More improvements in recommendation systems
This invention deals with more improvements in recommendation systems. Retailers want to grow their business and increase sales. One embodiment creates upsell by arranging similar items by price. Another embodiment creates similar items by calculating the number of shared attributes. A final embodiment shows how recently viewed items affect cross-sell, similar and personalized recommendations to make more relevant recommendations.
1. A method to automatically display dynamic web pages, which may be used for personalized crowd-based and genomic product recommendations in electronic form on a user's computing device with one or more similar items for a target item wherein each item includes a first attribute having multiple attribute values, and there is a catalog with other items also having multiple attribute values from the first attribute, the method comprising the steps of:
a. tracking items from the catalog that users view by monitoring click stream data from a website,
b. creating a first list of similar items by correlating the items that are viewed by common users, resulting in correlation values that are based on a number of common users, wherein the correlation values are generated using an efficient computer implementation that: (i) organizes the click stream data into a first array arranged by item index and a second array arranged by user index to reduce a time required to find the number of common users, (ii) stores the first and second arrays as jagged arrays to reduce a memory required to store the click stream data and further reduce the time required to find the number of common users, (iii) for each item, uses the first array to find the users that viewed that item, (iv) for each of these found users, uses the second array to update the number of common users for each item viewed by the found user, and (v) chooses and stores the K largest correlation values to further reduce search time and memory usage,
c. sorting the first list by correlation value,
d. creating a second list of similar items that calculates the number of the same attribute values of the first attribute between the target item and other items in the catalog,
e. sorting the second list by the number of the same attribute values,
f. keeping track of one or more recently viewed items from the website,
g. re-sorting the first or second list based on these recently viewed items,
h. responding in real-time via a web service with two or more similar items, wherein the two or more similar items include one or more items from the first list and one or more items from the second list, per a request from the user's computing device, such that the similar items can be different for different users based on each user's recently viewed items,
i. automatically displaying in real-time the two or more similar items to the user, wherein the display includes the target item in an electronic view with the similar items on the user's computing device which communicates through the web service,
wherein steps (a) through (e) are computationally complex and performed by a non-real-time train component running on a first computer that is separate from the user's computing device, and the non-real-time train component transforms click stream and first attribute data to similar items, steps (g) through (h) are performed by a real-time recommend component running on a second computer that is separate from the user's computing device and the first computer, the website runs on a third computer that is separate from the first computer the second computer and the user's computing device, and the real-time recommend component uses a table lookup to respond with the similar items; thus, separating the computational complexity from the response and enabling the real-time recommend component to reply in real-time.
2. The method of claim 1 wherein the steps also include determining the number of sales for said one or more similar items, and secondarily sorting the second list of said one or more similar items by said number of sales.
3. The method of claim 1 wherein a second attribute having multiple elements is also used, weighted less than said first attribute in sorting the second list, and if no similar items share both an element of said first attribute and an element of said second attribute, items sharing at least one element of said first attribute or one element of said second attribute are backfilled.
4. The method of claim 3 wherein the steps also include determining the number of sales for said one or more of the similar items, and weighted less than said first attribute and second attribute in sorting said at least two similar items.
5. The method of claim 1 wherein a maximum number of actions for any item is determined, and the sorting of the second list is accomplished by multiplying the number of shared first attributes times the maximum number of actions for any item.
6. The method of claim 1 wherein the one or more similar items are arranged into a list ordered by a price factor related to the price of each similar item.
7. The method of claim 6 wherein said price factor demotes the similar item based upon the ratio of the price of the similar items to the price of the target item, if the similar items is less than the price of said target item.
8. The method of claim 1 wherein the electronic view comprises a web page or email.
9. The method of claim 1 comprising the additional steps of calculating the similar items for the one or more of the recently viewed items using steps (a) through (c) in the non-real-time train component, and, in the real-time recommend component, performed before step (i), summing the correlation value if the similar item for the recently viewed item is the same as the similar item for the target item.
10. The method of claim 2 comprising the additional steps of calculating the similar items for the one or more of the recently viewed items using steps (d) through (e) in the non-real-time train component, and, in the real-time recommend component, performed before step (i), summing the sales if the similar item for the recently viewed item is the same as the similar item for the target item.
11. The method of claim 1 comprising the step, in the real-time recommend component, performed before step (i), decreasing the correlation value if the similar item for the target item is also the recently viewed item, and the items have been recently viewed multiple times.