Systems and methods for identifying candidates for item substitution

1. A system, comprising:

one or more processors; and

one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform:

receiving a test set comprising potential candidate items for substitution for a target item;

determining association scores for each of the potential candidate items in the test set;

determining one or more semantic similarity features of the potential candidate items in the test set to the target item in relation to semantic data for the target item and the potential candidate items;

determining one or more image similarity features of the potential candidate items in the test set to the target item in relation to a similarity of images of the potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN);

creating a substitutive probability model by determining a relative contribution of each of the association scores, the one or more semantic similarity features and the one or more image similarity features to a substitutive probability for the potential candidate items in the test set, with reference to a baseline set of the potential candidate items; and

determining substitutive probabilities for an evaluation set comprising candidate items for substitution for the target item by using the substitutive probability model, wherein the candidate items comprise at least one of the potential candidate items.

2. The system of claim 1 , wherein receiving the test set comprises:

receiving the potential candidate items that are identified as one or more of:

(i) one or more potential candidate items that have been viewed during a same browsing session as the target item by a same viewer, for a minimum number of browsing sessions; and

(ii) one or more of the potential candidate items belonging to a same category as the target item.

3. The system of claim 1 , wherein the baseline set of the potential candidate items comprises one or more of the potential candidate items that are manually identified as a substitute for the target item.

4. The system of claim 1 , wherein determining the association scores further comprises:

determining a viewed-also-viewed association score of the association scores by evaluating a formula:

P ( VAV )_ ij =(total number of browsing sessions in which item i and item j are both viewed)/(total number of browsing sessions in which item i is viewed and item j is not viewed),

where item i represents the target item and item j represents a potential candidate item.

5. The system of claim 1 , wherein determining the association scores further comprises:

determining a selected-also-selected association score of the association scores by evaluating a formula:

P ( SAS )_ ij =(total number of browsing sessions in which item i and item j are both selected)/(total number of browsing sessions in which item i is selected and item j is not selected),

where item i represents the target item and item j represents a potential candidate item.

6. The system of claim 1 , wherein determining the association scores further comprises:

determining a viewed-ultimately-selected association score of the association scores by evaluating a formula:

P ( VUS )_ ij =(total number of browsing sessions in which item j is selected and item i is viewed)/(total number of browsing sessions in which an item is selected and item i is viewed),

where item i represents the target item and item j represents a potential candidate item.

7. The system of claim 1 , wherein determining the one or more semantic similarity features of the potential candidate items in the test set to the target item comprises:

determining at least one semantic similarity feature in relation to semantic data for the target item and at least one of the potential candidate items that includes at least one of an item title, an item description, or one or more item categories.

8. The system of claim 1 , wherein determining the one or more semantic similarity features of the potential candidate items in the test set to the target item comprises:

evaluating semantic similarity according to a Latent Dirichlet Allocation (LDA) topic model.

9. The system of claim 1 , wherein creating the substitutive probability model by determining the relative contribution of each of the association scores, the one or more semantic similarity features and the one or more image similarity features to the substitutive probability for the potential candidate items in the test set, with reference to the baseline set of the potential candidate items, comprises:

performing a logistic regression to determine substitutive probability Pij represented by formula:

P

ij

=

1

1

+

e

-

(

β

0

+

∑

k

=

1

n

⁢

β

k

⁢

x

ijk

)

where

the x ijk 's each represent a viewed-also-viewed association score, a selected-also-selected score, a viewed-ultimately-selected score, the one or more semantic similarity features, and the one or more image similarity features, and

the β k 's each represent respective contributions of the viewed-also-viewed association score, the selected-also-selected score, the viewed-ultimately-selected score, the one or more semantic similarity features, and the one or more image similarity features, as determined via logistic regression with reference to the baseline set of the potential candidate items.

10. The system of claim 1 , wherein determining the substitutive probabilities comprises:

receiving the evaluation set of potential candidate items for substitution for the target item, the evaluation set comprising one or more of (a) potential candidate items in the test set and (b) non-test set potential candidate items that were not in the test set;

for any of the non-test set potential candidate items in the evaluation set that were not in the test set:

determining non-test set association scores for each of the non-test set potential candidate items, the non-test set association scores comprising:

(i) a viewed-also-viewed association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both viewed by a same viewer during a same browsing session;

(ii) a selected-also-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item was viewed and the non-test set potential candidate items were selected by a same viewer during a same browsing session;

determining one or more semantic similarity features of the non-test set potential candidate items to the target item in relation to semantic data for the target item and the non-test set potential candidate items; and

determining one or more image similarity features of the non-test set potential candidate items to the target item in relation to a similarity of images of the non-test set potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN); and

applying the substitutive probability model created using the test set of potential candidate items to evaluate the substitutive probabilities for one or more non-test set potential candidates of the evaluation set, using the association scores, the one or more semantic similarity features and the one or more image similarity features determined for non-test set potential candidate items in the evaluation set.

11. The system of claim 1 , wherein the association scores further comprise:

(i) a viewed-also-viewed association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item and one or more candidate items were both viewed by a same viewer during a same browsing session;

(ii) a selected-also-selected association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item and the potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item was viewed and the potential candidate items were selected by a same viewer during a same browsing session.

12. A method being implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:

receiving a test set comprising potential candidate items for substitution for a target item;

determining association scores for each of the potential candidate items in the test set;

determining one or more semantic similarity features of the potential candidate items in the test set to the target item in relation to semantic data for the target item and the potential candidate items;

determining one or more image similarity features of the potential candidate items in the test set to the target item in relation to a similarity of images of the potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN);

creating a substitutive probability model by determining a relative contribution of each of the association scores, the one or more semantic similarity features and the one or more image similarity features to a substitutive probability for the potential candidate items in the test set, with reference to a baseline set of the potential candidate items; and

determining substitutive probabilities for an evaluation set comprising candidate items for substitution for the target item by using the substitutive probability model, wherein the candidate items comprise at least one of the potential candidate items.

13. The method of claim 12 , wherein the baseline set of potential candidate items comprises one or more of the potential candidate items that are manually identified as substitutes for the target item.

14. The method of claim 12 , wherein determining the association scores further comprises:

determining a viewed-also-viewed association score of the association scores by evaluating a formula:

P ( VAV )_ ij =(total number of browsing sessions in which item i and item j are both viewed)/(total number of browsing sessions in which item i is viewed and item j is not viewed),

where item i represents the target item and item j represents a potential candidate item, and

determining a selected-also-selected association score of the association scores comprises by evaluating a formula:

P ( SAS )_ ij =(total number of browsing sessions in which item i and item j are both selected)/(total number of browsing sessions in which item i is selected and item j is not selected),

where item i represents the target item and item j represents a potential candidate item, and

determining a viewed-ultimately-selected association score of the association scores by evaluating a formula:

P ( VUS )_ ij =(total number of browsing sessions in which item j is selected and item i is viewed)/(total number of browsing sessions in which an item is selected and item i is viewed),

where item i represents the target item and item j represents a potential candidate item.

15. The method of claim 12 , wherein determining the one or more semantic similarity features of the potential candidate items in the test set to the target item comprises:

determining at least one of the one or more semantic similarity features in relation to the semantic data for the target item and at least one of the potential candidate items that includes at least one of an item title, an item description, or one or more item categories, and

evaluating semantic similarity of the semantic data according to a Latent Dirichlet Allocation (LDA) topic model.

16. The method of claim 12 , wherein creating the substitutive probability model by determining the relative contribution of each of the association scores, the one or more semantic similarity features and the one or more image similarity features to the substitutive probability for the potential candidate items in the test set, with reference to the baseline set of the potential candidate items, comprises:

performing a logistic regression to determine substitutive probability P ij represented by formula:

P

ij

=

1

1

+

e

-

(

β

0

+

∑

k

=

1

n

⁢

β

k

⁢

x

ijk

)

where

the x ijk 's each represent a viewed-also-viewed association score, a selected-also-selected score, a viewed-ultimately-selected score, the one or more semantic similarity features, and the one or more image similarity features, and

the β k 's each represent respective contributions of the viewed-also-viewed association score, the selected-also-selected score, the viewed-ultimately-selected score, the one or more semantic similarity features, and the one or more image similarity features, as determined via logistic regression with reference to the baseline set of the potential candidate items.

17. The method of claim 12 , wherein determining the substitutive probabilities comprises:

receiving the evaluation set of potential candidate items for substitution for the target item, the evaluation set comprising one or more of (a) potential candidate items in the test set and (b) non-test set potential candidate items that were not in the test set;

for any of the non-test set potential candidate items in the evaluation set that were not in the test set:

determining non-test set association scores for each of the non-test set potential candidate items, the non-test set association scores comprising:

(i) a viewed-also-viewed association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both viewed by a same viewer during a same browsing session;

(ii) a selected-also-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item was viewed and the non-test set potential candidate items were selected by a same viewer during a same browsing session;

determining one or more semantic similarity features of the non-test set potential candidate items to the target item in relation to semantic data for the target item and the non-test set potential candidate items; and

determining one or more image similarity features of the non-test set potential candidate items to the target item in relation to a similarity of images of the non-test set potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN); and

applying the substitutive probability model created using the test set of potential candidate items to evaluate the substitutive probabilities for one or more non-test set potential candidates of the evaluation set, using the association scores, the one or more semantic similarity features and the one or more image similarity features determined for non-test set potential candidate items in the evaluation set.

18. The method of claim 12 , wherein the association scores further comprise:

(i) a viewed-also-viewed association score for each of the potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the potential candidate items were both viewed by a same viewer during a same browsing session;

(ii) a selected-also-selected association score for each of the potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the potential candidate items that is determined in relation to a total number of browsing sessions in which the target item was viewed and the potential candidate items were selected by a same viewer during a same browsing session.

19. A method being implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:

receiving an evaluation set of potential candidate items for substitution for a target item, the evaluation set comprising one or more of potential candidate items in a test set and non-test set potential candidate items that were not in the test set;

for any non-test set potential candidate items in the evaluation set that were not in the test set:

determining non-test set association scores for each of the non-test set potential candidate items;

determining one or more semantic similarity features of the non-test set potential candidate items to the target item in relation to semantic data for the target item and the non-test set potential candidate items; and

determining one or more image similarity features of the non-test set potential candidate items to the target item in relation to a similarity of images of the non-test set potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN); and

applying a substitutive probability model created using the test set of potential candidate items to evaluate substitutive probabilities for one or more non-test set potential candidates of the evaluation set, using the non-test set association scores, the one or more semantic similarity features and the one or more image similarity features determined for the one or more non-test set potential candidate items in the evaluation set.

20. The method of claim 19 , wherein the substitutive probability model created using the test set of potential candidate items is created by:

receiving the test set comprising the potential candidate items for substitution for the target item; and

determining association scores for each of the potential candidate items in the test set, the association scores comprising:

(i) a viewed-also-viewed association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item and one or more candidate items were both viewed by a same viewer during a same browsing session; and

(ii) a selected-also-selected association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item and the potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the potential candidate items in the test set that is determined in relation to a total number of browsing sessions in which the target item was viewed and the potential candidate items were selected by a same viewer during a same browsing session;

determining one or more semantic similarity features of the potential candidate items in the test set to the target item in relation to semantic data for the target item and the potential candidate items;

determining one or more image similarity features of the potential candidate items in the test set to the target item in relation to a similarity of images of the potential candidate items to one or more images of the target item by evaluating image similarity using a convolutional neural network (CNN);

creating a substitutive probability model by determining a relative contribution of each of the association scores, the one or more semantic similarity features and the one or more image similarity features to a substitutive probability for the potential candidate items in the test set, with reference to a baseline set of the potential candidate items; and

determining substitutive probabilities for an evaluation set comprising candidate items for substitution for the target item by using the substitutive probability model, wherein the candidate items comprise at least one of the potential candidate items.

21. The method of claim 19 , wherein the non-test set association scores further comprise:

(i) a viewed-also-viewed association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both viewed by a same viewer during a same browsing session;

(ii) a selected-also-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item and the non-test set potential candidate items were both selected by a same viewer during a same browsing session; and

(iii) a viewed-ultimately-selected association score for each of the non-test set potential candidate items that is determined in relation to a total number of browsing sessions in which the target item was viewed and the non-test set potential candidate items were selected by a same viewer during a same browsing session.