 730 is configured to select recalled items from the candidate items according to the similarity between the representation vector of the target user feature and the representation vector of the candidate item feature, for example, to select a preset number of recalled items with the greatest similarity from the candidate items.

[0126] Figure 8 A schematic diagram showing a novelty item recall device according to some embodiments of the present disclosure.

[0127] like Figure 8 As shown, the novelty item recall device 800 includes: a memory 810 and a processor 820 coupled to the memory 810 , and the processor 820 is configured to execute the novelty item recall method in any of the aforementioned embodiments based on instructions stored in the memory 810 .

[0128] The device 800 may further include an input / output interface 830 , a network interface 840 , a storage interface 850 , etc. These interfaces 830 , 840 , 850 , the memory 810 , and the processor 820 may be connected via a bus 860 , for example.

[0129] Fig. 9 A schematic diagram showing a recommendation system according to some embodiments of the present disclosure.

[0130] like Fig. 9 As shown, the recommendation system 900 includes: novelty item recall devices 700, 800, and may also include a sorting device for sorting the recalled items, etc. The sorting device includes, for example, a rough sorting device 910 for performing a rough sorting, a fine sorting device 920 for performing a fine sorting, and a re-sorting device 930 for performing a re-sorting.

[0131] The embodiments of the present disclosure provide a non-transitory computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method for generating a novelty item recall model or the steps of a novelty item recall method are implemented.

[0132] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more non-transient computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code.

[0133] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0134] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0136] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method for generating a novelty item recall model, characterized in that: include: Construct multiple novel object recall tasks; Obtaining a novel sample corresponding to each novel item recall task to obtain multiple novel samples for multiple novel item recall tasks; Use a variety of novel samples to train the multi-task dual-tower model to generate a novel item recall model; The multi-task dual-tower model includes an item tower, a user tower and an output module. The item tower is configured to process item features in novelty samples and output corresponding representation vectors of the item features. The user tower is shared by the multiple novelty item recall tasks and is configured to process user features in each novelty sample and output corresponding representation vectors of the user features respectively. The output module is configured to calculate the similarity between the output of the item tower and the output of the user tower, and output it as a predicted value. The generation of the novelty item recall model includes: calculating the similarity between the representation vector corresponding to the user feature in each novelty sample output by the user tower and the representation vector corresponding to the item feature in each novelty sample output by the item tower as the predicted value of each novelty sample, calculating the loss of the novelty item recall task corresponding to each novelty sample according to the predicted value and label value of each novelty sample; and training the multi-task dual-tower model according to the loss of each novelty item recall task.

2. The method according to claim 1, characterized in that: Training the multi-task dual-tower model includes: Determine the total loss based on the losses in each novelty item recall task; The parameters of the multi-task dual-tower model are updated according to the total loss so as to train the multi-task dual-tower model.

3. The method according to claim 1, characterized in that The item tower includes an input module and an item feature learning module. The user tower includes the input module and the user feature learning module. in: The input module is configured to generate a first item characterization vector corresponding to an item feature or a first user characterization vector corresponding to a user feature; The item feature learning module is configured to perform dimensionality reduction processing on the first item representation vector to obtain a second item representation vector; The user feature learning module is configured to perform dimensionality reduction processing on the first user representation vector to obtain a second user representation vector.

4. The method according to claim 3, characterized in that The user feature learning module includes a user portrait feature learning unit and a user behavior feature learning unit. The user portrait feature learning unit is configured to perform mean processing on the first user portrait representation vector in the first user representation vector to obtain the second user portrait representation vector in the second user representation vector; The user behavior feature learning unit is configured to perform sequence feature learning on a first user behavior representation vector in the first user representation vector to obtain a second user behavior representation vector in the second user representation vector.

5. The method according to claim 3, characterized in that: The input module is configured to generate a first object image representation vector corresponding to an object image feature; The object feature learning module is configured to perform mean processing on the first object image representation vector to obtain a second object image representation vector.

6. The method according to claim 1, characterized in that Also includes: The recall ability of the trained multi-task dual-tower model for novel items is evaluated using evaluation indicators, and the multi-task dual-tower model that passes the evaluation is used as the novel item recall model. The evaluation indicator is used to characterize the probability that a first probability of predicting a novel positive sample is greater than a second probability of predicting a novel negative sample.

7. The method according to claim 1, characterized in that Each novelty sample is constructed based on business or implicit semantics. Constructing novel samples based on business includes: taking samples corresponding to the category of items affected by the user's current preset behavior that have not been affected by the user's preset behavior in the past n days as novel positive samples, negatively sampling samples corresponding to items not affected by the user's preset behavior as novel negative samples, or taking samples corresponding to the category of items affected by the user's current preset behavior that are not affected by the user's preset behavior in the past n days as novel negative samples, where n is a preset value; Constructing novel samples based on latent semantics includes: taking samples corresponding to items affected by the user's current preset behavior and whose novelty is greater than a certain threshold as novelty positive samples, negatively sampling samples corresponding to items not affected by the user's preset behavior as novelty negative samples, or taking samples corresponding to items affected by the user's current preset behavior and whose novelty is not greater than a certain threshold as novelty negative samples, and the novelty is used to characterize the novelty degree of the item to the user.

8. The method according to claim 7, characterized in that The novelty is characterized by the minimum distance between the first characterization vector of the item affected by the user's current preset behavior and each second characterization vector of each item affected by the user's historical preset behavior.

9. The method according to claim 7, characterized in that: The preset behavior includes at least one of click, exposure, or order placement; A category includes at least one item of different levels of categories or product words.

10. The method according to any one of claims 1 to 9, characterized in that: Also includes: The item tower of the novel item recall model is used to process the features of the candidate items to obtain a representation vector of the features of the candidate items; The target user features are processed using the user tower of the novelty item recall model to obtain a representation vector of the target user features; According to the similarity between the representation vector of the target user's features and the representation vector of the candidate item's features, the recalled items are selected from the candidate items.

11. The method according to claim 10, characterized in that The target user characteristics include the target user's latest behavior characteristics and portrait characteristics acquired in real time in response to the user's recall request.

12. The method according to claim 10, characterized in that The candidate items for recall include: A preset number of recalled items with the greatest similarity are selected from the candidate items.

13. A device for generating a novelty item recall model, comprising: Memory; and a processor coupled to the memory, wherein the processor is configured to execute the method for generating a novelty item recall model according to any one of claims 1 to 12 based on instructions stored in the memory.

14. A device for generating a novelty item recall model, comprising: A task construction module, configured to construct a plurality of novelty item recall tasks; A sample acquisition module is configured to acquire a novelty sample corresponding to each novelty item recall task to obtain a plurality of novelty samples for a plurality of novelty item recall tasks; A model generation module is configured to train the multi-task dual-tower model using a variety of novel samples to generate a novel item recall model; The multi-task dual-tower model includes an item tower and a user tower, the item tower is configured to process item features in novelty samples and output corresponding representation vectors of the item features, the user tower is shared by the multiple novelty item recall tasks, and is configured to process user features in each novelty sample and output corresponding representation vectors of the user features respectively, and the generation of the novelty item recall model includes: calculating the similarity between the representation vector corresponding to the user features in each novelty sample output by the user tower and the representation vector corresponding to the item features in each novelty sample output by the item tower as the predicted value of each novelty sample, calculating the loss of the novelty item recall task corresponding to each novelty sample according to the predicted value and label value of each novelty sample; and training the multi-task dual-tower model according to the loss of each novelty item recall task.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images