Integrated recommendation system capable of being regulated and controlled in real time
Technical Field
The invention relates to the technical field of recommendation systems, in particular to an integrated recommendation system capable of being regulated and controlled in real time.
Background
With the rapid development of the internet, people are facing massive information resources such as commodities, social users, etc., and thus personalized recommendation systems are growing and becoming increasingly important. For example, in a social scenario, how to match to interested and boring users may require consideration from multiple angles, such as feature information of users' double-shot, geographic location, user preference content, etc., which has led to increasing emphasis and development of research on recommendation technologies.
The current mainstream recommendation system service is mainly based on various algorithms, such as knowledge-based recommendation, content-based recommendation, algorithm model-based combined recommendation and the like. The traditional recommendation system is mostly a service framework based on a model, has low coupling with a portrait feature platform and a retrieval service, is overlong in recommendation link, and has poor integration of the data feature, the retrieval service and rules by the traditional platform although the recommendation model with good effect can obtain the content of interest of the user, so that great development cost is brought to the establishment of the enterprise recommendation system.
For example: chinese patent CN201711449141.1 discloses a method, a device, an intelligent terminal and a readable storage medium for pushing digital content, which utilize grading information graded based on selected characteristics of the digital content to be pushed, and perform rule ordering and filtering through a digital content screening strategy including a recommended quality intervention strategy, so that different grades of content are displayed for different users at a recommending end.
And the following steps: chinese patent CN201510184770.0 discloses a geographic location-based community service recommendation method, which improves recommendation efficiency and accuracy by using the geographic location-based recommendation method, but the method is only suitable for small-scale recommended service scenes, and in large-scale recommended scenes, personalized content cannot be provided for users well only by geographic location information.
In addition, in order to be compatible with content consumption requirements of users with different levels, most content recommendation platforms cannot set other rules such as grading thresholds for content release, and the existing recommendation systems often intervene through means such as reordering and cannot regulate and control strategies in real time.
In summary, most of the existing recommendation systems are single systems aiming at a certain direction or function and are expanded, the recommendation algorithm model, the rule intervention calculation strategy and the portrait platform feature depth cannot be combined, the regulation strategy cannot be performed in real time, and efficient and personalized physical examination cannot be brought to users in a large-scale recommendation scene.
Disclosure of Invention
In order to solve the problems, the invention provides an integrated recommendation system capable of being regulated and controlled in real time, so as to solve the problems that the existing recommendation system is single in performance and cannot be regulated and controlled in real time.
The invention adopts the following technical scheme:
an integrated recommendation system capable of being regulated and controlled in real time comprises an image feature acquisition module, a recommendation algorithm module, an index module, a real-time rule configuration module and a recommendation result return module;
The portrait characteristic acquisition module acquires portrait characteristics from a portrait characteristic platform FeatureServer and generates characteristic vectors;
The recommendation algorithm module recalls and/or sorts the feature vectors to calculate and obtain embedded vectors, and requests FAISSSERVER a retrieval platform to calculate a recommendation result TopK;
The real-time rule configuration module is used for acquiring dynamic ordering rules in real time, and carrying out rule ordering on the recommended result TopK according to the dynamic ordering rules so as to acquire a final personalized recommended result;
and the recommendation result returning module returns a final personalized recommendation result through the TF SERVING service system.
Further, the dynamic ordering rule is dynamically configured by an operator according to the service scene and the requirement.
Further, the real-time rule configuration module obtains the dynamic ordering rule in real time through EtcdUtil or KafkaUtil.
Further, the portrait characteristics platform FeatureServer includes Hbase and Redis components.
Further, the portrait features acquired by the portrait feature acquisition module comprise any one or more of offline features, real-time features and sequence features.
Further, the portrait characteristic acquisition module identifies the acquired portrait characteristic as a user portrait characteristic or an article portrait characteristic according to the application scene of the portrait characteristic.
Further, the recommendation algorithm model comprises a recall model and a sort model, the recall model and the sort model are respectively used for recalling and sorting the feature vectors, the recall model comprises any one of a LightGBM model, an I2I model or a U2I model, and the sort model is an FM sort model.
Further, the portrait characteristic acquisition module and the recommendation algorithm module communicate through GRPC communication protocol or Http service protocol.
Further, the TF SERVING service system uses GRPC communication protocol or Http service protocol for communication.
Further, the recommendation algorithm module requests FAISSSERVER the platform to calculate a recommendation result TopK. .
After the technical scheme is adopted, compared with the background technology, the invention has the following advantages:
1. The integrated recommendation system capable of being regulated and controlled in real time firstly extracts and calculates image features efficiently through the integrated image feature platform FeatureServer, then recalls and sorts the image features through the recommendation algorithm module, rapidly calculates similar users or articles to obtain a recommendation result TopK by using the massive vector retrieval recommendation service FAISSSERVER retrieval platform, and then regularly sorts the recommendation result TopK according to the dynamic sorting rule, thereby realizing real-time intervention regulation and control under different scenes to obtain a final personalized recommendation result.
2. The system integrates the portrait feature platform, the recommendation algorithm model and the rule intervention, combines the massive vector retrieval service with the real-time rule configuration module, realizes the configuration and real-time updating of rules, provides one-stop data service and recommendation service platform for enterprises, achieves the high integration of the recommendation platform module, avoids the problem of overlong service links of conventional recommendation systems, ensures that enterprises do not need to develop and iterate complex rule calculation functions, reduces the deployment and development cost of the enterprises, is not only suitable for quickly constructing recommendation system services for small-scale enterprises, but also can support large and medium-scale enterprise merchant user recommendation services, accurately positions potential clients and generates recommendations, and brings efficient and personalized recommendation services for users.
Drawings
FIG. 1 is a schematic diagram of a system according to the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Examples
As shown in FIG. 1, the integrated recommendation system capable of being regulated and controlled in real time comprises an image feature acquisition module, a recommendation algorithm module, an index module, a real-time rule configuration module and a recommendation result return module;
(1) The portrait characteristic acquisition module acquires portrait characteristics from a portrait characteristic platform FeatureServer and generates characteristic vectors;
Wherein the portrait characteristics platform FeatureServer includes Hbase and Redis components. The portrait features acquired by the portrait feature acquisition module comprise offline features, real-time features and sequence features.
The portrait characteristic acquisition module identifies the acquired portrait characteristic as a user portrait characteristic or an article portrait characteristic according to the application scene of the portrait characteristic. The item portrait refers to the content to be recommended in the application scene, for example, in a social recommendation system, the item refers to a recommended user, and the user portrait is the party requesting the recommendation; in the e-commerce recommendation scenario, the item refers to the item being recommended and the user refers to the buyer.
(2) The recommendation algorithm module recalls and sorts the feature vectors to calculate and obtain embedded vectors, and requests FAISSSERVER a retrieval platform to calculate a recommendation result TopK;
the recommendation algorithm model comprises a recall model and a sorting model, wherein the recall model and the sorting model are respectively used for recalling and sorting the feature vectors, the recall model comprises any one of a LightGBM model, an I2I model or a U2I model, and the sorting model is an FM sorting model.
The recall model and the order model are not necessary at the same time, one or a combination of the recall model and the order model can be selected according to the service scene, and it is noted that if the recall model and the order model are used at the same time, the recall model and the order model are calculated firstly and then the order model is calculated.
The recall and the sequencing are special words in a recommendation algorithm system, the recall is to roughly select candidate articles, and the sequencing is to further finely select articles obtained by recall. The recall algorithm model and the sorting algorithm model calculation method in the recommendation algorithm are not included in the invention, and are calculation methods used in the prior art.
FAISSSERVER the retrieval platform is a high-performance distributed mass vector similarity calculation service based on Faiss, and it can be understood that different business services in the embodiment of the invention construct different index services through FAISSSERVER, and the index services are obtained in real time through EtcdUtil or KafkaUtil.
(3) The real-time rule configuration module is used for acquiring dynamic ordering rules in real time, and carrying out rule ordering on the recommended result TopK according to the dynamic ordering rules so as to acquire a final personalized recommended result;
The real-time rule configuration module obtains the dynamic ordering rule in real time through EtcdUtil or KafkaUtil. In order to be compatible with content consumption requirements of users with different levels, most content recommendation platforms do not set classification thresholds and other rules for content release, and a strategy method for performing real-time intervention regulation is provided, in this embodiment, the dynamic ordering rules are obtained in real time through EtcdUtil to achieve real-time intervention regulation, specifically, the rules are dynamically configured by operators according to service requirements, and the dynamic ordering rules are dynamically configured by operators according to service scenes and requirements.
(4) The recommendation return module returns the final personalized recommendation recommend via TF SERVING service system.
And the portrait characteristic acquisition module and the recommendation algorithm module are communicated through a GRPC communication protocol. The TF SERVING service system communicates using a GRPC communications protocol. The GRPC communication protocol defines conventions for services and messages, and code can be generated from end-to-end, saving significant development time.
The system integrates the portrait feature platform, the recommendation algorithm model and the rule intervention, combines the massive vector retrieval service with the real-time rule configuration module, realizes the configuration and real-time updating of rules, provides one-stop data service and recommendation service platform for enterprises, achieves the high integration of the recommendation platform module, avoids the problem of overlong service links of conventional recommendation systems, and enables the enterprises not to need to develop and iterate complex rule calculation functions, thereby reducing the deployment and development cost of the enterprises, being applicable to the quick establishment of recommendation system services for small-scale enterprises, simultaneously being capable of supporting large and medium-scale enterprise merchant user recommendation services, accurately positioning potential clients and generating recommendations, and bringing high-efficiency and personalized recommendation services for users.
The present invention is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present invention are intended to be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.