I. Project Source Code
Project source code URL: (This project is not open source at present.)
II. Introduction
This is a new user entrance. If the user does not have a registered profile, then the interfaces of user prediction are called and the predicted users are displayed in the front-end. In that case, the user can quickly transform the predicted children data into children profile. In this way, the conversion rate of children’s profile can be improved, and the accuracy of children’s model prediction can be improved by combining online filling and off-line BI, and the range of recallable data sources can be expanded.
Data Flow: User’s children profile -> Mining new users -> Recall predicted children by accompany plan -> Update user profile ->… (Loop)
Data logic: Users who fill in high-quality potential children profile are regarded as seed users for new user prediction. By data analysis, it is found that as the growth of children, the commodities purchased by users are with obvious sequence characteristics. And the behavioral features can be extracted from the mother’s preference of commodities, categories, brand, product attributes in different time stages. It is obvious that the characteristics of commodities purchased by mother at their children’s different ages are different, so more potential users can be predicted by these characteristics.
III. Main Contents
This is a mini recommendation system, and the upgrade development is based on the original system, including online interface and offline model.
3.1 Upgrade of Online Interface
It includes predicting user’s age, ranking interface of children model, and displaying of different filter according to user’s specific pattern.
Label recommendation: After the first-level labels are ranked, the second-level label are ranked logically. (The second-level label is the sub-label of the first-level label displayed). And then whether the user has a subscription tag or not is recognized. If so, display the subscription tag in chronological order and the rest in default order.
Data flow: Commodity category -> User account -> Quality score of related category -> Summary of labels.
3.2 Upgrade of Offline User Feature Dimension
The 200 third-level categories were added, and the algorithm rules were formulated according to the different extended attributes of user’s age, gender and season, and the quality scores of candidate sets were calculated as the reason for recalling and recommending commodities.
The results are with the following dimensions: three-level classification, age (school age) label, gender label, attribute label, commodity id, ranking quality score of candidate set, as shown in the table below.
| Three-level categories | Age | Gender | Season | Extended attribute | sku-id | Quality score of ranking |
|---|---|---|---|---|---|---|
| 12345 | 72~83 months (6-year-old) | male | summer | 1 | 1234567890 | 4.19615 |
Based on the three-level categories of commodities, the discrimination degree of the goods in age, gender and season is determined to filter the data, and the sku quality score and the ranking of labels or three-level categories are determined.
Data recall: According to the parameters from the front-end of the children’s age, gender and season, the corresponding products can be found and displayed in groups with the different label attributes they bind.
Online filtering: There is mainly some inventory filtering.
IV. Enhanced Value
Scenarioalization of User Labels : The recall data sets of categories were expanded.
Precisionalization of commodity recommendation : refine all categories in age, gender and season, and the accuracy of commodity recommendation was improved.
Renovation of user profile : The user behaviors were transformed into user profile, so that more users are participated in the iteration of children model prediction.
By the end of the project, the application has owned one million active users, and the PV/UV value has been promoted.
V. References
[1]…
[2]…
[3]…