A personalized package recommendation method based on user history and feature factorization
By using factorization and deep learning models in the personalized fashion suit recommendation system, combining user historical behavior and suit compatibility, the problem of lack of targetedness and compatibility of recommendation results in the existing system is solved, and more accurate personalization is achieved Recommended fashion suits.
Patent Information
- Application Number
- CN202210828067.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-13
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-07-13
AI Technical Summary
The existing personalized fashion suit recommendation system fails to effectively utilize user historical behavior and feature factorization, resulting in a lack of targetedness and compatibility in recommendation results, and the model focuses too much on overall preferences and ignores the combination of products within the suit.
By extracting product visual features, using factorization to project user and product features into multiple implicit spaces, learning users' interests in different aspects of the set, and constructing user historical behavior representations, combining product compatibility and user relevance measurements within the set The matching score between the user and the suit is established to establish an end-to-end deep learning model.
It realizes more accurate and personalized fashion suit recommendations, which can better capture users' potential interests and suit characteristics and enhance the competitiveness of recommendation results.
Smart Images

Figure CN115146170B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of personalized fashion recommendation, and in particular to a method for recommending personalized fashion sets based on user historical behaviors and feature factorization, and has achieved competitive recommendation results on a commonly used fashion set recommendation data set. Background Art
[0002] Fashion recommendation system is a hot research topic in the recommendation field. Fashion-centric e-commerce and social networks have become a vibrant field. Currently, most recommendation algorithms are only oriented to fashion items, but users have an increasing demand for fashion set recommendations. This has inspired the development of accurate, high-quality, personalized fashion set recommendation systems. Set recommendation needs to solve the compatibility problem of the items within the set, and the recommended sets must meet the personalized needs of users. Excellent recommendation results provide convenience to users and can also bring considerable economic benefits to enterprises.
[0003] The most traditional method of recommendation system is collaborative filtering, which recommends suitable products to users by measuring the relationship between users, products, and users. In recent years, prediction models based on deep learning have achieved remarkable results. Compared with traditional linear models, deep learning-based methods enhance model capabilities by learning implicit feature interactions through nonlinear transformations, and have better learning capabilities for feature combinations by learning low-dimensional representations from high-dimensional sparse features.
[0004] In the recommendation system, users often do not express their intentions clearly, and there are only implicit user feedback signals, such as feedback information of clicks, purchases, collections, etc. Therefore, capturing user interests from user behavior is crucial for the recommendation results. In personalized single product recommendation, user historical behavior is usually used to enrich the feature expression of the user end; but in personalized package recommendation, the existing method does not use user historical behavior. The present invention enriches user representation by constructing a user history sequence. The user's interest expression should also be multifaceted. For example, user anchor points are used to characterize the user's interest in different types of packages, but this method initializes the user features obtained by model learning in a random manner, and does not establish a connection with the user's basic information. At the same time, the existing method measures the representation of the package as a whole, which will make the model too focused on learning the overall preference of the package. The features of the package should correspond to multiple different implicit spaces, such as shape, pattern, color, etc. Therefore, the implicit feature expression of the package is learned by factoring the features of the package. At the same time, the user end also adopts a factoring method to calculate the matching degree between the user and the package. For a package, the combination of each item in the package also affects the recommendation results to a certain extent. Therefore, when measuring the relationship between the user and the package, the compatibility of the internal products should also be considered. Summary of the invention
[0005] The present invention provides a personalized package recommendation method based on user history and feature factorization. This method extracts the visual features of the product, obtains the basic representation of the user with the user embedding vector, and then projects the product features and user features into multiple implicit spaces in a factorized manner, learns the user's interest in different aspects of the package, uses feature high-order combination and weighted pooling to aggregate the product features to obtain the package features, constructs the user history behavior to learn the user history representation, then aggregates the user embedding vector and the user history representation to obtain the user representation, and measures the matching score between the user and the package from two aspects: the compatibility of the products in the package and the relevance of the package to the user. This method incorporates user history behavior and feature factorization, and measures the matching score between the user and the package from two aspects: package compatibility and user personalization, and constructs an end-to-end deep learning model. Competitive recommendation results have been achieved on commonly used package personalized recommendation datasets.
[0006] A personalized package recommendation method based on user history and feature factorization includes the following steps:
[0007] Step (1): image data preprocessing;
[0008] Perform data augmentation on images in the fashion outfit recommendation dataset.
[0009] Step (2): Extract product visual features and obtain basic user representation through user embedding vectors. Project product features and user features into multiple implicit spaces through factorization, learn users’ interests in different aspects of the package, and make each implicit space different through distribution constraints.
[0010] Step (3): Aggregate product features through high-order feature combination and weighted pooling to obtain package representation.
[0011] Step (4): Sample the user's historical behavior from the set of sets that the user has interacted with to learn the user's historical representation, and then aggregate the user embedding vector and the user's historical representation to obtain an improved user representation.
[0012] Step (5): Measure the matching score between the user and the package from two aspects: the compatibility of the products in the package and the relevance of the package to the user.
[0013] Step (6): Build an end-to-end deep learning model according to the above steps, train the entire network on the training set, and test the performance of the final model on the test set.
[0014] Furthermore, the specific method of step (1) is as follows:
[0015] Obtain an existing fashion suit recommendation dataset, which consists of a training set and a test set. For the image data in the training set, flip the image horizontally with a probability of 50%. Finally, normalize the images in the training set and the test set to obtain the image data after data enhancement.
[0016] Furthermore, the specific method of step (2) is as follows:
[0017] 2-1. For suits, use the basic backbone network to extract the visual features of fashion items in the suit The combination of product attributes is regarded as an implicit factor. In order to learn the implicit k factors, the feature transformation matrix Project the product features into multiple different implicit spaces so that each space can implicitly represent a certain attribute of the product. The product is factorized and recorded as The factorized set is represented as Among them, n represents the dimension of the number of fashion items in the suit, k represents the dimension of the number of factors, and d represents the dimension of the visual features of the fashion items.
[0018] 2-2. For users, feature embedding is performed based on user data information to obtain user embedding vectors Similarly, through the feature transformation matrix Project the user features into multiple different implicit spaces and factorize the user features into
[0019] 2-3. In order to ensure the difference of the implicit space, the logarithmic determinant divergence loss is used to make the transformation matrix sparse.
[0020] Furthermore, the specific method of step (3) is as follows:
[0021] After obtaining the factorized representation of the products in the set, in order to learn the combination relationship between the products in the set, the multi-head attention mechanism is used to combine the product features, and the stacked self-attention function is used to learn the high-order combination relationship between the products. It can be expressed as follows:
[0022]
[0023] where σ(·) represents the feed-forward layer in row dimension, They represent self-attention, normalization and multi-head self-attention operations respectively, and H represents the set features after multi-head self-attention and residual normalization.
[0024] Finally, the high-level features of the suit are weighted and summed through the attention mechanism to obtain the overall expression of the suit.
[0025] Furthermore, the specific method of step (4) is as follows:
[0026] 4-1. The set of user-package pairs is denoted by Is user u i The associated set of packages, i.e. user u i The positive sample set in, For user u i The number of associated sets.
[0027] Since there is no time information of the user-suite interaction in the user-suite pair, i Related Set Collections Select N from the positive sample subset b The packages are used as the historical behavior expression of the user, denoted as In order to better express the user's behavioral intention, the sequence is used to filter out accidental or noisy user behavior data as much as possible, and the minimum variance is used as the basis for the sequence. A subset is selected from the dataset as the user behavior representation.
[0028] 4-2. User characteristics are embedded in their own information and user behavior sequence Aggregate representation. For a set The factorization and aggregation methods of the package features in steps (1) and (2) are used to perform the same operation on each package in the user's history sequence, and then average pooling is performed to obtain the user history representation Therefore, user characteristics It can be expressed as follows:
[0029]
[0030] Among them, η(·) represents the operation of high-order feature combination and weighted pooling, Represents an average pooling operation.
[0031] Furthermore, the specific method of step (5) is as follows:
[0032] The matching score between the user and the package is measured from two aspects: the correlation between the user characteristics and the package characteristics and the compatibility of the products in the package. Personalized score Compatibility score is calculated by the inner product of user features and package features. It is measured by the average inner product of the items in the set. Since the score weights of the two aspects are different, a learnable weight vector is used to adaptively learn the weights. It can be expressed as follows:
[0033]
[0034] Among them, M (p) and M (c) The learnable weight vectors representing the personalization score and the compatibility score, respectively, and They represent the factorized features of the pth and qth items in the jth package, The number of combinations of choosing 2 items from n items in a set.
[0035] Furthermore, the specific method of step (6) is as follows:
[0036] The end-to-end deep learning model, namely the personalized package recommendation network model, first extracts the visual features of the products through a convolutional neural network, obtains the basic user representation with the user embedding vector, and then projects the product features and user features into multiple implicit spaces through the feature factorization module. Then, the feature aggregation module performs high-order combination and weighted pooling of the package features to obtain the package representation, constructs the user historical behavior to learn the user historical representation, concatenates the user basic representation and the user historical representation to obtain the user representation, and finally calculates the compatibility score by the pairwise inner product of the products in the package, and obtains the correlation score by the row-dimensional inner product of the package representation and the user representation, and the sum of the two obtains the total matching score.
[0037] The personalized package recommendation network model is trained in the form of triples. i , the positive sample set is Randomly select a set of sets that are not positive samples from all sets O The set of suits is used as the negative sample set of the user So we get a training set consisting of a set of triple data It can be expressed as:
[0038]
[0039] Among them, the triple (u i ,p,q) means that package p is more popular with user u than package q i welcome.
[0040] The model aims to optimize the Bayesian ranking loss and the logarithmic determinant divergence loss. Therefore, the overall loss function of the model can be defined as follows:
[0041]
[0042] Among them, λ1 and λ2 are weight parameters, θ u and θ o are the network parameters that characterize users and sets, and ||·||1 represents the L1 norm.
[0043] During the training process, the parameters of each part of the network model are updated simultaneously, the final model is obtained and saved, and the training effect is tested on the test set.
[0044] Furthermore, the basic backbone network adopts AlexNet, VGG or ResNet to extract the visual features of the product.
[0045] Furthermore, in order to ensure the difference of implicit space, the logarithmic determinant divergence loss is used to make the transformation matrix sparse, and the constraints on user space and the constraints of the suit space It is expressed as follows:
[0046]
[0047] Where c is a constant, tr(·) is the rank of the matrix, logdet(·) is the logarithm of the matrix determinant, and T is the matrix transpose.
[0048] Beneficial effects of the present invention:
[0049] Based on the idea of constructing user historical behavior sequences and factorizing features, a method for personalized package recommendation is proposed. By constructing user historical behavior sequences, this method enriches the user's representation, making the recommendation results closer to the user's potential interests in historical behavior. By factorizing package features and user features, the correlation between users and packages is measured from multiple implicit spaces, making the recommendation results more able to reflect the user's interest in certain aspects of the package. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 It is a specific flow chart of an embodiment of the method of the present invention;
[0051] Figure 2 It is a schematic diagram of a network framework constructed in an embodiment of the method of the present invention.
[0052] Implementation details
[0053] The following is combined with Figure 1 , 2 The present invention is further described in detail.
[0054] A personalized package recommendation method based on user history and feature factorization, the steps are as follows:
[0055] Step (1): image data preprocessing;
[0056] Get a commonly used fashion suit recommendation dataset, which consists of a training set and a test set. In order to better extract the visual features of the products, it is necessary to perform conventional data enhancement operations on the images before the model starts training and testing.
[0057] Step (2): Feature factorization;
[0058] Extracting product visual features. Since products have multiple attributes, such as shape, color, pattern, etc., this method projects product features into multiple implicit spaces by factorization. Similarly, user interests correspond to product attributes, so after obtaining the basic representation of users with user embedding vectors, we also project them into multiple implicit spaces by factorization. And make each implicit space different by distribution constraints.
[0059] Step (3): suit characterization polymerization;
[0060] Since a suit contains multiple fashion items, after obtaining the factored features of the products, it is necessary to aggregate them to obtain the factored representation of the suit. This method uses Set Transformer to perform high-order combinations of product features and weighted pooling aggregation to obtain the suit representation.
[0061] Step (4): User representation aggregation;
[0062] Since the user's historical behavior records are not directly given in the dataset, we sample the user's historical behavior from the set of sets that the user has interacted with to learn the user's historical representation, and then aggregate the user embedding vector and the user's historical representation to obtain an improved user representation.
[0063] Step (5): correlation measurement;
[0064] Personalized package recommendation requires measuring the correlation between user characteristics and package characteristics. At the same time, the compatibility of each product within the package can also help to obtain better recommendation results. To this end, this method measures the matching score between users and packages from two aspects: the compatibility of products within the package and the correlation between the package and the user.
[0065] Step (6): Model training and testing;
[0066] According to the above steps, an end-to-end deep learning model is established, the entire network is trained and fine-tuned on the training set, and the performance of the final model is tested on the test set.
[0067] The image data preprocessing described in step (1) is specifically as follows:
[0068] Get a commonly used fashion suit recommendation dataset, which consists of a training set and a test set. For the 291*291 image data in the training set, flip the image horizontally with a probability of 50%. Finally, normalize the images in the training set and the test set to obtain the image data after data enhancement.
[0069] The feature factorization process described in step (2) is specifically as follows:
[0070] 2-1. Record the collection of fashion items as The set of packages is recorded as The user set is denoted as A set is made up of fashion items from multiple categories, including tops, bottoms, shoes, etc. Each set can be represented by a combination of items. Where N i Represents the number of all fashion items, N o Represents the number of all sets, N u Indicates the number of all users. Indicates the number of items in a set.
[0071] 2-2. For suits, use the basic backbone network to extract the visual features of fashion items in the suit Since products have multiple attributes, such as shape, color, pattern, etc., the combination of product attributes is regarded as an implicit factor. In order to learn the implicit k factors, the feature transformation matrix Project the product features into multiple different implicit spaces so that each space can implicitly represent a certain attribute of the product. The product is factorized and recorded as The factorized set can be expressed as Where n represents the dimension of the number of fashion items in the suit, k represents the dimension of the number of factors, and d represents the dimension of the visual features of the fashion items. The basic backbone network uses AlexNet, VGG or ResNet to extract the visual features of the goods.
[0072] 2-3. For users, feature embedding is performed based on user data information to obtain user embedding vectors Similarly, in order to learn the user's attention to various aspects of the package, the feature transformation matrix Project the user features into multiple different implicit spaces and factorize the user features into
[0073] 2-4. In order to ensure the difference of implicit space, the logarithmic determinant divergence loss is used to make the transformation matrix sparse, and the constraints on user space and the constraints of the suit space It can be expressed as follows:
[0074]
[0075] Where c is a constant, tr(·) is the rank of the matrix, logdet(·) is the logarithm of the matrix determinant, and T is the matrix transpose.
[0076] The suit characterization polymerization process described in step (3) is specifically as follows:
[0077] After obtaining the factorized representation of the products in the set, in order to learn the combination relationship between the features of the products in the set, the multi-head attention mechanism is used to combine the product features, and the stacked self-attention function is used to learn the high-order combination relationship between the products. It can be expressed as follows:
[0078]
[0079] where σ(·) represents the feed-forward layer in row dimension, They represent self-attention, normalization and multi-head self-attention operations respectively, and H represents the set features after multi-head self-attention and residual normalization.
[0080] Finally, the high-level features of the suit are weighted and summed through the attention mechanism to obtain the overall expression of the suit. It can be expressed as follows:
[0081]
[0082] in, represents the learnable weight parameter, and h represents the set feature weighted by the weight parameter s.
[0083] The user representation aggregation process described in step (4) is specifically as follows:
[0084] 4-1. The set of user-package pairs can be written as Is user u i The associated set of packages, i.e. user u i The positive sample set in, For user u i The number of associated sets.
[0085] Since the user-suite pair does not have the timestamp information of the user’s interaction with the suit, we i Related Sets Select N b The user's historical behavior sequence is constructed by using the package, denoted as In order to better express the user's behavioral intention, this method uses the minimum variance as the basis to filter out accidental or noisy user behavior data. A subset of sequences is selected as the user behavior representation.
[0086] 4-2. User features are embedded in user basic information and user behavior sequence For a set The same method as the package end is used, that is, after the feature is factored, the operation of formula 2 and formula 3 is used to aggregate the features. Therefore, the user feature It can be expressed as follows:
[0087]
[0088] Where η(·) represents the operation of formula (2) and formula (3), Represents an average pooling operation.
[0089] The correlation measurement process described in step (5) is specifically as follows:
[0090] The matching score between the user and the package is measured from two aspects: the correlation between the user characteristics and the package characteristics and the compatibility of the products in the package. Personalized score Compatibility score is calculated by the inner product of user features and package features. It is measured by the average inner product of the product pairs in the set. Since the score weights of the two aspects are different, a learnable weight vector is used to adaptively learn the weights. It can be expressed as follows:
[0091]
[0092] Among them, M (p) and M (c) The learnable weight vectors representing the personalization score and the compatibility score, respectively, and They represent the factorized features of the pth and qth items in the jth package, The number of combinations of choosing 2 items from n items in a set.
[0093] The construction of the deep learning model described in step (6) is specifically as follows:
[0094] The personalized package recommendation network model first extracts the visual features of the products through a convolutional neural network, obtains the basic user representation with the user embedding vector, and then projects the product features and user features into multiple implicit spaces through the feature factorization module. Then, the feature aggregation module performs high-order combination and weighted pooling of the package features to obtain the package representation, constructs the user historical behavior to learn the user historical representation, concatenates the user basic representation and the user historical representation to obtain the user representation, and finally calculates the compatibility score by the pairwise inner product of the products in the package, and obtains the correlation score by the row-dimensional inner product of the package representation and the user representation, and the sum of the two obtains the total matching score.
[0095] The personalized package recommendation network model is trained in the form of triples. i , the positive sample set is Randomly select a set of sets that are not positive samples from all sets O The set of suits is used as the negative sample set of the user So we get a training set consisting of a set of triple data It can be expressed as:
[0096]
[0097] Among them, the triple (u i ,p,q) means that package p is more popular with user u than package q i welcome.
[0098] The model aims to optimize the Bayesian ranking loss and the logarithmic determinant divergence loss. Therefore, the overall loss function of the model can be defined as follows:
[0099]
[0100] Among them, λ1 and λ2 are weight parameters, θ u and θ o are the network parameters that characterize users and sets, and ||·||1 represents the L1 norm.
[0101] During the training process, the parameters of each part of the network model are updated simultaneously, the final model is obtained and saved, and the training effect is tested on the test set. Example
[0102] In the first step, the image data is preprocessed as follows:
[0103] We use two datasets, Polyvore-630 and Polyvore-519, to verify our personalized outfit recommendation model. There are 630 users in the Polyvore-630 dataset, and the outfit contains a top, a bottom, and a pair of shoes; there are 519 users in the Polyvore-519 dataset, and the outfit contains two tops, a bottom, and a pair of shoes. Since the second top is missing, the first top is used to supplement the missing top. The image size in the dataset is 291*291. When training the model, these images are randomly flipped horizontally with a probability of 50%, and finally the image pixel values are normalized. When testing the model, the data processing process is similar to that during training, but no random horizontal flipping is required.
[0104] In the second step, the user's historical behavior sequence is constructed as follows:
[0105] The following is an example of the dataset Polyvore-630. We sample and construct the user's historical behavior sequence based on the visual features of the set. First, we use product images to train a classification network. This method uses the Alexnet network structure. The training set is all the images in the data set, which has three categories: tops, bottoms, and shoes. The trained network has an accuracy of more than 95% on the test set. Then use the network to extract the 2048-dimensional visual features of the product, and take the average value of the visual features of the products in the set as the feature of the set, and then calculate the variance of the set feature. We select the 10 items with the smallest variance as the historical behavior sequence of each user. We conducted an experiment to compare the effects of random selection (Random), selection with minimum variance (Minimum), and selection with maximum variance (Maximum) on the results. The results are shown in Table 1. It can be seen that the best recommendation results can be obtained when the minimum variance is selected.
[0106] Table 1 Ablation experiments on different ways of constructing historical user sequences
[0107]
[0108] Among them, Fashion Hashing Network (FHN) is the baseline model method, and FHN-B means adding user historical behavior representation to the baseline model FHN.
[0109] In the third step, the features are factorized and aggregated as follows:
[0110] The personalized package recommendation network model first uses the AlexNet structure to extract the 4096-dimensional visual features of the product, and then combines the 4096-dimensional features through two layers of full connection to reduce the dimension, obtaining 128-dimensional features to express the visual features of the product. For users, we represent the user ID as 128 dimensions through the embedding layer. Through experiments with different factorization numbers k, we found that the best result is when the number of factors k is 4. Apply the transformation matrix and Get user information embedding representation For the factorization of the package side, we can first perform a high-order combination of the product features to obtain the package features and then factorize them; we can also first factorize the product features of the package and then perform a high-order combination of the product features in each implicit space to obtain the factorized package features; at the same time, for the high-order combination of product features in each implicit space, we can share or not share the parameters. We experimented with these combination structures. In the end, we found that the best structure is to perform a high-order combination of the product factors and share the parameters in each implicit space.
[0111] Step 4: Model test results and comparison are as follows:
[0112] We compared the test results of our method with other models on the Polyvore_630 and Polyvore_519 datasets, and the results are shown in Table 3. We can see that enriching user representation with user historical behavior and using factorization to calculate the relevance between users and packages in multiple implicit spaces can improve the model recommendation results, and compared with other models, our method achieves the best results.
[0113] Table 2 Comparison of test results of different models
[0114]
[0115] Among them, SiameseNet (Siamese CNNs) represents a model structure that uses twin convolutional neural networks to extract product features and user features and then performs inner product pairwise inner product to obtain a matching score; Bi-LSTM (Bidirectional LSTM) represents a model structure that uses a bidirectional long short-term memory network to learn a package representation by treating the package as a sequence, and then performs inner product with the user representation to obtain a matching score; CSN (Conditional Similarity Networks) represents a model structure that maps the products in a package in pairs to a category-specific embedding space to learn product representations, and then performs inner product with the user representation to obtain a matching score; PFN (Personalized Fashion Network) represents a model structure that uses high-order combinations of product visual features to obtain package representations, while users are only represented by basic information embedding, and measures the matching scores between users and packages from two aspects: the compatibility of products in the package and the relevance of the package to the user; PFN-B is a model structure that adds user history representations to the PFN model structure; PFN-F is a model structure that uses factorization on the basis of the PFN model structure; PFN-BF simultaneously adds user history representations and applies factorization.
Claims
1. A personalized package recommendation method based on user history and feature factorization, characterized in that: The steps include: Step (1): image data preprocessing; Perform data augmentation on the images in the fashion outfit recommendation dataset; Step (2): Extract product visual features and obtain basic user representation through user embedding vectors; project product features and user features into multiple implicit spaces through factorization, learn users’ interests in different aspects of the package, and make each implicit space different through distribution constraints; Step (3): Aggregate product features through high-order feature combination and weighted pooling to obtain package representation; Step (4): Sampling the user's historical behavior from the set of sets that the user has interacted with to learn the user's historical representation, and then aggregating the user embedding vector and the user's historical representation to obtain an improved user representation; Step (5): Measure the matching score between the user and the package from two aspects: the compatibility of the products in the package and the relevance of the package to the user; Step (6): Build an end-to-end deep learning model according to the above steps, train the entire network on the training set, and test the performance of the final model on the test set; The specific method of step (2) is as follows: 2-1. For suits, use the basic backbone network to extract the visual features of fashion items in the suit The combination of product attributes is regarded as an implicit factor. In order to learn the implicit k factors, the feature transformation matrix Project the product features into multiple different implicit spaces so that each space can implicitly represent a certain attribute of the product. The product is factorized and recorded as The factorized set is represented as Where n represents the dimension of the number of fashion items in the suit, k represents the dimension of the number of factors, and d represents the dimension of the visual features of the fashion items; 2-2. For users, feature embedding is performed based on user data information to obtain user embedding vectors Through the feature transformation matrix Project the user features into multiple different implicit spaces and factorize the user features into 2-3. In order to ensure the difference of the implicit space, the logarithmic determinant divergence loss is used to make the transformation matrix sparse.
2. The personalized package recommendation method based on user history and feature factorization according to claim 1, characterized in that: Step (1) is as follows: An existing fashion suit recommendation dataset is obtained, which consists of a training set and a test set. For the image data in the training set, the images are horizontally flipped with a probability of 50%. Finally, the images in the training set and the test set are normalized to obtain the image data after data enhancement.
3. The personalized package recommendation method based on user history and feature factorization according to claim 2, characterized in that: The specific method of step (3) is as follows: After obtaining the factorized representation of the products in the set, in order to learn the combination relationship between the products in the set, the multi-head attention mechanism is used to combine the product features, and the stacked self-attention function is used to learn the high-order combination relationship between the products. Can be expressed as follows: where σ(·) represents the feed-forward layer in row dimension, They represent self-attention, normalization and multi-head self-attention operations respectively, and H represents the set features after multi-head self-attention and residual normalization; Finally, the high-level features of the suit are weighted and summed through the attention mechanism to obtain the overall expression of the suit.
4. The personalized package recommendation method based on user history and feature factorization according to claim 3, characterized in that: The specific method of step (4) is as follows: 4-1. The set of user-package pairs is denoted by Is user u i The associated set of packages, i.e. user u i The positive sample set in, For user u i The number of sets involved; Since there is no time information of the user-suite interaction in the user-suite pair, i Related Sets Select N from the positive sample subset b The packages are used as the historical behavior expression of the user, denoted as In order to better express the user's behavioral intention, the sequence is used to filter out accidental or noisy user behavior data as much as possible, and the minimum variance is used as the basis for the sequence. Select a subset from the dataset as the user behavior representation; 4-2. User characteristics are embedded in their own information and user behavior sequence Aggregate representation; for sets The factorization and aggregation methods of the package features in steps (1) and (2) are used to perform the same operation on each package in the user's history sequence, and then average pooling is performed to obtain the user history representation Therefore, user characteristics Can be expressed as follows: Among them, η(·) represents the operation of high-order feature combination and weighted pooling, Represents an average pooling operation.
5. The personalized package recommendation method based on user history and feature factorization according to claim 4, characterized in that: The specific method of step (5) is as follows: The matching score between the user and the package is measured from two aspects: the correlation between the user characteristics and the package characteristics and the compatibility of the products in the package. Personalized score Compatibility score is calculated by the inner product of user features and package features. It is measured by the average inner product of the product pairs in the set. Since the score weights of the two aspects are different, a learnable weight vector is used to adaptively learn the weights. It can be expressed as follows: Among them, M (p) and M (c) The learnable weight vectors representing the personalization score and the compatibility score, respectively, and They represent the factorized features of the pth and qth items in the jth package, The number of combinations of choosing 2 items from n items in a set.
6. The personalized package recommendation method based on user history and feature factorization according to claim 5, characterized in that: The specific method of step (6) is as follows: The end-to-end deep learning model, namely the personalized package recommendation network model, first extracts product visual features through a convolutional neural network, obtains the basic user representation with the user embedding vector, and then projects the product features and user features into multiple implicit spaces through the feature factorization module. Then, the package features are combined and weighted pooled through the feature aggregation module to obtain the package representation, construct the user history behavior to learn the user history representation, and concatenate the user basic representation and the user history representation to obtain the user representation. Finally, the compatibility score is calculated by the inner product of the pairs of products in the package, and the correlation score is obtained by the inner product of the row dimension of the package representation and the user representation. The sum of the two is the total matching score. The personalized package recommendation network model is trained in the form of triples; for user u i , the positive sample set is Randomly select a set of sets that are not positive samples from all sets O The set of suits is used as the negative sample set of the user So we get a training set consisting of a set of triple data Can be expressed as: Among them, the triple (u i ,p,q) means that package p is more popular with user u than package q i welcome; The model aims to optimize the Bayesian ranking loss and the logarithmic determinant divergence loss. Therefore, the overall loss function of the model is defined as follows: Among them, λ1 and λ2 are weight parameters, θ u and θ o are the network parameters that characterize users and sets, and ||·||1 represents the L1 norm; During the training process, the parameters of each part of the network model are updated simultaneously, the final model is obtained and saved, and the training effect is tested on the test set.
7. The personalized package recommendation method based on user history and feature factorization according to claim 1, characterized in that: The basic backbone network adopts AlexNet, VGG or ResNet to extract the visual features of the product.
8. The personalized package recommendation method based on user history and feature factorization according to claim 1, characterized in that: In order to ensure the difference of implicit space, the logarithmic determinant divergence loss is used to make the transformation matrix sparse, and the constraints on user space and the constraints of the suit space It is expressed as follows: Where c is a constant, tr(·) is the rank of the matrix, logdet(·) is the logarithm of the matrix determinant, and T is the matrix transpose.
Citation Information
Patent Citations
Personalized commodity recommendation method and personalized commodity recommendation system
CN108629665A
User personalized clothes matching method and device
CN110175253A
