CN-122022872-A - Knowledge-graph-based user behavior portrayal construction and marketing recommendation method

CN122022872ACN 122022872 ACN122022872 ACN 122022872ACN-122022872-A

Abstract

The invention discloses a knowledge graph-based user behavior portrayal construction and marketing recommendation method, and belongs to the field of data processing. The method comprises the steps of collecting multi-source original data of a user, constructing a space-time scene perception knowledge graph comprising a user entity, a commodity entity, a space entity and a scene entity, taking the user entity as a center to extract a micro behavior sequence and macro scene transfer characteristics, fusing and generating a user behavior characteristic sequence, outputting a dynamic interest vector of the user under a target space-time slice by utilizing a two-way long-short-term memory network and an attention mechanism, constructing a scene-driven dynamic user portrait, searching commodities matched with the target scene in the knowledge graph, calculating scene adaptation degree of the dynamic interest vector, and generating a marketing recommendation list. According to the invention, scene semantics are bound with user behavior depth, so that scene self-adaptive accurate recommendation is realized, and portrait dynamic sensing capability and recommendation conversion rate are remarkably improved.

Inventors

XU KUNPING
Deng Gangbiao

Assignees

广州普润惠泽网络科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260324

Claims (10)

1. A knowledge graph-based user behavior portrayal construction and marketing recommendation method is characterized by comprising the following steps: step 1, collecting multi-source original data of a target user in at least one historical time period; Step 2, performing entity extraction and relation extraction on the multi-source original data based on a preset entity type and a preset relation type to construct a space-time scene perception knowledge graph, wherein the entity type comprises a user entity, a commodity entity, a space entity and a scene entity, the space entity is a physical position with a geofence attribute which is identified according to global positioning system positioning data and Internet of things equipment signaling data, and the scene entity is a behavior situation label which is induced according to the space entity, user transaction data and the historical time period corresponding to behavior log data on a user line; Step 3, extracting multi-granularity user behavior characteristics from the space-time scene perception knowledge graph by taking the user entity as a center to generate a user behavior characteristic sequence fused with space-time scene semantics; Step 4, inputting the user behavior characteristic sequence into a two-way long-short-term memory network for processing, outputting a dynamic interest vector of a target user under any target space-time slice, and further constructing a dynamic user portrait based on scene driving; And 5, inputting the commodity entity to be recommended and the target space-time slice into the space-time scene perception knowledge graph again, combining the dynamic user portrait, calculating the scene adaptation degree of the commodity entity to be recommended and the dynamic user portrait, and generating a marketing recommendation list according to the scene adaptation degree.
2. The knowledge graph-based user behavior portrayal construction and marketing recommendation method according to claim 1 is characterized in that the multi-source raw data collected in the step 1 comprise user transaction data, user online behavior log data, internet of things equipment signaling data and global positioning system positioning data, wherein the user transaction data at least comprise transaction time, transaction amount and transaction commodity identification, the user online behavior log data at least comprise click stream data, page stay time and browsing commodity identification, the Internet of things equipment signaling data at least comprise equipment connection time and connected wireless access point identification, and the global positioning system positioning data at least comprise positioning time and longitude and latitude coordinates.
3. The knowledge graph-based user behavior portrayal construction and marketing recommendation method according to claim 1 is characterized in that in the space-time scene perception knowledge graph constructed in the step 2, the space entity is generated by clustering longitude and latitude coordinates in the global positioning system positioning data based on density to identify a geographic location area with a boundary, and meanwhile, combining coverage areas of the same wireless access point identifier in the Internet of things equipment signaling data to conduct boundary correction on the geographic location area to form the space entity with a first-stage space tag and a second-stage space longitude and latitude coordinate.
4. The knowledge graph-based user behavior portrait construction and marketing recommendation method according to claim 1 is characterized in that in the space-time scene perception knowledge graph constructed in the step 2, the scene entity is generated by performing combined coding on the space entity and the historical time period, and different behavior situation labels are generalized according to the fact that the historical time period belongs to a workday or a rest day and the difference of the historical time period belongs to a peak time period or a valley time period for the same space entity, and the behavior situation labels are defined as the scene entity.
5. The knowledge graph-based user behavior portrayal construction and marketing recommendation method according to claim 1, wherein in the spatiotemporal scene perception knowledge graph constructed in the step 2, the relationship type at least comprises a "purchase" relationship or a "browse" relationship between the user entity and the commodity entity, a "visit" relationship between the user entity and the space entity, an "adapt to" relationship between the commodity entity and the scene entity, and an "occurrence" relationship between the scene entity and the space entity, wherein time attributes are added to the "purchase" relationship, the "browse" relationship and the "visit" relationship, and the time attributes are used for recording time stamps or time periods of the occurrence of the relationships.
6. The knowledge-graph-based user behavior portrayal construction and marketing recommendation method according to claim 1, wherein the step 3 specifically comprises: Step 3.1, traversing the commodity entities of a first level which are directly connected with the user entity through a purchase relation or a browse relation, and sorting according to the time attribute added on the purchase relation or the browse relation to obtain microscopic behavior sequence characteristics; step 3.2, traversing the space entities directly connected with the user entities through a visit relation, associating the space entities with the corresponding scene entities, and constructing a transfer path of the user between the scene entities according to the time attribute added on the visit relation to obtain macroscopic scene transfer characteristics; And 3.3, aligning and splicing the microscopic behavior sequence features and the macroscopic scene transfer features on a time axis to form a multi-granularity user behavior feature sequence, wherein each behavior node in the user behavior feature sequence is associated with the corresponding commodity entity and the scene entity triggering the behavior.
7. The knowledge graph-based user behavior portrayal construction and marketing recommendation method according to claim 1, wherein in the step 4, the user behavior feature sequence is input into a bidirectional long-short-term memory network for processing, specifically, the commodity entity vector and the scene entity vector corresponding to each behavior node in the user behavior feature sequence are spliced, the commodity entity vector and the scene entity vector are used as the input of the bidirectional long-short-term memory network in each time step, the context information of the behavior sequence is captured through a forward layer and a backward layer of the bidirectional long-short-term memory network, and the outputs of the forward layer and the backward layer are spliced to generate a hidden state sequence fused with the context information.
8. The knowledge-graph-based user behavior portrayal construction and marketing recommendation method according to claim 7, wherein the outputting of the dynamic interest vector of the target user under any target space-time slice in the step 4 specifically comprises: step 4.1, taking a target scene entity corresponding to a preset target space-time slice as a query vector, calculating the attention score of each hidden state vector in the hidden state sequence and the query vector, and taking the attention score as an interest activation weight; Step 4.2, according to the interest activation weight, carrying out weighted summation on all hidden state vectors in the hidden state sequence to generate the dynamic interest vector of the target user under the target space-time slice; and 4.3, mapping the dynamic interest vector into probability distribution in a preset tag space through a full connection layer, selecting tags with probability exceeding a preset threshold as instant interest tags of the target user under the target space-time slice, and storing the instant interest tags in association with the target scene entity to form the scene-driving-based dynamic user portrait.
9. The knowledge-graph-based user behavior portrayal construction and marketing recommendation method according to claim 1, wherein the calculating in the step 5 of the scene adaptation degree between the commodity entity to be recommended and the dynamic user portrayal specifically comprises: Step 5.1, searching all commodity entities with the 'fit to' relation corresponding to the target scene entity in the space-time scene perception knowledge graph as candidate commodity sets; Step 5.2, respectively calculating cosine similarity between the embedded vector of each candidate commodity entity in the candidate commodity set and the dynamic interest vector generated in the step 4.2, and taking the cosine similarity as the scene adaptation degree of the candidate commodity entity; And 5.3, sorting candidate commodity entities in the candidate commodity set according to the sequence of the scene adaptation degree from high to low, selecting the candidate commodity entities with the front N positions to generate the marketing recommendation list, and pushing the marketing recommendation list to the user terminal of the target user.
10. The method for constructing and recommending user behavior portraits based on knowledge graph according to claim 9, wherein before pushing the marketing recommendation list to the user terminal of the target user in step 5.3, further comprising a scene trigger checking step of acquiring current global positioning system positioning data of the target user and current internet of things device signaling data in real time, identifying a real-time space entity where the target user is currently located, judging whether the real-time space entity is matched with a target scene entity corresponding to the target space-time slice, if the matching is successful, executing pushing operation, and if the matching is failed, suspending pushing and waiting for next scene trigger checking.

Description

Knowledge-graph-based user behavior portrayal construction and marketing recommendation method Technical Field The invention relates to the technical field of data processing, in particular to a knowledge-graph-based user behavior portrayal construction and marketing recommendation method. Background With the top-looking internet traffic bonus, online marketing has turned from rough traffic acquisition to refined stock user value mining. The accurate user behavior portrayal and personalized recommendation system is a core tool for achieving the aim. The existing recommendation technology is mainly divided into two types, namely recommendation based on collaborative filtering and recommendation based on a knowledge graph. Collaborative filtering-based recommendation methods rely on a user-item interaction matrix to make recommendations by computing similarities between users or items. However, such methods face severe data sparsity and cold start problems, and the recommendation is drastically reduced when new users or new items lack interactive records. In addition, it is difficult to mine the deep motivation behind the user's behavior, e.g., it knows that the user purchased "cell phones" and "treasures", but it cannot understand this because the user is going on business (scene) or just for standby (regular demand), resulting in serious homogenization of the recommendation results. In order to solve the above problems, knowledge graph technology is introduced in the industry. In the prior art, a knowledge graph is constructed by fusing static information such as attributes, categories and the like of articles, and path reasoning or feature learning is performed on the graph so as to enrich the recommended semantic information. For example, some methods generate recommendation candidates by constructing a user representation and performing a path search in combination with a knowledge graph. Still other methods are to fuse user features with item features by constructing social network graphs or item knowledge graphs to improve the accuracy of recommendation. However, after extensive analysis, it was found that the prior art still has significant drawbacks, first, behavior and scene fragility. The prior art mostly regards user behavior as isolated points, and even if time series are considered, the specific physical space (such as workplaces, home and tourist places) and social scenes (such as unions and parties) where the user behavior occurs cannot be quantified and fused into the user portrait. For example, a user frequently browses coffee shops in the workday noon (time) in an office area (space), which is quite different from the behavior intention of browsing coffee shops at home on weekends. The existing portrait cannot effectively distinguish interest drift caused by different 'space-time scenes'. Second, the statics and superficial layers of atlas semantics. The current knowledge graph is mostly built based on static data, and the entity relationship (such as 'belonging to' and 'similar') is single. When user interests change with a scene, static atlases cannot provide dynamic, scene-aware semantic support. This makes it difficult for the recommender system to understand the immediate, contextual needs of the user in a particular scenario, resulting in recommendations that, while relevant in item attributes, "disconcerting" in the context intent. Therefore, how to break the barriers between the user behavior data and the dynamic space-time scene and construct a user portrait and recommendation method capable of deeply sensing the scene and dynamically evolving is a technical problem to be solved currently. Disclosure of Invention Based on the above purpose, the invention provides a knowledge graph-based user behavior portrait construction and marketing recommendation method, which comprises the following steps: step 1, collecting multi-source original data of a target user in at least one historical time period; Step 2, performing entity extraction and relation extraction on the multi-source original data based on a preset entity type and a preset relation type to construct a space-time scene perception knowledge graph, wherein the entity type comprises a user entity, a commodity entity, a space entity and a scene entity, the space entity is a physical position with a geofence attribute which is identified according to global positioning system positioning data and Internet of things equipment signaling data, and the scene entity is a behavior situation label which is induced according to the space entity, user transaction data and the historical time period corresponding to behavior log data on a user line; Step 3, extracting multi-granularity user behavior characteristics from the space-time scene perception knowledge graph by taking the user entity as a center to generate a user behavior characteristic sequence fused with space-time scene semantics; Step 4, inputting the user behavior characteri