Dissociated commodity recommendation method, device and equipment and storage medium

Abstract

Embodiments of the present application relate to the technical field of data processing, in particular to a method and device for recommending disintegrated goods, equipment and storage medium. The purpose is to improve the accuracy of recommending goods to users. The method comprises: receiving information of a to-be-recommended good and historical click information of a user, the historical click information comprising a clicked good sequence, a non-clicked good sequence, a disliked good sequence and click time information; filtering the clicked good sequence and the non-clicked good sequence according to the disliked good sequence to obtain a representation of a good of interest to the user; screening the representation of the good of interest to the user according to the click time information of the user and the information of the to-be-recommended good to obtain a representation of a historical interest good of the user; classifying and aggregating the representation of the historical interest good to obtain a plurality of disintegrated representations of the user; and determining whether the to-be-recommended good is a good of interest to the user according to the plurality of disintegrated representations.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of data processing, in particular, to a dissociation commodity recommendation method and device, equipment and storage medium. BACKGROUND

[0002] Dissociation representation is to dissociate the user's intention to obtain multiple representations reflecting the user's different interests, the purpose is to mine the hidden factors behind the user's intention, which is an effective way to mine the user's intention and improves the accuracy and interpretability of the recommendation system. The existing recommendation system based on dissociation representation is based on the positive feedback information of the user, that is, the purchase or click information, to obtain the dissociation interest of the user, and then recommend commodities to the user.

[0003] In the prior art, only the dissociation interest of the user is obtained from the positive feedback of the user, and the obtained user interest is prone to homogenization and singleness, which affects the user click rate of commodity recommendation, and the prior art cannot handle the complex relationship between the dissociation interest of the user and multiple feedbacks in the multi-feedback scenario, nor can it handle the large amount of noise contained in multiple feedbacks, so it cannot be directly used in the multi-feedback scenario. The existing multi-feedback recommendation technology does not dissociate the user's interest, cannot well depict the user's overall interest, lacks interpretability, and has a low accuracy of recommended commodities. SUMMARY

[0004] Embodiments of the present application provide a dissociation commodity recommendation method, device, equipment and storage medium, which aims to improve the accuracy of recommending commodities to users.

[0005] The first aspect of the embodiments of the present application provides a dissociation commodity recommendation method, the method comprising:

[0006] receiving information of a commodity to be recommended and historical click information of a user, the historical click information comprising a clicked commodity sequence, a non-clicked commodity sequence, a disliked commodity sequence and click time information of the user;

[0007] filtering the clicked commodity sequence and the non-clicked commodity sequence according to the disliked commodity sequence to obtain a representation of a commodity of interest to the user;

[0008] screening the representation of the commodity of interest to the user according to the click time information of the user and the information of the commodity to be recommended to obtain a representation of a historical interest commodity of the user;

[0009] classifying and aggregating the representation of the historical interest commodity to obtain multiple dissociation representations of the user;

[0010] According to the multiple dissociation characteristics, it is judged whether the recommended commodity is the commodity of interest of the user.

[0011] Optionally, the method is implemented based on a commodity recommendation model, and a training step of the commodity recommendation model comprises:

[0012] A set composed of multiple groups of user information and corresponding commodity information is input into the commodity recommendation model as a training set.

[0013] The commodity recommendation model selects samples of corresponding difficulty in the training set according to a current learning state, adjusts a difficulty distribution of the samples at a corresponding rate, and obtains the trained commodity recommendation model after learning.

[0014] Optionally, the commodity recommendation model selects samples of corresponding difficulty in the training set according to a current learning state, and the method comprises:

[0015] The commodity recommendation model obtains a corresponding loss value after learning the samples in the training set.

[0016] The loss value is compared with a pre-set hyperparameter, and the difficulty of the sample learning is judged according to a comparison result.

[0017] According to the self parameter of the commodity recommendation model, samples of corresponding difficulty are determined to be learned.

[0018] Optionally, the clicked commodity sequence and the non-clicked commodity sequence are filtered according to the sequence of disliked commodities to obtain the representation of the commodity of interest of the user, and the method comprises:

[0019] The clicked commodity sequence, the non-clicked commodity sequence and the sequence of disliked commodities are respectively input into an encoder based on a multi-head attention mechanism for encoding to obtain the representation of the clicked commodity, the representation of the non-clicked commodity and the representation of the disliked commodity.

[0020] The representation of the disliked commodity is average-pooled to obtain a negative tendency representation of the user.

[0021] The representation of the clicked commodity and the representation of the non-clicked commodity are filtered based on the negative tendency representation to obtain the representation of the commodity of interest of the user.

[0022] Optionally, the representation of the clicked commodity and the representation of the non-clicked commodity are filtered based on the negative tendency representation to obtain the representation of the commodity of interest of the user, and the method comprises:

[0023] The representation of the clicked commodity and the representation of the non-clicked commodity are similarity calculated with the negative tendency representation.

[0024] According to the similarity calculation result, the representation of the clicked commodity and the representation of the non-clicked commodity are filtered to obtain the representation of the commodity of interest of the user.

[0025] Optionally, the representation of the commodity of interest of the user is screened according to the click time information of the user and information of the commodity to be recommended to obtain the representation of the historical interest commodity of the user, including:

[0026] According to the click time information, the representation of the commodity of interest is assigned a corresponding weight value;

[0027] According to the information of the commodity to be recommended, the representation of the commodity of interest is assigned a corresponding weight value;

[0028] The representation of the commodity of interest after the weight value is assigned is taken as the representation of the historical interest commodity of the user.

[0029] Optionally, the representation of the historical interest commodity is classified and aggregated to obtain a plurality of dissociation representations of the user, including:

[0030] The distance between the representation of the historical interest commodity and a plurality of interest prototypes is calculated to obtain a plurality of distance calculation results;

[0031] According to the plurality of distance calculation results, the representation of the historical interest commodity is aggregated around the plurality of interest prototypes to obtain the plurality of dissociation representations.

[0032] The second aspect of the embodiment of the application provides a dissociation commodity recommendation device, and the device includes:

[0033] An information receiving module is configured to receive information of a commodity to be recommended and historical click information of a user, wherein the historical click information includes a clicked commodity sequence, a non-clicked commodity sequence, a disliked commodity sequence, and click time information of the user;

[0034] A representation filtering module is configured to filter the clicked commodity sequence and the non-clicked commodity sequence according to the disliked commodity sequence to obtain a representation of a commodity of interest of the user.

[0035] A representation screening module is configured to screen the representation of the commodity of interest of the user according to the click time information of the user and the information of the commodity to be recommended to obtain a representation of a historical interest commodity of the user.

[0036] A representation aggregation module is configured to classify and aggregate the representation of the historical interest commodity to obtain a plurality of dissociation representations of the user.

[0037] The recommendation prediction module is configured to determine whether the recommended product is a product of interest of the user according to the plurality of disentangled representations.

[0038] Optionally, the method is implemented based on a product recommendation model, and a training step of the product recommendation model comprises:

[0039] A set composed of a plurality of groups of user information and corresponding product information is input into the product recommendation model as a training set.

[0040] The product recommendation model selects samples of corresponding difficulty in the training set for learning according to a current learning state, and adjusts a difficulty distribution of the samples at a corresponding rate, and after learning is completed, a trained product recommendation model is obtained.

[0041] Optionally, the product recommendation model selects samples of corresponding difficulty in the training set for learning according to a current learning state, and a training step of the product recommendation model comprises:

[0042] The product recommendation model obtains a corresponding loss value after learning the samples in the training set.

[0043] The loss value is compared with a pre-set hyperparameter, and difficulty of learning of the samples is determined according to a comparison result.

[0044] Corresponding difficulty of samples for learning is determined according to self-parameters of the product recommendation model.

[0045] Optionally, the representation filtering module comprises:

[0046] A sequence encoding submodule is configured to input the clicked product sequence, the unclicked product sequence and the disliked product sequence into an encoder based on a multi-head attention mechanism for encoding to obtain a representation of the clicked product, a representation of the unclicked product and a representation of the disliked product.

[0047] A negative representation obtaining submodule is configured to perform average pooling on the representation of the disliked product to obtain a negative inclination representation of the user.

[0048] A representation filtering submodule is configured to filter the representation of the clicked product and the representation of the unclicked product based on the negative inclination representation to obtain a representation of a product of interest of the user.

[0049] Optionally, the representation filtering submodule comprises:

[0050] A similarity calculation submodule is configured to calculate a similarity between the representation of the clicked product, the representation of the unclicked product and the negative inclination representation.

[0051] The interest commodity feature determination submodule is configured to filter the features of the clicked commodities and the features of the non-clicked commodities according to the similarity calculation result, and obtain the features of the interest commodities of the user.

[0052] Optionally, the feature screening module comprises:

[0053] The first feature screening submodule is configured to assign a corresponding weight to the features of the interest commodities according to the click time information.

[0054] The second feature screening submodule is configured to assign a corresponding weight to the features of the interest commodities according to the information of the commodities to be recommended.

[0055] The historical interest commodity feature determination submodule is configured to take the features of the interest commodities after the weight assignment as the historical interest commodity features of the user.

[0056] Optionally, the feature aggregation module comprises:

[0057] The distance calculation submodule is configured to calculate distances between the historical interest commodity features and a plurality of interest prototypes, and obtain a plurality of distance calculation results.

[0058] The feature aggregation submodule is configured to aggregate the historical interest commodity features according to the plurality of distance calculation results, and take the plurality of interest prototypes as the center to obtain the plurality of dissociated features.

[0059] The third aspect of the embodiments of the present application provides a readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the method of the first aspect of the present application.

[0060] The fourth aspect of the embodiments of the present application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of the method of the first aspect of the present application.

[0061] The application provides a dissociation commodity recommendation method, which comprises the following steps: receiving information of a commodity to be recommended and historical click information of a user, wherein the historical click information comprises a click commodity sequence, a non-click commodity sequence, a non-favorite commodity sequence and click time information of the user; filtering the click commodity sequence and the non-click commodity sequence according to the non-favorite commodity sequence to obtain a representation of a commodity of interest of the user; screening the representation of the commodity of interest of the user according to the click time information of the user and the information of the commodity to be recommended to obtain a representation of a historical interest commodity of the user; classifying and aggregating the representation of the historical interest commodity of the user to obtain a plurality of dissociation representations of the user; and judging whether the commodity to be recommended is the commodity of interest of the user according to the plurality of dissociation representations. The dissociation representation of the user is obtained through the multi-feedback data of the user, i.e. the click commodity sequence, the non-click commodity sequence and the non-favorite commodity sequence, and the interest of the user is accurately captured, so that the accuracy of commodity recommendation is improved and the click rate of the user is increased. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the application. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0063] Figure 1 FIG. 1 is a flowchart of a dissociation commodity recommendation method according to an embodiment of the application;

[0064] Figure 2 FIG. 2 is a dissociation recommendation flowchart and a controllable self-evaluation course learning schematic diagram according to an embodiment of the application;

[0065] Figure 3 FIG. 3 is a schematic diagram of a dissociation commodity recommendation device according to an embodiment of the application. DETAILED DESCRIPTION

[0066] The technical solutions of the embodiments of the application will be clearly and completely described in combination with the drawings of the embodiments of the application. Obviously, the described embodiments are only some embodiments of the application, rather than all the embodiments of the application. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0067] The embodiment of the application is realized based on a commodity recommendation model, which can be used in a webpage or an APP. The commodity recommendation model is used to determine whether a commodity to be recommended is a commodity of interest to a user. According to the determination result, it is determined whether the commodity is recommended to the user.

[0068] Reference Figure 1 , Figure 1 is a flowchart of a dissociation commodity recommendation method according to an embodiment of the application. As shown in Figure 1 , the method comprises the following steps:

[0069] S11: receiving information of a commodity to be recommended and historical click information of a user, wherein the historical click information comprises a click commodity sequence, a non-click commodity sequence, a disliked commodity sequence and click time information of the user.

[0070] In the embodiment, the information of the commodity to be recommended is a sequence of the commodity to be recommended, which comprises a name of the commodity to be recommended, a type of the recommended commodity, a use of the recommended commodity and the like. The click commodity sequence of the user is a sequence of all commodities clicked by the user when browsing a commodity list. The non-click commodity sequence of the user is a sequence of commodities not clicked by the user when browsing the commodity list. The disliked commodity sequence of the user is a sequence of commodities marked as disliked by the user when browsing the commodity list. The commodity sequence comprises a name of the commodity, a type of the commodity and a use of the commodity. The click time information comprises a time when the user clicks the commodity.

[0071] In the embodiment, the historical click information of the user is multi-feedback information of the user received by the commodity recommendation model. The historical click information can be recorded by a website or an APP on a click behavior of the user on commodities within a period of time.

[0072] In the embodiment, the commodity to be recommended can be a commodity that a webpage or an APP wants to recommend to the user when the user browses the webpage or the APP. The sequence of all commodities can be a sequence obtained by feature extraction of a word embedding network. The type of the commodity can be a clothing, an electronic product or the like.

[0073] For example, the commodity clicked by the user can be jeans, which belongs to the clothing category. The disliked commodity of the user can be a commodity on which the user clicks a dislike button or a commodity complained by the user.

[0074] S12: filtering the click commodity sequence and the non-click commodity sequence according to the disliked commodity sequence to obtain a representation of a commodity of interest to the user.

[0075] In this embodiment, the goods in the sequence of goods not liked by the user are obviously marked by the user as goods not liked, and thus can be used as negative samples for recommendation. The goods in the sequence of clicked goods and the sequence of unclicked goods can also contain goods not liked by the user, and the sequence of unclicked goods can also contain goods liked by the user. Therefore, the sequence of goods not liked by the user is used to filter the sequence of clicked goods and the sequence of unclicked goods, and the representations corresponding to the sequence of clicked goods and the sequence of unclicked goods are assigned weights. The specific steps are as follows:

[0076] S12-1: The sequence of clicked goods, the sequence of unclicked goods, and the sequence of goods not liked by the user are respectively input into an encoder based on a multi-head attention mechanism for encoding to obtain representations of the clicked goods, representations of the unclicked goods, and representations of the goods not liked by the user.

[0077] In this embodiment, the encoder based on the multi-head attention mechanism is used to convert the initialized sequence of input goods into representations of the goods. These representations can more accurately reflect the features of the goods, which is beneficial for the next step of filtering by the entire neural network model.

[0078] In this embodiment, when processing the initialized sequence of goods, the encoder based on the multi-head attention mechanism can pay more attention to the key parts of the sequence that reflect the features of the goods. Therefore, the extracted features can more accurately reflect the attributes of the goods, thereby ensuring the accuracy of interest mining.

[0079] S12-2: Average pooling is performed on the representations of the goods not liked by the user to obtain a negative tendency representation of the user.

[0080] In this embodiment, the goods not liked by the user are marked by the user as goods not liked. Therefore, the confidence of the representations of the goods not liked is high. By performing average pooling on these representations, a unified representation can be obtained, which is used as the negative tendency representation of the user.

[0081] In this embodiment, the average pooling is performed by a pooling layer in the goods recommendation model.

[0082] S12-3: Based on the negative tendency representation, the representations of the clicked goods and the unclicked goods are filtered to obtain a representation of goods of interest to the user.

[0083] In this embodiment, the representation of goods of interest to the user is obtained by filtering the representations of the clicked goods and the unclicked goods of the user based on the negative tendency representation. The representation of goods of interest to the user is the representation obtained after filtering. The specific steps include:

[0084] S12-3-1: Similarity calculation is performed between the representations of the clicked goods and the unclicked goods and the negative tendency representation.

[0085] S12-3-2: According to the similarity calculation result, the representation of the clicked commodity and the representation of the non-clicked commodity are filtered to obtain the representation of the commodity of interest of the user.

[0086] In this embodiment, the representation of all clicked commodities and the negative tendency representation are calculated for similarity, the representation of all non-clicked commodities and the negative tendency representation are calculated for similarity, and according to the similarity calculation result, the representation of the clicked commodity and the representation of the non-clicked commodity are filtered to obtain the representation of the commodity of interest of the user.

[0087] If the similarity between the representation corresponding to a commodity and the negative tendency representation is low, it means that the commodity corresponding to the representation is quite different from the commodity that the user does not like, and it is more likely to belong to the commodity that the user likes. If the similarity with the negative tendency representation is high, it means that the commodity corresponding to the representation is likely to belong to the commodity that the user does not like. For the representation with low similarity to the negative tendency representation, a higher weight is assigned to the representation, and the representation with higher weight has a greater impact on the final fused representation when the representations are fused. For the representation with high similarity to the negative tendency representation, a very low weight is assigned to the representation, and the representation has no impact on the final fused representation when the representations are fused. Each representation in the representation of the commodity of interest of multiple users is filtered from the representation of the clicked commodity and the representation of the non-clicked commodity, that is, each representation is assigned a weight according to the similarity calculation result. After sorting the representations, the weight of the representation of the commodity of interest of the user is high and the weight of the representation of the commodity that the user does not like is low after filtering.

[0088] For example, in the user historical click information, the clicked commodities are mobile phones, tablets, and televisions, among which the television is a mistaken click of the user, and the user actually does not need a television. The non-clicked commodities in the page are digital cameras, smart bracelets, electronic watches, and television sets, and the user accidentally clicks on the television again and marks it as not liked. After inputting the historical click information into the commodity recommendation network, the commodity recommendation network performs average pooling on the sequence corresponding to the television, filters the representation of the clicked commodity and the representation of the non-clicked commodity, and assigns a high weight to the representation corresponding to the mobile phone, the tablet, the digital camera, the smart bracelet, the smart watch, and the electronic watch, and assigns a very low weight to the representation corresponding to the television.

[0089] In this embodiment, the representation corresponding to the commodity is filtered based on the negative representation, that is, it is considered that there may be a commodity that the user does not like among the clicked commodities, and it is also considered that there may be a commodity of interest of the user among the non-clicked commodities, thereby reducing the impact of noise on the commodity recommendation effect from the feature level.

[0090] S13: Based on the user's click time information and the information of the products to be recommended, the representation of the user's interested products is filtered to obtain the representation of the user's historical interested products.

[0091] In this embodiment, the specific steps for filtering the representation of the user's interested products based on the user's click time information and the information of the products to be recommended, to obtain the representation of the user's historical interested products, are as follows:

[0092] S13-1: Based on the click time information, assign corresponding weights to the representation of the product of interest.

[0093] In this embodiment, the click time information is included in the user's historical click information. The representation of the user's interested products obtained after the previous step includes the representation of the clicked products. Based on the click time of the product corresponding to the representation, the representation is filtered, and the representation corresponding to the product whose click time is closer to the current time is given a higher weight, while the representation corresponding to the product whose click time is farther from the current time is given a very low weight. When performing representation fusion, the representation corresponding to the product whose click time is farther from the current time will not affect the fusion result.

[0094] In this embodiment, the product recommendation model assigns weights to the representation of each product based on the click time, thus avoiding the influence of products whose click time is far removed from the current time on the recommendation results.

[0095] S13-2: Based on the information of the products to be recommended, assign corresponding weights to the representations of the products of interest.

[0096] In this embodiment, the information of the product to be recommended includes the representation of the product to be recommended. When the representation of the product that the user is interested in has a significant difference from the representation of the product to be recommended, the representation is assigned a very low weight. When the representation is fused, the representation will not affect the fusion result. When the difference between the representation of the product and the representation of the product to be recommended is small, it means that the two products are relatively similar. When the representation is fused, the representation has a significant impact on the fusion result, which is beneficial to recommending products that the user is interested in.

[0097] S13-3: The representation of the interested products after weight assignment is used as the representation of the user's historical interested products.

[0098] S14: Classify and aggregate the representations of the historical interest items to obtain multiple deconstructed representations of the user.

[0099] In this embodiment, the specific steps for classifying and aggregating the representations of the historical interest items to obtain multiple dissociated representations of the user are as follows:

[0100] S14-1: Calculate the distance between the representation of the historical interest goods and a plurality of interest prototypes, to obtain a plurality of distance calculation results.

[0101] In this embodiment, the interest prototype refers to the representation of the category of the goods, and the distance between the representation of the historical interest goods and the interest prototype is the distance between them in the representation space.

[0102] In the representation space, the distance between two representations with high similarity is closer. The distance between the representation of the goods of interest and the interest prototype is calculated, and when the distance is less than a certain threshold, it means that the goods belong to the interest prototype.

[0103] For example, the category corresponding to the interest prototype can be a try-on clothing category, a snack category, a sports goods category, etc. When the goods are trousers, the representation of the goods is closer to the clothing category interest prototype, when the goods are biscuits, spicy strips, etc., the representation of the goods is closer to the snack category interest prototype, and when the goods are basketballs and footballs, the representation of the goods is closer to the sports goods category interest prototype.

[0104] S14-2: According to the plurality of distance calculation results, the representation of the historical interest goods is aggregated around the plurality of interest prototypes to obtain a plurality of dissociated representations.

[0105] In this embodiment, according to the plurality of distance calculation results, the distance between the representation of each historical interest goods and the interest prototype can be obtained. The representation of each historical interest goods has a closest interest prototype. The representation of the historical interest goods closest to the plurality of interest prototypes is aggregated to obtain a plurality of dissociated representations, which represent the representation of the same type of historical interest goods.

[0106] For example, when the goods are trousers, clothes, the representation of the goods is closer to the clothing category interest prototype, and the representation of the trousers and clothes is aggregated to obtain a dissociated representation, which reflects the user's preference for clothing type goods; when the goods are biscuits, spicy strips, etc., the representation of the goods is closer to the snack category interest prototype, and the representation corresponding to the biscuits and spicy strips is aggregated to obtain a dissociated representation, which reflects the user's preference for snack goods; when the goods are basketballs and footballs, the representation of the goods is closer to the sports goods category interest prototype, and the representation corresponding to the basketballs and footballs is aggregated to obtain a dissociated representation, which reflects the user's preference for sports goods.

[0107] S15: According to the plurality of dissociated representations, it is judged whether the goods to be recommended are the goods of interest of the user.

[0108] In this embodiment, after obtaining the plurality of dissociative representations, the similarity between the representation of the recommended commodity and the plurality of dissociative representations is calculated, and the result of the calculation indicates that the representation of the commodity to be recommended is close to which representation, and then it is determined that the commodity to be recommended belongs to which category. Then, according to the characteristics of the commodity and the information of the user, it is determined whether the commodity to be recommended is the commodity of interest of the user.

[0109] For example, the commodity to be recommended is a basketball shoe. The representation of the commodity is relatively close to the representation of the clothing category and is also relatively close to the representation of the sports supplies category. The information of the user indicates that the user is a male. Therefore, the commodity to be recommended is probably the commodity of interest of the user. The commodity is determined as the commodity of interest of the user. The information of the user can be input into the commodity recommendation model together with the historical click information of the user at the beginning.

[0110] In this embodiment, the dissociative representation of the user is obtained from the multi-feedback data of the user, the interest of the user is accurately captured, and the accuracy of the commodity recommendation is improved.

[0111] In another embodiment of the present application, the training step of the commodity recommendation model comprises:

[0112] S21: A set composed of a plurality of sets of user information and corresponding commodity information is input into the commodity recommendation model as a training set.

[0113] S22: The commodity recommendation model selects samples with corresponding difficulty in the training set for learning according to the current learning state, and adjusts the difficulty distribution of the samples at a corresponding rate. After learning is completed, the trained commodity recommendation model is obtained.

[0114] In this embodiment, the user information includes the ID information of the user, the historical click information of the user, and the like. The commodity information includes the name, category, and purpose of the commodity, and the like. The information also indicates whether the commodity is the commodity of interest of the user.

[0115] In the process of model training, the model can select samples with corresponding difficulty for learning according to the current learning state, i.e., the parameters currently obtained by the model. That is, samples with lower learning difficulty can be learned first, and then samples with higher learning difficulty can be learned. The difficulty distribution of the samples is adjusted at a corresponding rate, i.e., the samples in the entire training set are gradually learned at a certain rate, instead of being limited to samples with lower learning difficulty.

[0116] For example, the rate of difficulty adjustment can be to adjust the learning sample difficulty once every 10 training rounds, and finally gradually cover the learning samples to the entire training set to realize dynamic adjustment of the difficulty distribution of the samples.

[0117] The commodity recommendation model selects a sample of a corresponding difficulty in the training set according to a current learning state for learning, including:

[0118] S21-1: After the commodity recommendation model learns the sample in the training set, a corresponding loss value is obtained.

[0119] S21-2: The loss value is compared with a pre-set hyperparameter, and the difficulty of the sample learning is determined according to a comparison result.

[0120] S21-3: According to the self parameter of the commodity recommendation model, a sample of a corresponding difficulty is determined for learning.

[0121] In the embodiment, before the commodity recommendation model is trained, a hyperparameter is pre-set. The hyperparameter limits some parameters in the model, which will not be changed during training, for example, the dimension of each layer of the model and the expected loss value of the model. After the commodity recommendation model learns the sample, a loss value is obtained. The loss value is compared with the pre-set hyperparameter. If the loss value is similar to the value set in the hyperparameter, the weight of the learned sample is increased, and the sample is learned in more detail. When the sample is learned, the loss value is gradually selected to be different from the value set in the hyperparameter for learning.

[0122] For example, the expected loss value set in the hyperparameter is 0.5. The sample with a loss value of 0.5 is first learned. After learning, the sample with a loss value of 0.4, 0.6, 0.7, and 0.8 is learned. Thus, the learning of the entire training set is completed, and the trained commodity recommendation model is obtained.

[0123] In the embodiment, the difficulty of the learning sample is dynamically adjusted. The appropriate learning strategy is selected for the model as much as possible. The model is not affected by the noise and is not in a local optimal state. The difficulty of the data and the selection of the learning rate are set by the hyperparameter. The learning of various data sets can be conveniently applied without introducing additional training parameters. The efficiency of the model training and the effect of the trained model are guaranteed.

[0124] In another embodiment of the application, the trained commodity recommendation model is tested. If the test result of the commodity recommendation model is not ideal, that is, the success rate of the recommended commodity is not high, the hyperparameter of the model can be adaptively modified, and the commodity recommendation model is further trained.

[0125] In another embodiment of the application, the dissociation recommendation process and the controllable self-evaluation course learning diagram are combined to further illustrate the application.

[0126] As shown in Figure 2 ,Figure 2 is a dissociation recommendation process and a controllable self-evaluation course learning diagram proposed by an embodiment of the present application, wherein C, D and U represent the model of the encoder, d c and d u represent the weight of the vector, and F represents the user information.

[0127] As Figure 2 mentioned, in the interactive filtering dynamic routing module, it is divided into three steps of interest mining, intention aggregation and prediction.

[0128] In the interest mining step, the click sequence (1, 2, 3, 4, representing different goods respectively) is input into the C encoder to obtain the representation of the clicked goods of the user. Similarly, the representation of the disliked goods and the representation of the unclicked goods are obtained. Then the representation of the disliked goods is averaged and pooled to obtain the negative tendency. Based on the negative tendency, the clicked goods representation and the unclicked goods representation are filtered. As can be seen from the figure, in the clicked goods sequence, the background of representation 3 becomes lighter, representing that representation 3 is assigned a very low weight, and in the unclicked goods sequence, representation 4 is assigned a very low weight. After the time candidate goods attention module, the representation 4 of the unclicked sequence is assigned a very low weight.

[0129] In the intention aggregation step, the remaining representations are aggregated into intentions. Representation 1 and representation 2 in the clicked sequence are aggregated into one dissociation representation, representation 4 in the clicked sequence and representation 1 in the unclicked sequence are aggregated into one dissociation representation, and representation 3 in the unclicked sequence is aggregated into one dissociation representation, obtaining three dissociation representations.

[0130] In the prediction step, the similarity between the three dissociation representations and the representation of the goods is calculated, and then added, and combined with the user's information to obtain the final result, to determine whether the to-be-recommended goods are the goods of interest to the user.

[0131] In the controllable self-evaluation course learning module, the learning process of the entire model is subject to a Gaussian distribution. The figure is a three-dimensional coordinate axis. The difficulty of the learned sample expands constantly with the change of the training round until the entire training set is learned.

[0132] Based on the same inventive concept, an embodiment of the present application provides a dissociation goods recommendation device. Referring to Figure 3 , Figure 3 is a schematic diagram of a dissociation goods recommendation device 300 proposed by an embodiment of the present application.

[0133] As Figure 3 shown, the device comprises:

[0134] Information receiving module 301 is used to receive information about the products to be recommended and the user's historical click information. The historical click information includes the user's clicked product sequence, unclicked product sequence, disliked product sequence, and click time information.

[0135] The characterization filtering module 302 is used to filter the clicked product sequence and the unclicked product sequence based on the disliked product sequence to obtain a characterization of the user's interested products;

[0136] The characterization filtering module 303 is used to filter the characterizations of the user's interested products based on the user's click time information and the information of the products to be recommended, so as to obtain the characterizations of the user's historical interested products.

[0137] The characterization aggregation module 304 uses a characterization of the historical interest items to classify and aggregate them, thereby obtaining multiple deconstructed characterizations of the user.

[0138] The recommendation prediction module 305 is used to determine whether the product to be recommended is a product of interest to the user based on the multiple dissociated representations.

[0139] Optionally, the method is based on a product recommendation model, the training steps of which include:

[0140] The set of multiple sets of user information and their corresponding product information is used as the training set and input into the product recommendation model.

[0141] The product recommendation model selects samples of corresponding difficulty from the training set for learning based on the current learning state, and adjusts the difficulty distribution of the samples at a corresponding rate. After learning is completed, the trained product recommendation model is obtained.

[0142] Optionally, the product recommendation model selects samples of corresponding difficulty from the training set for learning based on the current learning state, including:

[0143] The product recommendation model learns from the samples in the training set and obtains the corresponding loss value.

[0144] The loss value is compared with the preset hyperparameters, and the difficulty of learning the sample is determined based on the comparison results.

[0145] Based on the parameters of the product recommendation model, samples of corresponding difficulty are determined for learning.

[0146] Optionally, the characterization filtering module includes:

[0147] The sequence coding submodule is configured to input the clicked commodity sequence, the unclicked commodity sequence and the disliked commodity sequence into an encoder based on a multi-head attention mechanism respectively for coding to obtain a clicked commodity representation, an unclicked commodity representation and a disliked commodity representation.

[0148] The negative representation obtaining submodule is configured to perform average pooling on the disliked commodity representation to obtain a negative inclination representation of the user.

[0149] The representation filtering submodule is configured to filter the clicked commodity representation and the unclicked commodity representation based on the negative inclination representation to obtain a representation of a commodity of interest of the user.

[0150] Optionally, the representation filtering submodule comprises:

[0151] The similarity calculation submodule is configured to perform similarity calculation on the clicked commodity representation, the unclicked commodity representation and the negative inclination representation.

[0152] The commodity of interest representation determining submodule is configured to filter the clicked commodity representation and the unclicked commodity representation according to a result of similarity calculation to obtain the representation of the commodity of interest of the user.

[0153] Optionally, the representation screening module comprises:

[0154] The first representation screening submodule is configured to assign a corresponding weight to the representation of the commodity of interest according to the click time information.

[0155] The second representation screening submodule is configured to assign a corresponding weight to the representation of the commodity of interest according to the information of the to-be-recommended commodity.

[0156] The historical interest commodity representation determining submodule is configured to take the representation of the commodity of interest after the weight assignment as a historical interest commodity representation of the user.

[0157] Optionally, the representation aggregation module comprises:

[0158] The distance calculation submodule is configured to calculate distances between the historical interest commodity representation and a plurality of interest prototypes to obtain a plurality of distance calculation results.

[0159] The representation aggregation submodule is configured to aggregate the historical interest commodity representation with the plurality of interest prototypes as centers according to the plurality of distance calculation results to obtain a plurality of dissociated representations.

[0160] Based on the same inventive concept, another embodiment of the present application provides a readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the dissociated commodity recommendation method according to any one of the above embodiments of the present application.

[0161] Based on the same inventive concept, another embodiment of the present application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor is executed to implement the steps of the dissociated commodity recommendation method according to any one of the above embodiments of the present application.

[0162] For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts are described in the part of the method embodiment.

[0163] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other.

[0164] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a device, or a computer program product. Therefore, the embodiments of the present application can be in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams according to the method, terminal device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the computer or other programmable data processing terminal device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The device that implements the function specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the function specified in one flow or multiple flows and / or blocks

[0166] These computer program instructions can also be stored in a computer-readable storage medium, which can guide the computer or other programmable data processing terminal device to work in a specific way, so that the instructions stored in the computer-readable storage medium produce a product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0167] These computer program instructions can also be loaded into computer or other programmable data processing terminal devices, so that a series of operation steps are performed on the computer or other programmable terminal devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable terminal devices provide a process for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0168] Although the preferred embodiments of the present application have been described, those skilled in the art who have the basic inventive concept can make further changes and modifications to the embodiments. Therefore, the appended claims are intended to include the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present application.

[0169] Finally, it should also be noted that, in this paper, relational terms such as first and second are used merely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or terminal device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or terminal device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or terminal device including the element.

[0170] The above provides a detailed introduction to the dissociation commodity recommendation method, device, equipment and storage medium provided by the present application. In this paper, specific examples are applied to explain the principles and implementation modes of the present application. The above description of the embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A deconstructed product recommendation method, the method comprising: Receive information about products to be recommended and the user's historical click information, including the user's clicked product sequence, unclicked product sequence, disliked product sequence, and click time information; Based on the disliked product sequence, the clicked product sequence and the unclicked product sequence are filtered to obtain a representation of the user's interested products, including: The clicked product sequence, the unclicked product sequence, and the disliked product sequence are respectively input into an encoder based on a multi-head attention mechanism for encoding to obtain the representations of clicked products, unclicked products, and disliked products. The user's negative tendency is obtained by average pooling the representation of disliked products. Based on the negative tendency representation, the representations of the clicked products and the non-clicked products are filtered to obtain the representations of the products that the user is interested in. Based on the click time information and the information of the products to be recommended, the characteristics of the products of interest are filtered to obtain the characteristics of the user's historical products of interest; The representations of the historical interest items are classified and aggregated to obtain multiple deconstructed representations of the user; Based on the multiple deconstructed representations, it is determined whether the product to be recommended is a product of interest to the user.

2. The method according to claim 1, characterized in that, The method is based on a product recommendation model, and the training steps of the product recommendation model include: The set of multiple sets of user information and their corresponding product information is used as the training set and input into the product recommendation model. The product recommendation model selects samples of corresponding difficulty from the training set for learning based on the current learning state, and adjusts the difficulty distribution of the samples at a corresponding rate. After learning is completed, the trained product recommendation model is obtained.

3. The method according to claim 2, characterized in that, The product recommendation model selects samples of corresponding difficulty from the training set for learning based on the current learning state, including: The product recommendation model learns from the samples in the training set and obtains the corresponding loss value. The loss value is compared with the preset hyperparameters, and the difficulty of learning the sample is determined based on the comparison results. Based on the parameters of the product recommendation model, samples of corresponding difficulty are determined for learning.

4. The method according to claim 1, characterized in that, Based on the negative tendency representation, the representations of clicked and unclicked products are filtered to obtain the representation of the products the user is interested in, including: The similarity between the representation of the clicked product and the representation of the unclicked product and the negative tendency representation is calculated. Based on the similarity calculation results, the representations of the clicked products and the unclicked products are filtered to obtain the representations of the products that the user is interested in.

5. The method according to claim 1, characterized in that, Based on the user's click time information and the information of the products to be recommended, the representation of the user's interested products is filtered to obtain the representation of the user's historical interested products, including: Based on the click time information, the representation of the product of interest is assigned a corresponding weight; Based on the information of the products to be recommended, the characteristics of the products of interest are assigned corresponding weights; The representation of the products of interest after weight assignment is used as the representation of the user's historical products of interest.

6. The method according to claim 1, characterized in that, The representations of the historical interest items are categorized and aggregated to obtain multiple deconstructed representations of the user, including: Calculate the distance between the representation of the historical interest item and multiple interest prototypes to obtain multiple distance calculation results; Based on the multiple distance calculation results, the representations of the historical interest products are aggregated with the multiple interest prototypes as the center to obtain the multiple dissociated representations.

7. A dissociative product recommendation device, characterized in that, The device includes: The information receiving module is used to receive information about the products to be recommended and the user's historical click information. The historical click information includes the user's clicked product sequence, unclicked product sequence, disliked product sequence, and click time information. The characterization filtering module is used to filter the clicked product sequence and the unclicked product sequence based on the disliked product sequence to obtain a characterization of the user's interested products, including: The clicked product sequence, the unclicked product sequence, and the disliked product sequence are respectively input into an encoder based on a multi-head attention mechanism for encoding to obtain the representations of clicked products, unclicked products, and disliked products. The user's negative tendency is obtained by average pooling the representation of disliked products. Based on the negative tendency representation, the representations of the clicked products and the non-clicked products are filtered to obtain the representations of the products that the user is interested in. The representation filtering module is used to filter the representations of the user's interested products based on the user's click time information and the information of the products to be recommended, so as to obtain the representations of the user's historical interested products. The representation aggregation module uses a representation of the historical interest items to classify and aggregate them, resulting in multiple deconstructed representations of the user. The recommendation prediction module is used to determine whether the product to be recommended is a product of interest to the user based on the multiple deconstructed representations.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

Images