Automatically generated search suggestions
Some implementations relate to methods and computer-readable media to automatically generate search suggestions. In some implementations, a method includes receiving search session data for a plurality of search sessions, each search session comprising a plurality of search terms input by a respective user. The method further includes, for each search session, identifying that a particular search term of the plurality of search terms is followed by a gameplay session of a particular game of a plurality of games of an online gaming platform, determining that the particular search term is a successful term when the particular search term is a last search term in the sequence, and in response to the determining, adding the search session data to a training corpus. The method further includes applying a machine learning algorithm to the training corpus to generate a plurality of embeddings of search terms.
1. A computer-implemented method comprising:
receiving search session data for a plurality of search sessions, each search session comprising a plurality of search terms input by a respective user, wherein the plurality of search terms of the search session are in a sequence;
for each search session,
identifying that a particular search term of the plurality of search terms is followed by a gameplay session of a particular game of a plurality of games of an online gaming platform;
determining that the particular search term is a successful term when the particular search term is a last search term in the sequence; and
in response to the determining, adding the search session data to a training corpus; and
applying a machine learning algorithm to the training corpus to generate a plurality of embeddings of search terms, each embedding corresponding to a particular search term of the plurality of search terms.
2. The computer-implemented method of claim 1 , wherein identifying that the particular search term of the plurality of search terms is followed by the gameplay session of the particular game comprises detecting that the user engaged in the gameplay session of the particular game for at least a threshold time duration.
3. The computer-implemented method of claim 1 , further comprising selecting a plurality of top embeddings, wherein the plurality of top embeddings is a subset of embeddings of search terms of the plurality of search terms.
4. The computer-implemented method of claim 3 , further comprising:
receiving search input that includes a fresh search term;
generating a query embedding based on the fresh search term;
comparing the query embedding with the plurality of top embeddings to obtain one or more search results, each search result indicative of a corresponding game of the plurality of games; and
providing the search results in response to the search input.
5. The computer-implemented method of claim 1 , wherein identifying that the particular search term of the plurality of search terms is followed by the gameplay session of the particular game comprises detecting that the user purchased an item associated with the particular game, provided a rating for the particular game, or customized an avatar in the particular game.
6. The computer-implemented method of claim 1 , wherein each term of the plurality of search terms is associated with a respective timestamp, and wherein a difference in timestamps associated with each pair of consecutive search terms is less than a threshold difference.
7. The computer-implemented method of claim 1 , wherein applying the machine learning algorithm comprises applying the machine learning algorithm to a plurality of n-grams constructed from the plurality of search terms.
8. The computer-implemented method of claim 1 , wherein applying the machine learning algorithm comprises shuffling the search terms associated with the search session prior to applying the machine learning algorithm.
9. The computer-implemented method of claim 1 , wherein each embedding is a multi-dimensional numerical feature vector associated with the search term.
10. The computer-implemented method of claim 1 , wherein each embedding has a dimension between about 100 and about 400.
11. The computer-implemented method of claim 1 , wherein determining that the particular search term is a successful term comprises determining that the particular search term is a last search term in the sequence and when no other search term of the plurality of search terms is followed by the gameplay session.
12. The computer-implemented method of claim 1 , wherein embeddings of search terms that are followed by a gameplay session of the same game have a cosine similarity greater than embeddings of search terms that are followed by gameplay sessions of different games.
13. A computer-implemented method comprising:
receiving a search term input by a user;
generating a query embedding for the search term input;
calculating a respective cosine similarity between the query embedding and a plurality of top embeddings;
identifying one or more top embeddings that are within a threshold cosine similarity from the query embedding; and
providing search results that include games associated with the one or more top embeddings.
14. The computer-implemented method of claim 13 , wherein providing search results comprises providing a user interface that includes the games associated with the one or more top embeddings sorted based on the cosine similarity.
15. The computer-implemented method of claim 13 , further comprising:
prior to generating the query embedding for the search term input, verifying that a click through rate associated with the search term input meets a predetermined threshold.
16. The computer-implemented method of claim 13 , wherein each query embedding is a multi-dimensional numerical feature vector associated with the search term input.
17. A non-transitory computer-readable medium comprising instructions that, responsive to execution by a processing device, causes the processing device to perform operations comprising:
receiving a search term input by a user;
generating a query embedding for the search term input;
calculating a respective cosine similarity between the query embedding and a plurality of top embeddings;
identifying one or more top embeddings that are within a threshold cosine similarity from the query embedding; and
providing search results that include games associated with the one or more top embeddings.
18. The non-transitory computer-readable medium of claim 17 , wherein providing search results comprising providing a user interface that includes the games associated with each of the one or more top embeddings sorted based on the cosine similarity.
19. The non-transitory computer-readable medium of claim 17 , wherein the operations further comprise:
prior to generating the query embedding for the search term input, verifying that a click through rate associated with the search term input meets a predetermined threshold.
20. The non-transitory computer-readable medium of claim 17 , wherein each query embedding is a multi-dimensional numerical feature vector associated with the search term input.