Methods and apparatuses for recommending seasonal items

1. A system comprising a computing device comprising at least one processor, the computing device coupled to at least one database, the computing device configured to:

obtain periodic sales data from the at least one database characterizing a number of purchases made of each item of a plurality of items in a specified period;

obtain periodic buyers data from the at least one database characterizing a number of unique customers of each item in the plurality of items in the specified period;

obtain periodic purchase data from the at least one database characterizing a count of purchases by each user in the specified period;

determine, using the at least one processor, an item embedding for each item in the plurality of items based on the periodic sales data and the periodic buyers data;

determine, using the at least one processor, an item seasonality embedding based on the periodic sales data and the periodic buyers data;

determine, using the at least one processor, a category seasonality embedding based on the periodic sales data and the periodic buyers data;

blend, using the at least one processor, the item seasonality embedding and the category seasonality embedding using a blending function that weighs the category seasonality embedding based on a number of items in a category to obtain a final item seasonality embedding for each item in the plurality of items;

determine, using the at least one processor, a user embedding for each user based on the periodic purchase data;

determine, using the at least one processor, a final user seasonality embedding for each user based on the periodic purchase data;

generate, using the at least one processor, final user-item scores by applying, for each item in the plurality of items and for each user, a trained machine learning model to a respective item embedding, a respective final item seasonality embedding, a respective user embedding, and a respective final user seasonality embedding, wherein the trained machine learning model is applied by:

generating an item matrix based on the respective item embedding and the respective final item seasonality embedding,

generating a user matrix based on the respective user embedding and the respective final user seasonality embedding,

computing a dot product of corresponding item and user vectors from the item matrix and the user matrix, respectively, to generate a final user-item score,

wherein the trained machine learning model is trained using the at least one processor based on Bayesian Personalized Ranking and matrix factorization using item pairs and a criterion/loss function; and

send, using the at least one processor, a recommendation to a user based on the final user-item scores.

2. The system of claim 1 , wherein the item embedding is a non-seasonal item embedding and the user embedding is a non-seasonal user embedding.

3. A method comprising:

obtaining periodic sales data from at least one database characterizing a number of purchases made of each item of a plurality of items in a specified period;

obtaining periodic buyers data from at least one database characterizing a number of unique customers of each item in the plurality of items in the specified period;

obtaining periodic purchase data from at least one database characterizing a count of purchases by each user in the specified period;

determining, using at least one processor, an item embedding for each item in the plurality of items based on the periodic sales data and the periodic buyers data;

determining, using the at least one processor, an item seasonality embedding based on the periodic sales data and the periodic buyers data;

determining, using the at least one processor, a category seasonality embedding based on the periodic sales data and the periodic buyers data;

blending, using the at least one processor, the item seasonality embedding and the category seasonality embedding using a blending function that weighs the category seasonality embedding based on a number of items in a category to obtain a final item seasonality embedding for each item in the plurality of items;

determining, using the at least one processor, a user embedding for each user based on the periodic purchase data;

determining, using the at least one processor, a final user seasonality embedding for each user based on the periodic purchase data;

generating, using the at least one processor, final user-item scores by applying, for each item in the plurality of items and for each user, a trained machine learning model to a respective item embedding, a respective final item seasonality embedding, a respective user embedding, and a respective final user seasonality embedding, wherein the trained machine learning model is applied by:

generating an item matrix based on the respective item embedding and the respective final item seasonality embedding,

generating a user matrix based on the respective user embedding and the respective final user seasonality embedding,

computing a dot product of corresponding item and user vectors from the item matrix and the user matrix, respectively, to generate a final user-item score,

wherein the trained machine learning model is trained using the at least one processor based on Bayesian Personalized Ranking and matrix factorization using item pairs and a criterion/loss function; and

sending, using the at least one processor, a recommendation to a user based on the final user-item scores.

4. The method of claim 3 , wherein the item embedding is a non-seasonal item embedding and the user embedding is a non-seasonal user embedding.

5. A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by at least one processor, cause a device to perform operations comprising:

obtaining periodic sales data from at least one database characterizing a number of purchases made of each item of a plurality of items in a specified period;

obtaining periodic buyers data from at least one database characterizing a number of unique customers of each item in the plurality of items in the specified period;

obtaining periodic purchase data from at least one database characterizing a count of purchases by each user in the specified period;

determining, using at least one processor, an item embedding for each item in the plurality of items based on the periodic sales data and the periodic buyers data;

determining, using the at least one processor, an item seasonality embedding based on the periodic sales data and the periodic buyers data;

determining, using the at least one processor, a category seasonality embedding based on the periodic sales data and the periodic buyers data;

blending, using the at least one processor, the item seasonality embedding and the category seasonality embedding using a blending function that weighs the category seasonality embedding based on a number of items in a category to obtain a final item seasonality embedding for each item in the plurality of items;

determining, using the at least one processor, a user embedding for each user based on the periodic purchase data;

determining, using the at least one processor, a final user seasonality embedding for each user based on the periodic purchase data;

generating, using the at least one processor, final user-item scores by applying, for each item in the plurality of items and for each user, a trained machine learning model to a respective item embedding, a respective final item seasonality embedding, a respective user embedding, and a respective final user seasonality embedding, wherein the trained machine learning model is applied by:

generating an item matrix based on the respective item embedding and the respective final item seasonality embedding,

generating a user matrix based on the respective user embedding and the respective final user seasonality embedding,

computing a dot product of corresponding item and user vectors from the item matrix and the user matrix, respectively, to generate a final user-item score,

wherein the trained machine learning model is trained using the at least one processor based on Bayesian Personalized Ranking and matrix factorization using item pairs and a criterion/loss function; and

sending, using the at least one processor, a recommendation to a user based on the final user-item scores.