(1)Headline today- guess what you're looking for
In terms of product form, guess the seven rows and two rows you want to search show 20 recommendations by default, which can be updated in a timely manner, the keywords are too long to display seven words by default, and the rest is represented by ellipses. Provides shielding operations, which can be hidden from display。
Today's headline recommendations recommend similar information to users based on past behavioral records, use content-based recommendations and direct content recommendations on the corresponding sub-pages, with the disadvantage that the high level of content match leads to a lower level of surprise of the recommended results and only increased dimensions can increase the recommended accuracy。
There is, however, a cold start-up problem, new users are not able to provide reliable recommended results, and, on the basis of data from existing user models, the various checklists that have been extracted have been supplemented and, for new users, the software is not able to obtain user preferences by searching for historical records。
"guess what you're looking for" will include some of the more recent news stories。
The figure below shows that the search history is for entertainment-related dramas or star information, but what you're looking for is still a recent, more popular event, not directly linked to the search history。
(2)TreasureRelevant recommendations - search finds
In product form, the five rows show 10 recommendations by default, which can be updated in a timely manner, provide a shielding operation, and hide away from display。
When the previous historical search records are removed, you can see a compact three rows, two or three rows, where you can see a recommendation for a previous search for historical content that can be updated in a timely manner, provide a shielding operation and hide away from display. Use synergetic filters to recommend, assuming that “you are likely to like what you like”。
Assuming that the search intent of the user can be expressed in a series of querys, then the same search intent will usually be expressed differently among people, assuming that a large number of users follow a certain pattern, then when an individual appears, we can recommend the right querys。
The main task is to find the users with the closest taste to your taste, so to predict what you might like in the light of his recent preferences. Such an approach could recommend items that are more diverse in content but are also of interest to users, who would be well advised to identify potential preferences。
As shown in the left chart below, according to the "historical search" lipstick, the "search finds" will be sent to you the limited editions of the home lipstick and the "l'oréal maple leaf" that may be searched by users like you。
However, there is also a cold start-up problem, which cannot be recommended to new users and may be of poor quality at the beginning of the system。
The electrician industry also often uses the recommendation based on the linkage rule, which is based on the linkage rule, with the purchased commodity as the subject of the rule。
As shown in the left chart below, the “history search” hair flaccid device, the “search discovery” that condenses current interests, will recommend to you multi-purpose blow-blowers, and the associated dig will reveal the relevance of different commodities in the marketing process。
Compared with the strategy of individualizing individual recommendations for each of the recommended scenarios, the treasure hunter is also included in an enhanced learning framework that integrates data from the entire link and responds to requests for recommendations for multiple isomeric scenarios。
Select the query strategy, according to the frequency query is searched for. Consider the search of query by the current query user, which can be used as the context for the user search and as a basis for our selection of query。
Using recommendations from query based on the ctr estimate, more consideration is given to the possibility that query may gain more hits with the same opportunity, and from the point of view of the search flow, more transformation can be achieved。
An estimate of the ctr of query, which is used in the model, is as follows: the character of the search word in relation to the recommended query; the characteristic of the search term in relation to the recommended query category; the characteristic of the candidate query static division in relation to the recommended query; the character of the recommended query in terms of the word; and the result page character in relation to the recommended query。
Depending on the content of the user search, it is possible to present an accurate image of the user。
The following left chart shows that the history search is mostly made up of women, so the search shows that the cosmetics or clothes are also used by women, and the right picture shows that the history search is carried out mostly in sports shoes and men's clothes, so the search also shows that the men's items are also sent。
The search for treasure found that the search term was relevant for the immediate future, the search term was relevant for the longer-term interest of the user, the search term was recommended for goods, shops and large v, the search found was more informative and the recommended content changed each time a page was refreshed。
The search found that there was a strong link to the historical search, but that there was no repetition of the content. The records of the two “fancy french sets” and “canadian goose snow curse” were previously searched, and the historical records were recently deleted。
Summary of analysis 1
Based on this analysis, it can be seen that the content of today's headlines is well recommended and that priority is given to the content of some of the themes in question if they are viewed on a permanent basis
But query's recommendation is not doing well, and the user search keyword doesn't affect what's going on in the headline "guess you want to search," and query's recommendation is not really guessing you want to search, but some real-time hotspot events from the system。
And the three messages at the bottom of the search box are selected from “guess what you want to search”, and the default search word is the first data for the three data below, with some repetition of content, with no obvious advantage except to click at a point that is closer to “guess what you want to search”。
Recommendation: further analysis of the historical records of user searches, sorting out the hot spots in "guess what you want to search" according to the user's interest in searching, would increase the number of further hits, otherwise "guess what you want to search" would be just a hot list。
2. Treasure hunting
The default search word will help users to further refine their search intentions, to find their preferred commodities more quickly and to save time and energy。
However, the words recommended in the search find will fluctuate considerably according to the user's latest search term, and the user may not buy the same commodity for a long time after having bought a commodity in the short term, thus leading to more locations found by the query recommendation, which does not meet the user's relevant interest。
Recommendation: the search needs of the resultable users are recommended in conjunction with the recent purchase order. If the user has purchased the same commodity, then the same commodity is not recommended, and it is possible to recommend different categories of items based on the same scene。
If there is a significant gap in the relevance of recent searches by users, there will be further prognosis recommendations, in addition to those related to the search, but the recommended fruit is not very good, and there will be words that are totally unrelated to the hobby, a direction that needs further optimization。
Achive
The content of the achilles default word search needs to be improved in comparison to the content of the treasure hunt, such as “the battle for love”, “the battle for love” and “the battle for love” and “the battle for love” in zhaokawa, which are almost identical, although updated in a timely manner but with little change in content, and the default word search has not been updated in real time on the basis of the content of a short search by the user, with only seven or eight fixed words being updated several times。
And guess what you want to do is you don't do a real-time update of the user's latest search, suggesting that a similar strategy can be used to make a real user search relevant。
