Obviously, recommendation systems tend to recommend products that have proven to be popular over time – where generally or among customers with certain taste profiles.
But popularity -based recommendations can miss Trending Products or product that, although they have yet reached high amounts of purchase, are rapidly increasing in popularity. Customers who make an inquiry today may feel the card change if they miss a new product that, two or three days from now, will prove to have been a game-chaming post in the product area.
We see that Amazon customers, when performing product queries, will not only receive a list of matches based on historical data, but also a list of trending matches so that they will have as much information as possible when making purchase decisions. Becuse We want to catch the trend as it happens – not after the fact when it appears in the data – we use time series that forecasts to predict that products will be in the near future.
We describe our method in a paper we presented at this year’s ACM Conference on Recommendation Systems (Recsys). First we define closely trend in the form of speed – The number of customer interactions with product pages at each time step – and acceleration – The speed at which speed changes from time steps to time steps. Then we offer a new machine learning scheme where we require a model in the task of the next-up recommendation so that it can learn product functions and correlates with interaction patterns. After predetermination, the model is fine -tuned on the task of predicting trends (acceleration speed).
To evaluate our model – tnertré – we compared it to three baselines: one was a simple Markov model that assumes a constant increase in transaction volume from once to the next (constant acceleration); One was an exponential moving-irritating model that predicts the acceleration at the next time step based on a weighted sum of accelerations in the last eight time steps; And one was a neural network trained on time series of acceleration rats.
We test all oven models on five data sets using two different measurements. We found that our model, with its additional knowledge of the relationship between product functions and sales patterns, surpassed the baseline neural model everywhere – by 15% and 16% on a dataset. Both neural models dramatically surpassed the heuristic base lines, indicating that changes in acceleration rats do not follow -linear patterns that can be seen in the data.
Learning Representation
The goal of our prior procedure is to learn the model to produce product representations that will be useful for the trend prediction task. The diampling is the product that the same subgroups of customers interact with will exhibit similar popularity trends over time. If the model learns to correlate product functions with specific subgroups, it can learn to correlate the same functions with specific trend patterns.
Accordingly, the prior task is to predict what product a given customer will interact with next, based on the customer’s interaction history. The model receives customers’ interaction history as input, and it learns to produce two vector representations (embedders): one of a customer’s taste and one of product functions.
After predetermination, the model is fine -tuned on the task of predicting future acceleration speed from previous acceleration speeds. Product Function Injury is still used, but the customer’s taste embedding is not. The diampling is that the flavor entrance has affected the product renting when they were trained together.
Time intervals
An important consideration in our training is the time interval that we evaluate a model’s performance. We want to train the model to predict acceleration speed – but acceleration speed over what time? An hour? A day? On the weekend?
We suspect there is a relationship between the time interval that we predict acceleration and the feasibility of learning a predictable model. If the interval is for shorts, the data is too noisy: for cases the acceleration speed may be flat or even negative in the first 15 minutes of the prediction period, even if they are very high in the next three hours. At the same time, if the time interval is too long, the increase in popularity at its end may be dead, so the overall acceleration speed artificially low.
When we train a trend prediction model, our goal is to find the sweet place between for shorts and long time interval – a number that can vary widely across data sets. In our experience, however, we found that the simplest of our baseline-heuristic model, which assumes constant acceleration development, is a good estimate of the form of the feasibility-versus-time interval curve to enable the choice of time interval. And because heuristics is so simple, the estimate can calculate effectively.
In our experimental setup, Tnerruck and all three baselines the ten products with the highest acceleration speeds of the predicted time interval, calculated using the constant acceleration heuristics. We evaluated the results using two measurements: the cumulative accelerations of the top 10 results, normalized accords for min-max scaling and trendy-normalized cumulative reinforcement (TNDCG) that evaluate the order of the 10 best results, which aside great weight for the top results.
Both neural models surpassed Nordisk de heuristic baselines, but to the trend gave the best results everywhere.