Relevance recommendation based on user behavior analysis

We will find that many sites have the content of the recommended function, not only like the consumer E-commerce class outstanding book recommendations, but also include interest class sites like watercress in watercress guess. This kind of function undoubtedly has the remarkable effect in helping the user discover the demand, promotes the commodity purchase and the service application. So how do you get this kind of recommendation? In fact, with the site data analysis is not unrelated, we can take a simple look at its principle and implementation.

Affiliate recommendations are divided into two categories in marketing:

Up Marketing: Provides higher value or other products or services to enhance the original function or purpose of the customer, based on past consumer preferences.

Cross-Marketing (Cross Marketing): Identify the customer's multiple needs from the customer's purchase behavior and sell the related products or services to them.

Upward marketing is based on a similar product line upgrade or optimize the product recommendation, and Cross-selling is based on similar but different types of products recommended. For a simple example, take a look at Apple's product line:

When you buy an ipod NaNO3, recommending upgrades Nano4, NANO5, or functionally similar itouch is called "Marketing", while recommending an iphone, Mac, or ipad is "cross-selling."

And the association recommendation can be divided into two kinds: the related recommendation based on product analysis and the related recommendation based on user analysis. The associated recommendation for product analysis is to find common ground by analyzing the characteristics of the product, such as the Web Analytics and the author of Web Analytics 2.0 are Avinash Kaushik, and the titles include Web Analytics, Are the Web analytics class books, but also may be the same publishing house ... Then the product based association can recommend web Analytics 2.0 to users who have purchased the Web Analytics. The recommendation based on user analysis is that by analyzing the user's historical behavior data, you may find that many of the users who purchased the Web Analytics also bought the book "The Elements of user experience," which can be recommended based on this discovery. This approach is the mining of association rules in Data Mining (association Rules), the classic case of which is Wal-Mart's beer and diapers story.

At present, many of the association recommendations are based on the product level, because the implementation is simpler (for the site, the product data is significantly less than the user's behavior data, and may be several orders of magnitude, so the analysis will be much lighter), based on the recommendation of the product more in the above mentioned two marketing means to achieve, more biased to the traditional "Push" marketing (individuals are less interested in this kind of marketing approach, especially "bundled sales").

Relevance recommendation based on user behavior analysis

Therefore, the individual is more inclined to the implementation based on user analysis, which is more conducive to discovering the potential needs of users, to help users better choose the products they need, and by the user to decide whether to buy, which is called "pull-style" marketing. By recommending products or services to users, stimulating the potential needs of users, and promoting consumer consumption, more in line with the "user-centric" concept. So the following main description of the user behavior analysis based on the relevance of the recommendation, whether you are E-commerce sites or any other type of site, in fact, you can achieve this function, as long as you have the following prerequisites:

1. Can effectively identify website users;

2. Retains the user's historical behavior data (click Stream Data (clickstream) or operational data (outcomes));

3. Of course, also need a good website data analyst.

Here take the E-commerce website as an example to illustrate the specific implementation of association rules. At present, most E-commerce sites provide user registration functions, and shopping users are generally based on the conditions of the completion of the login, so here for the user to identify the most effective identifier-user ID (on the user identification method, please refer to this article-site user identification); At the same time, the site will store all the user's shopping data in their own operating database, which provides a data base for user behavior analysis-user history shopping data. So to meet the above two conditions, we can proceed to analysis.

The principle of association rules is to realize that from all users ' shopping data (if the data is too large, you can choose a certain time interval, such as one year, one quarter, etc., to find the proportion of the number of people who bought the B commodity on the basis of a commodity, and when the ratio reached a predetermined target level, We think that there is a certain correlation between the two commodities, so when the user bought a product but did not buy B goods, we can recommend to such users B goods. The following figure:

From the figure above you can see that there are 3 sets involved: All users who have purchased a product, a collection of users who have purchased a, and a collection of users who have purchased B after a product has been purchased. Based on these 3 sets, 2 key metrics in association rule mining-support degrees (Support) and confidence (confidence) are calculated:

Support = number of purchases of A and B goods (set g)/number of all purchased goods (set U)

Confidence = number of people who purchased A and B (set g)/number of goods purchased a (set a)

With these two indicators, a minimum threshold, minimum support and minimum confidence, is required for the two indicators. Because in the user's purchase behavior, buys a product the user may not only buy the B product, also buys the C, D, E ... And so on a series of products, so we need to work out all of these combinations of support and confidence, only to meet such as support >0.2, reliability >0.6 of these commodity combinations can be considered to be relevant, it is recommended.

Of course, if your site is not an E-commerce site, you can also use the user to browse the site's click Stream data to implement the associated recommendation function. The same is based on user history behavior, such as browsing the page a users also browse the B page, watched a video users also watched B video, downloaded a file users also downloaded B file ...

Mining Association rules in data mining generally using Apriori algorithm based on frequent sets, is a relatively simple and effective algorithm, here is not specifically introduced, interested friends can go to check the data.

Some problems needing attention in the Analysis of association rules

Pay attention to the applicable scope and precondition of the association recommendation, not each kind of website is suitable or need to carry on the association recommendation;

Minimum support and minimum degree of execution need to be set according to the characteristics of the website operation, should not be high or low, the proposal is based on experiment or practice on the basis of continuous optimization, looking for a best tradeoff point.

It should be noted that in the association rules a commodity related to B goods, does not mean that B commodity and a commodity association also set up, because the confidence of the two algorithms are different, the associated direction can not be reversed.

Association rules analysis in the algorithm is not difficult, but it is really good on the site, in meeting the above 3 prerequisites on the basis of the need for continuous optimization algorithm, and more important is the need for the collaboration of various departments of the site to achieve.

Therefore, the association recommendation based on user behavior analysis is completely analyzed from the user's point of view, which is more in-depth and effective than comparing the products, and is more in line with the user's behavior habit, and is helpful to discover the latent demand of users.