/sellics-search-volume

Primary LanguageJava

###Assumptions

  1. Full search match based on original keyword is not representative (ex.: "iphone" returns "iphone" in the list);
  2. Only partial search will return trending information containing the relative search volumes (ex.: "ip" returns "iphone" in the list);
  3. The shortest substring of the original keyword will return the match with the biggest weight;
  4. The order is comparatively insignificant - that means: a) exact match based on original keyword doesn't show the trend; b) the order is changed depending on partial search length (so we have to take into account all possible partial searches);
  5. Prefix-matched words are misleading as they can point to another products, we should look at the match of partial search with original keyword.

###Algorithm

  1. Create an array of weights depending on how long is the partial search string. Assumption: The sum of all weights is 100, the biggest weight is for single-char partial search. Calculate number of chars for all partial searches (ex.: 3 chars + 2 chars + 1 chars = 3 * (3 + 1) / 2 ): int sumOfPartialCharCounts = maxPartialCharCount * (maxPartialCharCount + 1) / 2; Calculate weight per 1 char: double weightCoefficient = 100.0 / sumOfPartialCharCounts; Fill in the array of weights, so that 1 char search has the biggest weight. Example: for keyword "canon", partial searches ["c", "ca", "can", "cano"] the weight array is [40.0, 30.0, 20.0, 10.0]

  2. Get autocomplete list for each partial search, notice the position in the list. Example: for keyword "canon" partial searches would be "c", "ca", "can", "cano".

  3. Sum up partial search scores - this would be the final score. Example: for keyword "canon", partial searches ["c", "ca", "can", "cano"] and top position in the autocomplete list we would have:
    40.0 / 1 + 30.0 / 1 + 20.0 / 1 + 10.0 / 1 => 100.0 but usually it would be like: 40.0 / 4 + 30.0 / 3 + 20.0 / 2 + 10.0 / 1 for "c", "ca", "can", "cano" respectively.

In other words, the logic of this algorithm is: we take into account only search by part of the keyword. The less part of the original input is found - the more searched is the input. For instance, if I search "iphone charger" and get "iphone charger" - it is not surprising, but if I search the string "iphone" and still get "iphone charger" - it means this string is searched very often.

So, basically, we remove one character from the input and check, whether we have original input in the response or not.

The final score is sum of every partial search. The score of every partial search is based upon the assumptions:

  1. the smaller the input is, the more weight it has (the word is very trending);
  2. the place in the list returned - if it's higher, it has better search volume.

The weight is counted in such a way: 100 is the total weight. We found, how much weight we have per char position (we find number of positions like sum of natural numbers from 1 to length-1 )
Then we divide 100 by this number and have weightCoefficient per 1 position.

And multiply position and this number to have an array of weights, then use this information along with position in the autocomplete list for each partial search and sum all up.

###About the hint The hint is correct in the meaning:

  • exact match based on original keyword doesn't show the trend;
  • the order is changed depending on partial search length (so we have to take into account all possible partial searches);
  • Amazon could (potentially) raise the keyword in the list to increase its sales.

###How precise is solution It's potentially relatively precise because it takes into account weighted partial search with ranked match. Why? Because if the keyword is seen at the top position in the list, when only partial search is requested - it says about how often customers request this product. But it doesn't say anything about how it is positioned against other keywords, searched on the Amazon, as we have no information about it.