Feature engineering is used to extract user and item features. It can also be used to create specific algorithm features. Generally, feature engineering works as the input conditions for some algorithms.
A filtering rule is used to generate a recommendation filtering set, including a blacklist, a whitelist, and an attribute filtering list. A filtering rule enables users to filter items in the process of online inference.
A nearline strategy calibrates models or recommendation candidate sets that are generated offline or nearline, based on user data, item data, and user behaviors. The nearline flow is designed to guarantee instant data acquisition for better recommendation. It enables RES to quickly capture user behavior changes and update the recommendation results.
An online flow is used in the data split scenarios of A/B tests. Different online flow is tailored for different testing scenarios. The traffic can be distributed to different flows for A/B tests based on the traffic proportion.
An online service defines how analysis results are applied. In general, it is the inference service provided after an application is deployed online and offers APIs for external systems. RES has three online services, which are recommendation engine, text segmentation, and ranking.
A ranking strategy resorts the candidate sets generated by a retrieval or nearline strategy according to the used algorithm model, and finally generates a candidate set list for recommendation.
As a common application in the field of big data or AI, recommendation is another saying of "matching." That is, match an item with a person, no matter it is physical or virtual, or establish a relationship between an item and a person. Data, algorithm, and system play irreplaceable roles in recommendation. Data lays the foundation, algorithm acts as the core, and system provides the capability for recommendation.
A recommendation algorithm is used in offline computing for recommendation. Algorithms are the core of recommendation. Common algorithms include collaborative filtering recommendation and factorization-based recommendation.
A recommendation strategy refers to the whole configuration process from data source, data input, preprocessing, feature extraction, to algorithm parameters. A recommendation strategy relies on a recommendation algorithm and works over a large scope. A recommendation strategy is an end-to-end configuration process from the source data to the recommendation result whereas a recommendation algorithm is only a stage of the process (usually referred to as a training stage). A recommendation strategy consists of a retrieval strategy and a ranking strategy. Retrieval refers to a preliminary selection among a large number of items, removing those that are irrelevant to users and finally generating a personalized candidate set. Ranking refers to setting up a model based on multiple features for ranking candidate sets, for example, Click Through Rate (CTR).
Recommender System (RES) provides tailored media, short videos, and e-commerce recommendation services to help Internet enterprises build a recommender system in the right way. It can increase the click-through rate and retention rate to improve user experience.
A retrieval strategy is used to generate a recommendation candidate set, which matches a user from raw user data according to algorithms and rules.
Thank you for your score！Your feedback would help us improve the website.