Listwise autocomplete ranking with position debias

1 . A method, comprising:

receiving a user input indicative of a portion of a search query;

receiving a set of candidate queries based on the portion of the search query;

inputting, to a listwise click-through rate (CTR) prediction model that is trained using a shallow tower running parallel to one or more transformer layers to remove positional bias based on position information of respective candidate queries in the set of candidate queries, respective feature vectors corresponding to candidate queries of the set of candidate queries;

generating, by the listwise CTR prediction model, respective click probability prediction values corresponding to the candidate queries based on joint comparisons between the respective feature vectors;

generating respective ranking scores for the candidate queries based on a product of the respective click probability prediction values and respective reward per search values associated with the candidate queries; and

outputting at least one candidate query of the set of candidate queries based on an ordered set of candidate queries, wherein the ordered set of candidate queries is ordered based on the respective ranking scores and independent of the positional bias.

2 . The method of claim 1 , wherein the respective feature vectors include at least one attribute of the portion of the search query, at least one statistic associated with the portion of the search query, at least one attribute of the candidate queries, at least one statistic associated with the candidate queries or combinations thereof.

3 . The method of claim 1 , wherein the set of candidate queries are received based on a prefix match or a conjunctive match of each candidate query with the portion of the search query.

4 . The method of claim 1 , wherein the outputting comprises transmitting an indication of an autocomplete list including the at least one candidate query of the set of candidate queries to a device for display to a user.

5 . The method of claim 1 , wherein the at least one candidate query includes a subset of candidate queries each having a respective ranking score above a threshold ranking score.

6 . The method of claim 1 , wherein an autocomplete list includes the at least one candidate query of the set of candidate queries.

7 . A method, comprising:

receiving a user input indicative of a portion of a search query;

receiving a set of candidate queries based on the portion of the search query;

inputting, to a listwise click-through rate (CTR) prediction model that is trained using a shallow tower running parallel to one or more transformer layers to remove positional bias based on position information of respective candidate queries in the set of candidate queries, respective feature vectors corresponding to candidate queries of the set of candidate queries;

generating, by the listwise CTR prediction model, respective click probability prediction values corresponding to the candidate queries based on joint comparisons between the respective feature vectors, wherein the shallow tower is excluded from the listwise CTR prediction model to generate the respective click probability prediction values;

generating respective ranking scores for the candidate queries based on a product of the respective click probability prediction values and respective reward per search values associated with the candidate queries; and

outputting at least one candidate query of the set of candidate queries based on an ordered set of candidate queries, wherein the ordered set of candidate queries is ordered based on the respective ranking scores and independent of the positional bias.

8 . The method of claim 7 , wherein the respective feature vectors include at least one attribute of the portion of the search query, at least one statistic associated with the portion of the search query, at least one attribute of the candidate queries, at least one statistic associated with the candidate queries or combinations thereof.

9 . The method of claim 7 , wherein the set of candidate queries are received based on a prefix match or a conjunctive match of each candidate query with the portion of the search query.

10 . The method of claim 7 , wherein the outputting comprises transmitting an indication of an autocomplete list including the at least one candidate query of the set of candidate queries to a device for display to a user.

11 . The method of claim 7 , wherein the at least one candidate query includes a subset of candidate queries each having a respective ranking score above a threshold ranking score.

12 . The method of claim 7 , wherein an autocomplete list includes the at least one candidate query of the set of candidate queries.

13 . The method of claim 1 , wherein the listwise CTR prediction model includes the one or more transformer layers and one or more classification layers, and wherein the listwise CTR prediction model includes the shallow tower during training.

14 . The method of claim 13 , further comprising:

generating, by the listwise CTR prediction model and during the training of the listwise CTR prediction model, respective logits associated with a plurality of training candidate queries for training the listwise CTR prediction model, wherein the respective logits are based on a summation of a first output of the one or more transformer layers and a second output of the shallow tower;

providing the respective logits as input to the one or more classification layers to obtain training click probability prediction values associated with the plurality of training candidate queries; and

updating one or more parameters of the listwise CTR prediction model based on a loss calculated from the training click probability prediction values.

15 . The method of claim 14 , wherein the loss is based on a loss function that compares predicted probability distributions over the plurality of training candidate queries to ground truth labels derived from historical user interaction data, wherein the ground truth labels indicate whether respective training candidate queries of the plurality of training candidate queries are clicked.

16 . The method of claim 13 , wherein generating the respective click probability prediction values comprises:

processing the respective feature vectors through the one or more transformer layers to generate transformer output values; and

providing the transformer output values to the one or more classification layers to generate the respective click probability prediction values, wherein the shallow tower is excluded from the listwise CTR prediction model to generate the respective click probability prediction values.

17 . The method of claim 7 , wherein the listwise CTR prediction model includes the one or more transformer layers and one or more classification layers, and wherein the listwise CTR prediction model includes the shallow tower during training.

18 . The method of claim 17 , further comprising:

generating, by the listwise CTR prediction model and during the training of the listwise CTR prediction model, respective logits associated with a plurality of training candidate queries for training the listwise CTR prediction model, wherein the respective logits are based on a summation of a first output of the one or more transformer layers and a second output of the shallow tower;

providing the respective logits as input to the one or more classification layers to obtain training click probability prediction values associated with the plurality of training candidate queries; and

updating one or more parameters of the listwise CTR prediction model based on a loss calculated from the training click probability prediction values.

19 . The method of claim 18 , wherein the loss is based on a loss function that compares predicted probability distributions over the plurality of training candidate queries to ground truth labels derived from historical user interaction data, wherein the ground truth labels indicate whether respective training candidate queries of the plurality of training candidate queries are clicked.

20 . The method of claim 17 , wherein generating the respective click probability prediction values comprises:

processing the respective feature vectors through the one or more transformer layers to generate transformer output values; and

providing the transformer output values to the one or more classification layers to generate the respective click probability prediction values.