Shulou(Shulou.com)11/24 Report--

When shopping online, I found that I had already bought a basketball and recommended basketball to me at the bottom of the paid interface. Will I return it and buy it again? Why can't this recommendation algorithm be more intelligent? The author of this article discusses this issue and shares it with you.

I wonder if you can have such an experience: open an e-commerce software, after shopping, in the purchase of a successful page, still receive a lot of similar or even identical product recommendations.

Obviously just bought a product not long ago, in the future period of time, still on the home page of e-commerce, "forced" to browse the same product.

While you cross these "useless" product recommendations helplessly, you complain about the clumsiness of the e-commerce recommendation mechanism and think that it is a big BUG in the history of recommendation algorithm. I don't know what the product manager of the relevant business thinks. It has been 2022 for such a long time without optimization, and the recommendation does not even know which products they have bought and will not buy again.

The above experience is probably experienced by many people, so why is the recommendation efficiency so inefficient that when it has led to user dissatisfaction and reduced experience, e-commerce platforms repeatedly recommend similar products that users have purchased to users?

E-commerce platform recommendation strategy to understand the causes of a phenomenon, we might as well start from the nature behind this phenomenon, understand clearly what it is, and then think about why it is much easier.

First of all, I would like to introduce to you the current mainstream e-commerce recommendation strategy, which is mainly divided into item-based recommendation system, user-based recommendation system and group-based recommendation system.

The recommendation system based on items, mainly according to the items, from the items that users have purchased / collected, to calculate the correlation between other items and this commodity, so as to re-recommend and give priority to display the goods with high relevance.

This is also a conventional recommendation strategy adopted when the website obtains less user information. After all, when the user's portrait is not comprehensive, the probability of making a recommendation based on the product that the user is interested in will be much smaller.

This kind of recommendation strategy is very common in various e-commerce platforms. Users often see a lot of recommendations for similar products on the home page after buying / searching for a category of products.

But this strategy also has drawbacks: it is easy to form information cocoon rooms, and it is difficult to pry other potential interests of users.

The recommendation system based on users mainly takes people as the starting point, calculates the similarity between other users and users according to the behavior of users, and then recommends these products to the target users according to the favorite degree of similar users.

With such a recommendation mechanism, the coverage of recommended commodity categories will be further expanded to get rid of the phenomenon that it is limited to a few categories. Pinduoduo recommendation strategy takes into account this model, according to the user's behavior, to show users the goods of interest to people who are highly related to it, and to promote users to buy through the mode of "interest circle" and "small circle".

The recommendation system based on the group is mainly based on the characteristics of the group, abstracts the commonness of the group, marks and classifies the users, and then recommends to the group users the goods that most users are interested in in this group. at the same time, the goods of interest to similar groups are appropriately added to make recommendations.

The common group-based recommendation is that every time you sign up for a new account, the platform allows users to select the circle of interest, and then recommend the content and products that the group is interested in to the user based on the selected circle.

Whether it is based on the recommendation of users, groups, items, is to extract the commonness of the characters, and then according to the similarity, to recommend. Considering from this level, it is not difficult to understand why the recommendation system always recommends very similar products, because these products are highly relevant and are highly interested by users in the eyes of the recommendation system.

Then you may ask: why is the e-commerce recommendation system so stupid that it doesn't even do the most basic de-duplication? Did I buy a proposal ring and would I want to buy another proposal ring?

The reason for repeated recommendations is true, repeatedly recommending the same product not only bores users and reduces the sense of experience, but also has no improvement in terms of sales. After all, few people will buy a product one second and another "exactly the same" the next.

However, when a problem persists for a long time and does not change much, we might as well stop and think a little more about the reason behind it.

The high frequency of buying goods is the first reason for repeated recommendations.

When a user buys a sweater, he / she may choose to buy a few more sweaters because of the change of season.

When a user buys baby diapers, soon he / she will choose to buy diapers again, or even psychologically, place an order for another brand of diapers.

When a user buys an aromatherapy, he or she may be interested in aromatherapy and will continue to place orders for other brands of aromatherapy.

As can be seen from the above, users repeatedly buy a category of goods for a variety of reasons, perhaps for specific scenarios, for the frequency of the use of goods, for their own interests, but for users, they all have the characteristics of high frequency of purchase, and users are willing to repeat purchases.

High repurchase has always been the pursuit of e-commerce platform, if there is a product user has a tendency to buy repeatedly, then it is not difficult to understand that the platform is willing to lose part of the user experience in exchange for more potential triggering business opportunities.

The commercialization of the platform is the second reason for repeated recommendations.

Advertising business has always been the main revenue module of e-commerce platform, and this module is highly related to click-through rate and exposure rate.

People have a comparative mentality, even if users of this product have made it clear that they will not buy it in the short term, but out of curiosity and comparative psychology, there are always people who are willing to click on the recently purchased products page to view the details of other similar products and make a comparison. The more users click and the longer they stay, the higher the advertising benefit to the e-commerce platform.

Even if such behavior does not improve short-term sales, the platform can take advantage of this behavior to recommend high-quality merchants with more similar products to users, which not only earns advertising fees, but also helps merchants attract a group of target audiences. In addition, disguised species can also intensify competition among merchants and enable them to improve better services and products because of this recommendation mechanism.

The limitation of the algorithm is the third reason.

Even those who are most familiar with themselves cannot know all of a person's preferences, and they will misguess the minds of the people closest to them, let alone algorithms. The recommendation system can only guess what the user likes from the user's historical behavior (purchase, collection, like, etc.) and relevance, and cannot achieve 100% accuracy.

And it is also difficult to completely select the appropriate particle size to remove the weight of the commodity. For example, if a commodity is divided into three levels, such as basketball under the category of "Sports-Ball-Basketball", then basketball should not appear on the relevant recommendation page after the user has purchased the basketball. According to this logic, it should be reprocessed in the third type of target tag.

However, if you change the category, if the user buys earrings, the category becomes "accessories-earrings-earrings". According to this logic, after removing the earrings, the user may still receive recommendations for earclips and earrings. For users who no longer want to buy similar products, such weight removal is also ineffective.

Moreover, the categories are not completely inconsistent, and some categories will intersect (for example, cheese belongs to both dairy products and baking raw materials), so after the weight removal of one category, there is no guarantee that the goods still belong to other overlapping categories and can be completely filtered out.

In other words, it is difficult for the algorithm to remove duplicates completely effectively. And users are thousands of people and dynamic, their minds are difficult to predict, can not be reduplicated and recommended according to a unified standard, nor can they perfectly predict users' future behavior according to static historical data.

In the eyes of this user, things that have been bought (such as earrings) that they will no longer buy may be repeated in the eyes of another user. It may be possible to add some user behavior to consider (such as whether this person often shops and what kind of products he or she often buys), but because people are dynamic, it is difficult for the recommendation mechanism to perfectly distinguish what products are repeated in the eyes of users. What goods are high-frequency purchases in the eyes of users.

Uncontrollable user behavior is the fourth reason for repeated recommendations on the platform.

As a general rule, it is impossible for a user to buy a highly similar item within the same day after buying an item. Do you want the user to refund and place an order again?

But the "seven days without reason" refund makes this possible. Because of the new after-sales service, many users are used to buying more than one of the same kind of goods, and then choose the one that is most suitable for them to stay after comparison. This behavior makes it possible for users to buy a large number of similar products in a short period of time, and based on this habit, e-commerce platforms are also happy to recommend similar products, after all, sales look higher, display data will also look good.

And as pointed out in the third point, people are dynamic, and their behaviors and purposes are various. Things that you do not plan to buy today are likely to order similar goods again tomorrow because of the recommendations of your friends. These are uncontrollable, giving room for repeated recommendations.

How to improve the efficiency of repeated recommendation from a point of view, repeated recommendation of goods, both at the technical level is difficult to solve the problem, in the business, user level, there is indeed this kind of demand to buy goods purchased.

So, how to strike a balance between user experience and commercialization to improve recommendation efficiency?

On the client side, we need to consider the behavior factors of more users, spread out according to the repurchase probability as a dimension, and distinguish users.

For example, whether the user is a person who is keen on shopping and tends to return goods repeatedly; whether the user likes to collect a lot of similar goods; whether the user likes to buy a lot of similar goods and so on. For users who buy multiple similar products in a short period of time, the weight removal effect can be appropriately reduced, while for those users who move from searching / recommending to placing orders quickly and with low purchase frequency, the weight removal effect can be increased.

At the same time, give users a certain right of choice, based on the algorithm can not accurately guess the current situation of users' preferences, can add "do not like" function, so that users have the right to click "do not like" for goods, e-commerce platform can be based on feedback, constantly optimize the recommended product matrix.

On the commodity side, consider the category and purchase frequency of goods, and then repeat the recommendation logic for different categories of goods.

For example, women's wear, daily necessities, snacks, users buy frequently, and there is the possibility of buying multiple items at the same time, they can join the re-recommendation mechanism, put in more advertising display modules, promote users to buy, and improve the conversion rate.

If it is a product with a low frequency of use by users and a single piece can meet the demand, reduce the re-push frequency, such as glasses, mobile phones, portable battery, etc., after ordinary users buy, they will not buy again for a long time. At this time, more complementary products can be recommended to promote users to place orders.

In addition, in terms of commercialization, we also need to consider a suitable advertising model.

After all, if advertising is just put in, it has no effect, can not promote sales and sales, and it is difficult for the platform to make a long-term profit.

When placing an advertisement, you can collect the data of the original user clicking and repeating recommended products, and analyze which module users are more likely to click on similar products, whether it is in the information flow advertisement on the front page or in the recommendation page of a successful transaction. Or in the sharing module of the e-commerce platform?

Prescribe the right remedy to the case, re-recommend in the pages and nodes where users are more interested, increase the click rate of goods, and also reduce users' sense of experience due to repeated brushing to the same product.

Summing up the repeated recommendation strategy of the e-commerce platform, users find it silly and unbearable, but from the technology, commercial factors, and user behavior, it is not difficult to see that such a mechanism still exists, and it is not possible to cut it across the board. It is also difficult not to recommend similar purchased goods.

However, we can start from the repeated recommendation process, logic, optimization, so as to achieve more efficient recommendations, so that the e-commerce platform looks a little smarter.

The picture is from Unsplash and is based on CC0 protocol.

This article comes from the official account of Wechat: everyone is a product manager (ID:woshipm), author: Conlin

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.