AI customized furniture precision marketing system and method based on user portrait

Abstract

The invention belongs to the technical field of furniture marketing management, and discloses an AI customized furniture precision marketing system and method based on a user portrait. Comprising the steps of collecting user portrait feature vectors; constructing a user life scene model based on the home space image, the house type plane graph and the space size input information uploaded by the user; calculating a space adaptation degree and a style coordination degree of the furniture commodity and the target room, and generating a furniture candidate combination set on the premise of meeting a space constraint condition; inputting the user portrait feature vector, the user life scene model and the furniture candidate combination set into a multi-task recommendation model, predicting a purchase transformation probability and a display click intention of each furniture candidate combination in the current marketing scene, and outputting a sorted target recommendation combination; and on the basis of the pre-trained copywriting generation model, generating a personalized furniture marketing copywriting strongly related to the actual life scene of the user, and realizing AI customized furniture precision marketing based on the user portrait.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of furniture marketing management, more specifically, the present application relates to an AI customized furniture precision marketing system and method based on user portrait. BACKGROUND

[0002] With the deep integration of e-commerce and artificial intelligence technology, personalized recommendation has become the core means for retail industry to improve user experience and conversion efficiency. In the special category of furniture, personalized marketing faces more complex technical challenges than general commodity recommendation. Furniture is not only an independent commodity, but also a core element of living space and personal aesthetics and lifestyle. Therefore, under the background of deep integration of artificial intelligence and e-commerce, personalized recommendation technology has become the key to improving retail efficiency. However, in the high-value, low-frequency and experience-oriented consumer field of furniture, the existing technical architecture faces a fundamental problem that it is seriously out of touch with the user's real decision-making scene. The traditional model matches furniture as an isolated commodity with static labels such as style and price, completely ignoring the fact that furniture consumption is essentially the user's planning and construction of their personal living space. This cognitive bias causes the entire chain from algorithmic recommendation to marketing outreach to be suspended in the user's real life scene, forming a systematic scene break.

[0003] Specifically, the current technical system cannot establish and penetrate a unified user living scene model. This leads to the failure of the whole process from understanding, recommendation to communication. For example, when a user has a sofa purchase intention, since the system also cannot access the living room scene that should exist in the user's actual living scene, it leads to a one-sided recommendation result. The system may recommend a beautifully designed three-seater sofa to a user who loves Italian minimalist style, and highlight its head layer of cowhide and master design in marketing. However, since the recommendation engine is unaware that the user's living room is only 3 meters wide, the recommended goods cannot be physically accommodated, so only generic marketing copy can be generated, which cannot vividly depict "how this sofa can form a harmonious resting corner with your coffee table and carpet". As a result, the user receives a correct recommendation in terms of style preference, but the recommendation is ineffective in the living scene, and the promotional information is irrelevant to the user's situation. This scene break makes the recommendation degenerate from personalized to one-sided, and the marketing degenerates from precise to empty, seriously damaging user experience and business conversion rate. The lack of user living scene customization limits the impact and conversion efficiency.

[0004] In view of this, an AI customized furniture precision marketing system and method based on user portrait is designed. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purposes, the present application provides the following technical scheme: an AI customized furniture precision marketing method based on user portrait, comprising: A user portrait acquisition and fusion module is used to acquire user basic attribute data, historical behavior data and explicit preference data from a user terminal on an e-commerce platform server, and fuse them into a user portrait feature vector; A user life scene modeling module is used to construct a user life scene model based on home space image, house type plan and space size input information uploaded by the user; A furniture candidate combination generation module is used to calculate the space adaptation degree and style coordination degree of furniture goods and the target room according to the user life scene model and the size parameters, style labels and function labels of the furniture goods in the furniture goods library, and generate a furniture candidate combination set under the premise of meeting the space constraint condition; A furniture recommendation module is used to input the user portrait feature vector, the user life scene model and the furniture candidate combination set into a multi-task recommendation model to predict the purchase conversion probability and display click intention of each furniture candidate combination in the current marketing scene, and output a sorted target recommendation combination; A scene-driven marketing module is used to input space layout information, existing furniture information, style preference information and recommended goods information into a pre-trained copy generation model based on the user life scene model and the target recommendation combination, to generate personalized furniture marketing copy strongly related to the actual life scene of the user; A multi-channel reach module is used to push the target recommendation combination and the personalized furniture marketing copy to the user terminal.

[0006] Preferably, the fusion method of the user portrait feature vector comprises: The age range, family member structure, geographic location and residence type information of the user are acquired as basic attribute data; The browsing track, search keywords, collection and purchase record, historical order furniture category and price distribution of the user on the furniture page are acquired as historical behavior data; The preference intensity of the user for modern minimalist, Italian minimalist, Nordic and American style is inferred based on the data generated by the user in the style evaluation questionnaire, like or shield goods operation, as explicit preference data; The basic attribute data, historical behavior data and explicit preference data are fused by numerical encoding and weighting to generate a user portrait feature vector.

[0007] Preferably, the method for constructing the user life scene model comprises: Based on user-uploaded images of home spaces, the system identifies walls, floors, doors and windows, existing furniture outlines and major decorations, estimates the room's color tone and lighting direction, and obtains analysis results. Based on the floor plan, the length and width dimensions, door and window positions, and structural column positions of each room are extracted to generate a geometric layout diagram of the room. The system checks the consistency between the input space dimensions and the room's geometric layout, and issues a correction prompt to the user when a conflict occurs. Based on the analysis results of the home space image, the room geometric layout diagram, and the verified spatial size input information, a user living scene model is constructed in the form of a graph structure. In the graph structure, the nodes represent rooms and objects, and the edges represent containment relationships, adjacency relationships, and spatial constraint relationships.

[0008] Preferably, the method for generating the furniture candidate combination set includes: Select furniture items from the furniture product library that match the function of the target room, and extract the size, placement direction constraints, and disassembly and combination capabilities of the selected furniture items; In the user's life scenario model, for each candidate furniture item, calculate the set of feasible placement locations in the target room. The set of feasible placement locations includes spatial constraints that satisfy the distance from walls, doors and windows is not less than a preset safety interval and do not block the main passageway. For modular furniture that meets spatial constraints, a combination search is performed in the set of feasible placement locations. Combinations whose total length exceeds the reserved wall length or whose combination depth exceeds the available passage width are excluded, forming an initial set of candidate furniture combinations that meet the physical space adaptation conditions. Based on the explicit preference data in the user profile feature vector and the decoration style features extracted from the analysis results, the style coordination degree of each initial furniture candidate combination is calculated, and combinations with a style coordination degree lower than the preset style coordination degree threshold are removed to obtain the furniture candidate combination set.

[0009] Preferably, the multi-task recommendation model is deployed on an e-commerce platform server, based on a multi-tower neural network design, and trained using a pre-designed training sample set. The multi-task recommendation model consists of a user tower network, a product tower network, and a matching layer. The user tower network consists of an input layer and a feature fusion layer. The input layer receives two parallel feature inputs, one of which is the user profile feature vector. The input is the target room features extracted from the target room in the user's life scenario model. The feature fusion layer includes several fully connected layers and is activated using the ReLU function. The product pyramid network consists of an input layer and a feature abstraction layer. The input layer receives feature representations of a set of furniture candidate combinations, and the feature abstraction layer consists of several fully connected layers activated by the ReLU function. The matching layer is used to calculate the inner product of the user-scene joint representation vector and the combined representation vector, and then output the combined representation vector, and predict the purchase conversion probability and the click intention prediction value of each combination in the candidate combination set of furniture, re-rank all combinations according to the preset ranking strategy, and select the combination according to the rules, which is recorded as the target recommended combination after ranking.

[0010] Preferably, the training by the pre-designed training sample set comprises: Constructing a training sample set, wherein each training sample comprises a portrait feature vector of a historical user, a user life scene model feature extracted from the historical scene data of the user, a feature vector of a candidate combination of furniture that has been displayed to the user in history, and a real behavior label generated by the user to the combination; the real behavior label includes a click label for supervising the click intention prediction, and a purchase label for supervising the purchase conversion prediction; Inputting the user portrait feature vector and the target room feature in the user life scene model into the user tower network to obtain a user-scene joint representation vector, wherein the target room feature comprises geometric layout features and style features; Inputting the feature vector of the candidate combination of furniture into the commodity tower network to obtain a combination representation vector; The user-scene joint representation vector and the combination representation vector are matched and calculated by the matching layer to output the click intention prediction value and the purchase conversion probability prediction value respectively; The click prediction loss is calculated according to the click intention prediction value and the click label, and the conversion prediction loss is calculated according to the purchase conversion probability prediction value and the purchase label; the binary cross entropy loss function is used for both the click prediction loss and the conversion prediction loss, the total loss is obtained by weighted sum of the two losses, and the parameters of the user tower network, the commodity tower network and the matching layer are optimized by the back propagation algorithm until the model converges.

[0011] Preferably, the copy generation model is designed based on the Transformer structure and trained using a pre-designed domain corpus, comprising: An encoder is used to encode the room layout and style features extracted from the user life scene model and the furniture commodity features extracted from the target recommended combination into context representation sequences; wherein the room layout and style features include room size, existing furniture, wall color and lighting information, and the furniture commodity features include furniture name, style label and material characteristics; A decoder is configured to generate a marketing copy text according to a semantic structure of space depiction → combination relationship → lifestyle image → action guide, with the context representation sequence as a condition; wherein, the decoding process is forced to generate a text containing key room size information extracted from the user living scene model and main function description of furniture goods in the target recommendation combination through a constraint decoding mechanism.

[0012] Preferably, the method of training using a pre-designed domain corpus comprises: Collecting marketing copies corresponding to historical high-conversion furniture orders, description copies manually written for specific home scenes, and home design case texts to form an original marketing copy text set; Extracting and labeling the structured input features corresponding to each copy in the original text set to form training paired data; the structured input features include room layout and style features parsed from the scene described in the copy, and furniture product features parsed from the goods described in the copy; During the training process, a cross-entropy loss function is used to measure the difference between the model output text and the original marketing copy text, and a gradient descent optimization algorithm is used to iteratively update the model parameters to minimize the loss function, thereby completing the model training.

[0013] Preferably, the method of pushing the target recommendation combination and the personalized furniture marketing copy to the user terminal comprises: Packaging the target recommendation combination and the personalized furniture marketing copy into a marketing message supporting rich media display, and sending it to the user terminal through multiple touch channels; wherein, the multiple touch channels include APP push notification, in-site message, public number message and smart TV pop-up window; Providing an AR preview entry for the target recommendation combination on the APP or web page, allowing the user to superimpose a virtual furniture model in a real living room or bedroom environment based on the phone camera.

[0014] An AI customized furniture precision marketing method based on user portrait comprises: S1: Collecting user basic attribute data, historical behavior data and explicit preference data from the user terminal, and fusing them into a user portrait feature vector; S2: Based on the user uploaded home space image, house type plan and space size input information, a user living scene model is constructed; S3: According to the user living scene model and the size parameters, style labels and function labels of furniture goods in the furniture goods library, the spatial adaptability and style coordination of furniture goods and target rooms are calculated, and on the premise of meeting the spatial constraint condition, a furniture candidate combination set is generated; S4: input the user portrait feature vector, the user life scene model and the furniture candidate combination set into the multi-task recommendation model, predict the purchase conversion probability and display click intention of each furniture candidate combination in the current marketing scene, and output the ranked target recommendation combination; S5: based on the user life scene model and the target recommendation combination, input the space layout information, existing furniture information, style preference information and recommended commodity information into the copy generation model to generate a personalized furniture marketing copy strongly related to the actual life scene of the user; S6: push the target recommendation combination and the personalized furniture marketing copy to the user terminal.

[0015] The AI customized furniture precision marketing method based on the user portrait has the following technical effects and advantages: The scheme constructs a technical processing flow of data perception, scene understanding, scheme generation, intention prediction, content creation and immersion reach. All recommendations are strictly tested in physical space and style aesthetics, deeply bound to the personal life and specific family environment of the user, and extremely personalized and infectious, providing one-stop service from inspiration, scheme to visual verification, greatly shortening the decision path.

[0016] The user behavior data, visual space data, geometric drawing data and commodity knowledge graph data are deeply fused to construct a unified user life scene model. The spatial constraint solving of computer graphics and the style semantic calculation of artificial intelligence are combined to complete the double filtering of physical feasibility and aesthetic coordination before recommendation, from understanding space to generating scheme, to creating marketing content and providing AR experience, forming an end-to-end AI system driven by user personalized scene.

[0017] Through the structured feature input and constraint decoding mechanism, it is accurately applied to the highly vertical and rigorous furniture scene marketing copy generation, avoiding the illusion problem of general generation, and realizing the reliable landing of commercialization. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The system schematic diagram of the AI customized furniture precision marketing based on the user portrait of the application; Figure 2 The method schematic diagram of the AI customized furniture precision marketing based on the user portrait in the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0020] Please refer to Figure 1 and Figure 2 In the embodiments of the present application, an AI customized furniture precision marketing system based on user portrait includes: A user portrait acquisition and fusion module is configured to acquire user basic attribute data, historical behavior data and explicit preference data from a user terminal on an e-commerce platform server, and fuse the data into a user portrait feature vector; A user life scene modeling module is configured to construct a user life scene model based on home space image, house type plan and space size input information uploaded by a user, and realize expression of room structure, size constraint, existing furniture distribution and decoration style of an actual living space of the user; A furniture candidate combination generation module is configured to calculate spatial adaptation degree and style coordination degree of furniture goods and a target room according to the user life scene model and size parameters, style labels and function labels of the furniture goods in a furniture goods library, and generate a furniture candidate combination set under the premise of meeting the spatial constraint condition; A furniture recommendation module is configured to input the user portrait feature vector, the user life scene model and the furniture candidate combination set into a multi-task recommendation model, predict purchase conversion probability and display click intention of each furniture candidate combination in a current marketing scene, and output a sorted target recommendation combination; A scene-driven marketing module is configured to input spatial layout information, existing furniture information, style preference information and recommended goods information into a pre-trained copy generation model based on the user life scene model and the target recommendation combination, and generate an individualized furniture marketing copy strongly related to an actual life scene of the user; A multi-channel touch module is configured to push the target recommendation combination and the individualized furniture marketing copy to a user terminal.

[0021] In view of the problem that user preference representation is one-sided, sparse and implicit, and traditional portrait only relies on basic attributes or single behavior, and cannot accurately depict deep and explicit preferences of users for furniture style and space function, the present design integrates three types of heterogeneous data including basic attributes, historical behavior and explicit preference, constructs a user portrait feature vector with rich semantics, dynamically reflects aesthetic tendency, consumption ability and space use demand of the user, and provides accurate and comprehensive user-side input for subsequent individualized scene modeling and goods recommendation. The specific design is as follows: The user portrait feature vector fusion method comprises: Collecting age, family structure, geographic location and residence type information of the user (obtained through user registration information, address resolution or active selection. The information includes at least three dimensions of residential property, house structure and decoration state. Specifically, the residential property is, for example, an apartment, a villa, a loft, a dormitory, a rental house and a self-owned house; the house structure is, for example, an open room, a one-room-one-hall, a two-room-one-hall and a three-room-two-hall; and the decoration state is, for example, a raw house, a fine house and a renovated old house), as basic attribute data; Planting a buried point code at a key interaction node (such as page loading, staying, scrolling, clicking, search box input, collection button and purchase button) of a furniture page of an e-commerce platform to collect in real time the browsing track of the user on the furniture page, search keywords, collection and purchase records, and the distribution of furniture categories and prices in historical orders as historical behavior data. For the browsing track, the sequence of the user's access to the product detail page is recorded in time sequence, and the style label and category label of each product are extracted. The search keywords are normalized and classified, for example, a Nordic fabric sofa is parsed as {style: Nordic, material: fabric, category: sofa}. The collection and purchase records refer to the frequency of collection / purchase of goods of each category, price segment and style. The historical order analysis refers to the calculation of the average price of the purchased furniture, the category distribution (such as the proportion of living room furniture) and the style concentration (such as the proportion of modern minimalist style orders). The collected data is aggregated into a structured feature table with a user identifier as the primary key and a time window (such as the last 180 days) to form the historical behavior data.

[0022] Based on the data generated in the style evaluation questionnaire (showing the user pictures of furniture scenes containing different styles, asking the user to rate, such as 1-5 points, splicing the rating results corresponding to each style to directly generate an initial preference vector), the like or shield operation of the goods (extracting the style label of the goods liked by the user, counting the frequency of each style, calculating the proportion in all occurrence times, recording the style label of the goods shielded (or marked as "not interested") by the user, counting the frequency of each style being shielded, and also calculating the proportion in all occurrence times, and performing weighted summation on the two proportions with a preset weight to obtain a preference intensity vector for quantitatively representing the data generated in the like or shield operation of the goods), the preference intensity of the user for modern minimalist, Italian minimalist, Nordic, American style is inferred as explicit preference data (the initial preference vector and the preference intensity vector are weighted and fused, and normalized by a Softmax function to obtain the final user style preference intensity vector, denoted as explicit preference data, the dimension of the vector corresponds to the pre-defined style sequence, such as [modern minimalist, Italian minimalist, Nordic, American, Chinese]). The basic attribute data, historical behavior data and explicit preference data are fused by numerical encoding (unique array encoding for non-numerical labels) and weighting (weights can be pre-set according to historical data analysis) to generate a user portrait feature vector.

[0023] The core problem of the recommendation system is completely disconnected from the user's real physical space (scene break). The system has no knowledge of the actual size, layout, and existing environment of the user's room. The design builds a machine-understandable graph structure user living scene model to create a digital twin of the user's space. This model converts the abstract living room concept into a specific computable geometric space in the form of nodes and edges. This provides an indispensable scene-side basis for subsequent furniture space adaptation calculations. The specific design is as follows: The method for constructing the user living scene model comprises: Based on the user uploaded home space image, the wall, floor, door and window, existing furniture contour and main decoration are recognized, and the room color keynote and lighting direction are estimated to obtain the analysis result; specifically, the analysis result contains the following contents: Wall information: record all the walls identified in the image in the form of a list. Each wall element contains a unique identifier, contour information composed of a series of coordinate points, and the main color pattern, such as "light gray".

[0024] Floor information: record the identified floor area in the form of a list. Each floor element contains a unique identifier, inferred material properties, such as "wood floor", and the color pattern.

[0025] Door and window information: record all the doors and windows in the form of a list. Each door and window element contains a unique identifier, bounding box coordinates representing its position and size in the image, and the inferred orientation, such as "south".

[0026] Existing furniture information: record the existing furniture identified in the form of a list. Each piece of furniture contains its category, such as "sofa"; the bounding box coordinates in the image; and the inferred style attribute, such as "modern".

[0027] Overall color keynote: provide a general field to describe the overall color atmosphere of the room, such as "warm gray tone".

[0028] Lighting direction: provide a field to record the main natural light source direction inferred from the image.

[0029] Based on the house plan, the length and width of each room, the position of doors and windows, and the position of structural columns are extracted to generate a room geometry layout; the specific composition is as follows: Room List: The core is a list named rooms, where each element corresponds to an individual room.

[0030] Room Attributes: For each room object in the list, the key attributes include room identification (a functional type that uniquely identifies the room, such as living room), geometric outline (a list of ordered coordinate points that define the closed shape of the room. Each coordinate point is represented in [x, y] format, with numerical values usually in millimeters to accurately describe the size and shape of the room), door opening information (a list recording all doors in the room. Each door contains a unique identifier, location in the room outline coordinate system, and its hole width), window opening information (a list recording all windows in the room. Each window contains a unique identifier, location, and width), and structural column information (a list recording structural column obstacles in the room. Each structural column contains a unique identifier, location, and cross-sectional size).

[0031] Conduct consistency check between space size input information and room geometric layout diagram, and issue correction prompt to user when conflict occurs; Specifically, compare the user uploaded space size input information (e.g. user manually input living room length 5.2 meters) with the corresponding size automatically calculated from the room geometric layout diagram.

[0032] Verification rules include: a) If the difference between the two exceeds the preset threshold (e.g. ±150mm), it is marked as a hard conflict; b) If the aspect ratio of the corresponding room exceeds the threshold (e.g. 10%), it is marked as a soft conflict.

[0033] Conflict judgment: Any hard conflict or soft conflict of key dimensions will trigger the system's correction prompt process.

[0034] Correction process: In the user terminal's interactive interface (such as Web page or APP pop-up window), highlight the room and specific size that has a conflict; Provide visual comparison and list user input values and system analysis values; Provide users with operation options: ① adopt system analysis values, ② re-enter size, ③ re-upload or mark floor plan.

[0035] The user's choice will be the space size input information after verification.

[0036] Based on the analysis results of the home space image, the room geometric layout diagram and the verified space size input information, a user life scene model is constructed in a graph structure form, wherein a node of the graph structure represents a room and an object, and an edge represents a containing relationship, an adjacent relationship and a space constraint relationship. Specifically, the user life scene model is an attribute graph, and the specific structure definition is as follows: Node: Room node: each room, node attributes include room identification, function type, geometric contour, wall color, floor material, lighting level; Object node: each existing furniture, door, window and structural column, node attributes include object category, size, style, color, geometric bounding box and relative position coordinates in the room.

[0037] Edge: Containing relationship: used to connect a room node and an object node located inside the room.

[0038] Adjacent relationship: connecting two room nodes sharing a wall.

[0039] Space constraint relationship: connecting a furniture node and a wall node or a door and window node, and the edge attribute records the specific constraint type and parameter, The graph structure model is stored using a graph database structure.

[0040] In view of the double failure problems of physical unfeasibility and style inconsistency of the recommended results. The traditional recommendation may recommend a sofa that cannot be placed, or a tea table that conflicts with the existing decoration. The physical space constraint calculation is used for hard filtering, and the style coordination degree calculation is used for soft screening to generate a furniture candidate combination set. This is equivalent to completing a strict space adaptability and aesthetic consistency selection in the sea of goods in advance, ensuring that all feasible solutions above the pass are input into the recommendation model, greatly improving the efficiency and accuracy of subsequent recommendations. The specific design is as follows: The method for generating the furniture candidate combination set comprises: Selecting furniture goods matching the target room function from the furniture goods library, and extracting the size, placement direction constraint and detachable combination capability of the selected furniture goods; Specifically, the furniture goods library is a room function-furniture category mapping rule library maintained by the system. When the target room is a living room, the system automatically selects goods with product category tags belonging to the set {sofa, tea table, TV cabinet, side cabinet, leisure chair, carpet} from the furniture goods library; when the target room is a bedroom, the system selects goods with product category tags belonging to the set {bed, bedside table, wardrobe, dressing table} from the furniture goods library. The mapping relationship is predefined according to home common sense and business rules.

[0041] The following standardized attribute fields are directly extracted from the selected furniture goods records: Size: the length, width, and height values marked in the product detail page, with the unit being millimeters.

[0042] Placement direction constraint: a label field, such as needing to be placed against the wall, being able to be centered on all four sides, or having the short side against the wall.

[0043] Split and combination capability: refers to whether the furniture is modular in design and the connection and configuration rules of its modules. For non-modular furniture, this field is 0. This information is used to determine whether multiple single products can be arranged as a whole unit for space layout calculation.

[0044] In the user life scene model, for each candidate furniture product, calculate its set of feasible placement positions in the target room, where the set of feasible placement positions includes space constraints that meet the distance from the wall, door, and window not less than the preset safety interval, and do not block the main passage. Specifically, For the calculation of feasible placement positions, this process is simulated on the two-dimensional plane defined by the room geometry layout in the user life scene model. For each candidate furniture product, according to its size and placement direction constraint, one or more rectangles (or combined polygons) representing its outline are generated within the polygon outline of the target room. The furniture outline is systematically translated and rotated at a preset step size (such as 50 mm). For each possible position, collision detection is performed to check whether all hard constraints are met. If so, it is considered a feasible placement position, and all positions are counted to form a set, which is the set of feasible placement positions.

[0045] The preset safety interval is a set of rule values preset in the system, for example: The furniture and the wall need to maintain a gap of ≥50 mm (for outlets, skirting, or cleaning).

[0046] The furniture and the door and window opening edge need to maintain a gap of ≥300 mm (to ensure opening and passage).

[0047] The furniture and other existing fixed objects (such as structural columns) need to maintain a gap of ≥100 mm.

[0048] Space constraint of not blocking the main passage: refers to the predefinition of key passage paths in the user life scene model. For example, from the room entrance to the balcony, to the doorways of other rooms, a continuous passage with a width of not less than 600 mm is usually defined. When calculating the set of feasible placement positions, the furniture outline must not invade these predefined passage areas.

[0049] For combined furniture that meets the space constraints, search for combinations in the set of feasible placement positions, exclude combinations whose total length exceeds the reserved wall length or whose combined depth exceeds the available passage width, and form an initial set of furniture candidate combinations that meet the physical space adaptation conditions. Based on the explicit preference data in the user portrait feature vector, the style coordination degree of each initial furniture candidate combination is calculated with the decoration style features extracted from the analysis results, and the combinations with a style coordination degree lower than a preset style coordination threshold are removed to obtain a furniture candidate combination set. Specifically, the explicit preference data in the user portrait feature vector (such as [0.7, 0.2, 0.1, 0.0] corresponding to [modern minimalist, Italian minimalist, Nordic, American style]) is recorded as the user style preference intensity vector U. The decoration style features extracted from the analysis results (for example, it is judged as modern minimalist tendency according to the overall color tone and existing furniture style) are quantified as the room style vector R (such as [0.9, 0.1, 0.0, 0.0]). The style labels (such as modern minimalist) of all single products in an initial furniture candidate combination are counted and normalized to obtain the combination style vector F (such as [1.0, 0.0, 0.0, 0.0]). The weighted cosine similarity is used to calculate the coordination degree S = a x cosine_similarity(U, F) + β x cosine_similarity(R, F), wherein cosine_similarity(U, F) and cosine_similarity(R, F) represent the cosine similarity of U and F, and the cosine similarity of R and F, respectively, and a and β are preset weights (such as a = 0.6, β = 0.4), which represent different emphasis on user personal preference and existing environment coordination. The score S ranges between [0, 1].

[0050] A style coordination threshold (such as 0.6) is set. If the S value of any initial furniture candidate combination is lower than the threshold, it is considered that the style is not coordinated, and is removed from the set. The final furniture candidate combination set is a list, and each element in the list represents a feasible collocation scheme that passes the space and style double screening.

[0051] The core design goal of the multi-task recommendation model is to deeply integrate and match the discrete user portrait feature vector, user life scene model and furniture candidate combination in the same vector space, so as to realize the joint prediction of the user-scene-goods trinity intention. The specific description is as follows: Aiming at the problem that the traditional recommendation model intends to predict a single and ignores the scene context. The traditional model may only predict the click rate, and does not take the specific spatial scene where the user is as the core decision basis. A multi-tower neural network is designed, in which the user tower network fuses user portrait and scene features, the commodity tower network encodes combined information, the matching layer calculates the matching degree and simultaneously outputs the click intention and purchase conversion probability. The user-scene-commodity trinity joint deep matching is realized. The model can not only predict whether the user wants to see, but also predict whether the user wants to buy in a specific scene, and intelligently sort based on this. This makes the recommendation from the goods that the user may like to the overall solution that the user is very likely to purchase in his specific home, realizing the leap from general recommendation to scenario decision support. The specific design is as follows: The multi-task recommendation model is deployed on an e-commerce platform server, designed based on a multi-tower neural network, trained through a pre-designed training sample set, and composed of a user tower network, a commodity tower network, and a matching layer. The user tower network is the cognitive part of the model, and its core function is to fuse user personal attributes and the space environment it is in to generate a unified user-scene joint representation vector. This vector not only encodes who the user is (portrait), but also encodes what environment the user is currently in and what demand it faces (scene), including an input layer and a feature fusion layer. The input layer receives two parallel feature inputs, input one is a user portrait feature vector (e.g. 256 dimensions), and input two is the target room features extracted in the user life scene model for the target room. The feature fusion layer includes several fully connected layers (selected according to actual production needs, e.g. 2 layers, 3 layers or more), which concatenate the two input vectors and then perform nonlinear transformation and feature interaction through several fully connected layers (e.g. 2-3 layers). Use the ReLU function to activate; gradually fuse heterogeneous user information and scene information into a semantically unified user-scene joint representation vector (e.g. 128 dimensions).

[0052] The commodity tower network is the commodity part of the model, and its core function is to represent a complete furniture matching scheme, and encode the discrete characteristics of multiple single products in the furniture candidate combination into a whole combination representation vector, including an input layer and a feature abstraction layer. The input layer is used to receive a feature representation of a furniture candidate combination set, which is constructed by splicing the size parameters, style labels, material characteristics, price characteristics, etc. of each single product in the combination to form a long vector. If the combination contains multiple commodities, the feature vectors of the commodities are further serialized to form a fixed-dimensional combination input vector. The feature abstraction layer includes several fully connected layers (such as 2-3 layers) activated by ReLU function. The input vector is then nonlinearly transformed by the fully connected layer to learn the matching semantics and overall attributes within the commodity combination, and finally outputs a combination representation vector with the same dimension as the user-scene joint representation vector (e.g. 128 dimensions).

[0053] The matching layer is the decision part of the model, and its core function is to calculate the matching degree between the user-scene demand and the commodity combination supply, and to decompose the prediction of click and purchase of two different levels of user intention. The inner product of the user-scene joint representation vector and the combination representation vector is calculated to output the combination representation vector, and the purchase conversion probability and click intention prediction value of each combination in the furniture candidate combination set are predicted. According to the preset ranking strategy (such as weighting and summing the two scores according to a certain weight, or prioritizing the purchase conversion probability), all combinations are reordered, and the combination is selected according to the rules (such as the top one or several combinations), which is recorded as the target recommended combination after sorting. Specifically, the user-scene joint representation vector (denoted as u) and the combination representation vector (denoted as i) are received, and the inner product of the two is calculated to obtain a basic matching score: m_score = u T ·i, which comprehensively reflects the overall suitability of the combination to meet the user's personalized scene demand. On the basis of the matching score, two independent lightweight fully connected layers are connected in parallel, with the matching score as the main input and possibly supplemented by shallow features, to output a click intention prediction value (between 0 and 1), which estimates the probability of the user producing a click behavior after seeing the combination recommendation.

[0054] Similarly, the matching score is used as the main input to output a purchase conversion probability prediction value (between 0 and 1), which estimates the probability of the user completing the purchase after clicking. The two tasks share the underlying fusion representation (u and i), but learn different decision boundaries, consistent with the progressive psychological process of users from attention to interest to decision-making action; Scenario setting: User A, whose user profile feature vector is a 256-dimensional vector V_user. In the user life scenario model of his living room, the extracted target room features are: geometric layout features (area 20 m2, aspect ratio 1.25, etc. Coding) and style features (modern minimalist, 0.9), which together constitute the vector V_room. The system generates a candidate combination of furniture, including a modern minimalist sofa and a coffee table, whose combination feature vector is V_combo.

[0055] User tower network: concatenate V_user and V_room, pass through 3 fully connected layers (dimensions are 512, 256, 128 respectively, using ReLU activation), and finally output a 128-dimensional user-scene joint representation vector u.

[0056] Item tower network: input V_combo, pass through 2 fully connected layers (dimensions are 256, 128 respectively, using ReLU activation), and output a 128-dimensional combination representation vector i.

[0057] Matching layer: Calculate inner product: m_score = sum(u i) (element-wise multiplication and summation).

[0058] Input m_score into a single-layer fully connected layer, and output the click intention prediction value = 0.85 through the Sigmoid function.

[0059] Input m_score into another single-layer fully connected layer, and output the purchase conversion probability prediction value = 0.45 through the Sigmoid function.

[0060] Therefore, the model predicts that user A has an 85% probability of clicking to view the living room combination recommendation, but only a 45% probability of ultimately purchasing it. This may be because the combination, although style-matched (high click rate), is beyond the user's budget or other factors affect the purchase decision. The system can use these two scores to fine-tune the ranking and adjust the marketing strategy.

[0061] The method for training by using the pre-designed training sample set comprises the following steps: constructing a training sample set, wherein each training sample includes a portrait feature vector of a historical user, a user life scene model feature extracted from the user's historical scene data, a feature vector of a furniture candidate combination that has been shown to the user in the past, and a real behavior label of the user on the combination; the real behavior label includes a click label for supervised click intention prediction and a purchase label for supervised purchase conversion prediction; inputting the user portrait feature vector and the target room feature in the user life scene model into the user tower network to obtain a user-scene joint representation vector, wherein the target room feature includes geometric layout features and style features; inputting the feature vector of the furniture candidate combination into the commodity tower network to obtain a combination representation vector; performing matching calculation on the user-scene joint representation vector and the combination representation vector through the matching layer to respectively output a click intention prediction value and a purchase conversion probability prediction value; calculating a click prediction loss according to the click intention prediction value and the click label, and calculating a conversion prediction loss according to the purchase conversion probability prediction value and the purchase label; both the click prediction loss and the conversion prediction loss use a binary cross-entropy loss function, the total loss is obtained by weighted sum of the two losses, and the parameters of the user tower network, the commodity tower network and the matching layer are optimized through the back propagation algorithm until the model converges. The judgment standard is that on an independent validation set, the total loss function value does not decrease significantly for continuous multiple training periods (epochs). Use Adam optimizer for training.

[0062] In view of the problems of hollow marketing copy content and lack of scene empathy. General copy cannot stimulate users' specific imagination of the effect of furniture in my home. Based on the Transformer structure, the encoder encodes the structured scene and product features, and the decoder generates the copy according to the preset marketing logic under the constraint mechanism. Automatically generate highly personalized and strongly scene-related marketing copy. The copy can specifically describe how the recommended furniture integrates into the user's actual room size and existing environment, and convert product selling points into life scene experience, greatly improving the relevance, persuasiveness and conversion potential of marketing content. The core design goal of the pre-trained copy generation model is to automatically convert structured user life scene model information and target recommended combination information into a personalized marketing copy with strong scene empathy and sales guiding force. The specific design is as follows: The copy generation model is designed based on the Transformer structure, and is trained using a pre-designed domain corpus, including: An encoder is configured to encode room layout and style features extracted from the user's living scene model, and furniture item features extracted from the target recommendation combination, into a sequence of context representations; wherein the room layout and style features include room size, existing furniture, wall color and lighting information, and the furniture item features include furniture name, style label and material characteristics; A decoder is configured to generate a marketing copy text according to a semantic structure of space depiction → combination relationship → lifestyle image → action guidance, conditioned on the sequence of context representations; wherein the decoding process is forced to generate a text containing key room size information extracted from the user's living scene model, and main function descriptions of furniture items in the target recommendation combination, by a constrained decoding mechanism. For example: When decoding to a semantic position related to size, the system is forced to preferentially select a word or phrase containing a specific numerical value (e.g., 5.2 meters) extracted from the room size feature from a candidate word list.

[0063] When decoding to a part describing the function of furniture, the system is forced to mention the main function description (e.g., a smart sofa with USB ports) extracted from the furniture item features.

[0064] The method of training using a pre-designed domain corpus includes: Collecting marketing copies corresponding to historical high-conversion furniture orders, description copies written by artificial for specific home scenes, and home design case texts to form a set of original marketing copy texts; For each copy in the set of original texts, extract and label its corresponding structured input features to form training paired data; the structured input features include room layout and style features parsed from the scene described in the copy, and furniture item features parsed from the goods described in the copy; During the training process, a cross-entropy loss function is used to measure the difference between the model output text and the original marketing copy text, and a gradient descent optimization algorithm is used to iteratively update the model parameters to minimize the loss function, thereby completing the model training and making the text sequence generated by the model as close as possible to the corresponding original copy text in the paired data.

[0065] The method of pushing the target recommendation combination and the personalized furniture marketing copy to the user terminal includes: Pack the target recommendation combination and the personalized furniture marketing copy into a marketing message supporting rich media display, and send it to the user terminal through multiple touch channels respectively; wherein the multiple touch channels include APP push notification, in-site message, public number message and smart TV pop-up window; Provide an AR preview entry for the target recommendation combination on the APP or web page, allowing users to superimpose virtual furniture models in real living room or bedroom environments based on the phone camera.

[0066] Embodiment two, please refer to Figure 2 As shown in the embodiment, some of the descriptions are not detailed in embodiment 1, and a user portrait-based AI customized furniture precise marketing method is provided, comprising: S1: Collect user basic attribute data, historical behavior data and explicit preference data from the user terminal, and fuse them into a user portrait feature vector; S2: Based on the user uploaded home space image, house type plan and space size input information, a user life scene model is constructed; S3: According to the user life scene model and the size parameters, style label and function label of the furniture goods in the furniture goods library, the space adaptation degree and style coordination degree of the furniture goods and the target room are calculated, and under the premise of meeting the space constraint condition, a furniture candidate combination set is generated; S4: The user portrait feature vector, user life scene model and furniture candidate combination set are input into a multi-task recommendation model to predict the purchase conversion probability and display click intention of each furniture candidate combination in the current marketing scene, and the sorted target recommendation combination is output; S5: Based on the user life scene model and the target recommendation combination, the space layout information, existing furniture information, style preference information and recommended goods information are input into a copy generation model to generate personalized furniture marketing copy strongly related to the actual life scene of the user; S6: The target recommendation combination and the personalized furniture marketing copy are pushed to the user terminal.

[0067] The above only describes the preferred embodiments of the present application and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacement for part of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0068] It should be noted that the formula in the specification is a dimensionless value calculated, the formula is obtained by software simulation of a large number of data to obtain the most real situation, and the preset parameters and threshold values in the formula are set by those skilled in the art according to the actual situation.

[0069] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. An AI-powered, customized furniture precision marketing system based on user profiles, characterized in that: include: The user profile collection and fusion module is used to collect basic user attribute data, historical behavior data, and explicit preference data from user terminals on the e-commerce platform server, and fuse them into user profile feature vectors. The user life scenario modeling module is used to build user life scenario models based on user-uploaded home space images, floor plans, and space size input information; The furniture candidate combination generation module is used to calculate the spatial adaptability and style coordination between furniture products and target rooms based on the user's living scenario model and the size parameters, style tags and function tags of furniture products in the furniture product library, and generate a set of furniture candidate combinations under the premise of meeting spatial constraints. The furniture recommendation module is used to input user profile feature vectors, user life scenario models and furniture candidate combination sets into the multi-task recommendation model, predict the purchase conversion probability and display click intent of each furniture candidate combination in the current marketing scenario, and output the ranked target recommendation combination; The scenario-driven marketing module is used to combine user life scenario models with target recommendations, inputting spatial layout information, existing furniture information, style preference information and recommended product information into a pre-trained copy generation model to generate personalized furniture marketing copy that is strongly related to the user's actual life scenario. The multi-channel outreach module is used to push targeted recommendation combinations and personalized furniture marketing copy to user terminals.

2. The AI-based customized furniture precision marketing system based on user profiles as described in claim 1, characterized in that, The method for fusing user profile feature vectors includes: Collect users' age group, family structure, geographical location, and housing type information as basic attribute data; Collect users' browsing history, search keywords, favorites and add-to-cart records, and furniture category and price distribution in historical orders on furniture pages as historical behavioral data; Based on data generated by users in style assessment questionnaires, liking or blocking products, we infer the strength of users' preferences for modern minimalism, Italian minimalism, Nordic and American styles, as explicit preference data. Basic attribute data, historical behavior data, and explicit preference data are fused together through numerical encoding and weighting to generate user profile feature vectors.

3. The AI-based customized furniture precision marketing system based on user profiles according to claim 2, characterized in that, The method for constructing user life scenario models includes: Based on user-uploaded images of home spaces, the system identifies walls, floors, doors and windows, outlines of existing furniture, and major decorative elements, and estimates the room's color tone and lighting direction to obtain analytical results. Based on the floor plan, the length and width dimensions, door and window positions, and structural column positions of each room are extracted to generate a geometric layout diagram of the room. The system checks the consistency between the input space dimensions and the room's geometric layout, and issues a correction prompt to the user when a conflict occurs. Based on the analysis results of the home space image, the room geometric layout diagram, and the verified spatial size input information, a user living scene model is constructed in the form of a graph structure. In the graph structure, the nodes represent rooms and objects, and the edges represent containment relationships, adjacency relationships, and spatial constraint relationships.

4. The AI-based customized furniture precision marketing system based on user profiles according to claim 3, characterized in that, The method for generating the furniture candidate combination set includes: Select furniture items from the furniture product library that match the function of the target room, and extract the size, placement direction constraints, and disassembly and combination capabilities of the selected furniture items; In the user's life scenario model, for each candidate furniture item, calculate the set of feasible placement locations in the target room. The set of feasible placement locations includes spatial constraints that satisfy the distance from walls, doors and windows is not less than a preset safety interval and do not block the main passageway. For modular furniture that meets spatial constraints, a combination search is performed in the set of feasible placement locations. Combinations whose total length exceeds the reserved wall length or whose combination depth exceeds the available passage width are excluded, forming an initial set of candidate furniture combinations that meet the physical space adaptation conditions. Based on the explicit preference data in the user profile feature vector and the decoration style features extracted from the analysis results, the style coordination degree of each initial furniture candidate combination is calculated, and combinations with a style coordination degree lower than the preset style coordination degree threshold are removed to obtain the furniture candidate combination set.

5. The AI-based customized furniture precision marketing system based on user profiles according to claim 4, characterized in that, The multi-task recommendation model is deployed on an e-commerce platform server. Based on a multi-tower neural network design, it is trained using a pre-designed training sample set. The multi-task recommendation model consists of a user tower network, a product tower network, and a matching layer. The user tower network consists of an input layer and a feature fusion layer. The input layer receives two parallel feature inputs, one of which is the user profile feature vector. The input is the target room features extracted from the target room in the user's life scenario model. The feature fusion layer includes several fully connected layers and is activated using the ReLU function. The product pyramid network consists of an input layer and a feature abstraction layer. The input layer receives feature representations of a set of furniture candidate combinations, and the feature abstraction layer consists of several fully connected layers activated by the ReLU function. The matching layer is used to calculate the inner product of the user-scene joint representation vector and the combination representation vector, and then output the combination representation vector. It also predicts the purchase conversion probability and click intent prediction value of each combination in the furniture candidate combination set. According to the preset sorting strategy, all combinations are re-sorted, and combinations are selected according to the rules, which are denoted as the sorted target recommended combination.

6. The AI-based customized furniture precision marketing system based on user profiles according to claim 5, characterized in that, The method of training using a pre-designed training sample set includes: Construct a training sample set, wherein each training sample includes a profile feature vector of a historical user, a user life scenario model feature extracted from the user's historical scenario data, a feature vector of a candidate furniture combination that has been shown to the user in the past, and the user's real behavior label for the combination; the real behavior label includes a click label for supervising click intent prediction and a purchase label for supervising purchase conversion prediction. The user profile feature vector and the target room feature in the user life scene model are input into the user tower network to obtain the user-scene joint representation vector, where the target room features include geometric layout features and style features; The feature vectors of the furniture candidate combinations are input into the commodity tower network to obtain the combination representation vector; The user-scenario joint representation vector and the combined representation vector are matched and calculated by the matching layer, and the click intent prediction value and the purchase conversion probability prediction value are output respectively. Click prediction loss is calculated based on click intent prediction and click tag, and conversion prediction loss is calculated based on purchase conversion probability prediction and purchase tag. Both click prediction loss and conversion prediction loss use the binary cross-entropy loss function. The two losses are weighted and summed to obtain the total loss. The parameters of the user tower network, product tower network and matching layer are optimized through backpropagation algorithm until the model converges.

7. The AI-based customized furniture precision marketing system based on user profiles according to claim 6, characterized in that, The copy generation model is based on the Transformer architecture and is trained using a pre-designed domain corpus, including: The encoder is used to encode the room layout and style features extracted from the user's life scene model, as well as the furniture product features extracted from the target recommendation combination, into a context representation sequence; wherein, the room layout and style features include room size, existing furniture, wall color and lighting information, and the furniture product features include furniture name, style tag and material features; The decoder is used to generate marketing copy text based on the semantic structure of spatial description → combination relationship → lifestyle imagery → action guidance, with contextual representation sequence as the condition. The decoding process forces the generated text to include key room size information extracted from the user's life scenario model, as well as the main functional descriptions of furniture products in the target recommended combination, through a constraint decoding mechanism.

8. The AI-based customized furniture precision marketing system based on user profiles according to claim 7, characterized in that, The method of training using a pre-designed domain corpus includes: We collect marketing copy corresponding to historical high-conversion furniture orders, descriptive copy written by humans for specific home scenarios, and home design case texts to form a collection of original marketing copy texts; For each piece of copy in the original text set, extract and label its corresponding structured input features to form training pairing data; the structured input features include room layout and style features parsed from the scene described in the copy, and furniture product features parsed from the product described in the copy. During training, the cross-entropy loss function is used to measure the difference between the model output text and the original marketing copy text, and the gradient descent optimization algorithm is used to iteratively update the model parameters to minimize the loss function, thereby completing the model training.

9. The AI-based customized furniture precision marketing system based on user profiles according to claim 8, characterized in that, The method for pushing targeted recommended combinations and personalized furniture marketing copy to user terminals includes: The targeted recommendation package and personalized furniture marketing copy are packaged into a marketing message that supports rich media display and sent to user terminals through multiple channels, including APP push notifications, in-app messages, official account messages and smart TV pop-ups. Provide an AR preview entry for the target recommended combination on the APP or web page, allowing users to overlay virtual furniture models in the real living room or bedroom environment using their mobile phone camera.

10. A method for AI-driven customized furniture precision marketing based on user profiles, used to implement the AI-driven customized furniture precision marketing system based on user profiles as described in any one of claims 1 to 9, characterized in that, include: S1: Collect user basic attribute data, historical behavior data, and explicit preference data from user terminals, and integrate them into user profile feature vectors; S2: Based on user-uploaded images of home spaces, floor plans, and space size input information, construct a user's living scenario model; S3: Based on the user's life scenario model and the size parameters, style tags, and function tags of furniture products in the furniture product library, calculate the spatial adaptability and style coordination between the furniture products and the target room, and generate a set of candidate furniture combinations under the premise of meeting the spatial constraints; S4: Input the user profile feature vector, user life scenario model and furniture candidate combination set into the multi-task recommendation model, predict the purchase conversion probability and display click intent of each furniture candidate combination in the current marketing scenario, and output the ranked target recommendation combination; S5: Based on the combination of user life scenario model and target recommendation, input spatial layout information, existing furniture information, style preference information and recommended product information into copywriting generation model to generate personalized furniture marketing copy that is strongly related to the user's actual life scenario; S6: Push the target recommended combination and the personalized furniture marketing copy to the user terminal.

Images