CN-121998735-A - Intelligent hosting method for whole flow of E-commerce operation

CN121998735ACN 121998735 ACN121998735 ACN 121998735ACN-121998735-A

Abstract

The invention belongs to the field of artificial intelligence, and particularly relates to an electronic commerce operation full-flow intelligent hosting method. The method comprises the steps of establishing a global unified user identity system, collecting and standardizing user behavior event streams from application programs, applets and webpage ends in real time, wherein the user behavior event streams comprise page browsing, commodity clicking, shopping cart adding, stay time and rolling depth, and performing time stamp alignment and intention semantic fusion on multi-channel user behavior events based on the global unified user identity system to construct a user intention evolution sequence. According to the invention, the problem of cross-channel user identity splitting is thoroughly solved by constructing a global unified user identity identification system, and seamless behavior collection under anonymous and real-name states is realized; according to the method, the intention semantic enhancement mechanism is introduced, so that the original behavior event is converted into the enhancement event rich in commodity knowledge and context semantics, and the depth and accuracy of user intention characterization are obviously improved.

Inventors

KONG MING
LI YAN
ZANG MENG
LIU WENHUI

Assignees

深圳市聚商鼎力网络技术有限公司

Dates

Publication Date: 20260508
Application Date: 20260116

Claims (10)

1. The full-flow intelligent hosting method for the E-commerce operation is characterized by comprising the following steps of: Establishing a global unified user identity system; Collecting and standardizing user behavior event streams from application programs, applets and webpage ends in real time, wherein the user behavior event streams comprise page browsing, commodity clicking, shopping cart adding, searching initiation, order submitting, stay time and rolling depth; Performing time stamp alignment and intention semantic fusion on the multi-channel user behavior event based on the global unified user identity system to construct a user intention evolution sequence; Dynamically modeling the user intention evolution sequence by using a deep time sequence neural network to generate a user intention state vector at the current moment; and calling a recommendation strategy engine to generate a commodity or content recommendation list which is highly matched with the current intention based on the user intention state vector, and issuing the commodity or content recommendation list in real time through a target channel.
2. The method for intelligent escrow of e-commerce operations of claim 1, wherein the establishing a global uniform user identity hierarchy comprises: When a user accesses any front-end channel for the first time, if the user is logged in, global user numbers distributed by a platform are used as main identifications, and if the user is not logged in, temporary anonymous identifications are generated through combination of equipment characteristic hash values, network environment fingerprints, geographic position coarse-granularity codes and first access time stamps; When the user completes login operation in any subsequent channel, the temporary anonymous identifier and the global user number are subjected to binding mapping, and all collected behavior events under the temporary anonymous identifier are combined back to the corresponding global user number name.
3. The method for intelligent escrow of e-commerce operation complete flow according to claim 2, wherein the real-time collecting and standardized processing of the user behavior event stream comprises: respectively deploying lightweight behavior embedded point proxy modules at application programs, applets and webpage ends, wherein the embedded point proxy modules monitor user interaction events; Carrying out structured packaging on all the events according to a predefined unified event model, wherein the unified event model comprises event types, event occurrence time, trigger channel identifiers, unique codes of target objects, interaction intensity quantization values and context environment parameters; And transmitting the packaged event to a central event processing center in real time through an encrypted message queue.
4. The method for intelligent escrow of e-commerce operations of claim 3, wherein the performing time stamp alignment and intent semantic fusion on the multi-channel user behavior event based on the global unified user identity system to construct the user intent evolution sequence comprises: grouping event streams according to user identification carried in the event; strictly sequencing all events under the same user identifier according to the occurrence time of the events to form an original behavior time sequence chain; Performing intent semantic enhancement operation on each event in an original behavior time sequence chain to generate an enhanced behavior event, wherein the intent semantic enhancement operation comprises the steps of inquiring a commodity knowledge graph according to a target object unique code, acquiring a commodity class level to which the commodity belongs, a core attribute tag, an associated scene tag and a historical conversion rate index, and injecting semantic information into the original event; a sequence of user intent evolutions consisting of enhanced behavioral events is output.
5. The method for intelligent escrow of e-commerce operations of claim 4, wherein the dynamically modeling the sequence of evolution of user intent with the deep timing neural network to generate the user intent state vector at the current time comprises: Inputting a user intention evolution sequence to an encoder formed by stacking a plurality of layers of gating circulating units; Each enhanced behavior event is expressed as a high-dimensional dense vector, and the high-dimensional dense vector is obtained by linear projection after event type embedding, commodity semantic embedding, interaction strength value and environment parameter embedding and splicing; the encoder outputs hidden states at each time step and weight aggregates the historical hidden states through an attention mechanism to generate a user intent state vector at the current time.
6. The method for intelligently hosting the electronic commerce operation whole flow according to claim 5 is characterized in that the event type embedding is obtained by searching a pre-trained event type word list, the commodity semantic embedding is obtained by carrying out layering embedding and then weighting average on a class level, a core attribute tag and an associated scene tag, the interaction strength value is normalized and then used as scalar input, and the environment parameter embedding is used for mapping the context environment parameter to a fixed dimension vector through a multi-layer perceptron.
7. The method of claim 6, wherein invoking a recommendation policy engine to generate a recommendation list based on a user intent state vector comprises: Maintaining a plurality of candidate recommendation strategy templates, each template corresponding to a typical user intent pattern, including exploratory, price-specific, repurchase and impact; The cosine similarity between the current user intention state vector and the center vector of each strategy template is calculated, and the strategy template with the highest similarity is selected as the current activation strategy; the method comprises the steps that an activation strategy calls a bottom commodity recall pool, and the recall pool is jointly constructed by three layers of mechanisms of collaborative filtering, vector neighbor searching and rule filtering; Finely arranging and scoring the recall result according to the user intention state vector, wherein the scoring function is a weighted linear combination, and the weight is dynamically adjusted by an online learning module according to real-time conversion feedback; And outputting the ordered commodity list as a recommendation result.
8. The method for intelligent escrow of e-commerce operations of claim 7, further comprising a recommendation effect closed-loop optimization mechanism: the follow-up actions of the user on the recommended result are continuously monitored, and the follow-up actions comprise clicking, purchasing, ordering and giving up operations; The follow-up actions are used as strengthening signals to update the fine-ranking weight parameters in the online learning module; And injecting the subsequent behaviors into a user intention evolution sequence, and triggering an encoder to correct the user intention state vector in real time.
9. The method for intelligent escrow of electronic commerce operation complete process according to claim 8, wherein the constructing and maintaining of the commodity knowledge graph comprises: Extracting the total commodity information from the commodity main data system at regular intervals; Natural language processing is carried out on commodity titles, descriptions and attribute fields so as to extract entities and relations; constructing a hierarchical ontology structure based on a class system, and defining father-son relationships, peer-level mutual exclusion relationships and cross-class association relationships among class nodes; associating historical behavior statistical characteristics of each commodity node, including exposure click rate, purchase conversion rate and return rate; and supporting real-time increment updating, and automatically triggering local sub-graph reconstruction when new commodities are put on shelf or the attributes are changed.
10. The method for intelligently hosting an e-commerce operation complete process according to claim 9, wherein the central event processing center adopts a distributed stream processing architecture, and the core components of the central event processing center comprise an event parser, an identity normalizer, a semantic enhancer and a time sequence buffer; The event parser is configured to check and de-serialize an original event; The identity normalizer is used for inquiring the identity mapping table and replacing the temporary identifier with a global user number; The semantic enhancer is used for calling a commodity knowledge graph interface to complete event semantic injection; the time sequence buffer zone maintains a behavior event queue in a sliding time window for each user, the window length is set to 72 hours, and events exceeding the window are automatically removed.

Description

Intelligent hosting method for whole flow of E-commerce operation Technical Field The invention belongs to the field of artificial intelligence, and particularly relates to an electronic commerce operation full-flow intelligent hosting method. Background In the prior art, the e-commerce platform is generally provided with respective user behavior acquisition and recommendation logics at independent application programs, applets and web page ends respectively, and a unified identity identification system and behavior sequence alignment mechanism is lacking among channels, so that key behavior information such as historical browsing, clicking, purchasing, searching and the like of a user cannot be effectively inherited and fused when the user switches different terminals, and a recommendation model is forced to reinitialize user portraits in isolated contexts, thereby seriously weakening the relevance and timeliness of recommendation results. The traditional scheme is characterized in that coarse-grained association is carried out by depending on device fingerprints or temporary session identifications, complex interaction scenes of users in a cross-device mode, a cross-application mode and a cross-platform mode are difficult to deal with, and dynamic modeling capacity based on a user intention evolution path is not established, so that a recommendation strategy is disjointed with the real demands of the users. Disclosure of Invention The invention provides an electronic commerce operation full-flow intelligent hosting method, which aims to solve the technical problem that a personalized recommendation system cannot continuously identify user intention and further causes conversion rate reduction caused by cross-channel user behavior data fragmentation. In order to overcome the defects, the invention constructs a set of end-to-end cross-channel user intention continuity modeling and intelligent hosting system. The system is based on a universal unique user identity verification mechanism, forms a unified user intention evolution map covering full contact, full time period and full behavior through standardized access, time sequence alignment and semantic enhancement of multi-source heterogeneous behavior events, drives generation and execution of a dynamic self-adaptive recommendation strategy on the basis, and realizes fundamental transition from data splitting to intention continuity and static portrait to dynamic evolution. The method comprises the following steps of firstly establishing a global unified user identity system, secondly collecting and standardizing user behavior event streams from application programs, applets and web pages in real time, thirdly carrying out time stamp alignment and intention semantic fusion on multi-channel behavior events based on the unified identity, constructing a user intention evolution sequence, then carrying out dynamic modeling on the intention evolution sequence by utilizing a deep time sequence neural network to generate a user intention state vector at the current moment, and finally calling a recommendation strategy engine to generate a commodity or content recommendation list which is highly matched with the current intention based on the user intention state vector and carrying out real-time issuing through a target channel. The universal unified user identity identification system comprises a platform, a global user number, a temporary anonymous identifier, a network environment fingerprint, a geographic position coarse granularity code and a first access timestamp, wherein the global user number is distributed by the platform and is used as a main identification when a user accesses any front-end channel for the first time, the temporary anonymous identifier is generated through the combination of a device characteristic hash value, a network environment fingerprint, a geographic position coarse granularity code and a first access timestamp if the user is not logged in, and when the user completes login operation in any subsequent channel, the system automatically binds and maps the temporary anonymous identifier with the global user number, and backtracks and merges all collected behavior events under the temporary identifier to the corresponding global user number name, so that non-inductive identity normalization is realized. The method comprises the steps of collecting and standardizing user behavior event streams in real time, wherein a lightweight behavior embedded point proxy module is deployed at an application program, an applet and a webpage end respectively, the embedded point proxy module monitors user interaction events, the user interaction events comprise page browsing, commodity clicking, shopping cart adding, searching initiation, order submitting, stay time and rolling depth, all events are packaged in a structured mode according to a predefined unified event model, the unified event model comprises event types, event occurrence time, trig