CN-116975625-B - Recommendation model training method and device and electronic equipment

CN116975625BCN 116975625 BCN116975625 BCN 116975625BCN-116975625-B

Abstract

The invention discloses a recommendation model training method which comprises the steps of forming collected information into triplet data reflecting relation among entities, attribute values and attribute relations, taking the generated feature vectors reflecting the interrelation of elements in the triplet data as training targets, and carrying out first-order training on a vector conversion model to obtain a first-order vector conversion model, wherein the feature vectors corresponding to the attribute values are expressed as box vectors, the box vectors are called attribute box vectors, the feature vectors corresponding to the entities are expressed as point vectors, and the point vectors are called entity point vectors. The recommendation model obtained through training in the processing process can realize that the mutual relation expression among elements in the triplet data is clearer and accords with the characteristics of different elements, and further, the recommendation model can obtain a third-order vector conversion model capable of accurately expressing the user interest feature vector through further training, and the model can accurately recommend to a user.

Inventors

XU ZEZHONG
Qu Kencen
CHEN HUAJUN
ZHANG WEN
DAI ZELIN
CHEN QIANG
GUO WEI
XIONG FEIYU

Assignees

浙江天猫技术有限公司

Dates

Publication Date: 20260508
Application Date: 20230517

Claims (10)

1. A recommendation model training method, comprising: The method comprises the steps of collecting information, forming three-tuple data reflecting the relation among entities, attribute values and attribute relations, wherein the collected information is automatically collected information of a recommendation model, the sources of the collected information at least comprise shopping application platform information and commodity information, the collected information is converted in a knowledge graph mode to obtain vectorization characteristic expression which can be identified by a vector conversion model, when the collected information is vectorized, different types of data in the collected information are expressed in different modes according to the information data types in the collected information, wherein a characteristic vector corresponding to the attribute values is expressed as a box vector, the box vector is called an attribute box vector, the characteristic vector corresponding to the entity is expressed as a point vector, the point vector is called an entity point vector, the entity at least comprises users or commodities in the collected information, and the attribute values are used for describing characteristics or attributes of the entity and can represent the characteristics of interests of a wide range of a target user; And taking the generated feature vector reflects the interrelation of each element in the triplet data as a training target, and performing first-order training on the vector conversion model to obtain a first-order vector conversion model serving as a recommendation model, wherein the recommendation model is used for recommending objects to target users.
2. The recommended model training method of claim 1, wherein the method further comprises: obtaining initial feature vectors corresponding to the elements by using the first-order vector conversion model; And training the first-order vector conversion model according to a training target to obtain a second-order vector conversion model serving as a recommendation model, wherein the attribute box vector is generated by each attribute value corresponding to the entity, the mapped attribute box vector is obtained by carrying out relation mapping operation according to the attribute relation between the entity and the attribute value, and the entity point vector falls into or approaches to an intersection area formed by the intersection of the mapped attribute box vectors of each relevant attribute value of the entity.
3. The recommended model training method of claim 2, wherein the method further comprises: obtaining behavior data of a target user; according to the behavior data of the target user, selecting part of entities as training entities and the rest of entities as verification entities from the entities related to the behavior of the corresponding target user; Acquiring a target intersection region formed by intersections of the mapped attribute box vectors, which are obtained by performing relation mapping operation on each related attribute value of the target training entity according to the attribute relation, by adopting the second-order vector conversion model; Performing first operation processing on each target intersection region of each target training entity of the target user to generate a first-order interest feature vector corresponding to the target user, wherein the first-order interest feature vector is expressed as a box vector; and training the second-order vector conversion model by taking the region of the box vector, in which the verification entity falls into or approaches to the expression of the first-order interest feature vector, as a training target, so as to obtain a third-order vector conversion model serving as a recommendation model.
4. The recommendation model training method according to claim 1, wherein the triplet data comprises a first element, a second element and a third element, wherein the second element is an attribute relationship element, the first element and the third element can be an entity element or an attribute value element respectively, and the formed triplet data type comprises any one or more of a first entity-attribute relationship-second entity triplet data, a first attribute value-attribute relationship-second attribute value triplet data and an entity-attribute relationship-attribute value triplet data.
5. The method for training a recommendation model according to claim 4, wherein the training target is a training target that reflects correlations between elements in the triplet data with the generated feature vector, and the first-order vector conversion model is obtained by performing first-order training on the vector conversion model, and the training target includes that a vector corresponding to a first element in triplet data obtained by the vector conversion model can be obtained by performing a transfer operation between vectors corresponding to the attribute relationships according to a type of the triplet data.
6. The method according to claim 4, wherein if the type of the triplet data is first entity-attribute relationship-second entity triplet data, the step of using the generated feature vector to reflect the interrelation of the elements in the triplet data as a training target is required to achieve a training target for the type of triplet data, wherein the generated point vector corresponding to the first entity can obtain the second entity through a relationship mapping operation of the feature vector corresponding to the attribute relationship, or approach the second entity and reach a preset threshold range.
7. The recommended model training method of claim 6, wherein for achieving the training objective, for each set of first entity-attribute relationship-second entity triplet data, the following processes are performed: Obtaining a first entity point vector corresponding to the first entity and a second entity point vector corresponding to the second entity in the first entity-attribute relationship-second entity triplet data; performing relation mapping processing on the first entity point vector and the feature vector corresponding to the attribute relation to obtain a mapped first entity point vector; Calculating the distance between the mapped first entity point vector and the mapped second entity point vector according to a first distance function; judging whether the distance accords with a training target according to the distance between the mapped first entity point vector and the mapped second entity point vector and a preset training standard; After the first entity-attribute relation-second entity triplet data of each group are processed, whether the data of each group meet the training target or not is used as a basis for adjusting the vector conversion model.
8. A recommendation model training device, comprising: A processing unit configured to form the gathered information into triplet data reflecting the relationship of the entity, the attribute value, the attribute relationship with each other; the method comprises the steps of collecting information in a recommendation model, wherein the collected information is information automatically collected by the recommendation model, the sources of the collected information at least comprise shopping application platform information and commodity information, the collected information is converted in a knowledge graph mode to obtain vectorization feature expression identifiable by a vector conversion model, when the collected information is vectorized, different types of data in the collected information are represented in different modes according to information data types in the collected information, a feature vector corresponding to an attribute value is expressed as a box vector, the box vector is called an attribute box vector, the feature vector corresponding to the entity is expressed as a point vector, the point vector is called an entity point vector, the entity at least comprises users or commodities in the collected information, the attribute value is used for describing features or attributes of the entity, and the feature value can represent wide interesting features of a target user; The construction unit is configured to perform first-order training on the vector conversion model by taking the generated feature vector reflecting the interrelation of each element in the triplet data as a training target, so as to obtain a first-order vector conversion model serving as a recommendation model.
9. An electronic device is characterized by comprising a processor and a memory, wherein, The memory is for storing one or more computer instructions, wherein the one or more computer instructions are executable by the processor to implement the method of any of claims 1-7.
10. A computer readable storage medium having stored thereon one or more computer instructions executable by a processor to implement the method of any of claims 1-7.

Description

Recommendation model training method and device and electronic equipment Technical Field The invention relates to the field of recommendation systems, in particular to a recommendation model training method, a recommendation model training device, electronic equipment and a computer storage medium. Background With the rapid development of internet technology, the number of network information resources increases exponentially, and when users face massive data information, the problem of information overload often occurs. In this case, the recommendation model, as an information filtering system, can present information of interest to the user in time, and thus is touted. However, as the network information resources and the number of users are continuously increased, the task complexity and the technical processing difficulty of the recommendation model are also increased, so that the recommendation model needs to be continuously trained and optimized. In the existing recommendation model training process, the recommendation algorithm is often relied on to optimize the recommendation model, or the model optimization is performed by focusing on different model use data, such as user historical preference data and recommendation object data. The recommended model obtained through the process generally lacks the capability of accurately describing the interests of the user, and has low working efficiency and poor recommending effect of model learning training results. Therefore, the recommended model training method with high accuracy and good recommendation effect for model learning training results is a technical problem to be solved urgently. Disclosure of Invention In view of the above-mentioned drawbacks of the prior art, embodiments of the present application provide a recommendation model training method and apparatus, an electronic device, and a computer storage medium, so as to solve the problem that the model learning training result of the existing recommendation model training method cannot accurately describe the user's interests. The application provides a recommendation model training method, which comprises the steps of forming collected information into triplet data reflecting relation among entities, attribute values and attribute relations, taking the generated feature vectors reflecting the relation among elements in the triplet data as training targets, and carrying out first-order training on a vector conversion model to obtain a first-order vector conversion model, wherein the feature vectors corresponding to the attribute values are expressed as box vectors, the box vectors are called attribute box vectors, and the feature vectors corresponding to the entities are expressed as point vectors, and the point vectors are called entity point vectors. The method comprises the steps of obtaining an initial feature vector corresponding to each element by using a first-order vector conversion model, training the first-order vector conversion model according to a training target to obtain a second-order vector conversion model, wherein the attribute box vector is generated by each attribute value corresponding to an entity, the relation mapping operation is carried out according to the attribute relation between the entity and the attribute value to obtain a mapped attribute box vector, and the entity point vector falls into or approaches an intersection area formed by the intersection of the mapped attribute box vectors of each related attribute value of the entity. The method comprises the steps of obtaining behavior data of a target user, selecting part of entities as training entities and the rest of entities as verification entities from all entities related to behaviors of the target user according to the behavior data of the target user, obtaining target intersection areas formed by intersections of mapped attribute box vectors obtained after relation mapping operation is carried out on all relevant attribute values of the target training entities according to attribute relations by adopting the second-order vector conversion model, carrying out first operation processing on all the target intersection areas of all the target training entities of the target user, generating primary interest feature vectors corresponding to the target user, expressing the primary interest feature vectors as box vectors, and training the second-order vector conversion model by taking the areas, in which the verification entities fall into or are close to the box vectors expressed by the primary interest feature vectors, as training targets to obtain a third-order vector conversion model. Optionally, the triplet data comprises a first element, a second element and a third element, wherein the second element is an attribute relation element, the first element and the third element can be entity elements or attribute value elements respectively, and the formed triplet data type comprises any one or more of a fi