Product recommendation method and device, medium and program product

Abstract

The invention provides a product recommendation method and device, a medium and a program product. The product recommendation method disclosed by the invention comprises the following steps: acquiring a user portrait feature, and a first category feature and a second category feature of a product; determining a first recommendation weight of the product based on the user portrait feature and the first category feature of the product; determining an indirect score, a direct score and a product reference score of the user on the product based on the user portrait features and the first category features of the product; and respectively carrying out weighted summation on the indirect score, the direct score, the product reference score and the second category feature by utilizing the first recommendation weight to obtain a product recommendation score, so as to recommend or not recommend a product to the user based on the product recommendation score.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and more particularly to a product recommendation method, device, medium, and program product. Background Art

[0002] With the development of network technology, there are more and more channels for product promotion. For example, product information can be sent to users in a targeted manner or through professional application software.

[0003] However, some products have numerous attributes, and users' considerations are complex and diverse. In this context, accurately matching users with products is challenging. For example, in the real estate market, new home market information is complex, encompassing multiple dimensions such as location, unit type, price, developer, and surrounding amenities. This makes it difficult for users to quickly and effectively select properties that meet their needs. Existing product recommendation systems rely on a single algorithm. Once a user profile is determined, the recommended product types are highly similar, failing to meet users' dynamic and diverse selection needs. Summary of the Invention

[0004] The present disclosure provides a product recommendation method, device, medium, and program product.

[0005] According to a first aspect of the present disclosure, a product recommendation method is provided. The method specifically includes: obtaining user profile features, first category features, and second category features; determining a first recommendation weight for a product based on the user profile features and the first category features; determining an indirect rating, a direct rating, and a product reference rating for the product by the user based on the user profile features and the first category features; and obtaining a product recommendation score by weightedly summing the indirect rating, direct rating, product reference rating, and the second category features using the first recommendation weight, thereby recommending or not recommending the product to the user based on the product recommendation score.

[0006] Based on the above, when a user actively searches for a product or the server proactively recommends a product to a user, the server must find the user's most likely favorite product based on their needs and recommend it to them. When making recommendations, the product recommendation score is calculated based on multiple dimensions. These dimensions include indirect scores, direct scores, product reference scores, and a comprehensive calculation of second-category features. To achieve more accurate calculation results—that is, to make the product recommendation score more relevant to current user needs—each of these dimensions can be assigned a corresponding recommendation sub-weight (derived from the first recommendation weight). Because the first recommendation weight is calculated based on user profile features and first-category features, changes in the user profile features and / or first-category features will also affect the corresponding first recommendation weight and sub-weights, dynamically adjusting the product recommendation score. Based on this approach, the first recommendation weight is dynamically adjusted in real time based on user profile features and first-category features, resulting in a more accurate product recommendation score that meets the user's latest needs.

[0007] According to at least one embodiment of the present disclosure, a first recommendation weight is calculated based on user portrait features and first category features, including: splicing the user portrait features with the first category features to obtain a spliced feature vector; and performing a linear transformation on the spliced feature vector based on a multi-layer perceptron to obtain the first recommendation weight in vector form.

[0008] According to at least one embodiment of the present disclosure, a linear transformation is performed on the concatenated feature vector based on a multi-layer perceptron to obtain a first recommendation weight in vector form, including: inputting the concatenated feature vector into multiple hidden layers of the multi-layer perceptron and outputting a feature representation; inputting the feature representation into an output layer with a specified dimension in the multi-layer perceptron to obtain a first recommendation result including multiple recommendation sub-weights; wherein the output layer includes the same number of neurons as the recommendation sub-weights.

[0009] According to at least one embodiment of the present disclosure, an indirect score, a direct score, and a product reference score are calculated based on user portrait features and first category features, including: searching for similar users based on user portrait features, and determining similar user portrait features; and using a collaborative filtering algorithm, user portrait features, and similar user portrait features to calculate an indirect score of the user for the product; using a deep learning model, user portrait features, and first category features to calculate a direct score of the user for the product; searching for similar products based on the first category features, and determining the third category features of similar products; and using a similarity algorithm, user portrait features, and first category features to calculate a product reference score.

[0010] According to at least one embodiment of the present disclosure, a collaborative filtering algorithm is used, as well as user portrait features and similar user portrait features to calculate an indirect rating of a user for a product, including: calculating user similarity based on the user portrait features of the target user and the user portrait features of similar users; wherein the user portrait features include: at least one of the user's browsing time for the product, collection status, and consultation content; based on the comparison result of the user similarities, selecting a similar user with the greatest user similarity as at least one neighbor user; and obtaining an indirect rating by comprehensively calculating the average rating of the target product by neighbor users, the product of the user similarity and the average rating of the target product, and the ratio of the average rating of all products by neighbor users to the user similarity.

[0011] According to at least one embodiment of the present disclosure, the first recommendation weight is split and respectively integrated with the indirect score, direct score and product reference score, and then combined with the second category feature to calculate the product recommendation score to generate a product recommendation result based on the product recommendation score, including: integrating the first recommendation sub-weight obtained by splitting based on the first recommendation weight with the indirect score to obtain a first score result; integrating the second recommendation sub-weight obtained by splitting based on the first recommendation weight with the direct score to obtain a second score result; integrating the third recommendation sub-weight obtained by splitting based on the first recommendation weight with the product reference score to obtain a third score result; integrating the fourth recommendation sub-weight obtained by splitting based on the first recommendation weight with the second category feature to obtain a fourth score result; and calculating the product recommendation score using the first score result, the second score result, the third score result, and the fourth score result.

[0012] According to at least one embodiment of the present disclosure, after calculating the product recommendation score and generating the product recommendation result according to the product recommendation score, it also includes: sorting the recommended products according to the size of the product recommendation score to obtain a first recommendation result; in response to an update request of the user portrait feature and / or the first category feature, calculating a second recommendation weight based on the updated subsequent user portrait feature and / or the first category feature; judging the subordinate relationship between the previous user portrait feature and / or the first category feature and the updated subsequent user portrait feature and / or the first category feature; if the subsequent user portrait feature and / or the first category feature is subordinate to the previous user portrait feature and / or the first category feature, then calculating the updated product recommendation score and the second recommendation result based on the first recommendation result and the second recommendation weight; if the subsequent user portrait feature and / or the first category feature is in a parallel relationship with the previous user portrait feature and / or the first category feature, or the previous user portrait feature and / or the first category feature is subordinate to the subsequent user portrait feature and / or the first category feature, then calculating the updated product recommendation score based on the second recommendation weight, and recommending or not recommending the product to the user based on the updated product recommendation score.

[0013] According to a second aspect of the present disclosure, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, so that the processor executes the method described in the first aspect of any embodiment of the present disclosure.

[0014] According to a third aspect of the present disclosure, a readable storage medium is provided, in which execution instructions are stored. When the execution instructions are executed by a processor, they are used to implement the method described in the first aspect of any embodiment of the present disclosure.

[0015] According to a fourth aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the method according to the first aspect of any embodiment of the present disclosure is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings illustrate exemplary embodiments of the present disclosure and together with the description serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0017] Figure 1 A flowchart of a product recommendation method provided by the present disclosure.

[0018] Figure 2 This is a schematic block diagram of the structure of a product recommendation device according to one embodiment of the present disclosure.

[0019] Figure 3 A schematic diagram of the product recommendation process for illustrating the present disclosure.

[0020] Figure 4 The present invention is a block diagram showing the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present disclosure is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are intended only to illustrate the relevant content and are not intended to limit the present disclosure. It should also be noted that, for ease of description, only the portions relevant to the present disclosure are shown in the accompanying drawings.

[0022] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present disclosure can be combined with each other. The technical solution of the present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0023] With the development of network technology, there are more and more ways to promote products. Furthermore, the product categories are diverse and complex, and user needs are also dynamic and changing. For example, in the promotion of new homes, the new home market information is complex, including multiple dimensions such as location, unit type, price, developer, and surrounding facilities. This makes it difficult for users to quickly and effectively select properties that meet their needs. Existing new home recommendation systems often use a single recommendation algorithm. These methods fail to fully consider the dynamic behavior of users during the new home purchase process and the diversity of new home characteristics, resulting in inaccurate recommendations and a poor user experience.

[0024] Furthermore, user behavior and product preferences are dynamic. For example, when initially selecting a new home, a user may be primarily interested in properties in a specific location. However, after conducting a site visit, they may become more concerned with price and choose a wider range of locations, rather than being limited to a single location. In this case, existing single-source recommendation algorithms are unable to accurately recommend the products desired by users. Therefore, a solution that accurately recommends products to users is urgently needed.

[0025] Figure 1 This is a flow chart of a product recommendation method provided by the present disclosure. Figure 1 The method shown includes steps 101 to 104. The method can be executed by an electronic device such as a server (a local server or a cloud server).

[0026] Specifically, Figure 1 The method shown includes: Step 101: Obtain user portrait features, first category features, and second category features.

[0027] The user portrait features mentioned here can include basic features of the user, such as the user's age, occupation, name, etc., and can also include features associated with the user and the product, such as the number of times the user browses product details, the type of product the user browses, the product the user collects, the product the user consults, the keywords the user searches for, and the user's stay time. It should be noted that user portrait features are recorded in time series, and a half-life of the decay function is set for each feature in order to highlight recent user portrait features, so that recent user portrait features play a more important role in subsequent scoring work. It can be seen that the basic features of users are usually fixed, but some features associated with products can change dynamically. For example, the types of products browsed by users, the types and quantities of products collected by users, etc. will change over time.

[0028] The first category of features mentioned here may include the basic features of the product. For example, taking a new house as an example, the corresponding basic features may include: the name of the property, geographical location, apartment type, area, property rights, supporting schools, supporting transportation, medical resources; as well as the developer, property management company, floor area ratio, greening rate, number of buildings, number of households, number of parking spaces, etc.

[0029] The second-category features mentioned here may include the property's sales stage (including the customer accumulation period, launch period, and sales period), promotional activities, and pre-sale status. These second-category features change over time. These changes can directly or indirectly alter user interest in a product. For example, increased discounts on new homes may attract users to the property and lead them to place an order. Since second-category features are not persistent or stable, they cannot be used as a factor in calculating the primary recommendation weight.

[0030] Step 102: Determine a first recommendation weight of the product based on the user portrait features and the first category features.

[0031] When calculating the first recommendation weight, the user portrait features and the first category features are comprehensively considered. When the user portrait features and / or the first category features change, the first recommendation weight can be adjusted in a timely manner, thereby affecting the recommendation results. For example, a user previously browsed properties in District C1, City B1, Province A, and recently started browsing properties in District C2, City B1, Province A. The user's browsing history has changed, resulting in changes in the user portrait features. The first recommendation weight will be recalculated. Similarly, if the first category features change, for example, the remaining apartment types are only small ones, which cannot meet the user's needs, the first recommendation weight also needs to be recalculated.

[0032] It should be noted here that when dynamically calculating the first recommendation weight, if the changed user portrait features and / or first category features have continuity, that is, the features are further refined based on the earlier version of the user portrait features and / or first category features (for example, the earlier features only mentioned the real estate project, and the latest features clarified the apartment types of the real estate project), then when calculating the first recommendation weight, it is necessary to inherit the previous recommendation results to further optimize the recommendation results.

[0033] If the user profile features and / or first category features that have changed during the dynamic calculation of the first recommendation weight are not continuous, that is, the user profile features and / or first category features of the earlier version are completely different from the user profile features and / or first category features of the latest version. For example, if the user's focus city switches from City A to City B, and there is a significant change in the geographical area, then the earlier first recommendation weight and recommendation results will not be inherited. The first recommendation weight and corresponding recommendation results need to be recalculated.

[0034] Step 103: Determine the user's indirect rating, direct rating, and product reference rating for the product based on the user portrait features and the first category features.

[0035] In practice, to more accurately recommend products to users, we must consider not only the user themselves but also other users. Because a single user can only provide limited information about their user profile and characteristics, referencing other users can provide a more accurate understanding of their evaluations and preferences for a particular product. Simply put, a good product (for example, a high-quality, cost-effective new home in a high-quality school district) will be liked by many similar users, while a poor product will often be disliked by many users. This means leveraging user profile characteristics and other similar user characteristics to calculate indirect ratings across multiple users, resulting in a more comprehensive, objective, and accurate rating.

[0036] The direct ratings mentioned here refer to the user's preference for a specific product. When calculating user preferences for products, we use a combination of user profile features and first-category features for feature extraction, and then use algorithms such as deep learning to calculate the user's preference for a specific product.

[0037] In actual applications, when recommending products to users, multiple related products are often recommended to improve the success rate of recommendations. For example, if the user is currently focusing on property i, then other adjacent properties j can be found within the same area, and the similarity between properties i and j can be calculated. When the calculated product reference score exceeds the similarity threshold, properties i and j are considered to be relatively similar. When recommending property i, alternative property j can also be recommended to the user. Furthermore, when calculating the similarity using the first-category features of two products, the first-category features can also be used to calculate the differences and find the advantages of each product. When recommending products, the advantages of each product can be displayed, making it easier for users to intuitively understand and compare multiple similar products.

[0038] Step 104: Use the first recommendation weight to weight and sum the indirect score, direct score, product reference score, and second category features to obtain a product recommendation score, so as to recommend or not recommend the product to the user based on the product recommendation score.

[0039] After calculating using the aforementioned methods, we can obtain indirect scores, direct scores, and product reference scores, which respectively represent the general user preference for the product, the preference of specific users for a specific product, and the recommendation of other products related to the specific product. Furthermore, because products sometimes have personalized features, such as when a product is on sale, these second-category features can be taken into account as a factor in calculating the product recommendation score.

[0040] To ensure that the final product recommendations are more in line with user expectations, when calculating the product recommendation score, corresponding recommendation sub-weights are assigned to the previously calculated indirect score, direct score, product reference score, and the acquired second-category features. As mentioned above, the first recommendation weight can be dynamically adjusted based on the user profile features and the first-category features. This means that the recommendation sub-weights derived from the first recommendation weight are also dynamically adjusted based on the user profile features and the first-category features.

[0041] Based on the above disclosure, it can be seen that when a user actively searches for a product or a server actively recommends a product to a user, the server must find the user's most likely favorite product based on the user's needs and recommend it to the user. When making recommendations, the product recommendation score is calculated based on multiple dimensions. These dimensions include indirect scores, direct scores, product reference scores, and a comprehensive calculation of second-category features. To achieve more accurate calculation results, that is, to make the product recommendation score more consistent with current user needs, corresponding recommendation sub-weights (derived from the first recommendation weight) are assigned to these dimensions. Since the first recommendation weight is calculated based on user profile features and first-category features, when the user profile features and / or first-category features change, the corresponding first recommendation weight and recommendation sub-weights will also change accordingly, thereby dynamically adjusting the product recommendation score. Based on this approach, the first recommendation weight is dynamically adjusted in real time based on the user profile features and first-category features, making the final calculated product recommendation score more accurate and in line with the user's latest needs.

[0042] In one or more embodiments of the present disclosure, a first recommendation weight is calculated based on user portrait features and first category features, including: splicing the user portrait features with the first category features to obtain a spliced feature vector; performing a linear transformation on the spliced feature vector based on a multi-layer perceptron to obtain the first recommendation weight in vector form.

[0043] In practical applications, before concatenating user profile features with first-category features, the feature matrix needs to be transformed. The user profile feature matrix typically contains multiple types of user information, such as age, gender, geographic location, and spending habits. The data types and value ranges of these features may vary. In order to process these features within a unified framework, normalization is often required. For example, the age feature may range from 1 to 100 years old, while the spending amount feature may range from 0 to tens of thousands of yuan. Through normalization, the values of these features can be mapped to a fixed interval, such as [0, 1] or [-1, 1]. This eliminates the impact of different dimensions between different features.

[0044] If user profile features or first-category features contain categorical features, such as gender (male or female) or location (different cities), one-hot encoding is often used to convert them. For example, gender, originally a simple categorical label, becomes a two-dimensional vector after one-hot encoding. For example, male can be represented as [1,0] and female as [0,1]. This converts non-numeric features into numerical ones, making subsequent mathematical operations possible.

[0045] After normalizing and encoding these features, feature concatenation is performed. Specifically, the processed user profile feature matrix and the first-category feature matrix are concatenated in a specific order. For example, assuming the user profile feature matrix is Hu_t and the first-category feature matrix is Fi, they can be concatenated column-wise to obtain a new concatenated feature vector X, i.e., X = [Hu_t, Fi]. This creates a concatenated feature vector that integrates various user and product features, providing the data foundation for subsequent recommendation calculations.

[0046] Next, the concatenated feature vectors are fed into a multilayer perceptron (MLP) for processing to calculate the output. The MLP is a feedforward artificial neural network model consisting of an input layer, multiple hidden layers, and an output layer. The following example uses an MLP with one input layer, two hidden layers, and one output layer to detail the structure and function of each layer.

[0047] The input layer is the starting layer of the multilayer perceptron. The number of neurons in it corresponds to the dimension of the concatenated feature vector. For example, if the concatenated feature vector has n elements, then the input layer has n neurons. These neurons are responsible for receiving the input concatenated feature vector and passing its information to the next hidden layer.

[0048] The first hidden layer usually performs linear transformation and feature extraction on the input information. In this layer, each neuron receives input from all neurons in the input layer and generates its own output through weighted summation and activation function processing.

[0049] For example, nonlinear transformation: z1=W1*X+b1, activation function: ReLU(z1)=max(0,z1).

[0050] Here, X is the eigenvector of the input layer (also known as the concatenated eigenvector), W1 is the weight matrix between the input layer and the first hidden layer, b1 is the bias vector, and ReLU is the activation function. The activation function introduces nonlinearity into the neural network, effectively addressing the vanishing gradient problem and enabling the neural network to better train and learn complex nonlinear relationships.

[0051] The second hidden layer further extracts features and integrates information from the output of the first hidden layer. Its calculation process is similar to that of the first hidden layer: Non-linear transformation: z2=W2*ReLU(z1)+b2, activation function: ReLU(z2)=max(0,z2).

[0052] Among them, ReLU(z1) is the output of the first hidden layer, W2 is the weight matrix between the first hidden layer and the second hidden layer, and b2 is the bias vector.

[0053] After processing through the two hidden layers, the information in the concatenated feature vector has been fully extracted and transformed, resulting in a more representative feature representation. Finally, these features are passed to the output layer, which processes them according to the specific task requirements to obtain the first recommendation weight in vector form.

[0054] At the output layer, a linear combination is typically used to calculate the output. The calculation formula is as follows: Linear transformation: z = W * ReLU(z²) + b. Here, the bias vector is b, the weight matrix is W, and the output of the second hidden layer is ReLU(z²). This calculation yields the first recommendation weight in vector form.

[0055] To sum up, by effectively processing and splicing the user portrait features and the first category features, and using the multi-layer perceptron for linear transformation and nonlinear feature extraction, we finally obtained the first recommendation weight vector that can accurately reflect the user preferences. Moreover, this first recommendation weight is determined by the user portrait features and the first category features. Therefore, when the user portrait features and / or the first category features change dynamically over time, the first recommendation weight will also change dynamically.

[0056] In one or more embodiments of the present disclosure, the linear transformation of the concatenated feature vector based on the multi-layer perceptron to obtain the first recommendation weight in vector form includes: inputting the concatenated feature vector into multiple hidden layers of the multi-layer perceptron to output feature representation; inputting the feature representation into an output layer with a specified dimension in the multi-layer perceptron to obtain a first recommendation result including multiple recommendation sub-weights; wherein the output layer includes multiple neurons.

[0057] As mentioned above, after splitting the first recommendation weight, four sub-weights are needed. In other words, four neurons in the output layer are used to output the corresponding sub-weights. These four sub-weights strictly correspond to the indirect rating, direct rating, product reference rating, and second-category features, giving each sub-weight a specific function.

[0058] For example, if the user profile feature Hu_t indicates frequent recent inquiries, after updating the user profile feature, the recommendation weight z_β (the second recommendation sub-weight in deep learning) re-output by the multi-layer perceptron (MLP) can be increased, as the deep learning model DL(u,i) excels at capturing temporal behavior. If the first-category feature Fi emphasizes property location information, z_γ (the third recommendation sub-weight for product similarity) may be increased. The multi-layer perceptron (MLP) uses training sample data (such as user profile feature training samples and first-category feature training samples) to learn the mapping between features and the four recommendation sub-weights, optimizing recommendation accuracy.

[0059] In one alternative, the calculated first recommendation weight is a matrix vector. Assuming this matrix vector is a one-dimensional matrix, it can be directly split into four parts. For example, Z = [2.0, 1.5, 0.5, 1.0] is converted via softmax to α ≈ 0.455, β ≈ 0.276, γ ≈ 0.102, and δ ≈ 0.167, which determine the linear weighted proportions of each recommendation function in R(u,i). The maximum z_α indicates the highest contribution of CF(u,i), which is dynamically adjusted as Hu_t changes. It should be noted that the Softmax function used here converts a real number vector Z into a probability distribution, such that the probability value of each category lies between 0 and 1, and the sum of all category probabilities is 1. This means that the output of the softmax function can be directly interpreted as the probability of the classification. The converted probabilities are assigned as recommendation sub-weights to the corresponding scores and similarities.

[0060] The MLP mentioned here is pre-trained, and the training method is as follows: You need to prepare a dataset containing product recommendation scores and four parameters (indirect score, direct score, product reference score, and second category features). The dataset should contain multiple samples, each with a label (for example, product recommendation score) and four feature values.

[0061] A multilayer perceptron (MLP) is trained using the prepared dataset. During training, the MLP learns how to adjust its weights (including those in the output layer) to minimize the loss function. This typically involves optimization techniques such as backpropagation and gradient descent. The MLP also learns the mapping between features and recommender weights during training.

[0062] In one or more embodiments of the present disclosure, indirect scores, direct scores, and product reference scores are calculated based on user portrait features and first category features, including: searching for similar users based on the user portrait features and determining similar user portrait features; using a collaborative filtering algorithm, as well as user portrait features and first category features to calculate an indirect score of the user for the product; using a deep learning model, as well as user portrait features and first category features to calculate a direct score of the user for the product; searching for similar products based on the first category features and determining the third category features of similar products; and using a similarity algorithm, as well as user portrait features and third category features of similar products to calculate a product reference score.

[0063] The following will describe various scoring and similarity calculation processes through specific examples.

[0064] Direct ratings are a quantitative representation of a specific user's preference for a specific product, and can be calculated using a deep learning model. It's important to note that the deep learning model here is also pre-trained, using a training set consisting of user profile feature samples, first-category feature samples, and direct ratings.

[0065] Taking a specific scenario as an example, the embedding layer in a deep learning model is used to perform dimensionality reduction on user profile features (e.g., user ID, age, gender, historical behavior, and other feature encoding information) and first-category features (e.g., product attributes related to new homes, real estate projects, or vehicles, such as area, floor plan, location, vehicle model, and configuration). Embedding technology maps high-dimensional sparse features into a low-dimensional dense vector space, resulting in reduced-dimensional features for the corresponding users and products.

[0066] Subsequently, Long Short-Term Memory (LSTM) networks are used to extract temporal features from the user's reduced-dimensional vectors. LSTM, with its unique gating mechanism, effectively processes long-term dependencies in sequential data, thereby accurately capturing temporal information that reflects the temporal characteristics of user behavior.

[0067] The extracted user behavior temporal features are then concatenated with the product dimensionality reduction features. The concatenated feature vectors are fed into multiple fully connected layers. Through the fully connected connections between neurons in these layers, nonlinear transformations are performed on the features, mining deep-level feature combinations and relationships between features.

[0068] The nonlinear transformation output from the fully connected layer is fed into the Sigmoid activation function. Based on its mathematical formula, the Sigmoid function constrains the input value to the range (0, 1). Ultimately, it generates a predicted preference score for the user for a product (e.g., a property), taking values between 0 and 1, representing the user's preference.

[0069] To optimize the model, the binary cross entropy loss function is used to measure the difference between the predicted results and the true labels. The model parameters are updated through the back propagation algorithm to improve the model prediction accuracy.

[0070] A product reference score quantifies the degree of similarity between other products and the target product of the current user's interest. In practical applications, its calculation often relies on a specific similarity algorithm. Cosine similarity is one common algorithm, and the minimum edit distance can also be used to calculate the similarity between two products.

[0071] Cosine similarity is based on the mathematical properties of vectors and measures similarity by calculating the cosine of the angle between two vectors. The value range is -1 to 1. Values closer to 1 indicate higher reference ratings for the two products; values closer to -1 indicate lower similarity; and a value of 0 indicates no similarity between the two products.

[0072] Taking real estate analysis as an example, cosine similarity is used to calculate the attribute similarity between properties i and j. The formula is: S(i,j)=(F_i*F_j) / (||F_i||*||F_j||). F_i and F_j are the feature vectors of properties i and j, respectively. * represents the vector dot product, which multiplies the corresponding elements and then sums them; |||| represents the vector modulus, which is the square root of the sum of the squares of the vector elements. It should be noted that when searching for property j to calculate similarity with property i, you can first define an area. For example, the properties within an area with a radius of 2 kilometers, centered on property i, are used as the properties for the similarity calculation (in practice, there can be multiple properties j).

[0073] The feature vector F encompasses several key attributes of a property, including price (reflecting its market positioning and economic cost), area (indicating the size of the living space), unit type (relating to the layout and functionality of the living space), and surrounding amenities (representing the property's geographical advantages and living convenience). Using these attributes, the feature vector constructed using the cosine similarity formula can quantitatively assess the similarity between different properties, providing a valuable reference for accurate product recommendations in the real estate sector. For example, products with a reference score greater than a similarity threshold can be prioritized for recommendation, or products ranked by reference score can be prioritized for recommendation based on the top three rankings.

[0074] Based on the same idea, the collaborative filtering algorithm is used, as well as user portrait features and similar user portrait features to calculate the user's indirect rating of the product, including: calculating user similarity based on the user portrait features of the target user and the similar user portrait features; wherein, the user portrait features include: at least one of the user's browsing time, collection status, and consultation content of the product; based on the comparison results of user similarity, the similar user with the greatest user similarity is selected as at least one neighbor user; the user's indirect rating of the product is obtained by comprehensively calculating the average rating of the target product by neighbor users, the product of user similarity and the average rating of the target product, and the ratio of the average rating of all products by neighbor users to user similarity.

[0075] In real-world applications, each user's profile features are limited, and their evaluations of products (e.g., real estate listings) are also limited. Therefore, to better discover other neighboring users and help users better discover other products (e.g., real estate listings), we can calculate the similarity between users and thus comprehensively calculate the neighboring users' ratings of the target product (e.g., real estate listings), which is an indirect rating.

[0076] For example, collaborative filtering (CF(u,i)): uses a user-based collaborative filtering algorithm, and the specific implementation is as follows.

[0077] Calculate user similarity: Use cosine similarity to calculate the similarity sim(u,v) between user u and other users v: sim(u,v)=(R_u*R_v) / (||R_u||*||R_v||).

[0078] R_u and R_v represent the rating vectors for users u and v, respectively. The elements of the rating vectors are composed of weighted values of user profile features (e.g., browsing behavior, number of views, browsing duration, product additions, product inquiries, and number of inquiries). (For example, browsing duration is proportionally mapped to a score of 1-5, with additions scoring 4 points and inquiries scoring 5 points). These reflect the user's dynamic needs in new home purchases (e.g., location preferences, apartment type interests, etc.). * denotes the dot product of the vectors; || || denotes the modulus of the vector. The elements of the rating vector are a direct representation of the user's preference for the property (e.g., the weighted values of browsing, adding to favorites, and inquiries).

[0079] When selecting neighbor user v, the K users with the highest similarity to user u's needs are selected from all users who follow the same or similar properties as user u as their neighbor user set N(u). Only users with sim(u,v) > 0 are considered as neighbors. Once neighbor users are selected, they serve the following purposes in the subsequent rating calculation: They provide a preference reference. The ratings R_v,i of neighbor user N(u) reflect the interests of groups with similar needs and serve as the basis for neighbor user u's predictions for unrated properties. Bias adjustment is performed by calculating the bias of neighbor user v's ratings by subtracting the average value (R_v) from the average value (R_u). This bias is then weighted and integrated to adjust the average value (R_u). For personalized predictions, the similarity sim(u,v) is used as a weight to ensure that more similar neighbors contribute more, generating a CF(u,i) that meets user u's needs.

[0080] Predicted rating: Use the following formula to predict the rating of user u for property i: CF(u,i)=mean value (R_u) + Among them: v∈N(u): indicates that neighbor user v belongs to the set N(u), and neighbor user v has a rating record for property i.

[0081] R_v,i: represents the rating of neighbor user v for property i.

[0082] Average value (R_v): represents the average rating of neighbor user v for his rated properties.

[0083] Average (R_u): represents the average rating of user u for the properties he / she has rated, and serves as the benchmark rating for predicting user u's rating of property i (since user u has not yet rated property i).

[0084] If the user u has not rated any property, the average value (R_u) is the global average rating of all properties on the platform.

[0085] It should be noted that if no user in N(u) has rated property i, then CF(u,i) = average value (R_u).

[0086] Through the above method, based on the user portrait characteristics and the ratings of other neighbor users on a specific property, it can help predict, correct and improve the user's rating of the property.

[0087] In one or more embodiments of the present disclosure, the product recommendation score is obtained by weighting and summing the indirect score, direct score, product reference score and second category feature respectively using the first recommendation weight, including: fusing the first recommendation sub-weight obtained by splitting based on the first recommendation weight with the indirect score to obtain a first score result. Fusing the second recommendation sub-weight obtained by splitting based on the first recommendation weight with the direct score to obtain a second score result. Fusing the third recommendation sub-weight obtained by splitting based on the first recommendation weight with the product reference score to obtain a third score result. Fusing the fourth recommendation sub-weight obtained by splitting based on the first recommendation weight with the second category feature to obtain a fourth score result. The product recommendation score is calculated using the first score result, the second score result, the third score result, and the fourth score result to generate a product recommendation result based on the product recommendation score.

[0088] When calculating product recommendation scores, we comprehensively consider multiple factors, including indirect scores, direct scores, product reference scores, and second-category features. This score is then used to generate product recommendations. Furthermore, to ensure more accurate product recommendation scores, dynamically adjustable recommendation sub-weights are assigned to each of these factors. The following example illustrates the calculation process for the integrated recommendation sub-weights.

[0089] The calculation formula for the product recommendation score is as follows: R(u,i)=α×CF(u,i)+β×DL(u,i)+γ×S(i,j)+δ×P(i).

[0090] Where R(u,i) is the interest rating of user u for property i.

[0091] CF(u,i): Indirect score calculated by collaborative filtering algorithm, based on user similarity or listing similarity.

[0092] DL(u,i): Direct scores predicted by the deep learning model, capturing complex nonlinear relationships.

[0093] S(i,j): The score of the product reference rating based on the static attributes of the property.

[0094] P(i): represents the second category features. For example, in the real estate scenario, the second category features can be features related to the particularity of the new home business (for example, new home promotions, etc.).

[0095] α, β, γ, δ: Recommendation sub-weights, which can be dynamically adjusted and optimized based on user portrait features and first category features.

[0096] Through the above solution, by assigning dynamically adjustable recommendation sub-weights to multiple dimensional parameters used in the product recommendation score calculation, the current latest status of the product can be better captured, so that the final calculated score is more in line with the user's current latest needs.

[0097] In one or more disclosed embodiments, after calculating the product recommendation score to generate the product recommendation result according to the product recommendation score, it also includes: sorting the recommended products according to the size of the product recommendation score to obtain a first recommendation result; in response to an update request for the user portrait feature and / or the first category feature, calculating a second recommendation weight based on the updated subsequent user portrait feature and / or the first category feature; judging the subordinate relationship between the previous user portrait feature and / or the first category feature and the updated subsequent user portrait feature and / or the first category feature; if the subsequent user portrait feature and / or the first category feature is subordinate to the previous user portrait feature and / or the first category feature, then calculating the updated product recommendation score and the second recommendation result based on the first recommendation result and the second recommendation weight; if the subsequent user portrait feature and / or the first category feature is in a parallel relationship with the previous user portrait feature and / or the first category feature, or the previous user portrait feature and / or the first category feature is subordinate to the subsequent user portrait feature and / or the first category feature, then calculating the updated product recommendation score based on the second recommendation weight, and recommending or not recommending the product to the user based on the updated product recommendation score.

[0098] In actual applications, when the user portrait features and / or the first category features change, the rule for updating the first recommendation weight should be selected based on the feature relationship before and after the change.

[0099] If the changed or newly added feature is in a subordinate relationship with the previous feature, that is, the new feature belongs to the previous feature, it means that the user or product wants to further refine the previous feature.

[0100] Specifically, if a user portrait feature changes or is newly added, determine whether the newly added or changed user portrait feature belongs to the existing user portrait feature. If so, it means that the newly added or changed later user portrait feature is subordinate to the earlier user portrait feature. For example, an earlier user portrait feature is that the user searches for the type of new house type, and a later user portrait feature is that the user searches for a three-bedroom apartment. It can be seen that the newly added later user portrait feature is subordinate to the earlier user portrait feature. In this case, the first recommendation result calculated previously has a certain reference value, and the second recommendation result can be calculated based on the first recommendation result. In other words, making detailed recommendations based on the first recommendation result so that the recommendation results have an inheritance effect can effectively further improve the recommendation accuracy.

[0101] If the first category feature changes or is newly added, determine whether the newly added or changed first category feature belongs to the existing first category feature. If so, it means that the newly added or changed later first category feature is subordinate to the earlier first category feature. For example, an earlier first category feature is to support pre-ordered products, and a later first category feature is to offer a 5% discount on pre-orders. It can be seen that the newly added later first category feature is subordinate to the earlier first category feature. In this case, the first recommendation result calculated previously has a certain reference value, and the second recommendation result can be calculated based on the first recommendation result. In other words, making detailed recommendations based on the first recommendation result makes the recommendation results have an inheritance effect, which can effectively further improve the recommendation accuracy.

[0102] If the changed or newly added feature is in a parallel relationship with the previous feature, it means that the user or product wants to make changes to the previous feature.

[0103] Specifically, if a user profile feature changes or is newly added, determine whether the newly added or changed user profile feature is in a parallel relationship with the existing user profile feature. For example, if a previous user profile feature shows that the user is searching for a two-bedroom apartment, and a subsequent user profile feature shows that the user is searching for a three-bedroom apartment, then it can be determined that the newly added user profile feature is in a parallel relationship with the previous user profile feature. In this case, the first recommendation result calculated previously is not of reference value, and the third recommendation result needs to be recalculated using the subsequent user profile feature and the first category feature.

[0104] If a first-category feature changes or is newly added, determine whether the newly added or changed first-category feature is in a parallel relationship with the existing first-category feature. For example, if an earlier first-category feature supports a pre-ordered product, and a later first-category feature does not support a pre-ordered product, then the newly added later first-category feature is in a parallel relationship with the earlier first-category feature. In this case, the previously calculated first recommendation result is of no reference value, and the third recommendation result needs to be recalculated using the later user profile features and the first-category features.

[0105] It should be noted that the second category features, first category features, and user profile features mentioned here can be selected and adjusted according to actual needs. In different scenarios, the selected first category features and second category features are different. This is only for illustration and does not constitute a limitation of the technical solution disclosed herein.

[0106] Based on any of the above embodiments, the present disclosure also provides a product recommendation device. Figure 2 This is a schematic block diagram of the structure of a product recommendation device according to an embodiment of the present disclosure. Figure 2 As shown, the product recommendation method device includes: an acquisition module 21, which is used to obtain user portrait features, first category features and second category features. A first calculation module 22 is used to determine the first recommendation weight of the product based on the user portrait features and the first category features. A second calculation module 23 is used to determine the user's indirect rating, direct rating, and product reference rating of the product based on the user portrait features and the first category features of the product. A fusion calculation module 24 is used to use the first recommendation weight to weight and sum the indirect rating, direct rating, product reference rating and the second category features respectively to obtain a product recommendation score, so as to recommend the product to the user or not based on the product recommendation score.

[0107] The first calculation module 22 is used to splice the user portrait feature with the first category feature to obtain a spliced feature vector; and perform a linear transformation on the spliced feature vector based on a multi-layer perceptron to obtain a first recommendation weight in vector form.

[0108] The first computing module 22 is used to input the concatenated feature vector into multiple hidden layers of a multi-layer perceptron and output a feature representation; input the feature representation into an output layer with a specified dimension in the multi-layer perceptron to obtain a first recommendation result containing multiple recommendation sub-weights; wherein the output layer contains the same number of neurons as the recommendation sub-weights.

[0109] The second calculation module 23 is used to find similar users based on the user portrait features and determine similar user portrait features; use the collaborative filtering algorithm, as well as the user portrait features and similar user portrait features to calculate the user's indirect rating of the product; use the deep learning model, as well as the user portrait features and the first category features to calculate the user's direct rating of the product; find similar products based on the first category features and determine the third category features of similar products; use the similarity algorithm, as well as the user portrait features and the third category features of similar products to calculate the product reference rating.

[0110] The second calculation module 23 is used to calculate user similarity based on the user portrait features of the target user and the portrait features of similar users; wherein the user portrait features include: at least one of the user's browsing time for the product, collection status, and consultation content; based on the comparison result of the user similarity, a similar user with the greatest user similarity is selected as at least one neighbor user; the user's indirect score for the product is obtained by comprehensively calculating the average score of the neighbor users for the target product, the product of the user similarity and the average score of the target product, and the ratio of the average score of the neighbor users for all products to the user similarity.

[0111] The fusion calculation module 24 is used to fuse the first recommendation sub-weight obtained by splitting the first recommendation weight with the indirect score to obtain a first score result; fuse the second recommendation sub-weight obtained by splitting the first recommendation weight with the direct score to obtain a second score result; fuse the third recommendation sub-weight obtained by splitting the first recommendation weight with the product reference score to obtain a third score result; fuse the fourth recommendation sub-weight obtained by splitting the first recommendation weight with the second category feature to obtain a fourth score result; and use the first score result, the second score result, the third score result, and the fourth score result to calculate the product recommendation score.

[0112] Optionally, it also includes an update calculation module 25, which is used to sort the recommended products according to the size of the product recommendation score to obtain a first recommendation result; in response to an update request for the user portrait feature and / or the first category feature, calculate the second recommendation weight based on the updated subsequent user portrait feature and / or the first category feature; judge the subordinate relationship between the previous user portrait feature and / or the first category feature and the updated subsequent user portrait feature and / or the first category feature; if the subsequent user portrait feature and / or the first category feature is subordinate to the previous user portrait feature and / or the first category feature, then calculate the updated product recommendation score and the second recommendation result based on the first recommendation result and the second recommendation weight; if the subsequent user portrait feature and / or the first category feature is in a parallel relationship with the previous user portrait feature and / or the first category feature, or the previous user portrait feature and / or the first category feature is subordinate to the subsequent user portrait feature and / or the first category feature, then calculate the updated product recommendation score based on the second recommendation weight, and recommend the product to the user or not based on the updated product recommendation score.

[0113] The implementation process of the functions and effects of each module in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0114] It should be noted here that the various artificial intelligence models such as the deep learning model used in the above scheme are trained using specific training samples, so that the artificial intelligence models mentioned above can achieve corresponding functions.

[0115] For ease of understanding, this disclosure will illustrate the implementation process of this solution through a specific embodiment. Figure 3 This is a schematic diagram of the product recommendation process for this disclosure. This embodiment takes the new house recommendation scenario as an example. Figure 3 As you can see, the calculation requires both user profile features and first-category features (i.e., new home features Fi). User profile features are derived from the user data source (i.e., user interactions with the product, including both online and offline interactions), while first-category features (i.e., real estate features) are derived from the real estate data source (i.e., basic attributes of new homes, real estate information, surrounding facilities, sales information, etc.).

[0116] User profile features (Hu_t) can include at least one of the following: browsing property details (view time, view count), clicking on floor plans (click house type), viewing a mockup (view model), adding a property to a favorite (favorite), consulting a property consultant (consult), joining a group buying tour (join), comparing properties (compare), searching for keywords (search keywords), and dwelling time (dwell time). Other features are also acceptable and can be adjusted as needed. These user profile features are recorded in time series and weighted using a time decay function to emphasize the importance of recent behavior. For example, an exponential decay function can be used: w(t) = exp(-λt), where t is the time difference and λ is the decay coefficient.

[0117] The first category of features (also known as new home features (Fi)) includes basic attributes such as the property name, location (latitude and longitude), price, unit type, area, renovation status, and ownership period. Property information includes the developer, property management company, floor area ratio, greening rate, total number of buildings, total number of households, and number of parking spaces. Surrounding facilities include educational resources, medical resources, transportation conditions, commercial facilities, and environmental amenities.

[0118] The second category of features (i.e., product personalized features P(i)) includes: launch time, delivery time, promotional activities, subscription status, sales stage (customer accumulation period, launch period, sales period), etc.

[0119] After obtaining the above features, the user portrait feature vector Hu_t and the first category feature vector Fi are spliced into a spliced feature vector X=[Hu_t,Fi].

[0120] Multilayer Perceptron (MLP): This process is performed using an MLP network with two hidden layers. The first layer performs a linear transformation: z1 = W1 * X + b1, with an activation function: ReLU(z1) = max(0,z1). The second layer performs a linear transformation: z2 = W2 * ReLU(z1) + b2, with an activation function: ReLU(z2) = max(0,z2). The output layer performs a linear transformation: z = W * ReLU(z2) + b. The output z is divided into four components: z_α, z_β, z_γ, and z_δ, each corresponding to one of the four weights. The softmax function is then used for normalization: α = softmax(z_α), β = softmax(z_β), γ = softmax(z_γ), and δ = softmax(z_δ). This yields the four recommended sub-weights.

[0121] Among them, W1, W2, W, b1, b2 and b are the weight matrix and bias vector that need to be learned through training. The softmax formula is as follows: . zi represents z_α, z_β, z_γ and z_δ.

[0122] Next, the collaborative filtering algorithm is used to calculate the indirect score CF(u,i), that is, to predict the interest score of user u in property i. The input here is the user portrait feature Hu_t and the first category feature Fi of property i, and the output is CF(u,i).

[0123] Use the deep learning model to calculate the direct score DL (u, i), that is, to predict the user's preference for the property, input the user portrait feature Hu_t, and output DL (u, i).

[0124] The similarity algorithm is used to calculate the product reference score (S). The input is the first category feature of property i and the first category feature of property j, and the output is the product reference score S(i, j).

[0125] The second category of features for new home businesses includes: sales stage, discount information, subscription status, etc. These are non-persistent features that will appear or disappear at a specific time point.

[0126] Next, the calculated score, similarity, second category features and four recommendation sub-weights are combined and input into the product recommendation score R (u, i).

[0127] Furthermore, the products can be sorted according to the product recommendation scores, and the top-ranked products can be generated as recommendation results, which can be sent to users to display the product recommendation results.

[0128] The execution subject of the product recommendation method in the specific embodiment of the present disclosure may be an electronic device such as a server (including a local server or a cloud server).

[0129] Therefore, based on any of the above embodiments, the present disclosure further provides an electronic device, which can execute the product recommendation method of any of the embodiments described above in the present disclosure.

[0130] Figure 4 The present invention is a block diagram showing the structure of an electronic device according to an embodiment of the present invention.

[0131] The hardware structure of the electronic device 1000 can be implemented using a bus architecture. The bus architecture can include any number of interconnecting buses and bridges, depending on the specific application and overall design constraints of the hardware. The bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. The bus 1100 can also connect various other circuits 1400 such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0132] Bus 1100 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component Architecture (EISA) bus. Buses can be classified as address buses, data buses, control buses, and the like. For ease of illustration, this figure shows only one connecting line, but this does not imply that there is only one bus or only one type of bus.

[0133] The present disclosure also provides a readable storage medium having a computer program stored therein, which is used to implement the above-mentioned method when the computer program is executed by a processor. "Readable storage medium" can be any device that can contain, store, communicate, propagate or transmit a program for use in an instruction execution system, device or equipment or in combination with these instruction execution systems, devices or equipment. More specific examples of readable storage media include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and editable read-only memory (EPROM or flash memory), an optical fiber device, and a portable read-only memory (CDROM), etc.

[0134] The present disclosure also provides a computer program product. The method of the present disclosure can be implemented in whole or in part using software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, the process or function of the present disclosure is performed in whole or in part.

[0135] A computer program or instruction can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instruction can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any accessible medium or a data storage device such as a server or data center that integrates one or more accessible media. The accessible medium can be a magnetic medium such as a floppy disk, hard disk, or magnetic tape; an optical medium such as a digital video disk; or a semiconductor medium such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or can include both volatile and non-volatile types of storage media.

[0136] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing method device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing method device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0138] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0140] In the description of this specification, the description with reference to the terms "one embodiment / method", "some embodiments / methods", "example", "specific example", or "some examples" means that the specific features, structures, or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment / method or example. Moreover, the specific features, structures, or characteristics described may be combined in a suitable manner in any one or more embodiments / methods or examples. In addition, those skilled in the art may combine and combine different embodiments / methods or examples described in this specification and the features of different embodiments / methods or examples, unless they are contradictory.

[0141] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. Throughout the present disclosure, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0142] Those skilled in the art will appreciate that the above embodiments are merely intended to clearly illustrate the present disclosure and are not intended to limit the scope of the present disclosure. Other changes or modifications may be made based on the above disclosure, and such changes or modifications are still within the scope of the present disclosure.

Claims

1. A product recommendation method, characterized in that: The method comprises: Obtain user portrait features, product first category features, and second category features; Determining a first recommendation weight for the product based on the user portrait feature and the first category feature of the product; Determining the user's indirect rating, direct rating, and product reference rating for the product based on the user portrait features and the first category features of the product; The first recommendation weight is used to weight and sum the indirect score, the direct score, the product reference score and the second category feature to obtain a product recommendation score, so as to recommend or not recommend a product to the user based on the product recommendation score.

2. The method according to claim 1, characterized in that The calculating a first recommendation weight based on the user portrait feature and the first category feature includes: Concatenate the user portrait feature with the first category feature to obtain a concatenated feature vector; Perform a linear transformation on the concatenated feature vector based on a multi-layer perceptron to obtain the first recommendation weight in vector form.

3. The method according to claim 2, characterized in that The performing a linear transformation on the concatenated feature vector based on a multi-layer perceptron to obtain the first recommendation weight in vector form includes: Inputting the concatenated feature vector into multiple hidden layers of the multi-layer perceptron and outputting feature representation; The feature representation is input into an output layer with a specified dimension in the multilayer perceptron to obtain a first recommendation result including multiple recommendation sub-weights; wherein the output layer includes the same number of neurons as the recommendation sub-weights.

4. The method according to claim 1, wherein The indirect score, direct score, and product reference score calculated based on the user portrait features and the first category features include: Find similar users based on the user portrait features and determine similar user portrait features; use a collaborative filtering algorithm and the user portrait features and similar user portrait features to calculate the user's indirect rating of the product; Calculating the user's direct rating of the product using a deep learning model, the user portrait features, and the first category features; Similar products are searched based on the first category feature, and the third category feature of the similar products is determined; and the product reference score is calculated using a similarity algorithm and the first category feature and the third category feature of the similar products.

5. The method according to claim 4, characterized in that The indirect rating of the user on the product is calculated by using a collaborative filtering algorithm and the user portrait features and similar user portrait features, including: Calculating user similarity based on the target user's user profile features and similar user profile features; wherein the user profile features include: at least one of the user's browsing time for the product, collection status, and consultation content; According to the user similarity comparison result, a similar user with the greatest user similarity is selected as at least one neighbor user; The indirect rating of the user on the product is obtained by comprehensively calculating the average rating of the target product by the neighbor users, the product of the user similarity and the average rating of the target product, and the ratio of the average rating of all products by the neighbor users to the user similarity.

6. The method according to claim 1, characterized in that The method of using the first recommendation weight to weight and sum the indirect score, the direct score, the product reference score, and the second category feature to obtain a product recommendation score includes: fusing a first recommendation sub-weight obtained by splitting the first recommendation weight with the indirect score to obtain a first score result; fusing a second recommendation sub-weight obtained by splitting the first recommendation weight with the direct score to obtain a second score result; Merging a third recommendation sub-weight obtained by splitting the first recommendation weight with the product reference score to obtain a third score result; fusing a fourth recommendation sub-weight obtained by splitting the first recommendation weight with the second category feature to obtain a fourth scoring result; The product recommendation score is calculated using the first scoring result, the second scoring result, the third scoring result, and the fourth scoring result.

7. The method according to claim 1, characterized in that After calculating the product recommendation score and generating the product recommendation result according to the product recommendation score, the process further includes: Sort the recommended products according to the product recommendation scores to obtain a first recommendation result; In response to a request to update the user portrait feature and / or the first category feature, calculating a second recommendation weight based on the updated subsequent user portrait feature and / or the first category feature; Determining the subordinate relationship between the previous user profile features and / or first category features and the updated later user profile features and / or first category features; If the subsequent user profile feature and / or first category feature is subordinate to the previous user profile feature and / or first category feature, then calculating an updated product recommendation score and a second recommendation result based on the first recommendation result and the second recommendation weight; If the subsequent user portrait feature and / or first category feature are in a parallel relationship with the prior user portrait feature and / or first category feature, or the prior user portrait feature and / or first category feature are subordinate to the subsequent user portrait feature and / or first category feature, then the updated product recommendation score is calculated based on the second recommendation weight, and the product is recommended to the user or not based on the updated product recommendation score.

8. An electronic device, characterized in that: include: a memory storing execution instructions; as well as A processor, wherein the processor executes the execution instructions stored in the memory, so that the processor executes the method according to any one of claims 1 to 7.

9. A readable storage medium, characterized in that The readable storage medium stores execution instructions, which are used to implement the method according to any one of claims 1 to 7 when executed by a processor.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images