Hema is a technologically innovative company, and it is also a consumption-driven company that returns to consumer values: buy it, buy it well, buy it conveniently, buy it with confidence, and buy it with confidence. happy.
Hema includes Hema Fresh, X Member Store, Hema Chaoyun, Hema Neighborhood and other business models. The core business model is online and offline integration, the fastest The O2O (Hema Fresh Food) model that delivers products to your home in 30 minutes.
Hema Ma selects global quality products and pursues the ultimate freshness; combined with category characteristics and consumer shopping experience expectations, it chooses the most efficient business model for different categories. Hema Fresh’s sales account for 60% to 70%, making it the core category. This category is characterized by high timeliness expected by users, and is very suitable for businesses like Hema Fresh that open stores near users.
Before considering what supply chain model to use, logistics costs and inventory costs need to be balanced. For the same quantity of goods, the entire logistics cost is much lower than that of parcel logistics, and the cold chain difference is even more obvious (delivery timeliness and scale effects are not considered for the time being). On the contrary, the more dispersed the inventory, the greater the uncertainty in demand. If the consumer demand cannot be accurately grasped, the inventory will be placed in the store in advance, which will lead to high stockouts and losses, that is, high inventory costs.
In order to achieve efficient supply and demand matching, multiple business models require supporting supply chain support. Hema's multiple business models are integrated at the backend and share trunk networks and inventory as much as possible, greatly improving resource utilization and improving supply chain efficiency.
Supply chain algorithm is essentially based on traditional supply chain methods, using the power of algorithms and data to improve efficiency.
1. E-commerce industry supply chain algorithm positioning
# #E-commerce industry algorithms are mainly divided into basic algorithms and business algorithms.
Basic algorithms include images, speech, text, etc. E-commerce business algorithms are intuitively divided into three types. First, from the front end (traffic, user * product dimension), it is not much different from video websites and information websites, using search, advertising, and recommendation methods; second, from the middle end (retail , commodity dimension), including commodity, price and inventory prediction and decision-making algorithms; third, from the back-end (logistics, order dimension), including fulfillment, warehousing, distribution and other optimization algorithms. The broad supply chain algorithm includes mid-end (retail and commodity forecasting and decision-making) and back-end (logistics and order optimization), while the narrow supply chain algorithm only includes mid-end algorithms.
The "bullwhip effect" refers to a phenomenon of demand variation amplification in the supply chain, which makes the information flow When transmitted from the final client to the original supplier, information sharing cannot be effectively realized, causing the information to be distorted and gradually amplified, resulting in greater and greater fluctuations in demand information. The amplification effect of this information distortion looks like a swing on the graph. The bullwhip effect is therefore vividly called the "bullwhip effect".
Retailers are the enterprises closest to consumers, best able to perceive and grasp consumer needs, and most capable of responding to market fluctuations through data and algorithms.
Retailing is essentially the matching of people and goods. In the daily operation of the enterprise, first of all, the company will launch a sales plan and strategy. The sales plan determines the supply, and the supply capacity determines the inventory. The short-term sales ceiling is determined through inventory and fulfillment capacity, thereby regulating sales. Next The sales plan of the wheel refers to the above historical information to achieve a complete closed loop of commercial operation.
High-quality products are Hema’s consistent pursuit. In order to satisfy customers' ultimate freshness experience, Hema products have a very short shelf life and even provide a one-day-only series of daily fresh products that only meet daily needs, covering daily consumer categories such as milk, vegetables, meat and poultry, etc., becoming an industry leader. Benchmark. Instant fresh food service is very in line with consumer demand, but it also places high demands on the supply chain.
A shorter sales period means that out-of-stock or loss is more likely to occur. In this regard, the Hema supply chain algorithm considers Factors such as weather, seasonality, holidays, product substitution, marketing activities, online and offline displays, etc. have been used to build a series of high-precision demand forecast models with Hema characteristics, and the inventory is optimized through the simulation system to achieve a highly automated ordering system. , significantly reducing labor costs while optimizing inventory indicators.
In terms of technical depth and innovation, the spatiotemporal heterogeneous graph neural network model has been successfully introduced into product sales forecasting needs, which has solved the problem of excessive information loss in complex marketing activities and significantly improved the prediction accuracy.
In terms of algorithm effect, the Hema prediction algorithm won the championship and runner-up in the main data set of the Alibaba Group time series prediction competition. Results, overall accuracy is high and versatility is strong.
In terms of business results, the adoption rate of fresh food ordering system algorithm recommendations has stabilized at over 96%, and ordering efficiency has increased by 70% , the loss rate is reduced by 30%, and the out-of-stock rate is reduced by 25%. In terms of industry influence, Hema’s algorithm solution integrating forecasting, inventory, price and control was successfully shortlisted for the 2022 Franz Edelman Outstanding Achievement Award Finals. 
Hema The iteration path of the sales forecast algorithm is divided into four stages: simple model, machine learning model, deep time series model, and spatiotemporal graph network model. The simple model is close to business understanding and ensures coverage of all SKUs; the machine learning model is relatively stable and has fewer outliers, but relies heavily on feature engineering and has poor timing scalability; the deep timing model does not rely on feature factories, has less original timing information missing, and has strong timing scalability. ; The spatio-temporal graph network model considers the correlation between commodities and the influence between samples.
When forecasting, the main information for forecasting is product sales, considering Since activities have an impact on product sales, a graph relationship can be established between products and activities. In addition, other factors that have an impact on product sales can also be included in the graph model. For this model, in each time window, a A heterogeneous picture of product sales and various characteristics is drawn. During the calculation process, algorithms such as GraphSage and GATNE are first used to extract graph information on each time slice, thereby obtaining point update information, and then passing it down to obtain the entire time series information. In practical applications, graph information is auxiliary information, and the main information is still timing information; at the same time, timing models can be replaced with different timing models according to different scenarios.
The inventory model is mainly divided into There are two parts, the first is the inventory of a single SKU, and the second is the summary of multiple warehouses. A unique scenario for new retail. The safety reserve refers to when online and offline share inventory. Online customers place orders first and then take the physical items, while offline customers take the physical items first and then place orders. This will cause online The problem is that the goods ordered first by online customers are taken away by offline users, resulting in online orders being unable to be fulfilled. Therefore, in such a scenario, it is necessary to control parameters. When the inventory is less than a certain value, sales will not be carried out online to prevent the risk of being unable to ship in time.
is making order quantity decisions The process is mainly divided into three steps. The first is to split the business goals, and determine whether a certain product should focus on ensuring out-of-stock or loss protection based on differences in business goals and scenarios. The second step is to infer inventory batches. For example, in our warehouse There are 10 items, 5 of which are close-to-date products. If the order is not ordered in time, there may be out-of-stock due to the expiration of the close-to-date products, so inventory batch inference is required; the third step is the calculation of optimal order quantity, which supports various The model is used to estimate order quantities.
Due to the short shelf life of fresh products, although the sales forecast and inventory model will maximize the model performance, they still There are cases where a small amount of goods are ordered too much or too little, especially when ordering too much will cause a lot of losses. Therefore, by establishing a dynamic inventory control system, we can monitor sales in real time and update forecast results, conduct inventory warnings, and adjust inventory through online promotions, APP traffic tilt, and offline store discounts to try to avoid losses caused by inventory backlog. The difficulty of this system lies in the joint regulation of flow and price, which requires global optimization of high-frequency flow decisions and low-frequency price decisions.
A1: The industry usually uses weighted MAPE, that is, accuracy = 1-(total error of the product pool/total true sales)*100%. This error is weighted. For example, if the actual sales volume of a certain product is large, the error will be large and the contribution to the global error will be large, which is in line with business understanding.
A2: Mainly refer to the idea of this space-time isomorphic graph paper, combined with its own scene, using heterogeneous graphs. Huang Y, Bi H K, Li Z, et al. STGAT: Modeling spatial-temporal interactions for human trajectory prediction[C]//Proceedings of the IEEE International Conference on Computer Vision. 2019: 6272-6281.
#A3: It is not the key to use a simple model, the main thing is that it can cover all SKUs and be used for the bottom-up strategy. Common ones include simple moving average method, same period last week, etc.
The above is the detailed content of Hema supply chain algorithm in practice. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
