A Global Ranking Method Based on the Interaction Between Commodities

⏵ 1. Preface

Traditional search and sorting methods cannot take into account the mutual impact of displayed products. Similarly, traditional methods for estimating ctr and cvr for a single product are based on the assumption that the ctr and cvr of the product are not affected by other products displayed simultaneously (we call them display context). In fact, the display context of a product can affect users' click or purchase decisions: If the products around a product are similar to it, but the price is more convenient than it, the probability of users buying it is not high; Conversely, if the surrounding products are more expensive than it, the probability of users buying it will greatly increase.

If we break the assumption that traditional sorting models show that context has no impact, how can we sort them? For this reason, we propose for the first time a global ordering method that considers the interaction between products. We describe e-commerce ordering as a global optimization problem, with the goal of reflecting user satisfaction with the transaction volume of the product: GMV (Gross Merchandise Volume). To be precise, the optimization goal of global sorting is to maximize the mathematical expectations of GMV. To calculate the mathematical expectation of GMV, it is necessary to know the transaction probability of a product, which is influenced by each other. Therefore, we propose a transaction probability estimation model that considers the interaction between products.

First, we propose a global feature extension approach that adds the impact of other products to the probability estimation model in the form of global features when estimating the transaction probability of a product, taking into account the impact of other products during estimation. Then, we further propose an RNN model to accurately consider the impact of product ordering on transaction probability. By using the RNN model, we turn e-commerce ranking into a sequence generation problem, and use the beam search algorithm to find a better ranking. We conducted a large number of experiments on the Taobao wireless main search platform, and achieved a 5% improvement in GMV compared to the current Taobao wireless main search algorithm.

⏵ 2. Global sorting method

The input of the global sorting stage is the N items to be sorted, and the output is the sorting result of the N items.

We use S=(1,..., N) to represent the top ‑ N product sequence of the basic sorting output;

O represents the full array set of S;

Represents an array of S;

In addition, it is used to indicate the position of commodity 1. JPG in the sorting o, that is;

Represents the display context of item i in sorting o. Specific definitions will be discussed later based on the situation;

Indicates the value of the product.

We define the target as the GMV generated by this search, so given o, there is

The ultimate goal of global ranking is to find a ranking with the greatest expected return:

Finding this optimal ranking requires solving two problems:

Question 1. How to accurately estimate this probability;

Problem 2. Searching is a combinatorial optimization problem, and the time complexity of violent search is N! We need to find a reasonable and efficient approximation algorithm.

Problem 1 is the key to improving the effect. The more accurate the estimation, the more obvious the effect of the subsequent combination optimization, and conversely, it will lead to the meaningless combination optimization in the second step.

Theoretically, 1. JPG is a complete display context for i. However, if the display context is considered fully, it is difficult to reduce the complexity of combinatorial optimization in Problem 2. In order to solve Problem 2 more efficiently, it is necessary to simplify 1. JPG appropriately. According to the degree of simplification of 1. JPG pairs, we divide global sorting models into two categories, which are described below.

2.1 Global Feature Extension

The first type of model only uses the commodity collection information of the display context, without delving into the order of the actual display context, namely, 1.JPG. Intuitively, the first type of model assumes that when determining whether to purchase product i, the user only remembers that he has seen all the product sets S, regardless of the order in which the products appear in S.

This situation is equivalent to the user knowing their selection set, comparing and selecting the products they most want to buy from it. We refer to the characteristics of product i itself as local characteristics 1. JPG. At this point, the local features of commodity 1. JPG can be compared with the local features of other candidate commodities to obtain the results of comparing the characteristics of this commodity with other commodities in the candidate set, and the results of these comparisons can be added to the prediction as features containing global information.

Taking the price characteristics as an example, we will rank S according to the price dimension, and uniformly normalize the price order of commodity i to a value between (the most expensive price becomes 1, and the most convenient price becomes 0). This price characteristic extends the global price characteristics of commodity i.

By using the above method, each dimension can be expanded according to its position to obtain corresponding global features. Finally, we will splice the local and global features as features of the commodity i that are considered to display the context. At this time, it is shown that context is added to the model through feature expansion to help the model predict the transaction probability.

Under the assumption of the first type of model, the combinatorial optimization of Problem 2 becomes very simple. The product set is static, so the transaction probabilities of each product can be calculated independently. However, the ranking of product i is not taken into account in the DNN calculation, and a correction should be made based on the position bias, namely. For solving Problem 2, we do not need to know the specific value, but only need to know that the bias decreases sequentially with the position from front to back, because at this time, as long as the goods are sorted from high to low, it is certain to obtain the most profitable ranking.

2.2 Generation of sorting sequence

The second type of model not only considers the product set information that displays the context, but also accurately considers the order of the products that precede i. Similar to the first type of model, the product set information that displays the context is added to the features of each product by extending the global features, that is. However, unlike the first type of model, the second type of model also considers the real order in front of i when calculating, that is, problem 1 becomes a sequential probability estimation problem. The most intuitive method is to use RNN to calculate.

Unlike the first type of model, because we consider the order of the products before i, changes in the order of the preceding products will affect the transaction probability of the products. At this point, it is not possible to calculate the income of each product separately, as the income of product i is affected by the products that rank ahead of it; "Before the final ordering is determined, it is impossible to know what the previous ordering of product i was, and our goal is to determine such an optimal ordering.". At the same time, it is precisely because we only consider the order of products that precede i, allowing us to sort them step by step from front to back, each step selecting the products in the current position.

To facilitate understanding, let's first look at a simple case: quantile sorting. The so-called quantized sorting refers to ranking the product with the largest income first, then recalculating the income of the remaining products based on the condition above the first product, and selecting the product with the largest income to rank second, and so on. Clearly, quantized sorting is a greedy algorithm. The beam search algorithm can be understood as an extension of the greedy algorithm. At each step of the search, the beam search will retain the sequence with the greatest benefit, until the last step, and then return the best sequence.

The original RNN model does not solve the problem of long distance dependence: When we calculate the transaction probability of products in the 20th position, the top 4 products have almost no impact. As the products in the front row are the first to be seen by users, they generally have a deep impression, and their impact on the products in the rear should not be ignored. In order to enable the model to take into account long distances and dependence on sequencing locations, we designed a new attachment mechanism to be added to the RNN network. By introducing the presence and embedding the location, the model can automatically learn how to adjust the presence of different locations based on the data to obtain better prediction results. Intuitively, the structure diagram of the model is as follows:

⏵ 3. Experimental effect

3.1 Estimation of transaction probability

Where DNN is a baseline that only uses as a feature. ReDNN is a DNN global ordering model using global features, miRNN is an RNN global ordering model, and miRNN+attention is an RNN model that adds our attention mechanism.

3.2 Taobao Wireless Main Search Online A/B Test

Using the miRNN and miRNN+attention models for sorting, the algorithm's time is and, respectively, where is the number of items being sorted and is the beam size parameter of the beam search algorithm. For a large-scale search platform such as Taobao Search, the complexity is obviously too high. "However, we can only sort the top N items based on the baseline sorting. As long as N is not obtained significantly, the computational cost can be borne.". At the same time, because the top products have the greatest impact on the effect, the benefits of reordering these products are also relatively large.

We compared the GMV growth (response model effect) and latency (response computing burden) growth curves of the search engine for various global ranking methods proposed in this paper under different N and k conditions:

Summary of final results: GMV and search engine Latency want to contribute to the growth of baseline DNN:

⏵ 4. Conclusion

We first proposed a global ranking method for e-commerce that considers the interaction between products, and achieved significant results in Taobao main search. At present, the biggest problem is that the RNN method is more effective, but it brings too much computational burden. How to reduce the computational load will be an important research direction.