E-commerce Scenario Cold Start Recommendation Method Combining Meta-Learning and Causal Inference

Abstract

The present invention relates to the technical field of e-commerce, specifically to a cold start recommendation method for e-commerce scenarios that combines meta-learning and causal inference. The method includes: obtaining user and product feature data and performing vectorized representation; analyzing the interactive causal relationship between user features and product features, constructing a causal graph, and quantifying the causal influence factor of user behavior on product recommendation; constructing a cold start recommendation model based on the meta-learning framework, and transforming the cold start problem into a cross-task transfer learning problem; training the meta-model to learn the general patterns between different tasks; using the influence factor obtained by causal inference as a constraint condition to optimize the objective function of the meta-learning recommendation model; designing a personalized recommendation strategy, and generating personalized cold start recommendation results by combining causal inference constraints. The method of the present invention comprehensively utilizes meta-learning and causal inference, effectively solves the cold start problem in e-commerce scenarios, and improves the personalization and interpretability of recommendations.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce, and specifically to a cold-start recommendation method for e-commerce scenarios that combines meta-learning and causal inference. Background Art

[0002] When facing the cold-start problem, existing e-commerce recommendation systems still have some limitations and deficiencies. Traditional recommendation methods mainly rely on users' historical interaction data. For new users or new products, there is a lack of sufficient interaction information, resulting in low recommendation quality and personalization. Although some studies have attempted to use auxiliary information of users and products to alleviate the cold-start problem, they often do not consider the internal relationship and influence mechanism between user characteristics and product characteristics, ignoring the causal relationship between them, making the interpretability and rationality of the recommendation results insufficient.

[0003] In addition, existing cross-domain recommendation methods aim to use the knowledge of the source domain to assist the recommendation task of the target domain. However, they usually assume that the user preferences and behavior patterns between different domains are consistent, ignoring the differences in user decision-making mechanisms under different scenarios, resulting in limited effects of transfer learning. At the same time, although existing recommendation methods can learn general knowledge from historical tasks and quickly adapt to new tasks, they mainly focus on the transfer ability of the model and do not fully utilize domain knowledge and causal relationships to guide the recommendation process, making the robustness to be improved.

[0004] Therefore, there is an urgent need for a cold-start recommendation method that can comprehensively utilize multi-source data, consider causal mechanisms, adapt to different scenarios, and learn robust representations to better solve the cold-start problem in e-commerce scenarios and provide more personalized and high-quality recommendation services.

[0005] In view of this, the present invention proposes a cold-start recommendation method for e-commerce scenarios that combines meta-learning and causal inference. Summary of the Invention

[0006] To achieve the above object, the present invention provides a cold-start recommendation method for e-commerce scenarios that combines meta-learning and causal inference. The specific technical solution is as follows: The cold-start recommendation method for e-commerce scenarios that combines meta-learning and causal inference includes:

[0007] Step 1: Obtain user feature data and product feature data, and vectorize the extracted user and product features;

[0008] Step 2: Analyze the interaction causal relationship between user feature data and product feature data, construct a causal graph in combination with domain knowledge, and use a causal inference method to quantify the causal influence factor of user behavior on product recommendation;

[0009] Step 3: Build a product recommendation model for cold start in the e-commerce sales scenario based on the meta-learning framework, and transform the cold start problem into a cross-task transfer learning problem;

[0010] Step 4: Train the meta-model based on historical task data so that the meta-model learns the general patterns between different tasks and adapts to the cold start tasks in the e-commerce sales scenario;

[0011] Step 5: Use the influencing factors obtained from causal inference as constraint conditions to optimize the objective function of the meta-learning recommendation model;

[0012] Step 6: Design personalized recommendation strategies for the specific causal relationships of different users or products, and generate personalized cold start recommendation results in combination with the constraint conditions of causal inference.

[0013] Preferably, obtain user feature data, including collecting the demographic features x d of users, b behavioral features x s and social features x

[0014] ; Obtain product feature data, including collecting the content features y c of products, t category features y s and statistical features y

[0015] Preferably, build a multi-layer perceptron MLP model, using user features and product features as inputs to learn the embedding representations e u and e i ;

[0016] Obtain the user embedding matrix and the product embedding matrix through the multi-layer perceptron MLP model, where d is the embedding dimension, |U| is the set of users, and |I| is the set of products; each user u and product i correspond to a d-dimensional embedding vector e u and e i respectively.

[0017] Preferably, use the PC algorithm based on the graph model for causal relationship discovery and causal graph construction; the steps of the PC algorithm are as follows:

[0018] Step a: Initialize a complete undirected graph G, where the node set V contains all user feature nodes V u and product feature nodes V i ;

[0019] Step b: For any two nodes v u , v iFor all \(v \in V\), determine whether there is an edge between them through conditional independence tests;

[0020] Step c: Assume the conditional set is If \(v\) u and \(v\) i are independent given \(S\), then delete the edge between \(v\) u and \(v\) i ;

[0021] Step d: Determine the threshold \(\alpha\) of the conditional independence test according to statistical knowledge;

[0022] Step e: Repeat steps b and c until the conditional independence tests for all node pairs are completed;

[0023] Step f: According to the V-structure and domain knowledge, orient some of the edges in the undirected graph \(G\) to obtain a partially directed acyclic graph;

[0024] Step g: Utilize the properties of the acyclic graph to orient the remaining undirected edges to obtain a fully directed acyclic graph, i.e., obtain the causal graph \(G\).

[0025] Preferably, adopt the causal effect estimation of the frontier set. For each node \(v_0\in V\) in the causal graph \(G\), calculate the causal effect of each node on all its children nodes;

[0026] Let the set of all children nodes of node \(v_0\) be \(Ch(v_0)\). For each child node \(v_0'\in Ch(v_0)\): Find all directed paths from \(v_0\) to \(v_0'\), denoted as \(P_1, P_2, \cdots, P\) K , where \(K\) is the number of paths;

[0027] For each path \(P\) k , \(k = 1, 2, \cdots, K\), calculate the causal strength \(CE(P\) k ); where \(w(v_0, v_0')\) represents the causal strength of the edge \((v_0, v_0')\);

[0028] Calculate the causal effect \(CAE(v_0, v_0')\) of \(v\) o on \(v_0'\):

[0029] Integrate the causal effects of all child nodes to obtain the total causal effect \(TCAE(v_0)\) of node \(v_0\):

[0030] Introduce a weight factor pair \(\alpha\) to weight the causal effect of user-item interaction to obtain the weighted causal effect \(CAE\) w (v u , v i );

[0031] CAE w (v u ,v i ) = α a CAE a (v u ,v i ) + α b CAE b (v u ,v i ) + α c CAE c (v u ,v i )

[0032] Among them, CAE a (v u ,v i ), CAE b (v u ,v i ), CAE c (v u ,v i ) respectively represent the causal effects in three interaction modes of click, purchase, and rating in the e-commerce transaction scenario, and α a , α b , and α c represent the influencing factors of click, purchase, and rating in the e-commerce transaction scenario;

[0033] Quantify the causal influence factor, and calculate the causal influence factor β(v u , v i ) of user feature v u on commodity feature v i based on the weighted causal effect:

[0034]

[0035] Integrate the causal influence factors β(v u , v i ) of all user features v u on all commodity features v i to generate a causal influence factor matrix V u is all user feature nodes, and V i is all commodity feature nodes.

[0036] Preferably, transform the cold start recommendation problem in the e-commerce scenario into a cross-task transfer learning problem; obtain T historical recommendation tasks Each task contains a support set S t and a query set Q t, where \(t\in[1,T]\);

[0037] Support set \(S\) t contains user-item interaction data, denoted as where denotes the feature vector of the \(l\)th z user-item pair, represents whether the user is interested in the item, and \(L\) Z is the size of the support set; Query set \(Q\) t contains the user-item pairs to be predicted, denoted as where \(L\) C is the size of the query set.

[0038] Preferably, adopt the model parameter meta-learning MAML framework to learn the initialization recommendation model parameters adapted to different tasks;

[0039] Use historical task data to train the meta-model, learn the general patterns between different tasks, and adapt to new cold-start tasks; adopt the following meta-training process:

[0040] Randomly initialize the parameters \(\theta\). For each historical task compute the gradient of the loss function on the support set \(S\) t ; use the gradient descent method to update the parameters \(\theta\) to obtain the task-specific parameters \(\theta'\): t

[0041] Compute the loss value of the task-specific model t on the query set \(Q\) ; update the initialization parameters \(\theta\), and repeat the meta-model training process until the initialization parameters \(\theta\) converge or reach the preset number of iterations.

[0042] Preferably, use the influence factor matrix obtained by causal inference as a constraint term to optimize the objective function of the meta-learning recommendation model;

[0043] Design personalized recommendation strategies for the specific causal relationships of different users or items; for user \(u\), according to the user feature vector \(e\) u and the causal influence factor matrix \(B\), compute the preference weight vector \(w\) ui of the user for different item features:

[0044]

[0045] where, is the \(u_0\)th element of the user feature vector \(e\) u , \(u_0\in[1,u]\), and \(C\) ui represents the interaction score between user \(u\) and item \(i\).

[0046] Preferably, combined with the constraint conditions of causal inference, personalized cold-start recommendation results are generated;

[0047] For the new user u', the user feature vector e u' and the product feature vector e i are used. Through the adapted recommendation model f θ' , predict the preference scores of the new user for all products:

[0048] According to the predicted scores and the personalized preference weight vector w u'i , calculate the product recommendation score r u'i for the final cold-start scenario:

[0049]

[0050] where ξ ∈ [0, 1] is a balance factor, is the v v -th element of the product feature vector y o ; Sort the products according to the cold-start recommendation score r u'i to generate a personalized cold-start product recommendation list for the new user u'.

[0051] An e-commerce scenario cold-start recommendation system that combines meta-learning and causal inference, which is used to implement the e-commerce scenario cold-start recommendation method that combines meta-learning and causal inference, includes: a data collection module, a causal analysis module, a meta-learning module, and a cold-start recommendation module;

[0052] The data collection module is used to obtain user feature data and product feature data, and vectorize the extracted user and product features;

[0053] The causal analysis module is used to analyze the interaction causal relationship between user feature data and product feature data, construct a causal graph in combination with domain knowledge, and use a causal inference method to quantify the causal influence factor of user behavior on product recommendation;

[0054] The meta-learning module constructs a product recommendation model for cold start in the e-commerce sales scenario based on the meta-learning framework, and transforms the cold-start problem into a cross-task transfer learning problem; trains the meta-model based on historical task data, enables the meta-model to learn the general patterns between different tasks, and adapts to the cold-start task in the e-commerce sales scenario;

[0055] The cold start recommendation module is used to use the influencing factors obtained by causal inference as constraint conditions to optimize the objective function of the meta-learning recommendation model; for the specific causal relationships of different users or products, design personalized recommendation strategies, and generate personalized cold start recommendation results in combination with the constraint conditions of causal inference.

[0056] An electronic device includes: a processor and a memory. Among them, a computer program that can be called by the processor is stored in the memory; the processor executes the e-commerce scenario cold start recommendation method combining meta-learning and causal inference by calling the computer program stored in the memory.

[0057] A computer-readable storage medium stores instructions. When the instructions run on a computer, the computer executes the e-commerce scenario cold start recommendation method combining meta-learning and causal inference.

[0058] The beneficial effects of the present invention: By obtaining and vectorizing the multi-dimensional features of users and products, the present invention mines the intrinsic semantic information, provides high-quality input for subsequent analysis and modeling, and improves the recommendation accuracy and personalization.

[0059] The present invention introduces causal inference to analyze user-product interactions, quantifies the impact of user behavior on recommendations, reveals the user decision-making logic, provides interpretability for recommendations, and enhances persuasion and trust.

[0060] The present invention uses meta-learning to construct a cold start recommendation model, transforms the cold start into cross-task transfer learning, learns general knowledge from historical tasks, quickly adapts to new scenarios, and alleviates the cold start problem.

[0061] The present invention trains a meta-model to learn the commonalities between tasks, enables the meta-model to flexibly adapt to new cold start tasks, reduces data requirements and training costs, and improves practicality and generalization ability.

[0062] The present invention introduces causal inference influencing factors into the meta-learning optimization objective, takes into account the causal constraints of user-product interactions, and generates recommendation results with stronger causal consistency and higher rationality.

[0063] The present invention designs a recommendation strategy according to the user's personalized causal relationship, generates personalized recommendations in combination with causal constraints, respects the user's causal preferences, and provides considerate and targeted recommendation services. Description of the Drawings

[0064] Figure 1 It is a flowchart of the e-commerce scenario cold start recommendation method combining meta-learning and causal inference provided by the present invention;

[0065] Figure 2 It is a structural diagram of the e-commerce scenario cold start recommendation system combining meta-learning and causal inference provided by the present invention. Detailed implementation manners

[0066] To better understand the present invention, more detailed descriptions will be made for various aspects of the present invention with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of exemplary implementation manners of the present invention, and do not limit the scope of the present invention in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0067] In the accompanying drawings, for the convenience of illustration, the sizes, dimensions and shapes of the elements have been slightly adjusted. The accompanying drawings are only examples and are not drawn strictly to scale. As used herein, terms such as "substantially", "about" and similar terms are used as terms indicating approximation, rather than terms indicating degree, and are intended to illustrate the inherent deviations in measured or calculated values that would be recognized by those of ordinary skill in the art. Additionally, in the present invention, the order of description of each step process does not necessarily represent the order in which these processes occur in actual operation, unless there are clear other limitations or can be deduced from the context.

[0068] It should also be understood that expressions such as "comprising", "including", "having", "containing" and / or "including" are open-ended rather than closed-ended expressions in this specification, which means that there are the stated features, elements and / or components, but do not exclude the existence of one or more other features, elements, components and / or their combinations. In addition, when an expression such as "at least one of..." appears after a list of listed features, it modifies the entire list of features, rather than just a single element in the list. In addition, when describing the implementation manners of the present invention, the use of "may" means "one or more implementation manners of the present invention". And the term "exemplary" is intended to refer to an example or illustration.

[0069] Unless otherwise defined, all terms used herein (including engineering terms and scientific and technical terms) have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention pertains. It should also be understood that, unless clearly stated in the present invention, words defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the related art, and should not be interpreted in an idealized or overly formal sense.

[0070] It should be noted that, without conflict, the implementation manners in the present invention and the features in the implementation manners can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the implementation manners.

[0071] Example 1

[0072] Refer to Figure 1, which is the first embodiment of the present invention, provides an e-commerce scenario cold start recommendation method that combines meta-learning and causal inference.

[0073] Step 1: Obtain user feature data and product feature data, and vectorize the extracted user and product features.

[0074] Obtain user feature data, including collecting the user's demographic features x d , behavioral features x b and social features x s , where the demographic features x d include gender, age, and occupation, and are represented as an n d dimensional vector The behavioral features x b include browsing history, purchase records, and rating data, and are represented as an n b dimensional vector The social features x s include friendship and interaction records, and are represented as an n s dimensional vector

[0075] Obtain product feature data, including collecting the content features y c , category features y t and statistical features y s , where the content features y c include title, description, and pictures, and are represented as an m c dimensional vector The category features y t include product classification and brand, and are represented as an m t dimensional vector The statistical features y s include sales volume and rating, and are represented as an m s dimensional vector

[0076] Construct a multi-layer perceptron MLP model, with user features and product features as inputs, and learn their embedding representations e u and e i ; Optimize the loss function where f MLP is the MLP network function, W is the network parameter, and λ is the regularization coefficient.

[0077] Obtain the user embedding matrix and the product embedding matrix where d is the embedding dimension; each user u and item i corresponds to a d-dimensional embedding vector e u and e i . The constructed embedding vectors contain the compressed semantic information of the original features and are used for subsequent analysis and modeling.

[0078] Step 1 effectively captures the internal features of users and items by obtaining multi-dimensional feature data of users and items and learning their low-dimensional semantic representations, laying a foundation for subsequent causal analysis and recommendation modeling.

[0079] Step 2: Analyze the interactive causal relationship between user feature data and item feature data, construct a causal graph by combining domain knowledge, and use causal inference methods to quantify the causal impact factor of user behavior on item recommendation.

[0080] Use the PC algorithm based on the graph model to discover causal relationships and construct a causal graph; the basic steps of the PC algorithm are as follows:

[0081] Step a: Initialize a complete undirected graph G, where the node set V contains all user feature nodes V u and item feature nodes V i ;

[0082] Step b: For any two nodes v u , v i ∈V, determine whether there is an edge between them through conditional independence testing;

[0083] Step c: Assume the conditioning set is If v u and v i are independent given S, then delete the edge between v u and v i ;

[0084] Step d: Determine the threshold α of the conditional independence test according to statistical knowledge; usually take 0.05 or 0.01;

[0085] Step e: Repeat steps b and c until the conditional independence tests for all node pairs are completed;

[0086] Step f: According to the V-structure and domain knowledge, orient some of the edges in the undirected graph G to obtain a partially directed acyclic graph;

[0087] Step g: Use the properties of the acyclic graph to orient the remaining undirected edges to obtain a fully directed acyclic graph, that is, obtain the causal graph G.

[0088] Adopt the causal effect estimation of the frontier set. For each node v0 ∈ V in the causal graph G, calculate the causal effect of each node on all its children nodes.

[0089] Let the set of all child nodes of node v0 be Ch(v0). For each child node v0' ∈ Ch(v0): Find all directed paths from v0 to v0', denoted as P1, P2,..., PK, K , where K is the number of paths.

[0090] For each path Pk, k = 1, 2,..., K, calculate the causal strength CE(Pk): k , where w(v0, v0') represents the causal strength of the edge (v0, v0'), estimated by a data-driven method. k ): where w(v0, v0') represents the causal strength of the edge (v0, v0'), estimated by a data-driven method.

[0091] Calculate the causal effect CAE(v0, v0') of v0 on v0': o

[0092] Integrate the causal effects of all child nodes to obtain the total causal effect TCAE(v0) of node v0:

[0093] Introduce a weight factor α to weight the causal effect of user-item interaction to obtain the weighted causal effect CAE(v0, v0'): w (v0, v0'): u , v0'): i ):

[0094] CAE(v0, v0') = α CAE(v0, v0') + α CAE(v0, v0') + α CAE(v0, v0') w (v0, v0') u , v0') i ) = α a CAE a (v0, v0') u , v0') i ) + α b CAE b (v0, v0') u , v0') i ) + α c CAE c (v0, v0') u , v0') i )

[0095] where CAE(v0, v0'), CAE(v0, v0'), CAE(v0, v0') respectively represent the causal effects in the three interaction modes of click, purchase, and rating in the e-commerce transaction scenario, α a (v0, v0') u , v0') i ), CAE b (v0, v0') u , v0') i ), CAE c (v0, v0') u , v0') i ) respectively represent the causal effects in the three interaction modes of click, purchase, and rating in the e-commerce transaction scenario, α a 、α​b and α c represent the influencing factors of clicks, purchases, and ratings in the e-commerce transaction scenario.

[0096] Quantify the causal influence factors and calculate the user feature v based on the weighted causal effect u for the item feature v i of the causal influence factor β(v u , v i ):

[0097]

[0098] Integrate all the causal influence factors β(v u of all user features v i for all item features v u , v i ) to generate a causal influence factor matrix for subsequent optimization of the recommendation model, where V u is the set of all user feature nodes and V i is the set of all item feature nodes.

[0099] Step 2 uses causal inference methods to analyze the interactive causal relationship between user features and item features, quantifies the influencing factors of user behavior on item recommendations, reveals the internal causal mechanism of user-item interactions, and provides an interpretable basis for personalized recommendations.

[0100] Step 3: Build a cold-start item recommendation model in the e-commerce sales scenario based on the meta-learning framework and transform the cold-start problem into a cross-task transfer learning problem.

[0101] Transform the cold-start recommendation problem in the e-commerce scenario into a cross-task transfer learning problem; obtain T historical recommendation tasks Each task contains a support set S t and a query set Q t , where t ∈ [1, T].

[0102] The support set S t contains user-item interaction data, denoted as where represents the feature vector of the l-th user-item pair, z and indicates whether the user is interested in the item, and L Z is the size of the support set; the query set Q t contains the user-item pairs to be predicted, denoted as where L C is the size of the query set.

[0103] Step 3 constructs a cold start recommendation model based on the meta - learning framework, transforms the cold start problem into a cross - task transfer learning problem, enables the model to quickly adapt to new e - commerce scenarios, effectively alleviates the cold start problem, and improves the generalization ability of the recommendation system.

[0104] Step 4: Train the meta - model based on historical task data, enabling the meta - model to learn the general patterns between different tasks and adapt to the cold start tasks in the e - commerce sales scenario.

[0105] Adopt the model - parameter meta - learning MAML framework to learn the initial recommendation model parameters suitable for different tasks; the goal of the parameter meta - learning MAML framework is to find the initial parameters.

[0106] Use historical task data Train the meta - model to learn the general patterns between different tasks and adapt to new cold start tasks; adopt the following meta - training process:

[0107] Randomly initialize the parameter θ. For each historical task On the support set S t Calculate the gradient of the loss function:

[0108]

[0109] where, f θ is the recommendation model with parameter θ, is the cross - entropy loss function, and use the gradient descent method to update the parameter θ to obtain the task - specific parameter θ t ':

[0110] where, α is the learning rate; calculate the loss value of the task - specific model t on the query set Q :

[0111]

[0112] Update the initial parameter θ, and repeat the meta - model training process until the initial parameter θ converges or reaches the preset number of iterations.

[0113] By training the meta - model on historical task data in Step 4, learning the general patterns between different tasks, the meta - model can quickly adapt to new cold start tasks, greatly reducing the training cost in new scenarios and improving the flexibility and practicality of the recommendation system.

[0114] Step 5: Use the influence factors obtained from causal inference as constraint conditions to optimize the objective function of the meta - learning recommendation model.

[0115] The influence factor matrix obtained from causal inference As a constraint item, optimize the objective function ψ'(f θ ):

[0116]

[0117] where ψ(f θ ) is the objective function of the recommendation model with parameter θ, μ is the balance factor, and θ ui is the feature parameter connecting the u-th user feature and the i-th item in the recommendation model, and |·|2 represents the L2 norm.

[0118] Design personalized recommendation strategies for specific causal relationships of different users or items; for user u, calculate the preference weight vector w u of the user for different item features according to the user feature vector e ui and the causal influence factor matrix B:

[0119]

[0120] where is the u0-th element of the user feature vector e u , u0 ∈ [1, u], and C ui represents the interaction score between user u and item i.

[0121] Step 5 takes the influence factors obtained by causal inference as constraint conditions to optimize the objective function of the meta-learning recommendation model, enabling the model to consider the causal relationship of user-item interaction while learning the general recommendation pattern, and improving the interpretability and rationality of the recommendation results.

[0122] Step 6: Design personalized recommendation strategies for specific causal relationships of different users or items, and generate personalized cold-start recommendation results in combination with the constraint conditions of causal inference.

[0123] For the new user u', use the user feature vector e u' and the item feature vector e i , and through the adapted recommendation model f θ' , predict the preference scores of the new user for all items:

[0124] According to the predicted scores and the personalized preference weight vector w u'i , calculate the final item recommendation score r u'i in the cold-start scenario:

[0125]

[0126] where ξ ∈ [0, 1] is the balance factor, For the commodity feature vector y v is the v o th element; according to the cold start recommendation score r u'i sort the commodities to generate a personalized cold start commodity recommendation list for the new user u'.

[0127] Step 6 designs personalized recommendation strategies for different users, generates personalized cold start recommendation lists by combining the constraints of causal inference, fully considers the personalized preferences and causal relationships of users, and improves the accuracy and satisfaction of recommendation results in cold start scenarios.

[0128] Embodiment 2

[0129] Referring to Figure 2 , the second embodiment of the present invention provides a cold start recommendation system for e-commerce scenarios that combines meta-learning and causal inference.

[0130] The system includes: a data acquisition module, a causal analysis module, a meta-learning module, and a cold start recommendation module.

[0131] The data acquisition module is used to obtain user feature data and commodity feature data, and vectorize the extracted user and commodity features.

[0132] The causal analysis module is used to analyze the interaction causal relationship between user feature data and commodity feature data, construct a causal graph by combining domain knowledge, and adopt a causal inference method to quantify the causal influence factor of user behavior on commodity recommendation.

[0133] The meta-learning module constructs a cold start commodity recommendation model for e-commerce sales scenarios based on the meta-learning framework, and transforms the cold start problem into a cross-task transfer learning problem; trains the meta-model based on historical task data to enable the meta-model to learn the general patterns between different tasks and adapt to the cold start task in e-commerce sales scenarios.

[0134] The cold start recommendation module is used to use the influence factor obtained by causal inference as a constraint condition to optimize the objective function of the meta-learning recommendation model; design personalized recommendation strategies for the specific causal relationships of different users or commodities, and generate personalized cold start recommendation results by combining the constraint conditions of causal inference.

[0135] Embodiment 3

[0136] The present invention also provides an electronic device. The electronic device may include one or more processors and one or more memories. Among them, computer-readable code is stored in the memory, and when the computer-readable code is run by one or more processors, it can execute the above-mentioned cold start recommendation method for e-commerce scenarios that combines meta-learning and causal inference.

[0137] The method or system according to an embodiment of the present invention can also be implemented by means of the architecture of the electronic device of the present invention.

[0138] The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc.

[0139] The storage device in the electronic device, such as ROM or a hard disk, can store the e-commerce scenario cold start recommendation method provided by the present invention, which combines meta-learning and causal inference.

[0140] The e-commerce scenario cold start recommendation method that combines meta-learning and causal inference includes: obtaining user feature data and product feature data, and vectorizing the extracted user and product features; analyzing the interaction causal relationship between the user feature data and the product feature data, constructing a causal graph in combination with domain knowledge, and using a causal inference method to quantify the causal influence factor of user behavior on product recommendation; constructing a cold start product recommendation model for the e-commerce sales scenario based on the meta-learning framework, and transforming the cold start problem into a cross-task transfer learning problem; training a meta-model based on historical task data to enable the meta-model to learn the general patterns between different tasks and adapt to the cold start task in the e-commerce sales scenario; using the influence factor obtained by causal inference as a constraint condition to optimize the objective function of the meta-learning recommendation model; designing personalized recommendation strategies for the specific causal relationships of different users or products, and generating personalized cold start recommendation results in combination with the constraint conditions of causal inference.

[0141] Furthermore, the electronic device may further include a user interface. Of course, the architecture of the present invention is only exemplary. When implementing different devices, one or more components of the electronic device disclosed in the present invention may be omitted according to actual needs.

[0142] Example 4

[0143] The present invention also discloses a computer-readable storage medium.

[0144] Computer-readable instructions are stored on the computer-readable storage medium.

[0145] When the computer-readable instructions are run by a processor, the e-commerce scenario cold start recommendation method according to the embodiment of the present invention described with reference to the above drawings can be executed.

[0146] The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory, etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. Additionally, according to an embodiment of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program.

[0147] For example, the present invention provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be run by a processor to execute instructions corresponding to the method steps provided by the present invention, such as: obtaining user feature data and commodity feature data, and vectorizing the extracted user and commodity features; analyzing the interaction causal relationship between the user feature data and the commodity feature data, constructing a causal graph by combining domain knowledge, and using a causal inference method to quantify the causal impact factor of user behavior on commodity recommendation; constructing a cold-start commodity recommendation model in an e-commerce sales scenario based on a meta-learning framework, and transforming the cold-start problem into a cross-task transfer learning problem; training a meta-model based on historical task data to enable the meta-model to learn the general patterns between different tasks and adapt to the cold-start task in the e-commerce sales scenario; using the impact factor obtained by causal inference as a constraint condition to optimize the objective function of the meta-learning recommendation model; designing personalized recommendation strategies for the specific causal relationships of different users or commodities, and generating personalized cold-start recommendation results by combining the constraint conditions of causal inference.

[0148] When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present invention are executed. The method and apparatus, device of the present invention can be implemented in many ways. For example, the method and apparatus, device of the present invention can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware.

[0149] The above order of steps for the method is only for illustration, and the steps of the method of the present invention are not limited to the above specifically described order, unless otherwise specifically stated.

[0150] In addition, in some embodiments, the present invention can also be implemented as a program recorded in a recording medium, and these programs include machine-readable instructions for implementing the method according to the present invention. Thus, the present invention also covers a recording medium storing a program for executing the method according to the present invention.

[0151] Furthermore, in the above technical solutions provided in the embodiments of the present invention, the parts that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive elaboration.

[0152] The specific embodiments described above further elaborate in detail the objectives, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only the specific embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A cold start recommendation method for e-commerce scenarios combining meta-learning and causal inference, characterized in that: include: Step 1: Obtain user feature data and product feature data, and quantize the extracted user and product features; Step 2: Analyze the interactive causal relationship between user feature data and product feature data, build a causal graph based on domain knowledge, and use causal inference methods to quantify the causal influence factors of user behavior on product recommendations; Step 3: Based on the meta-learning framework, a cold-start product recommendation model is constructed in the e-commerce sales scenario, and the cold-start problem is transformed into a cross-task transfer learning problem; Step 4: Train the metamodel based on historical task data so that the metamodel can learn common patterns between different tasks and adapt to cold start tasks in e-commerce sales scenarios; Step 5: Use the influencing factors obtained by causal inference as constraints to optimize the objective function of the meta-learning recommendation model; Step 6: Design personalized recommendation strategies for specific causal relationships between different users or products, and generate personalized cold start recommendation results based on the constraints of causal inference; The impact factor matrix obtained by causal inference As a constraint, V u For all user feature nodes, V i Optimize the objective function of the meta-learning recommendation model for all product feature nodes; Design personalized recommendation strategies based on the specific causal relationships of different users or products; For user u, according to user u feature vector e u and the causal influence factor matrix B, calculate the user's preference weight vector w for different product features ui : in, is the user feature vector e u The u0th element of C, u0∈[1,u], ui represents the interaction score between user u and product i; Combined with the constraints of causal inference, personalized cold start recommendation results are generated; For new user u', use the user feature vector e u' and product feature vector e i , through the adapted recommendation model f θ' , predict the preference scores of new users for all products: Based on the prediction score and personalized preference weight vector w u'i , calculate the final cold start scenario product recommendation score r u'i : Among them, ξ∈[0,1] is the balance factor, is the product feature vector y v No.v o elements; based on the cold start recommendation score r u'i Sort the products and generate a personalized cold-start product recommendation list for new user u'.

2. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 1 is characterized in that: Obtain user characteristic data, including collecting user demographic characteristics x d , behavioral characteristics x b and social features x s ; Obtain product feature data, including collecting product content features c , category feature y t and the statistical characteristics y s .

3. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 2 is characterized in that: Construct a multi-layer perceptron MLP model based on user characteristics and product features As input, learn the embedding representation e u and e i ; The user embedding matrix is ​​obtained through the multi-layer perceptron MLP model and the product embedding matrix Where d is the embedding dimension, |U| is the user set, and |I| is the product set; each user u and product i corresponds to a d-dimensional embedding vector e u and e i .

4. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 3 is characterized in that: The PC algorithm based on the graph model is used to discover causal relationships and construct causal graphs; the steps of the PC algorithm are as follows: Step a: Initialize a completely undirected graph G, where the node set V contains all user feature nodes V u and product feature node V i ; Step b: For any two nodes v u ,v i ∈V, determine whether they have edges through conditional independence test; Step c: Assume that the condition set is If v u and v i Given S, if they are independent, then delete v u and v i The edge between Step d: Determine the threshold α of the conditional independence test based on statistical knowledge; Step e: Repeat steps b and c until the conditional independence test of all node pairs is completed; Step f: According to the V-structure and domain knowledge, direct some edges in the undirected graph G to obtain a partial directed acyclic graph; Step g: Using the properties of acyclic graphs, direct the remaining undirected edges to obtain a completely directed acyclic graph, that is, the causal graph G.

5. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 4 is characterized in that: Using the causal effect estimation of the frontier set, for each node v0∈V in the causal graph G, calculate the causal effect of each node on all child nodes; Let the set of all child nodes of node v0 be Ch(v0), for each child node v0'∈Ch(v0): find all directed paths from v0 to v0', denoted by P1, P2, ..., P K , where K is the number of paths; For each path P k ,k=1,2,...,K,calculate the causal strength CE(P k ): Among them, w(v0,v0') represents the causal strength of the edge (v0,v0'); Calculate v o Causal effect on v0' CAE(v0,v0'): Combining the causal effects of all child nodes, we get the total causal effect TCAE(v0) of node v0: The weight factor is introduced to weight the causal effect of α on the user-product interaction, and the weighted causal effect CAE is obtained. w (v u ,v i ): CAE w (v u ,v i )=α a CAE a (v u ,v i )+α b CAE b (v u ,v i )+α c CAE c (v u ,v i ) Among them, CAE a (v u ,v i ),CAE b (v u ,v i ),CAE c (v u ,v i ) represent the causal effects of the three interactive modes of click, purchase and rating in e-commerce transaction scenarios, α a , α b and α c Indicates the factors affecting clicks, purchases, and ratings in e-commerce transaction scenarios; Quantify the causal influence factor and calculate the user feature v based on the weighted causal effect u For product features v i The causal influence factor β(v u ,v i ): All user features v u For all product features v i The causal influence factor β(v u ,v i ), integrate and generate the causal influence factor matrix 6. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 5 is characterized in that: Convert the cold start recommendation problem in the e-commerce scenario into a cross-task transfer learning problem; obtain T historical recommendation tasks Each task Contains a support set S t and a query set Q t , t∈[1,T]; Support Set S t Contains user-item interaction data, represented as in Indicates the first z feature vector of a user-item pair, Indicates whether the user is interested in the product, L Z is the size of the support set; the query set Q t Contains the user-item pair that needs to be predicted, expressed as Where L C is the size of the query set.

7. The e-commerce scenario cold start recommendation method combining meta-learning and causal inference according to claim 6 is characterized in that: Adopt the model parameter meta-learning MAML framework to learn the initialization recommendation model parameters that are suitable for different tasks; Using historical task data Train the meta-model to learn common patterns between different tasks and adapt to new cold start tasks; the following meta-training process is used: Randomly initialize the parameters θ, for each historical task In support set S t Calculate the gradient of the loss function on Use the gradient descent method to update the parameter θ to obtain the task-specific parameter θ t ': In the query set Q t Compute task-specific models The loss value of Update the initialization parameter θ and repeat the meta-model training process until the initialization parameter θ converges or reaches the preset number of iterations.

8. An e-commerce scenario cold start recommendation system combining meta-learning and causal inference, which is used to implement the e-commerce scenario cold start recommendation method combining meta-learning and causal inference as described in any one of claims 1 to 7, characterized in that: include: Data collection module, causal analysis module, meta-learning module and cold start recommendation module; The data collection module is used to obtain user feature data and product feature data, and quantize the extracted user and product features; The causal analysis module is used to analyze the interactive causal relationship between user feature data and product feature data, construct a causal graph based on domain knowledge, and use causal inference methods to quantify the causal influence factors of user behavior on product recommendations; The meta-learning module builds a cold-start product recommendation model in an e-commerce sales scenario based on a meta-learning framework, and transforms the cold-start problem into a cross-task transfer learning problem; Train the metamodel based on historical task data to enable the metamodel to learn common patterns between different tasks and adapt to cold start tasks in e-commerce sales scenarios; The cold start recommendation module is used to optimize the objective function of the meta-learning recommendation model by taking the influencing factors obtained by causal inference as constraints; design personalized recommendation strategies for the specific causal relationships of different users or commodities, and generate personalized cold start recommendation results in combination with the constraints of causal inference.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the e-commerce scenario cold start recommendation method combining meta-learning and causal inference as described in any one of claims 1 to 7 by calling the computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: Instructions are stored, and when the instructions are executed on a computer, the computer executes the e-commerce scenario cold start recommendation method combining meta-learning and causal inference as described in any one of claims 1 to 7.

Images